From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: by sourceware.org (Postfix, from userid 1118) id B59023858D28; Thu, 5 Jan 2023 18:16:39 +0000 (GMT) DKIM-Filter: OpenDKIM Filter v2.11.0 sourceware.org B59023858D28 DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gcc.gnu.org; s=default; t=1672942599; bh=q/zT/CVNz6Ym740pCC+2Pz3Jo9HujFcm3UwnHBQXA3w=; h=From:To:Subject:Date:From; b=FLNICyGGEQcLU/GKP4SEUCUbexDJ3MCMKhjT+vqZDlbxYdLvJxhqH0M3zMXlUAVEa 24MasIPaUIlHnjoiq/qT2w7BKQpgAfhjMCrQLmJaTBqIf3+pgBC4PBq1CJQj6KA7fg uLyKQyrhkCwUa354n7F5b19W1wXAMbGrnqi7TyFI= MIME-Version: 1.0 Content-Transfer-Encoding: 7bit Content-Type: text/plain; charset="utf-8" From: John David Anglin To: gcc-cvs@gcc.gnu.org, libstdc++-cvs@gcc.gnu.org Subject: [gcc r13-5036] Fix atomic operations on PA-RISC 2.0 processors. X-Act-Checkin: gcc X-Git-Author: John David Anglin X-Git-Refname: refs/heads/master X-Git-Oldrev: 4413365616e8c6024d1ff4e23309e5012ee33b9f X-Git-Newrev: 9807c31af91326988d083436483f9577296e1f9a Message-Id: <20230105181639.B59023858D28@sourceware.org> Date: Thu, 5 Jan 2023 18:16:39 +0000 (GMT) List-Id: https://gcc.gnu.org/g:9807c31af91326988d083436483f9577296e1f9a commit r13-5036-g9807c31af91326988d083436483f9577296e1f9a Author: John David Anglin Date: Thu Jan 5 18:15:18 2023 +0000 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 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. Diff: --- libstdc++-v3/config/cpu/hppa/atomicity.h | 29 +++++++++++++++++++---------- 1 file changed, 19 insertions(+), 10 deletions(-) diff --git a/libstdc++-v3/config/cpu/hppa/atomicity.h b/libstdc++-v3/config/cpu/hppa/atomicity.h index bb997e70c1d..658073537a4 100644 --- a/libstdc++-v3/config/cpu/hppa/atomicity.h +++ b/libstdc++-v3/config/cpu/hppa/atomicity.h @@ -25,6 +25,15 @@ #include #include +/* 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"); }