12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232 |
- /*
- * Distributed under the Boost Software License, Version 1.0.
- * (See accompanying file LICENSE_1_0.txt or copy at
- * http://www.boost.org/LICENSE_1_0.txt)
- *
- * Copyright (c) 2009 Helge Bahmann
- * Copyright (c) 2013 Tim Blechmann
- * Copyright (c) 2014 Andrey Semashev
- */
- /*!
- * \file atomic/detail/ops_gcc_ppc.hpp
- *
- * This header contains implementation of the \c operations template.
- */
- #ifndef BOOST_ATOMIC_DETAIL_OPS_GCC_PPC_HPP_INCLUDED_
- #define BOOST_ATOMIC_DETAIL_OPS_GCC_PPC_HPP_INCLUDED_
- #include <cstddef>
- #include <boost/memory_order.hpp>
- #include <boost/atomic/detail/config.hpp>
- #include <boost/atomic/detail/storage_type.hpp>
- #include <boost/atomic/detail/operations_fwd.hpp>
- #include <boost/atomic/detail/ops_gcc_ppc_common.hpp>
- #include <boost/atomic/capabilities.hpp>
- #ifdef BOOST_HAS_PRAGMA_ONCE
- #pragma once
- #endif
- namespace boost {
- namespace atomics {
- namespace detail {
- // The implementation below uses information from this document:
- // http://www.rdrop.com/users/paulmck/scalability/paper/N2745r.2010.02.19a.html
- /*
- Refer to: Motorola: "Programming Environments Manual for 32-Bit
- Implementations of the PowerPC Architecture", Appendix E:
- "Synchronization Programming Examples" for an explanation of what is
- going on here (can be found on the web at various places by the
- name "MPCFPE32B.pdf", Google is your friend...)
- Most of the atomic operations map to instructions in a relatively
- straight-forward fashion, but "load"s may at first glance appear
- a bit strange as they map to:
- lwz %rX, addr
- cmpw %rX, %rX
- bne- 1f
- 1:
- That is, the CPU is forced to perform a branch that "formally" depends
- on the value retrieved from memory. This scheme has an overhead of
- about 1-2 clock cycles per load, but it allows to map "acquire" to
- the "isync" instruction instead of "sync" uniformly and for all type
- of atomic operations. Since "isync" has a cost of about 15 clock
- cycles, while "sync" hast a cost of about 50 clock cycles, the small
- penalty to atomic loads more than compensates for this.
- Byte- and halfword-sized atomic values are implemented in two ways.
- When 8 and 16-bit instructions are available (in Power8 and later),
- they are used. Otherwise operations are realized by encoding the
- value to be represented into a word, performing sign/zero extension
- as appropriate. This means that after add/sub operations the value
- needs fixing up to accurately preserve the wrap-around semantic of
- the smaller type. (Nothing special needs to be done for the bit-wise
- and the "exchange type" operators as the compiler already sees to
- it that values carried in registers are extended appropriately and
- everything falls into place naturally).
- The register constraint "b" instructs gcc to use any register
- except r0; this is sometimes required because the encoding for
- r0 is used to signify "constant zero" in a number of instructions,
- making r0 unusable in this place. For simplicity this constraint
- is used everywhere since I am to lazy to look this up on a
- per-instruction basis, and ppc has enough registers for this not
- to pose a problem.
- */
- template< bool Signed >
- struct operations< 4u, Signed > :
- public gcc_ppc_operations_base
- {
- typedef typename make_storage_type< 4u >::type storage_type;
- typedef typename make_storage_type< 4u >::aligned aligned_storage_type;
- static BOOST_CONSTEXPR_OR_CONST std::size_t storage_size = 4u;
- static BOOST_CONSTEXPR_OR_CONST bool is_signed = Signed;
- static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- fence_before(order);
- __asm__ __volatile__
- (
- "stw %1, %0\n\t"
- : "+m" (storage)
- : "r" (v)
- );
- }
- static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT
- {
- storage_type v;
- if (order == memory_order_seq_cst)
- __asm__ __volatile__ ("sync" ::: "memory");
- if ((static_cast< unsigned int >(order) & (static_cast< unsigned int >(memory_order_consume) | static_cast< unsigned int >(memory_order_acquire))) != 0u)
- {
- __asm__ __volatile__
- (
- "lwz %0, %1\n\t"
- "cmpw %0, %0\n\t"
- "bne- 1f\n\t"
- "1:\n\t"
- "isync\n\t"
- : "=&r" (v)
- : "m" (storage)
- : "cr0", "memory"
- );
- }
- else
- {
- __asm__ __volatile__
- (
- "lwz %0, %1\n\t"
- : "=&r" (v)
- : "m" (storage)
- );
- }
- return v;
- }
- static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lwarx %0,%y1\n\t"
- "stwcx. %2,%y1\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "+Z" (storage)
- : "b" (v)
- : "cr0"
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE bool compare_exchange_weak(
- storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT
- {
- int success;
- fence_before(success_order);
- __asm__ __volatile__
- (
- "li %1, 0\n\t"
- "lwarx %0,%y2\n\t"
- "cmpw %0, %3\n\t"
- "bne- 1f\n\t"
- "stwcx. %4,%y2\n\t"
- "bne- 1f\n\t"
- "li %1, 1\n\t"
- "1:\n\t"
- : "=&b" (expected), "=&b" (success), "+Z" (storage)
- : "b" (expected), "b" (desired)
- : "cr0"
- );
- if (success)
- fence_after(success_order);
- else
- fence_after(failure_order);
- return !!success;
- }
- static BOOST_FORCEINLINE bool compare_exchange_strong(
- storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT
- {
- int success;
- fence_before(success_order);
- __asm__ __volatile__
- (
- "li %1, 0\n\t"
- "0: lwarx %0,%y2\n\t"
- "cmpw %0, %3\n\t"
- "bne- 1f\n\t"
- "stwcx. %4,%y2\n\t"
- "bne- 0b\n\t"
- "li %1, 1\n\t"
- "1:\n\t"
- : "=&b" (expected), "=&b" (success), "+Z" (storage)
- : "b" (expected), "b" (desired)
- : "cr0"
- );
- if (success)
- fence_after(success_order);
- else
- fence_after(failure_order);
- return !!success;
- }
- static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lwarx %0,%y2\n\t"
- "add %1,%0,%3\n\t"
- "stwcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lwarx %0,%y2\n\t"
- "sub %1,%0,%3\n\t"
- "stwcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lwarx %0,%y2\n\t"
- "and %1,%0,%3\n\t"
- "stwcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lwarx %0,%y2\n\t"
- "or %1,%0,%3\n\t"
- "stwcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lwarx %0,%y2\n\t"
- "xor %1,%0,%3\n\t"
- "stwcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE bool test_and_set(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT
- {
- return !!exchange(storage, (storage_type)1, order);
- }
- static BOOST_FORCEINLINE void clear(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT
- {
- store(storage, 0, order);
- }
- };
- #if defined(BOOST_ATOMIC_DETAIL_PPC_HAS_LBARX_STBCX)
- template< bool Signed >
- struct operations< 1u, Signed > :
- public gcc_ppc_operations_base
- {
- typedef typename make_storage_type< 1u >::type storage_type;
- typedef typename make_storage_type< 1u >::aligned aligned_storage_type;
- static BOOST_CONSTEXPR_OR_CONST std::size_t storage_size = 1u;
- static BOOST_CONSTEXPR_OR_CONST bool is_signed = Signed;
- static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- fence_before(order);
- __asm__ __volatile__
- (
- "stb %1, %0\n\t"
- : "+m" (storage)
- : "r" (v)
- );
- }
- static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT
- {
- storage_type v;
- if (order == memory_order_seq_cst)
- __asm__ __volatile__ ("sync" ::: "memory");
- if ((static_cast< unsigned int >(order) & (static_cast< unsigned int >(memory_order_consume) | static_cast< unsigned int >(memory_order_acquire))) != 0u)
- {
- __asm__ __volatile__
- (
- "lbz %0, %1\n\t"
- "cmpw %0, %0\n\t"
- "bne- 1f\n\t"
- "1:\n\t"
- "isync\n\t"
- : "=&r" (v)
- : "m" (storage)
- : "cr0", "memory"
- );
- }
- else
- {
- __asm__ __volatile__
- (
- "lbz %0, %1\n\t"
- : "=&r" (v)
- : "m" (storage)
- );
- }
- return v;
- }
- static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lbarx %0,%y1\n\t"
- "stbcx. %2,%y1\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "+Z" (storage)
- : "b" (v)
- : "cr0"
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE bool compare_exchange_weak(
- storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT
- {
- int success;
- fence_before(success_order);
- __asm__ __volatile__
- (
- "li %1, 0\n\t"
- "lbarx %0,%y2\n\t"
- "cmpw %0, %3\n\t"
- "bne- 1f\n\t"
- "stbcx. %4,%y2\n\t"
- "bne- 1f\n\t"
- "li %1, 1\n\t"
- "1:\n\t"
- : "=&b" (expected), "=&b" (success), "+Z" (storage)
- : "b" (expected), "b" (desired)
- : "cr0"
- );
- if (success)
- fence_after(success_order);
- else
- fence_after(failure_order);
- return !!success;
- }
- static BOOST_FORCEINLINE bool compare_exchange_strong(
- storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT
- {
- int success;
- fence_before(success_order);
- __asm__ __volatile__
- (
- "li %1, 0\n\t"
- "0: lbarx %0,%y2\n\t"
- "cmpw %0, %3\n\t"
- "bne- 1f\n\t"
- "stbcx. %4,%y2\n\t"
- "bne- 0b\n\t"
- "li %1, 1\n\t"
- "1:\n\t"
- : "=&b" (expected), "=&b" (success), "+Z" (storage)
- : "b" (expected), "b" (desired)
- : "cr0"
- );
- if (success)
- fence_after(success_order);
- else
- fence_after(failure_order);
- return !!success;
- }
- static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lbarx %0,%y2\n\t"
- "add %1,%0,%3\n\t"
- "stbcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lbarx %0,%y2\n\t"
- "sub %1,%0,%3\n\t"
- "stbcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lbarx %0,%y2\n\t"
- "and %1,%0,%3\n\t"
- "stbcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lbarx %0,%y2\n\t"
- "or %1,%0,%3\n\t"
- "stbcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lbarx %0,%y2\n\t"
- "xor %1,%0,%3\n\t"
- "stbcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE bool test_and_set(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT
- {
- return !!exchange(storage, (storage_type)1, order);
- }
- static BOOST_FORCEINLINE void clear(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT
- {
- store(storage, 0, order);
- }
- };
- #else // defined(BOOST_ATOMIC_DETAIL_PPC_HAS_LBARX_STBCX)
- template< >
- struct operations< 1u, false > :
- public operations< 4u, false >
- {
- typedef operations< 4u, false > base_type;
- typedef base_type::storage_type storage_type;
- static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lwarx %0,%y2\n\t"
- "add %1,%0,%3\n\t"
- "rlwinm %1, %1, 0, 0xff\n\t"
- "stwcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lwarx %0,%y2\n\t"
- "sub %1,%0,%3\n\t"
- "rlwinm %1, %1, 0, 0xff\n\t"
- "stwcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- };
- template< >
- struct operations< 1u, true > :
- public operations< 4u, true >
- {
- typedef operations< 4u, true > base_type;
- typedef base_type::storage_type storage_type;
- static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lwarx %0,%y2\n\t"
- "add %1,%0,%3\n\t"
- "extsb %1, %1\n\t"
- "stwcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lwarx %0,%y2\n\t"
- "sub %1,%0,%3\n\t"
- "extsb %1, %1\n\t"
- "stwcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- };
- #endif // defined(BOOST_ATOMIC_DETAIL_PPC_HAS_LBARX_STBCX)
- #if defined(BOOST_ATOMIC_DETAIL_PPC_HAS_LHARX_STHCX)
- template< bool Signed >
- struct operations< 2u, Signed > :
- public gcc_ppc_operations_base
- {
- typedef typename make_storage_type< 2u >::type storage_type;
- typedef typename make_storage_type< 2u >::aligned aligned_storage_type;
- static BOOST_CONSTEXPR_OR_CONST std::size_t storage_size = 2u;
- static BOOST_CONSTEXPR_OR_CONST bool is_signed = Signed;
- static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- fence_before(order);
- __asm__ __volatile__
- (
- "sth %1, %0\n\t"
- : "+m" (storage)
- : "r" (v)
- );
- }
- static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT
- {
- storage_type v;
- if (order == memory_order_seq_cst)
- __asm__ __volatile__ ("sync" ::: "memory");
- if ((static_cast< unsigned int >(order) & (static_cast< unsigned int >(memory_order_consume) | static_cast< unsigned int >(memory_order_acquire))) != 0u)
- {
- __asm__ __volatile__
- (
- "lhz %0, %1\n\t"
- "cmpw %0, %0\n\t"
- "bne- 1f\n\t"
- "1:\n\t"
- "isync\n\t"
- : "=&r" (v)
- : "m" (storage)
- : "cr0", "memory"
- );
- }
- else
- {
- __asm__ __volatile__
- (
- "lhz %0, %1\n\t"
- : "=&r" (v)
- : "m" (storage)
- );
- }
- return v;
- }
- static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lharx %0,%y1\n\t"
- "sthcx. %2,%y1\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "+Z" (storage)
- : "b" (v)
- : "cr0"
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE bool compare_exchange_weak(
- storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT
- {
- int success;
- fence_before(success_order);
- __asm__ __volatile__
- (
- "li %1, 0\n\t"
- "lharx %0,%y2\n\t"
- "cmpw %0, %3\n\t"
- "bne- 1f\n\t"
- "sthcx. %4,%y2\n\t"
- "bne- 1f\n\t"
- "li %1, 1\n\t"
- "1:\n\t"
- : "=&b" (expected), "=&b" (success), "+Z" (storage)
- : "b" (expected), "b" (desired)
- : "cr0"
- );
- if (success)
- fence_after(success_order);
- else
- fence_after(failure_order);
- return !!success;
- }
- static BOOST_FORCEINLINE bool compare_exchange_strong(
- storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT
- {
- int success;
- fence_before(success_order);
- __asm__ __volatile__
- (
- "li %1, 0\n\t"
- "0: lharx %0,%y2\n\t"
- "cmpw %0, %3\n\t"
- "bne- 1f\n\t"
- "sthcx. %4,%y2\n\t"
- "bne- 0b\n\t"
- "li %1, 1\n\t"
- "1:\n\t"
- : "=&b" (expected), "=&b" (success), "+Z" (storage)
- : "b" (expected), "b" (desired)
- : "cr0"
- );
- if (success)
- fence_after(success_order);
- else
- fence_after(failure_order);
- return !!success;
- }
- static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lharx %0,%y2\n\t"
- "add %1,%0,%3\n\t"
- "sthcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lharx %0,%y2\n\t"
- "sub %1,%0,%3\n\t"
- "sthcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lharx %0,%y2\n\t"
- "and %1,%0,%3\n\t"
- "sthcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lharx %0,%y2\n\t"
- "or %1,%0,%3\n\t"
- "sthcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lharx %0,%y2\n\t"
- "xor %1,%0,%3\n\t"
- "sthcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE bool test_and_set(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT
- {
- return !!exchange(storage, (storage_type)1, order);
- }
- static BOOST_FORCEINLINE void clear(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT
- {
- store(storage, 0, order);
- }
- };
- #else // defined(BOOST_ATOMIC_DETAIL_PPC_HAS_LHARX_STHCX)
- template< >
- struct operations< 2u, false > :
- public operations< 4u, false >
- {
- typedef operations< 4u, false > base_type;
- typedef base_type::storage_type storage_type;
- static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lwarx %0,%y2\n\t"
- "add %1,%0,%3\n\t"
- "rlwinm %1, %1, 0, 0xffff\n\t"
- "stwcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lwarx %0,%y2\n\t"
- "sub %1,%0,%3\n\t"
- "rlwinm %1, %1, 0, 0xffff\n\t"
- "stwcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- };
- template< >
- struct operations< 2u, true > :
- public operations< 4u, true >
- {
- typedef operations< 4u, true > base_type;
- typedef base_type::storage_type storage_type;
- static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lwarx %0,%y2\n\t"
- "add %1,%0,%3\n\t"
- "extsh %1, %1\n\t"
- "stwcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "lwarx %0,%y2\n\t"
- "sub %1,%0,%3\n\t"
- "extsh %1, %1\n\t"
- "stwcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- };
- #endif // defined(BOOST_ATOMIC_DETAIL_PPC_HAS_LHARX_STHCX)
- #if defined(BOOST_ATOMIC_DETAIL_PPC_HAS_LDARX_STDCX)
- template< bool Signed >
- struct operations< 8u, Signed > :
- public gcc_ppc_operations_base
- {
- typedef typename make_storage_type< 8u >::type storage_type;
- typedef typename make_storage_type< 8u >::aligned aligned_storage_type;
- static BOOST_CONSTEXPR_OR_CONST std::size_t storage_size = 8u;
- static BOOST_CONSTEXPR_OR_CONST bool is_signed = Signed;
- static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- fence_before(order);
- __asm__ __volatile__
- (
- "std %1, %0\n\t"
- : "+m" (storage)
- : "r" (v)
- );
- }
- static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT
- {
- storage_type v;
- if (order == memory_order_seq_cst)
- __asm__ __volatile__ ("sync" ::: "memory");
- if ((static_cast< unsigned int >(order) & (static_cast< unsigned int >(memory_order_consume) | static_cast< unsigned int >(memory_order_acquire))) != 0u)
- {
- __asm__ __volatile__
- (
- "ld %0, %1\n\t"
- "cmpd %0, %0\n\t"
- "bne- 1f\n\t"
- "1:\n\t"
- "isync\n\t"
- : "=&b" (v)
- : "m" (storage)
- : "cr0", "memory"
- );
- }
- else
- {
- __asm__ __volatile__
- (
- "ld %0, %1\n\t"
- : "=&b" (v)
- : "m" (storage)
- );
- }
- return v;
- }
- static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "ldarx %0,%y1\n\t"
- "stdcx. %2,%y1\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "+Z" (storage)
- : "b" (v)
- : "cr0"
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE bool compare_exchange_weak(
- storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT
- {
- int success;
- fence_before(success_order);
- __asm__ __volatile__
- (
- "li %1, 0\n\t"
- "ldarx %0,%y2\n\t"
- "cmpd %0, %3\n\t"
- "bne- 1f\n\t"
- "stdcx. %4,%y2\n\t"
- "bne- 1f\n\t"
- "li %1, 1\n\t"
- "1:"
- : "=&b" (expected), "=&b" (success), "+Z" (storage)
- : "b" (expected), "b" (desired)
- : "cr0"
- );
- if (success)
- fence_after(success_order);
- else
- fence_after(failure_order);
- return !!success;
- }
- static BOOST_FORCEINLINE bool compare_exchange_strong(
- storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT
- {
- int success;
- fence_before(success_order);
- __asm__ __volatile__
- (
- "li %1, 0\n\t"
- "0: ldarx %0,%y2\n\t"
- "cmpd %0, %3\n\t"
- "bne- 1f\n\t"
- "stdcx. %4,%y2\n\t"
- "bne- 0b\n\t"
- "li %1, 1\n\t"
- "1:\n\t"
- : "=&b" (expected), "=&b" (success), "+Z" (storage)
- : "b" (expected), "b" (desired)
- : "cr0"
- );
- if (success)
- fence_after(success_order);
- else
- fence_after(failure_order);
- return !!success;
- }
- static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "ldarx %0,%y2\n\t"
- "add %1,%0,%3\n\t"
- "stdcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "ldarx %0,%y2\n\t"
- "sub %1,%0,%3\n\t"
- "stdcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "ldarx %0,%y2\n\t"
- "and %1,%0,%3\n\t"
- "stdcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "ldarx %0,%y2\n\t"
- "or %1,%0,%3\n\t"
- "stdcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
- {
- storage_type original, result;
- fence_before(order);
- __asm__ __volatile__
- (
- "1:\n\t"
- "ldarx %0,%y2\n\t"
- "xor %1,%0,%3\n\t"
- "stdcx. %1,%y2\n\t"
- "bne- 1b\n\t"
- : "=&b" (original), "=&b" (result), "+Z" (storage)
- : "b" (v)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
- );
- fence_after(order);
- return original;
- }
- static BOOST_FORCEINLINE bool test_and_set(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT
- {
- return !!exchange(storage, (storage_type)1, order);
- }
- static BOOST_FORCEINLINE void clear(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT
- {
- store(storage, 0, order);
- }
- };
- #endif // defined(BOOST_ATOMIC_DETAIL_PPC_HAS_LDARX_STDCX)
- BOOST_FORCEINLINE void thread_fence(memory_order order) BOOST_NOEXCEPT
- {
- if (order != memory_order_relaxed)
- {
- #if defined(__powerpc64__) || defined(__PPC64__)
- if (order != memory_order_seq_cst)
- __asm__ __volatile__ ("lwsync" ::: "memory");
- else
- __asm__ __volatile__ ("sync" ::: "memory");
- #else
- __asm__ __volatile__ ("sync" ::: "memory");
- #endif
- }
- }
- BOOST_FORCEINLINE void signal_fence(memory_order order) BOOST_NOEXCEPT
- {
- if (order != memory_order_relaxed)
- #if defined(__ibmxl__) || defined(__IBMCPP__)
- __fence();
- #else
- __asm__ __volatile__ ("" ::: "memory");
- #endif
- }
- } // namespace detail
- } // namespace atomics
- } // namespace boost
- #endif // BOOST_ATOMIC_DETAIL_OPS_GCC_PPC_HPP_INCLUDED_
|