re PR target/52941 (SH Target: Add support for movco.l / movli.l atomics on SH4A)
authorOleg Endo <olegendo@gcc.gnu.org>
Tue, 29 May 2012 20:26:41 +0000 (20:26 +0000)
committerOleg Endo <olegendo@gcc.gnu.org>
Tue, 29 May 2012 20:26:41 +0000 (20:26 +0000)
PR target/52941
* config/sh/predicates.md (atomic_arith_operand,
atomic_logical_operand): New predicates.
* config/sh/sh.c (sh_option_override): Check atomic options.
* config/sh/sh.h (TARGET_ANY_ATOMIC, UNSUPPORTED_ATOMIC_OPTIONS,
UNSUPPORTED_HARD_ATOMIC_CPU): New macros.
(DRIVER_SELF_SPECS): Use UNSUPPORTED_ATOMIC_OPTIONS and
UNSUPPORTED_HARD_ATOMIC_CPU.
* config/sh/sync.md: Update description comments.
(I12): New mode iterator.
(fetchop_predicate, fetchop_constraint): New code attributes.
(atomic_compare_and_swapsi_hard, atomic_compare_and_swap<mode>_hard,
atomic_exchangesi_hard, atomic_exchange<mode>_hard,
atomic_fetch_<fetchop_name>si_hard,
atomic_fetch_<fetchop_name><mode>_hard,
atomic_fetch_nandsi_hard, atomic_fetch_nand<mode>_hard,
atomic_<fetchop_name>_fetchsi_hard,
atomic_<fetchop_name>_fetch<mode>_hard,
atomic_nand_fetchsi_hard, atomic_nand_fetch<mode>_hard,
atomic_test_and_set_hard): New insns.
(atomic_compare_and_swap<mode>_soft, atomic_exchange<mode>_soft,
atomic_fetch_<fetchop_name><mode>_soft, atomic_fetch_nand<mode>_soft,
atomic_<fetchop_name>_fetch<mode>_soft, atomic_nand_fetch<mode>_soft,
atomic_test_and_set_soft): Use same formatting for the first line of
the asm block as in new insns above.
(atomic_compare_and_swap<mode>, atomic_exchange<mode>,
atomic_fetch_<fetchop_name><mode>, atomic_<fetchop_name>_fetch<mode>,
atomic_test_and_set): Integrate new *_hard insns into expanders.
* config/sh/sh.opt (mhard-atomic): New option.
* doc/invoke.texi (SH Options): Document it.

From-SVN: r187987

gcc/ChangeLog
gcc/config/sh/predicates.md
gcc/config/sh/sh.c
gcc/config/sh/sh.h
gcc/config/sh/sh.opt
gcc/config/sh/sync.md
gcc/doc/invoke.texi

index e2ebe2f54ca05717ca3b57c2f23f8be1e64c2144..26382f61cabca5e22a618a5d6a41eb686ad90645 100644 (file)
@@ -1,3 +1,36 @@
+2012-05-29  Oleg Endo  <olegendo@gcc.gnu.org>
+
+       PR target/52941
+       * config/sh/predicates.md (atomic_arith_operand,
+       atomic_logical_operand): New predicates.
+       * config/sh/sh.c (sh_option_override): Check atomic options.
+       * config/sh/sh.h (TARGET_ANY_ATOMIC, UNSUPPORTED_ATOMIC_OPTIONS,
+       UNSUPPORTED_HARD_ATOMIC_CPU): New macros.
+       (DRIVER_SELF_SPECS): Use UNSUPPORTED_ATOMIC_OPTIONS and
+       UNSUPPORTED_HARD_ATOMIC_CPU.
+       * config/sh/sync.md: Update description comments.
+       (I12): New mode iterator.
+       (fetchop_predicate, fetchop_constraint): New code attributes.
+       (atomic_compare_and_swapsi_hard, atomic_compare_and_swap<mode>_hard,
+       atomic_exchangesi_hard, atomic_exchange<mode>_hard,
+       atomic_fetch_<fetchop_name>si_hard,
+       atomic_fetch_<fetchop_name><mode>_hard,
+       atomic_fetch_nandsi_hard, atomic_fetch_nand<mode>_hard,
+       atomic_<fetchop_name>_fetchsi_hard,
+       atomic_<fetchop_name>_fetch<mode>_hard,
+       atomic_nand_fetchsi_hard, atomic_nand_fetch<mode>_hard,
+       atomic_test_and_set_hard): New insns.
+       (atomic_compare_and_swap<mode>_soft, atomic_exchange<mode>_soft,
+       atomic_fetch_<fetchop_name><mode>_soft, atomic_fetch_nand<mode>_soft,
+       atomic_<fetchop_name>_fetch<mode>_soft, atomic_nand_fetch<mode>_soft,
+       atomic_test_and_set_soft): Use same formatting for the first line of
+       the asm block as in new insns above.
+       (atomic_compare_and_swap<mode>, atomic_exchange<mode>,
+       atomic_fetch_<fetchop_name><mode>, atomic_<fetchop_name>_fetch<mode>,
+       atomic_test_and_set): Integrate new *_hard insns into expanders.
+       * config/sh/sh.opt (mhard-atomic): New option.
+       * doc/invoke.texi (SH Options): Document it.
+
 2012-05-29  Meador Inge  <meadori@codesourcery.com>
 
        * c-decl.c (c_push_function_context): Always create a new language
index c6d0d464f08c847e7cd5fa034c3b159ada4071b5..f75675ec0964cb74f5b82364c0f56d785a408032 100644 (file)
     }
   return 0;
 })
+
+;; The atomic_* operand predicates are used for the atomic patterns.
+;; Depending on the particular pattern some operands can be immediate
+;; values.  Using these predicates avoids the usage of 'force_reg' in the
+;; expanders.
+(define_predicate "atomic_arith_operand"
+  (ior (match_code "subreg,reg")
+       (and (match_test "satisfies_constraint_I08 (op)")
+           (match_test "mode != QImode")
+           (match_test "mode != HImode")
+           (match_test "TARGET_SH4A_ARCH"))))
+
+(define_predicate "atomic_logical_operand"
+  (ior (match_code "subreg,reg")
+       (and (match_test "satisfies_constraint_K08 (op)")
+           (match_test "mode != QImode")
+           (match_test "mode != HImode")
+           (match_test "TARGET_SH4A_ARCH"))))
+
index 4a2e8f43d8968c189ec4b4fbeb9b590844872dc3..3c9a7e9398cfb48da0903f82878f33d134041409 100644 (file)
@@ -882,6 +882,14 @@ sh_option_override (void)
   /* This target defaults to strict volatile bitfields.  */
   if (flag_strict_volatile_bitfields < 0 && abi_version_at_least(2))
     flag_strict_volatile_bitfields = 1;
+
+  /* Make sure that only one atomic mode is selected and that the selection
+     is valid for the current target CPU.  */
+  if (TARGET_SOFT_ATOMIC && TARGET_HARD_ATOMIC)
+    error ("-msoft-atomic and -mhard-atomic cannot be used at the same time");
+  if (TARGET_HARD_ATOMIC && ! TARGET_SH4A_ARCH)
+    error ("-mhard-atomic is only available for SH4A targets");
+
 }
 \f
 /* Print the operand address in x to the stream.  */
index a6fac9484c2443ed44e128fa3606461c9d7ea286..b5f20f90558aad74fa93c5f19422a5e4c35ec674 100644 (file)
@@ -172,6 +172,9 @@ do { \
   (TARGET_SH1 && ! TARGET_SH2E && ! TARGET_SH5 \
    && ! (TARGET_HITACHI || sh_attr_renesas_p (FUN_DECL)))
 
+/* Nonzero if either soft or hard atomics are enabled.  */
+#define TARGET_ANY_ATOMIC (TARGET_SOFT_ATOMIC | TARGET_HARD_ATOMIC)
+
 #ifndef TARGET_CPU_DEFAULT
 #define TARGET_CPU_DEFAULT SELECT_SH1
 #define SUPPORT_SH1 1
@@ -433,7 +436,20 @@ do { \
 "%{m2a*:%eSH2a does not support little-endian}}"
 #endif
 
-#define DRIVER_SELF_SPECS UNSUPPORTED_SH2A
+#define UNSUPPORTED_ATOMIC_OPTIONS \
+"%{msoft-atomic:%{mhard-atomic:%e-msoft-atomic and -mhard-atomic cannot be \
+used at the same time}}"
+
+#if TARGET_CPU_DEFAULT & MASK_SH4A
+#define UNSUPPORTED_HARD_ATOMIC_CPU ""
+#else
+#define UNSUPPORTED_HARD_ATOMIC_CPU \
+"%{!m4a*:%{mhard-atomic:%e-mhard-atomic is only available for SH4A targets}}"
+#endif
+
+#undef DRIVER_SELF_SPECS
+#define DRIVER_SELF_SPECS UNSUPPORTED_SH2A, UNSUPPORTED_ATOMIC_OPTIONS,\
+                         UNSUPPORTED_HARD_ATOMIC_CPU
 
 #define ASSEMBLER_DIALECT assembler_dialect
 
index 3ab2c51be4a220a2202c6ca6865e5338e232e9b3..6a78d04ed109c71ad9b72906b9150e83c75cae1e 100644 (file)
@@ -321,7 +321,11 @@ Follow Renesas (formerly Hitachi) / SuperH calling conventions
 
 msoft-atomic
 Target Report Var(TARGET_SOFT_ATOMIC)
-Use software atomic sequences supported by kernel
+Use gUSA software atomic sequences
+
+mhard-atomic
+Target Report Var(TARGET_HARD_ATOMIC)
+Use hardware atomic sequences
 
 menable-tas
 Target Report RejectNegative Var(TARGET_ENABLE_TAS)
index 258e048f3c7bb61878459d4d532cb8c3ebaceba2..79cd765d87f2be1395d18da18157a025e9199d78 100644 (file)
 ;;
 ;; Atomic integer operations for the Renesas / SuperH SH CPUs.
 ;;
+;; On SH CPUs atomic integer operations can be done either in 'software' or
+;; in 'hardware', where true hardware support was introduced with the SH4A.
+;; In addition to that all SH CPUs support the 'tas.b' instruction, which
+;; can be optionally used to implement the 'atomic_test_and_set' builtin.
+;;
+;; tas.b atomic_test_and_set (-menable-tas)
+;;
+;; Depending on the particular hardware configuration, usage of the 'tas.b'
+;; instruction might be undesired or even unsafe.  Thus, it has to be
+;; enabled by the user explicitely.  If it is not enabled, the
+;; 'atomic_test_and_set' builtin is implemented either with hardware or with
+;; software atomics, depending on which is enabled.  It is also possible to
+;; enable the 'tas.b' instruction only, without enabling support for the 
+;; other atomic operations.
+;;
+;;
+;; Hardware Atomics (-mhard-atomic, SH4A only)
+;;
+;; Hardware atomics implement all atomic operations using the 'movli.l' and
+;; 'movco.l' instructions that are availble on SH4A.  On multi-core hardware
+;; configurations hardware atomics is the only safe mode.
+;; However, it can also be safely used on single-core configurations.
+;; Since these instructions operate on SImode memory only, QImode and HImode
+;; have to be emulated with SImode and subreg masking, which results in
+;; larger code.
+;;
+;;
+;; Software Atomics (-msoft-atomic)
+;;
 ;; On single-core systems there can only be one execution context running
 ;; at a given point in time.  This allows the usage of rewindable atomic
 ;; sequences, which effectively emulate locked-load / conditional-store
-;; operations.
+;; operations.  This requires complementary support in the interrupt / 
+;; exception handling code (e.g. kernel) and does not work safely on multi-
+;; core configurations.
+;;
 ;; When an execution context is interrupted while it is an atomic
 ;; sequence, the interrupted context's PC is rewound to the beginning of
 ;; the atomic sequence by the interrupt / exception handling code, before
 ;; For correct operation the atomic sequences must not be rewound after
 ;; they have passed the write-back instruction.
 ;;
-;; The current implementation is limited to QImode, HImode and SImode 
+;; The current atomic support is limited to QImode, HImode and SImode 
 ;; atomic operations.  DImode operations could also be implemented but
 ;; would require some ABI modifications to support multiple-instruction
 ;; write-back.  This is because SH1/SH2/SH3/SH4 does not have a DImode
 ;; store instruction.  DImode stores must be split into two SImode stores.
 ;;
-;; For some operations it would be possible to use insns with an immediate
-;; operand such as add #imm,Rn.  However, since the original value before
-;; the operation also needs to be available, this is not so handy.
+;; On single-core SH4A CPUs software atomic aware interrupt / exception code
+;; is actually compatible with user code that utilizes hardware atomics.
+;; Since SImode hardware atomic sequences are more compact on SH4A they are
+;; always used, regardless of the selected atomic mode.
 
 (define_c_enum "unspec" [
   UNSPEC_ATOMIC
 ])
 
 (define_mode_iterator I124 [QI HI SI])
+(define_mode_iterator I12 [QI HI])
 
 (define_mode_attr i124suffix [(QI "b") (HI "w") (SI "l")])
 (define_mode_attr i124extend_insn [(QI "exts.b") (HI "exts.w") (SI "mov")])
 (define_code_attr fetchop_name
   [(plus "add") (minus "sub") (ior "or") (xor "xor") (and "and")])
 
+(define_code_attr fetchop_predicate
+  [(plus "atomic_arith_operand") (minus "register_operand")
+   (ior "atomic_logical_operand") (xor "atomic_logical_operand")
+   (and "atomic_logical_operand")])
+
+(define_code_attr fetchop_constraint
+  [(plus "rI08") (minus "r") (ior "rK08") (xor "rK08") (and "rK08")])
+
+;;------------------------------------------------------------------------------
+;; comapre and swap
+
 (define_expand "atomic_compare_and_swap<mode>"
   [(match_operand:SI 0 "register_operand" "")          ;; bool success output
    (match_operand:I124 1 "register_operand" "")                ;; oldval output
    (match_operand:I124 2 "memory_operand" "")          ;; memory
-   (match_operand:I124 3 "register_operand" "")                ;; expected input
-   (match_operand:I124 4 "register_operand" "")                ;; newval input
+   (match_operand:I124 3 "atomic_arith_operand" "")    ;; expected input
+   (match_operand:I124 4 "atomic_arith_operand" "")    ;; newval input
    (match_operand:SI 5 "const_int_operand" "")         ;; is_weak
    (match_operand:SI 6 "const_int_operand" "")         ;; success model
    (match_operand:SI 7 "const_int_operand" "")]                ;; failure model
-  "TARGET_SOFT_ATOMIC && !TARGET_SHMEDIA"
+  "TARGET_ANY_ATOMIC && !TARGET_SHMEDIA"
 {
-  rtx addr;
+  rtx addr = force_reg (Pmode, XEXP (operands[2], 0));
+  rtx old_val = gen_lowpart (SImode, operands[1]);
+  rtx exp_val = operands[3];
+  rtx new_val = operands[4];
+  rtx atomic_insn;
+
+  if (TARGET_HARD_ATOMIC || (TARGET_SH4A_ARCH && <MODE>mode == SImode))
+      atomic_insn = gen_atomic_compare_and_swap<mode>_hard (old_val, addr,
+                                                           exp_val, new_val);
+  else
+      atomic_insn = gen_atomic_compare_and_swap<mode>_soft (old_val, addr,
+                                                           exp_val, new_val);
+  emit_insn (atomic_insn);
 
-  addr = force_reg (Pmode, XEXP (operands[2], 0));
-  emit_insn (gen_atomic_compare_and_swap<mode>_soft
-            (gen_lowpart (SImode, operands[1]), addr, operands[3],
-             operands[4]));
   if (<MODE>mode == QImode)
     emit_insn (gen_zero_extendqisi2 (gen_lowpart (SImode, operands[1]),
                                     operands[1]));
   DONE;
 })
 
+(define_insn "atomic_compare_and_swapsi_hard"
+  [(set (match_operand:SI 0 "register_operand" "=&r")
+       (unspec_volatile:SI
+         [(mem:SI (match_operand:SI 1 "register_operand" "r"))
+          (match_operand:SI 2 "arith_operand" "rI08")
+          (match_operand:SI 3 "arith_operand" "rI08")]
+         UNSPECV_CMPXCHG_1))
+   (set (mem:SI (match_dup 1))
+       (unspec_volatile:SI [(const_int 0)] UNSPECV_CMPXCHG_2))
+   (set (reg:SI T_REG)
+       (unspec_volatile:SI [(const_int 0)] UNSPECV_CMPXCHG_3))
+   (clobber (reg:SI R0_REG))]
+  "TARGET_ANY_ATOMIC && TARGET_SH4A_ARCH"
+{
+  return "\r0: movli.l @%1,r0"         "\n"
+        "      cmp/eq  %2,r0"          "\n"
+        "      bf{.|/}s        0f"     "\n"
+        "      mov     r0,%0"          "\n"
+        "      mov     %3,r0"          "\n"
+        "      movco.l r0,@%1"         "\n"
+        "      bf      0b"             "\n"
+        "0:";
+}
+  [(set_attr "length" "14")])
+
+(define_insn "atomic_compare_and_swap<mode>_hard"
+  [(set (match_operand:SI 0 "register_operand" "=&r")
+       (unspec_volatile:SI
+         [(mem:I12 (match_operand:SI 1 "register_operand" "r"))
+          (match_operand:I12 2 "register_operand" "r")
+          (match_operand:I12 3 "register_operand" "r")]
+         UNSPECV_CMPXCHG_1))
+   (set (mem:I12 (match_dup 1))
+       (unspec_volatile:I12 [(const_int 0)] UNSPECV_CMPXCHG_2))
+   (set (reg:SI T_REG)
+       (unspec_volatile:SI [(const_int 0)] UNSPECV_CMPXCHG_3))
+   (clobber (reg:SI R0_REG))
+   (clobber (match_scratch:SI 4 "=&r"))
+   (clobber (match_scratch:SI 5 "=&r"))
+   (clobber (match_scratch:SI 6 "=1"))]
+  "TARGET_HARD_ATOMIC && TARGET_SH4A_ARCH"
+{
+  return "\r   mov     #-4,%5"                 "\n"
+        "      <i124extend_insn>       %2,%4"  "\n"
+        "      and     %1,%5"                  "\n"
+        "      xor     %5,%1"                  "\n"
+        "      add     r15,%1"                 "\n"
+        "      add     #-4,%1"                 "\n"
+        "0:    movli.l @%5,r0"                 "\n"
+        "      mov.l   r0,@-r15"               "\n"
+        "      mov.<i124suffix>        @%1,%0" "\n"
+        "      mov.<i124suffix>        %3,@%1" "\n"
+        "      cmp/eq  %4,%0"                  "\n"
+        "      bf{.|/}s        0f"             "\n"
+        "      mov.l   @r15+,r0"               "\n"
+        "      movco.l r0,@%5"                 "\n"
+        "      bf      0b"                     "\n"
+        "0:";
+}
+  [(set_attr "length" "30")])
+
 (define_insn "atomic_compare_and_swap<mode>_soft"
   [(set (match_operand:SI 0 "register_operand" "=&u")
        (unspec_volatile:SI
    (clobber (reg:SI R1_REG))]
   "TARGET_SOFT_ATOMIC && !TARGET_SHMEDIA"
 {
-  return "mova 1f,r0"                          "\n"
+  return "\r   mova    1f,r0"                  "\n"
         "      <i124extend_insn>       %2,%4"  "\n"
         "      .align 2"                       "\n"
         "      mov     r15,r1"                 "\n"
 }
   [(set_attr "length" "20")])
 
+;;------------------------------------------------------------------------------
+;; read - write - return old value
+
 (define_expand "atomic_exchange<mode>"
   [(match_operand:I124 0 "register_operand" "")                ;; oldval output
    (match_operand:I124 1 "memory_operand" "")          ;; memory
-   (match_operand:I124 2 "register_operand" "")                ;; newval input
+   (match_operand:I124 2 "atomic_arith_operand" "")    ;; newval input
    (match_operand:SI 3 "const_int_operand" "")]                ;; memory model
-  "TARGET_SOFT_ATOMIC && !TARGET_SHMEDIA"
+  "TARGET_ANY_ATOMIC && !TARGET_SHMEDIA"
 {
   rtx addr = force_reg (Pmode, XEXP (operands[1], 0));
-  emit_insn (gen_atomic_exchange<mode>_soft
-            (operands[0], addr, operands[2]));
+  rtx val = operands[2];
+  rtx atomic_insn;
+
+  if (TARGET_HARD_ATOMIC || (TARGET_SH4A_ARCH && <MODE>mode == SImode))
+      atomic_insn = gen_atomic_exchange<mode>_hard (operands[0], addr, val);
+  else
+      atomic_insn = gen_atomic_exchange<mode>_soft (operands[0], addr, val);
+
+  emit_insn (atomic_insn);
+
   if (<MODE>mode == QImode)
     emit_insn (gen_zero_extendqisi2 (gen_lowpart (SImode, operands[0]),
                                     operands[0]));
   DONE;
 })
 
+(define_insn "atomic_exchangesi_hard"
+  [(set (match_operand:SI 0 "register_operand" "=&r")
+       (mem:SI (match_operand:SI 1 "register_operand" "r")))
+   (set (mem:SI (match_dup 1))
+       (unspec:SI
+         [(match_operand:SI 2 "arith_operand" "rI08")] UNSPEC_ATOMIC))
+   (clobber (reg:SI R0_REG))]
+  "TARGET_ANY_ATOMIC && TARGET_SH4A_ARCH"
+{
+  return "\r0: movli.l @%1,r0"         "\n"
+        "      mov     r0,%0"          "\n"
+        "      mov     %2,r0"          "\n"
+        "      movco.l r0,@%1"         "\n"
+        "      bf      0b";
+}
+  [(set_attr "length" "10")])
+
+(define_insn "atomic_exchange<mode>_hard"
+  [(set (match_operand:I12 0 "register_operand" "=&r")
+       (mem:I12 (match_operand:SI 1 "register_operand" "r")))
+   (set (mem:I12 (match_dup 1))
+       (unspec:I12
+         [(match_operand:I12 2 "register_operand" "r")] UNSPEC_ATOMIC))
+   (clobber (reg:SI R0_REG))
+   (clobber (match_scratch:SI 3 "=&r"))
+   (clobber (match_scratch:SI 4 "=1"))]
+  "TARGET_HARD_ATOMIC && TARGET_SH4A_ARCH"
+{
+  return "\r   mov     #-4,%3"                 "\n"
+        "      and     %1,%3"                  "\n"
+        "      xor     %3,%1"                  "\n"
+        "      add     r15,%1"                 "\n"
+        "      add     #-4,%1"                 "\n"
+        "0:    movli.l @%3,r0"                 "\n"
+        "      mov.l   r0,@-r15"               "\n"
+        "      mov.<i124suffix>        @%1,%0" "\n"
+        "      mov.<i124suffix>        %2,@%1" "\n"
+        "      mov.l   @r15+,r0"               "\n"
+        "      movco.l r0,@%3"                 "\n"
+        "      bf      0b";
+}
+  [(set_attr "length" "24")])
+
 (define_insn "atomic_exchange<mode>_soft"
   [(set (match_operand:I124 0 "register_operand" "=&u")
        (mem:I124 (match_operand:SI 1 "register_operand" "u")))
    (clobber (reg:SI R1_REG))]
   "TARGET_SOFT_ATOMIC && !TARGET_SHMEDIA"
 {
-  return "mova 1f,r0"                          "\n"
+  return "\r   mova    1f,r0"                  "\n"
         "      .align 2"                       "\n"
         "      mov     r15,r1"                 "\n"
         "      mov     #(0f-1f),r15"           "\n"
 }
   [(set_attr "length" "14")])
 
+;;------------------------------------------------------------------------------
+;; read - add|sub|or|and|xor|nand - write - return old value
+
 (define_expand "atomic_fetch_<fetchop_name><mode>"
   [(set (match_operand:I124 0 "register_operand" "")
        (match_operand:I124 1 "memory_operand" ""))
    (set (match_dup 1)
        (unspec:I124
          [(FETCHOP:I124 (match_dup 1)
-            (match_operand:I124 2 "register_operand" ""))]
+            (match_operand:I124 2 "<fetchop_predicate>" ""))]
          UNSPEC_ATOMIC))
    (match_operand:SI 3 "const_int_operand" "")]
-  "TARGET_SOFT_ATOMIC && !TARGET_SHMEDIA"
+  "TARGET_ANY_ATOMIC && !TARGET_SHMEDIA"
 {
-  rtx addr;
+  rtx addr = force_reg (Pmode, XEXP (operands[1], 0));
+  rtx atomic_insn;
+
+  if (TARGET_HARD_ATOMIC || (TARGET_SH4A_ARCH && <MODE>mode == SImode))
+    atomic_insn = gen_atomic_fetch_<fetchop_name><mode>_hard (operands[0], addr,
+                                                             operands[2]);
+  else
+      atomic_insn = gen_atomic_fetch_<fetchop_name><mode>_soft (operands[0],
+                                                               addr,
+                                                               operands[2]);
+  emit_insn (atomic_insn);
 
-  addr = force_reg (Pmode, XEXP (operands[1], 0));
-  emit_insn (gen_atomic_fetch_<fetchop_name><mode>_soft
-            (operands[0], addr, operands[2]));
   if (<MODE>mode == QImode)
     emit_insn (gen_zero_extendqisi2 (gen_lowpart (SImode, operands[0]),
                                     operands[0]));
   DONE;
 })
 
+(define_insn "atomic_fetch_<fetchop_name>si_hard"
+  [(set (match_operand:SI 0 "register_operand" "=&r")
+       (mem:SI (match_operand:SI 1 "register_operand" "r")))
+   (set (mem:SI (match_dup 1))
+       (unspec:SI
+         [(FETCHOP:SI (mem:SI (match_dup 1))
+            (match_operand:SI 2 "<fetchop_predicate>" "<fetchop_constraint>"))]
+         UNSPEC_ATOMIC))
+   (clobber (reg:SI R0_REG))]
+  "TARGET_ANY_ATOMIC && TARGET_SH4A_ARCH"
+{
+  return "\r0: movli.l @%1,r0"         "\n"
+        "      mov     r0,%0"          "\n"
+        "      <fetchop_name>  %2,r0"  "\n"
+        "      movco.l r0,@%1"         "\n"
+        "      bf      0b";
+}
+  [(set_attr "length" "10")])
+
+(define_insn "atomic_fetch_<fetchop_name><mode>_hard"
+  [(set (match_operand:I12 0 "register_operand" "=&r")
+       (mem:I12 (match_operand:SI 1 "register_operand" "r")))
+   (set (mem:I12 (match_dup 1))
+       (unspec:I12
+         [(FETCHOP:I12 (mem:I12 (match_dup 1))
+            (match_operand:I12 2 "<fetchop_predicate>" "<fetchop_constraint>"))]
+         UNSPEC_ATOMIC))
+   (clobber (reg:SI R0_REG))
+   (clobber (match_scratch:SI 3 "=&r"))
+   (clobber (match_scratch:SI 4 "=1"))]
+  "TARGET_HARD_ATOMIC && TARGET_SH4A_ARCH"
+{
+  return "\r   mov     #-4,%3"                 "\n"
+        "      and     %1,%3"                  "\n"
+        "      xor     %3,%1"                  "\n"
+        "      add     r15,%1"                 "\n"
+        "      add     #-4,%1"                 "\n"
+        "0:    movli.l @%3,r0"                 "\n"
+        "      mov.l   r0,@-r15"               "\n"
+        "      mov.<i124suffix>        @%1,r0" "\n"
+        "      mov     r0,%0"                  "\n"
+        "      <fetchop_name>  %2,r0"          "\n"
+        "      mov.<i124suffix>        r0,@%1" "\n"
+        "      mov.l   @r15+,r0"               "\n"
+        "      movco.l r0,@%3"                 "\n"
+        "      bf      0b";
+}
+  [(set_attr "length" "28")])
+
 (define_insn "atomic_fetch_<fetchop_name><mode>_soft"
   [(set (match_operand:I124 0 "register_operand" "=&u")
        (mem:I124 (match_operand:SI 1 "register_operand" "u")))
    (clobber (reg:SI R1_REG))]
   "TARGET_SOFT_ATOMIC && !TARGET_SHMEDIA"
 {
-  return "mova 1f,r0"                          "\n"
+  return "\r   mova    1f,r0"                  "\n"
         "      .align 2"                       "\n"
         "      mov     r15,r1"                 "\n"
         "      mov     #(0f-1f),r15"           "\n"
    (set (match_dup 1)
        (unspec:I124
          [(not:I124 (and:I124 (match_dup 1)
-            (match_operand:I124 2 "register_operand" "")))]
+                    (match_operand:I124 2 "atomic_logical_operand" "")))]
          UNSPEC_ATOMIC))
    (match_operand:SI 3 "const_int_operand" "")]
-  "TARGET_SOFT_ATOMIC && !TARGET_SHMEDIA"
+  "TARGET_ANY_ATOMIC && !TARGET_SHMEDIA"
 {
-  rtx addr;
+  rtx addr = force_reg (Pmode, XEXP (operands[1], 0));
+  rtx atomic_insn;
+
+  if (TARGET_HARD_ATOMIC || (TARGET_SH4A_ARCH && <MODE>mode == SImode))
+    atomic_insn = gen_atomic_fetch_nand<mode>_hard (operands[0], addr,
+                                                   operands[2]);
+  else
+    atomic_insn = gen_atomic_fetch_nand<mode>_soft (operands[0], addr,
+                                                   operands[2]);
+
+  emit_insn (atomic_insn);
 
-  addr = force_reg (Pmode, XEXP (operands[1], 0));
-  emit_insn (gen_atomic_fetch_nand<mode>_soft
-            (operands[0], addr, operands[2]));
   if (<MODE>mode == QImode)
     emit_insn (gen_zero_extendqisi2 (gen_lowpart (SImode, operands[0]),
                                     operands[0]));
   DONE;
 })
 
+(define_insn "atomic_fetch_nandsi_hard"
+  [(set (match_operand:SI 0 "register_operand" "=&r")
+       (mem:SI (match_operand:SI 1 "register_operand" "r")))
+   (set (mem:SI (match_dup 1))
+       (unspec:SI
+         [(not:SI (and:SI (mem:SI (match_dup 1))
+                  (match_operand:SI 2 "logical_operand" "rK08")))]
+         UNSPEC_ATOMIC))
+   (clobber (reg:SI R0_REG))]
+  "TARGET_ANY_ATOMIC && TARGET_SH4A_ARCH"
+{
+  return "\r0: movli.l @%1,r0"         "\n"
+        "      mov     r0,%0"          "\n"
+        "      and     %2,r0"          "\n"
+        "      not     r0,r0"          "\n"
+        "      movco.l r0,@%1"         "\n"
+        "      bf      0b";
+}
+  [(set_attr "length" "12")])
+
+(define_insn "atomic_fetch_nand<mode>_hard"
+  [(set (match_operand:I12 0 "register_operand" "=&r")
+       (mem:I12 (match_operand:SI 1 "register_operand" "r")))
+   (set (mem:I12 (match_dup 1))
+       (unspec:I12
+         [(not:I12 (and:I12 (mem:I12 (match_dup 1))
+                   (match_operand:I12 2 "logical_operand" "rK08")))]
+         UNSPEC_ATOMIC))
+   (clobber (reg:SI R0_REG))
+   (clobber (match_scratch:SI 3 "=&r"))
+   (clobber (match_scratch:SI 4 "=1"))]
+  "TARGET_HARD_ATOMIC && TARGET_SH4A_ARCH"
+{
+  return "\r   mov     #-4,%3"                 "\n"
+        "      and     %1,%3"                  "\n"
+        "      xor     %3,%1"                  "\n"
+        "      add     r15,%1"                 "\n"
+        "      add     #-4,%1"                 "\n"
+        "0:    movli.l @%3,r0"                 "\n"
+        "      mov.l   r0,@-r15"               "\n"
+        "      mov.<i124suffix>        @%1,r0" "\n"
+        "      mov     r0,%0"                  "\n"
+        "      and     %2,r0"                  "\n"
+        "      not     r0,r0"                  "\n"
+        "      mov.<i124suffix>        r0,@%1" "\n"
+        "      mov.l   @r15+,r0"               "\n"
+        "      movco.l r0,@%3"                 "\n"
+        "      bf      0b";
+}
+  [(set_attr "length" "30")])
+
 (define_insn "atomic_fetch_nand<mode>_soft"
   [(set (match_operand:I124 0 "register_operand" "=&u")
        (mem:I124 (match_operand:SI 1 "register_operand" "u")))
    (clobber (reg:SI R1_REG))]
   "TARGET_SOFT_ATOMIC && !TARGET_SHMEDIA"
 {
-  return "mova 1f,r0"                          "\n"
+  return "\r   mova    1f,r0"                  "\n"
         "      mov     r15,r1"                 "\n"
         "      .align 2"                       "\n"
         "      mov     #(0f-1f),r15"           "\n"
 }
   [(set_attr "length" "20")])
 
+;;------------------------------------------------------------------------------
+;; read - add|sub|or|and|xor|nand - write - return new value
+
 (define_expand "atomic_<fetchop_name>_fetch<mode>"
   [(set (match_operand:I124 0 "register_operand" "")
        (FETCHOP:I124
          (match_operand:I124 1 "memory_operand" "")
-         (match_operand:I124 2 "register_operand" "")))
+         (match_operand:I124 2 "<fetchop_predicate>" "")))
    (set (match_dup 1)
        (unspec:I124
          [(FETCHOP:I124 (match_dup 1) (match_dup 2))]
          UNSPEC_ATOMIC))
    (match_operand:SI 3 "const_int_operand" "")]
-  "TARGET_SOFT_ATOMIC && !TARGET_SHMEDIA"
+  "TARGET_ANY_ATOMIC && !TARGET_SHMEDIA"
 {
-  rtx addr;
+  rtx addr = force_reg (Pmode, XEXP (operands[1], 0));
+  rtx atomic_insn;
+
+  if (TARGET_HARD_ATOMIC || (TARGET_SH4A_ARCH && <MODE>mode == SImode))
+    atomic_insn = gen_atomic_<fetchop_name>_fetch<mode>_hard (operands[0], addr,
+                                                             operands[2]);
+  else
+    atomic_insn = gen_atomic_<fetchop_name>_fetch<mode>_soft (operands[0], addr,
+                                                             operands[2]);
+  emit_insn (atomic_insn);
 
-  addr = force_reg (Pmode, XEXP (operands[1], 0));
-  emit_insn (gen_atomic_<fetchop_name>_fetch<mode>_soft
-            (operands[0], addr, operands[2]));
   if (<MODE>mode == QImode)
     emit_insn (gen_zero_extendqisi2 (gen_lowpart (SImode, operands[0]),
                                     operands[0]));
   DONE;
 })
 
+(define_insn "atomic_<fetchop_name>_fetchsi_hard"
+  [(set (match_operand:SI 0 "register_operand" "=&z")
+       (FETCHOP:SI
+         (mem:SI (match_operand:SI 1 "register_operand" "r"))
+         (match_operand:SI 2 "<fetchop_predicate>" "<fetchop_constraint>")))
+   (set (mem:SI (match_dup 1))
+       (unspec:SI
+         [(FETCHOP:SI (mem:SI (match_dup 1)) (match_dup 2))]
+         UNSPEC_ATOMIC))]
+  "TARGET_ANY_ATOMIC && TARGET_SH4A_ARCH"
+{
+  return "\r0: movli.l @%1,%0"         "\n"
+        "      <fetchop_name>  %2,%0"  "\n"
+        "      movco.l %0,@%1"         "\n"
+        "      bf      0b";
+}
+  [(set_attr "length" "8")])
+
+(define_insn "atomic_<fetchop_name>_fetch<mode>_hard"
+  [(set (match_operand:I12 0 "register_operand" "=&r")
+       (FETCHOP:I12
+         (mem:I12 (match_operand:SI 1 "register_operand" "r"))
+         (match_operand:I12 2 "<fetchop_predicate>" "<fetchop_constraint>")))
+   (set (mem:I12 (match_dup 1))
+       (unspec:I12
+         [(FETCHOP:I12 (mem:I12 (match_dup 1)) (match_dup 2))]
+         UNSPEC_ATOMIC))
+
+   (clobber (reg:SI R0_REG))
+   (clobber (match_scratch:SI 3 "=&r"))
+   (clobber (match_scratch:SI 4 "=1"))]
+  "TARGET_HARD_ATOMIC && TARGET_SH4A_ARCH"
+{
+  return "\r   mov     #-4,%3"                 "\n"
+        "      and     %1,%3"                  "\n"
+        "      xor     %3,%1"                  "\n"
+        "      add     r15,%1"                 "\n"
+        "      add     #-4,%1"                 "\n"
+        "0:    movli.l @%3,r0"                 "\n"
+        "      mov.l   r0,@-r15"               "\n"
+        "      mov.<i124suffix>        @%1,r0" "\n"
+        "      <fetchop_name>  %2,r0"          "\n"
+        "      mov.<i124suffix>        r0,@%1" "\n"
+        "      mov     r0,%0"                  "\n"
+        "      mov.l   @r15+,r0"               "\n"
+        "      movco.l r0,@%3"                 "\n"
+        "      bf      0b";
+}
+  [(set_attr "length" "28")])
+
 (define_insn "atomic_<fetchop_name>_fetch<mode>_soft"
   [(set (match_operand:I124 0 "register_operand" "=&u")
        (FETCHOP:I124
    (clobber (reg:SI R1_REG))]
   "TARGET_SOFT_ATOMIC && !TARGET_SHMEDIA"
 {
-  return "mova 1f,r0"                          "\n"
+  return "\r   mova    1f,r0"                  "\n"
         "      mov     r15,r1"                 "\n"
         "      .align 2"                       "\n"
         "      mov     #(0f-1f),r15"           "\n"
   [(set (match_operand:I124 0 "register_operand" "")
        (not:I124 (and:I124
          (match_operand:I124 1 "memory_operand" "")
-         (match_operand:I124 2 "register_operand" ""))))
+         (match_operand:I124 2 "atomic_logical_operand" ""))))
    (set (match_dup 1)
        (unspec:I124
          [(not:I124 (and:I124 (match_dup 1) (match_dup 2)))]
          UNSPEC_ATOMIC))
    (match_operand:SI 3 "const_int_operand" "")]
-  "TARGET_SOFT_ATOMIC && !TARGET_SHMEDIA"
+  "TARGET_ANY_ATOMIC && !TARGET_SHMEDIA"
 {
-  rtx addr;
+  rtx addr = force_reg (Pmode, XEXP (operands[1], 0));
+  rtx atomic_insn;
+
+  if (TARGET_HARD_ATOMIC || (TARGET_SH4A_ARCH && <MODE>mode == SImode))
+    atomic_insn = gen_atomic_nand_fetch<mode>_hard (operands[0], addr,
+                                                   operands[2]);
+  else
+    atomic_insn = gen_atomic_nand_fetch<mode>_soft (operands[0], addr,
+                                                   operands[2]);
+  emit_insn (atomic_insn);
 
-  addr = force_reg (Pmode, XEXP (operands[1], 0));
-  emit_insn (gen_atomic_nand_fetch<mode>_soft
-            (operands[0], addr, operands[2]));
   if (<MODE>mode == QImode)
     emit_insn (gen_zero_extendqisi2 (gen_lowpart (SImode, operands[0]),
                                     operands[0]));
   DONE;
 })
 
+(define_insn "atomic_nand_fetchsi_hard"
+  [(set (match_operand:SI 0 "register_operand" "=&z")
+       (not:SI (and:SI (mem:SI (match_operand:SI 1 "register_operand" "r"))
+                       (match_operand:SI 2 "logical_operand" "rK08"))))
+   (set (mem:SI (match_dup 1))
+       (unspec:SI
+         [(not:SI (and:SI (mem:SI (match_dup 1)) (match_dup 2)))]
+         UNSPEC_ATOMIC))]
+  "TARGET_ANY_ATOMIC && TARGET_SH4A_ARCH"
+{
+  return "\r0: movli.l @%1,%0"         "\n"
+        "      and     %2,%0"          "\n"
+        "      not     %0,%0"          "\n"
+        "      movco.l %0,@%1"         "\n"
+        "      bf      0b";
+}
+  [(set_attr "length" "10")])
+
+(define_insn "atomic_nand_fetch<mode>_hard"
+  [(set (match_operand:I12 0 "register_operand" "=&r")
+       (not:I12 (and:I12 (mem:I12 (match_operand:SI 1 "register_operand" "r"))
+                         (match_operand:I12 2 "logical_operand" "rK08"))))
+   (set (mem:I12 (match_dup 1))
+       (unspec:I12
+         [(not:I12 (and:I12 (mem:I12 (match_dup 1)) (match_dup 2)))]
+         UNSPEC_ATOMIC))
+   (clobber (reg:SI R0_REG))
+   (clobber (match_scratch:SI 3 "=&r"))
+   (clobber (match_scratch:SI 4 "=1"))]
+  "TARGET_HARD_ATOMIC && TARGET_SH4A_ARCH"
+{
+  return "\r   mov     #-4,%3"                 "\n"
+        "      and     %1,%3"                  "\n"
+        "      xor     %3,%1"                  "\n"
+        "      add     r15,%1"                 "\n"
+        "      add     #-4,%1"                 "\n"
+        "0:    movli.l @%3,r0"                 "\n"
+        "      mov.l   r0,@-r15"               "\n"
+        "      mov.<i124suffix>        @%1,r0" "\n"
+        "      and     %2,r0"                  "\n"
+        "      not     r0,%0"                  "\n"
+        "      mov.<i124suffix>        %0,@%1" "\n"
+        "      mov.l   @r15+,r0"               "\n"
+        "      movco.l r0,@%3"                 "\n"
+        "      bf      0b";
+}
+  [(set_attr "length" "28")])
+
 (define_insn "atomic_nand_fetch<mode>_soft"
   [(set (match_operand:I124 0 "register_operand" "=&u")
        (not:I124 (and:I124
    (clobber (reg:SI R1_REG))]
   "TARGET_SOFT_ATOMIC && !TARGET_SHMEDIA"
 {
-  return "mova 1f,r0"                          "\n"
+  return "\r   mova    1f,r0"                  "\n"
         "      .align 2"                       "\n"
         "      mov     r15,r1"                 "\n"
         "      mov     #(0f-1f),r15"           "\n"
 }
   [(set_attr "length" "18")])
 
+;;------------------------------------------------------------------------------
+;; read - test against zero - or with 0x80 - write - return test result
+
 (define_expand "atomic_test_and_set"
   [(match_operand:SI 0 "register_operand" "")          ;; bool result output
    (match_operand:QI 1 "memory_operand" "")            ;; memory
    (match_operand:SI 2 "const_int_operand" "")]                ;; model
-  "(TARGET_SOFT_ATOMIC || TARGET_ENABLE_TAS) && !TARGET_SHMEDIA"
+  "(TARGET_ANY_ATOMIC || TARGET_ENABLE_TAS) && !TARGET_SHMEDIA"
 {
   rtx addr = force_reg (Pmode, XEXP (operands[1], 0));
 
     emit_insn (gen_tasb (addr));
   else
     {
-      rtx val;
-
-      val = gen_int_mode (targetm.atomic_test_and_set_trueval, QImode);
+      rtx val = gen_int_mode (targetm.atomic_test_and_set_trueval, QImode);
       val = force_reg (QImode, val);
-      emit_insn (gen_atomic_test_and_set_soft (addr, val));
+
+      if (TARGET_HARD_ATOMIC)
+         emit_insn (gen_atomic_test_and_set_hard (addr, val));
+      else
+         emit_insn (gen_atomic_test_and_set_soft (addr, val));
     }
 
   /* The result of the test op is the inverse of what we are
    (clobber (reg:SI R1_REG))]
   "TARGET_SOFT_ATOMIC && !TARGET_ENABLE_TAS && !TARGET_SHMEDIA"
 {
-  return "mova 1f,r0"                  "\n"
+  return "\r   mova    1f,r0"          "\n"
         "      .align 2"               "\n"
         "      mov     r15,r1"         "\n"
         "      mov     #(0f-1f),r15"   "\n"
 }
   [(set_attr "length" "16")])
 
+(define_insn "atomic_test_and_set_hard"
+  [(set (reg:SI T_REG)
+       (eq:SI (mem:QI (match_operand:SI 0 "register_operand" "r"))
+              (const_int 0)))
+   (set (mem:QI (match_dup 0))
+       (unspec:QI [(match_operand:QI 1 "register_operand" "r")] UNSPEC_ATOMIC))
+   (clobber (reg:SI R0_REG))
+   (clobber (match_scratch:SI 2 "=&r"))
+   (clobber (match_scratch:SI 3 "=&r"))
+   (clobber (match_scratch:SI 4 "=0"))]
+  "TARGET_HARD_ATOMIC && !TARGET_ENABLE_TAS && TARGET_SH4A_ARCH"
+{
+  return "\r   mov     #-4,%2"         "\n"
+        "      and     %0,%2"          "\n"
+        "      xor     %2,%0"          "\n"
+        "      add     r15,%0"         "\n"
+        "      add     #-4,%0"         "\n"
+        "0:    movli.l @%2,r0"         "\n"
+        "      mov.l   r0,@-r15"       "\n"
+        "      mov.b   @%0,%3"         "\n"
+        "      mov.b   %1,@%0"         "\n"
+        "      mov.l   @r15+,r0"       "\n"
+        "      movco.l r0,@%2"         "\n"
+        "      bf      0b"             "\n"
+        "      tst     %3,%3";
+}
+  [(set_attr "length" "26")])
+
index e9b3b4328c2f6cb96d088f22e0147497fdae9ae3..efab5168d4202e659d511ef847d4b62f0b08ffa9 100644 (file)
@@ -884,7 +884,7 @@ See RS/6000 and PowerPC Options.
 -mprefergot  -musermode -multcost=@var{number} -mdiv=@var{strategy} @gol
 -mdivsi3_libfunc=@var{name} -mfixed-range=@var{register-range} @gol
 -mindexed-addressing -mgettrcost=@var{number} -mpt-fixed @gol
--maccumulate-outgoing-args -minvalid-symbols -msoft-atomic @gol
+-maccumulate-outgoing-args -minvalid-symbols -msoft-atomic -mhard-atomic @gol
 -mbranch-cost=@var{num} -mcbranchdi -mcmpeqdi -mfused-madd -mpretend-cmove @gol
 -menable-tas}
 
@@ -18054,10 +18054,22 @@ which is incompatible with the SH ABI@.
 Generate GNU/Linux compatible gUSA software atomic sequences for the atomic
 built-in functions.  The generated atomic sequences require support from the 
 interrupt / exception handling code of the system and are only suitable for
-single-core systems.  They will not perform correctly on multi-core systems.
+single-core systems.  They will not operate correctly on multi-core systems.
 This option is enabled by default when the target is @code{sh-*-linux*}.
+When the target is SH4A, this option will also partially utilize the hardware
+atomic instructions @code{movli.l} and @code{movco.l} to create more
+efficient code.
 For details on the atomic built-in functions see @ref{__atomic Builtins}.
 
+@item -mhard-atomic
+@opindex hard-atomic
+Generate hardware atomic sequences for the atomic built-in functions.  This
+is only available on SH4A and is suitable for multi-core systems.  Code
+compiled with this option will also be compatible with gUSA aware
+interrupt / exception handling systems.  In contrast to the
+@option{-msoft-atomic} option this will only use the instructions
+@code{movli.l} and @code{movco.l} to create atomic sequences.
+
 @item -menable-tas
 @opindex menable-tas
 Generate the @code{tas.b} opcode for @code{__atomic_test_and_set}.