diff options
| author | John David Anglin <danglin@gcc.gnu.org> | 2023-01-05 18:15:18 +0000 |
|---|---|---|
| committer | John David Anglin <danglin@gcc.gnu.org> | 2023-01-05 18:15:18 +0000 |
| commit | 9807c31af91326988d083436483f9577296e1f9a (patch) | |
| tree | 309a400a027d826d8a5c5769255e5d50eec47b57 | |
| parent | 4413365616e8c6024d1ff4e23309e5012ee33b9f (diff) | |
| download | gcc-9807c31af91326988d083436483f9577296e1f9a.zip gcc-9807c31af91326988d083436483f9577296e1f9a.tar.gz gcc-9807c31af91326988d083436483f9577296e1f9a.tar.bz2 | |
Fix atomic operations on PA-RISC 2.0 processors.
PA-RISC 2.0 supports out-of-order execution for loads and stores.
Thus, we need to synchonize memory accesses.
This change revises the lock releases in __exchange_and_add and
__atomic_add to use an ordered store with release semantics. We
also use an ordered load in the inner spin loop.
We use the "ldcw,co" instruction instead of "ldcw" when compiled
for PA 2.0. Most PA 2.0 processors are coherent and can execute
the ldcw instruction in cache for improved performance.
Finally, the inner spin loop is revised to immediately branch to
the ldcw instruction when it detects the lock is free.
2023-01-05 John David Anglin <danglin@gcc.gnu.org>
libstdc++-v3/ChangeLog:
* config/cpu/hppa/atomicity.h (_PA_LDCW_INSN): Define.
(__exchange_and_add): Use _PA_LDCW_INSN. Use ordered store for
lock release. Revise loop.
(__atomic_add): Likewise.
| -rw-r--r-- | libstdc++-v3/config/cpu/hppa/atomicity.h | 29 |
1 files changed, 19 insertions, 10 deletions
diff --git a/libstdc++-v3/config/cpu/hppa/atomicity.h b/libstdc++-v3/config/cpu/hppa/atomicity.h index bb997e7..6580735 100644 --- a/libstdc++-v3/config/cpu/hppa/atomicity.h +++ b/libstdc++-v3/config/cpu/hppa/atomicity.h @@ -25,6 +25,15 @@ #include <bits/c++config.h> #include <ext/atomicity.h> +/* Perform ldcw operation in cache when possible. */ +#ifndef _PA_LDCW_INSN +# ifdef _PA_RISC2_0 +# define _PA_LDCW_INSN "ldcw,co" +# else +# define _PA_LDCW_INSN "ldcw" +# endif +#endif + namespace __gnu_cxx _GLIBCXX_VISIBILITY(default) { _GLIBCXX_BEGIN_NAMESPACE_VERSION @@ -51,19 +60,19 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION int tmp; volatile int& lock = _Atomicity_lock<0>::_S_atomicity_lock; - __asm__ __volatile__ ("ldcw 0(%1),%0\n\t" + __asm__ __volatile__ (_PA_LDCW_INSN " 0(%1),%0\n\t" "cmpib,<>,n 0,%0,.+20\n\t" - "ldw 0(%1),%0\n\t" - "cmpib,= 0,%0,.-4\n\t" + "ldw,ma 0(%1),%0\n\t" + "cmpib,<> 0,%0,.-12\n\t" "nop\n\t" - "b,n .-20" + "b,n .-12" : "=&r" (tmp) : "r" (&lock) : "memory"); result = *__mem; *__mem = result + __val; - __asm__ __volatile__ ("stw %1,0(%0)" + __asm__ __volatile__ ("stw,ma %1,0(%0)" : : "r" (&lock), "r" (tmp) : "memory"); return result; } @@ -75,18 +84,18 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION int tmp; volatile int& lock = _Atomicity_lock<0>::_S_atomicity_lock; - __asm__ __volatile__ ("ldcw 0(%1),%0\n\t" + __asm__ __volatile__ (_PA_LDCW_INSN " 0(%1),%0\n\t" "cmpib,<>,n 0,%0,.+20\n\t" - "ldw 0(%1),%0\n\t" - "cmpib,= 0,%0,.-4\n\t" + "ldw,ma 0(%1),%0\n\t" + "cmpib,<> 0,%0,.-12\n\t" "nop\n\t" - "b,n .-20" + "b,n .-12" : "=&r" (tmp) : "r" (&lock) : "memory"); *__mem += __val; - __asm__ __volatile__ ("stw %1,0(%0)" + __asm__ __volatile__ ("stw,ma %1,0(%0)" : : "r" (&lock), "r" (tmp) : "memory"); } |