On Sat, 2012-03-03 at 10:31 -0800, Richard Henderson wrote:
> On 03/02/2012 10:11 AM, Richard Henderson wrote:
> > I'm in the process of sanity testing this on x86_64 with trueval set to
> > 0x80.
> > Jakub, ok for 4.7 branch if it passes?
> >
> > * optabs.c (expand_atomic_test_and_set): Honor
> > atomic_test_and_set_trueval even when atomic_test_and_set
> > optab is not in use.
>
> I've committed this patch to mainline. I still think it ought to
> go onto the 4.7 branch...
>
Attached is a slightly modified version of the patch from
http://gcc.gnu.org/ml/gcc-patches/2012-03/msg00085.html
I have removed the signed char weirdo and adjusted the comment above
TARGET_ATOMIC_TEST_AND_SET_TRUEVAL accordingly.
Tested by compiling some test functions that use __atomic_test_and_set /
__GCC_ATOMIC_TEST_AND_SET_TRUEVAL with various SH atomic option
combinations and looking at the output asm.
OK to apply to trunk?
Richard, could you also please take the
TARGET_ATOMIC_TEST_AND_SET_TRUEVAL hunk from this patch for the 4.7
branch?
Cheers,
Oleg
2012-03-04 Oleg Endo <[email protected]>
* config/sh/sh.h (TARGET_ATOMIC_TEST_AND_SET_TRUEVAL): New hook.
* config/sh/sync.md (atomic_test_and_set): New expander.
(tasb, atomic_test_and_set_soft): New insns.
* config/sh/sh.opt (menable-tas): New option.
* doc/invoke.texi (SH Options): Document it.
Index: gcc/doc/invoke.texi
===================================================================
--- gcc/doc/invoke.texi (revision 184877)
+++ gcc/doc/invoke.texi (working copy)
@@ -887,7 +887,8 @@
-mdivsi3_libfunc=@var{name} -mfixed-range=@var{register-range} @gol
-madjust-unroll -mindexed-addressing -mgettrcost=@var{number} -mpt-fixed @gol
-maccumulate-outgoing-args -minvalid-symbols -msoft-atomic @gol
--mbranch-cost=@var{num} -mcbranchdi -mcmpeqdi -mfused-madd -mpretend-cmove}
+-mbranch-cost=@var{num} -mcbranchdi -mcmpeqdi -mfused-madd -mpretend-cmove @gol
+-menable-tas}
@emph{Solaris 2 Options}
@gccoptlist{-mimpure-text -mno-impure-text @gol
@@ -17823,6 +17824,15 @@
This option is enabled by default when the target is @code{sh-*-linux*}.
For details on the atomic built-in functions see @ref{__atomic Builtins}.
+@item -menable-tas
+@opindex menable-tas
+Generate the @code{tas.b} opcode for @code{__atomic_test_and_set}.
+Notice that depending on the particular hardware and software configuration
+this can degrade overall performance due to the operand cache line flushes
+that are implied by the @code{tas.b} instruction. On multi-core SH4A
+processors the @code{tas.b} instruction must be used with caution since it
+can result in data corruption for certain cache configurations.
+
@item -mspace
@opindex mspace
Optimize for space instead of speed. Implied by @option{-Os}.
Index: gcc/config/sh/sh.h
===================================================================
--- gcc/config/sh/sh.h (revision 184877)
+++ gcc/config/sh/sh.h (working copy)
@@ -2473,4 +2473,10 @@
/* FIXME: middle-end support for highpart optimizations is missing. */
#define high_life_started reload_in_progress
+/* The tas.b instruction sets the 7th bit in the byte, i.e. 0x80.
+ This value is used by optabs.c atomic op expansion code as well as in
+ sync.md. */
+#undef TARGET_ATOMIC_TEST_AND_SET_TRUEVAL
+#define TARGET_ATOMIC_TEST_AND_SET_TRUEVAL 0x80
+
#endif /* ! GCC_SH_H */
Index: gcc/config/sh/sync.md
===================================================================
--- gcc/config/sh/sync.md (revision 184877)
+++ gcc/config/sh/sync.md (working copy)
@@ -404,3 +404,61 @@
"1: mov r1,r15";
}
[(set_attr "length" "18")])
+
+(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"
+{
+ rtx addr = force_reg (Pmode, XEXP (operands[1], 0));
+
+ if (TARGET_ENABLE_TAS)
+ emit_insn (gen_tasb (addr));
+ else
+ {
+ rtx val = force_reg (QImode,
+ gen_int_mode (TARGET_ATOMIC_TEST_AND_SET_TRUEVAL,
+ QImode));
+ emit_insn (gen_atomic_test_and_set_soft (addr, val));
+ }
+
+ /* The result of the test op is the inverse of what we are
+ supposed to return. Thus invert the T bit. The inversion will be
+ potentially optimized away and integrated into surrounding code. */
+ emit_insn (gen_movnegt (operands[0]));
+ DONE;
+})
+
+(define_insn "tasb"
+ [(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 [(const_int 128)] UNSPEC_ATOMIC))]
+ "TARGET_ENABLE_TAS && !TARGET_SHMEDIA"
+ "tas.b @%0"
+ [(set_attr "insn_class" "co_group")])
+
+(define_insn "atomic_test_and_set_soft"
+ [(set (reg:SI T_REG)
+ (eq:SI (mem:QI (match_operand:SI 0 "register_operand" "u"))
+ (const_int 0)))
+ (set (mem:QI (match_dup 0))
+ (unspec:QI [(match_operand:QI 1 "register_operand" "u")] UNSPEC_ATOMIC))
+ (clobber (match_scratch:QI 2 "=&u"))
+ (clobber (reg:SI R0_REG))
+ (clobber (reg:SI R1_REG))]
+ "TARGET_SOFT_ATOMIC && !TARGET_ENABLE_TAS && !TARGET_SHMEDIA"
+{
+ return "mova 1f,r0" "\n"
+ " .align 2" "\n"
+ " mov r15,r1" "\n"
+ " mov #(0f-1f),r15" "\n"
+ "0: mov.b @%0,%2" "\n"
+ " mov.b %1,@%0" "\n"
+ "1: mov r1,r15" "\n"
+ " tst %2,%2";
+}
+ [(set_attr "length" "16")])
+
Index: gcc/config/sh/sh.opt
===================================================================
--- gcc/config/sh/sh.opt (revision 184877)
+++ gcc/config/sh/sh.opt (working copy)
@@ -323,6 +323,10 @@
Target Report Mask(SOFT_ATOMIC)
Use software atomic sequences supported by kernel
+menable-tas
+Target Report RejectNegative Var(TARGET_ENABLE_TAS)
+Use tas.b instruction for __atomic_test_and_set
+
mspace
Target RejectNegative Alias(Os)
Deprecated. Use -Os instead
