123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556 |
- /*
- * 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) 2012 Tim Blechmann
- * Copyright (c) 2014 - 2018 Andrey Semashev
- */
- /*!
- * \file atomic/detail/ops_gcc_x86_dcas.hpp
- *
- * This header contains implementation of the double-width CAS primitive for x86.
- */
- #ifndef BOOST_ATOMIC_DETAIL_OPS_GCC_X86_DCAS_HPP_INCLUDED_
- #define BOOST_ATOMIC_DETAIL_OPS_GCC_X86_DCAS_HPP_INCLUDED_
- #include <boost/cstdint.hpp>
- #include <boost/memory_order.hpp>
- #include <boost/atomic/detail/config.hpp>
- #include <boost/atomic/detail/storage_type.hpp>
- #include <boost/atomic/detail/string_ops.hpp>
- #include <boost/atomic/capabilities.hpp>
- #ifdef BOOST_HAS_PRAGMA_ONCE
- #pragma once
- #endif
- namespace boost {
- namespace atomics {
- namespace detail {
- // Note: In the 32-bit PIC code guarded with BOOST_ATOMIC_DETAIL_X86_ASM_PRESERVE_EBX below we have to avoid using memory
- // operand constraints because the compiler may choose to use ebx as the base register for that operand. At least, clang
- // is known to do that. For this reason we have to pre-compute a pointer to storage and pass it in edi. For the same reason
- // we cannot save ebx to the stack with a mov instruction, so we use esi as a scratch register and restore it afterwards.
- // Alternatively, we could push/pop the register to the stack, but exchanging the registers is faster.
- // The need to pass a pointer in edi is a bit wasteful because normally the memory operand would use a base pointer
- // with an offset (e.g. `this` + offset). But unfortunately, there seems to be no way around it.
- #if defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG8B)
- template< bool Signed >
- struct gcc_dcas_x86
- {
- typedef typename make_storage_type< 8u >::type storage_type;
- typedef typename make_storage_type< 8u >::aligned aligned_storage_type;
- typedef uint32_t BOOST_ATOMIC_DETAIL_MAY_ALIAS aliasing_uint32_t;
- static BOOST_CONSTEXPR_OR_CONST bool full_cas_based = true;
- static BOOST_CONSTEXPR_OR_CONST bool is_always_lock_free = true;
- static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
- {
- if (BOOST_LIKELY((((uint32_t)&storage) & 0x00000007) == 0u))
- {
- #if defined(__SSE__)
- typedef float xmm_t __attribute__((__vector_size__(16)));
- xmm_t xmm_scratch;
- __asm__ __volatile__
- (
- #if defined(__AVX__)
- "vmovq %[value], %[xmm_scratch]\n\t"
- "vmovq %[xmm_scratch], %[storage]\n\t"
- #elif defined(__SSE2__)
- "movq %[value], %[xmm_scratch]\n\t"
- "movq %[xmm_scratch], %[storage]\n\t"
- #else
- "xorps %[xmm_scratch], %[xmm_scratch]\n\t"
- "movlps %[value], %[xmm_scratch]\n\t"
- "movlps %[xmm_scratch], %[storage]\n\t"
- #endif
- : [storage] "=m" (storage), [xmm_scratch] "=x" (xmm_scratch)
- : [value] "m" (v)
- : "memory"
- );
- #else
- __asm__ __volatile__
- (
- "fildll %[value]\n\t"
- "fistpll %[storage]\n\t"
- : [storage] "=m" (storage)
- : [value] "m" (v)
- : "memory"
- );
- #endif
- }
- else
- {
- #if defined(BOOST_ATOMIC_DETAIL_X86_ASM_PRESERVE_EBX)
- __asm__ __volatile__
- (
- "xchgl %%ebx, %%esi\n\t"
- "movl %%eax, %%ebx\n\t"
- "movl (%[dest]), %%eax\n\t"
- "movl 4(%[dest]), %%edx\n\t"
- ".align 16\n\t"
- "1: lock; cmpxchg8b (%[dest])\n\t"
- "jne 1b\n\t"
- "xchgl %%ebx, %%esi\n\t"
- :
- : "a" ((uint32_t)v), "c" ((uint32_t)(v >> 32)), [dest] "D" (&storage)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "edx", "memory"
- );
- #else // defined(BOOST_ATOMIC_DETAIL_X86_ASM_PRESERVE_EBX)
- __asm__ __volatile__
- (
- "movl %[dest_lo], %%eax\n\t"
- "movl %[dest_hi], %%edx\n\t"
- ".align 16\n\t"
- "1: lock; cmpxchg8b %[dest_lo]\n\t"
- "jne 1b\n\t"
- : [dest_lo] "=m" (storage), [dest_hi] "=m" (reinterpret_cast< volatile aliasing_uint32_t* >(&storage)[1])
- : [value_lo] "b" ((uint32_t)v), "c" ((uint32_t)(v >> 32))
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "eax", "edx", "memory"
- );
- #endif // defined(BOOST_ATOMIC_DETAIL_X86_ASM_PRESERVE_EBX)
- }
- }
- static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order) BOOST_NOEXCEPT
- {
- storage_type value;
- if (BOOST_LIKELY((((uint32_t)&storage) & 0x00000007) == 0u))
- {
- #if defined(__SSE__)
- typedef float xmm_t __attribute__((__vector_size__(16)));
- xmm_t xmm_scratch;
- __asm__ __volatile__
- (
- #if defined(__AVX__)
- "vmovq %[storage], %[xmm_scratch]\n\t"
- "vmovq %[xmm_scratch], %[value]\n\t"
- #elif defined(__SSE2__)
- "movq %[storage], %[xmm_scratch]\n\t"
- "movq %[xmm_scratch], %[value]\n\t"
- #else
- "xorps %[xmm_scratch], %[xmm_scratch]\n\t"
- "movlps %[storage], %[xmm_scratch]\n\t"
- "movlps %[xmm_scratch], %[value]\n\t"
- #endif
- : [value] "=m" (value), [xmm_scratch] "=x" (xmm_scratch)
- : [storage] "m" (storage)
- : "memory"
- );
- #else
- __asm__ __volatile__
- (
- "fildll %[storage]\n\t"
- "fistpll %[value]\n\t"
- : [value] "=m" (value)
- : [storage] "m" (storage)
- : "memory"
- );
- #endif
- }
- else
- {
- // Note that despite const qualification cmpxchg8b below may issue a store to the storage. The storage value
- // will not change, but this prevents the storage to reside in read-only memory.
- #if defined(BOOST_ATOMIC_DETAIL_X86_NO_ASM_AX_DX_PAIRS)
- uint32_t value_bits[2];
- // We don't care for comparison result here; the previous value will be stored into value anyway.
- // Also we don't care for ebx and ecx values, they just have to be equal to eax and edx before cmpxchg8b.
- __asm__ __volatile__
- (
- "movl %%ebx, %%eax\n\t"
- "movl %%ecx, %%edx\n\t"
- "lock; cmpxchg8b %[storage]\n\t"
- : "=&a" (value_bits[0]), "=&d" (value_bits[1])
- : [storage] "m" (storage)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
- );
- BOOST_ATOMIC_DETAIL_MEMCPY(&value, value_bits, sizeof(value));
- #else // defined(BOOST_ATOMIC_DETAIL_X86_NO_ASM_AX_DX_PAIRS)
- // We don't care for comparison result here; the previous value will be stored into value anyway.
- // Also we don't care for ebx and ecx values, they just have to be equal to eax and edx before cmpxchg8b.
- __asm__ __volatile__
- (
- "movl %%ebx, %%eax\n\t"
- "movl %%ecx, %%edx\n\t"
- "lock; cmpxchg8b %[storage]\n\t"
- : "=&A" (value)
- : [storage] "m" (storage)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
- );
- #endif // defined(BOOST_ATOMIC_DETAIL_X86_NO_ASM_AX_DX_PAIRS)
- }
- return value;
- }
- static BOOST_FORCEINLINE bool compare_exchange_strong(
- storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order, memory_order) BOOST_NOEXCEPT
- {
- #if defined(__clang__)
- // Clang cannot allocate eax:edx register pairs but it has sync intrinsics
- storage_type old_expected = expected;
- expected = __sync_val_compare_and_swap(&storage, old_expected, desired);
- return expected == old_expected;
- #elif defined(BOOST_ATOMIC_DETAIL_X86_ASM_PRESERVE_EBX)
- bool success;
- #if defined(BOOST_ATOMIC_DETAIL_ASM_HAS_FLAG_OUTPUTS)
- __asm__ __volatile__
- (
- "xchgl %%ebx, %%esi\n\t"
- "lock; cmpxchg8b (%[dest])\n\t"
- "xchgl %%ebx, %%esi\n\t"
- : "+A" (expected), [success] "=@ccz" (success)
- : "S" ((uint32_t)desired), "c" ((uint32_t)(desired >> 32)), [dest] "D" (&storage)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
- );
- #else // defined(BOOST_ATOMIC_DETAIL_ASM_HAS_FLAG_OUTPUTS)
- __asm__ __volatile__
- (
- "xchgl %%ebx, %%esi\n\t"
- "lock; cmpxchg8b (%[dest])\n\t"
- "xchgl %%ebx, %%esi\n\t"
- "sete %[success]\n\t"
- : "+A" (expected), [success] "=qm" (success)
- : "S" ((uint32_t)desired), "c" ((uint32_t)(desired >> 32)), [dest] "D" (&storage)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
- );
- #endif // defined(BOOST_ATOMIC_DETAIL_ASM_HAS_FLAG_OUTPUTS)
- return success;
- #else // defined(BOOST_ATOMIC_DETAIL_X86_ASM_PRESERVE_EBX)
- bool success;
- #if defined(BOOST_ATOMIC_DETAIL_ASM_HAS_FLAG_OUTPUTS)
- __asm__ __volatile__
- (
- "lock; cmpxchg8b %[dest]\n\t"
- : "+A" (expected), [dest] "+m" (storage), [success] "=@ccz" (success)
- : "b" ((uint32_t)desired), "c" ((uint32_t)(desired >> 32))
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
- );
- #else // defined(BOOST_ATOMIC_DETAIL_ASM_HAS_FLAG_OUTPUTS)
- __asm__ __volatile__
- (
- "lock; cmpxchg8b %[dest]\n\t"
- "sete %[success]\n\t"
- : "+A" (expected), [dest] "+m" (storage), [success] "=qm" (success)
- : "b" ((uint32_t)desired), "c" ((uint32_t)(desired >> 32))
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
- );
- #endif // defined(BOOST_ATOMIC_DETAIL_ASM_HAS_FLAG_OUTPUTS)
- return success;
- #endif // defined(BOOST_ATOMIC_DETAIL_X86_ASM_PRESERVE_EBX)
- }
- 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
- {
- return compare_exchange_strong(storage, expected, desired, success_order, failure_order);
- }
- static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
- {
- #if defined(BOOST_ATOMIC_DETAIL_X86_ASM_PRESERVE_EBX)
- #if defined(BOOST_ATOMIC_DETAIL_X86_NO_ASM_AX_DX_PAIRS)
- uint32_t old_bits[2];
- __asm__ __volatile__
- (
- "xchgl %%ebx, %%esi\n\t"
- "movl (%[dest]), %%eax\n\t"
- "movl 4(%[dest]), %%edx\n\t"
- ".align 16\n\t"
- "1: lock; cmpxchg8b (%[dest])\n\t"
- "jne 1b\n\t"
- "xchgl %%ebx, %%esi\n\t"
- : "=a" (old_bits[0]), "=d" (old_bits[1])
- : "S" ((uint32_t)v), "c" ((uint32_t)(v >> 32)), [dest] "D" (&storage)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
- );
- storage_type old_value;
- BOOST_ATOMIC_DETAIL_MEMCPY(&old_value, old_bits, sizeof(old_value));
- return old_value;
- #else // defined(BOOST_ATOMIC_DETAIL_X86_NO_ASM_AX_DX_PAIRS)
- storage_type old_value;
- __asm__ __volatile__
- (
- "xchgl %%ebx, %%esi\n\t"
- "movl (%[dest]), %%eax\n\t"
- "movl 4(%[dest]), %%edx\n\t"
- ".align 16\n\t"
- "1: lock; cmpxchg8b (%[dest])\n\t"
- "jne 1b\n\t"
- "xchgl %%ebx, %%esi\n\t"
- : "=A" (old_value)
- : "S" ((uint32_t)v), "c" ((uint32_t)(v >> 32)), [dest] "D" (&storage)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
- );
- return old_value;
- #endif // defined(BOOST_ATOMIC_DETAIL_X86_NO_ASM_AX_DX_PAIRS)
- #else // defined(BOOST_ATOMIC_DETAIL_X86_ASM_PRESERVE_EBX)
- #if defined(__MINGW32__) && ((__GNUC__+0) * 100 + (__GNUC_MINOR__+0)) < 407
- // MinGW gcc up to 4.6 has problems with allocating registers in the asm blocks below
- uint32_t old_bits[2];
- __asm__ __volatile__
- (
- "movl (%[dest]), %%eax\n\t"
- "movl 4(%[dest]), %%edx\n\t"
- ".align 16\n\t"
- "1: lock; cmpxchg8b (%[dest])\n\t"
- "jne 1b\n\t"
- : "=&a" (old_bits[0]), "=&d" (old_bits[1])
- : "b" ((uint32_t)v), "c" ((uint32_t)(v >> 32)), [dest] "DS" (&storage)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
- );
- storage_type old_value;
- BOOST_ATOMIC_DETAIL_MEMCPY(&old_value, old_bits, sizeof(old_value));
- return old_value;
- #elif defined(BOOST_ATOMIC_DETAIL_X86_NO_ASM_AX_DX_PAIRS)
- uint32_t old_bits[2];
- __asm__ __volatile__
- (
- "movl %[dest_lo], %%eax\n\t"
- "movl %[dest_hi], %%edx\n\t"
- ".align 16\n\t"
- "1: lock; cmpxchg8b %[dest_lo]\n\t"
- "jne 1b\n\t"
- : "=&a" (old_bits[0]), "=&d" (old_bits[1]), [dest_lo] "+m" (storage), [dest_hi] "+m" (reinterpret_cast< volatile aliasing_uint32_t* >(&storage)[1])
- : "b" ((uint32_t)v), "c" ((uint32_t)(v >> 32))
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
- );
- storage_type old_value;
- BOOST_ATOMIC_DETAIL_MEMCPY(&old_value, old_bits, sizeof(old_value));
- return old_value;
- #else // defined(BOOST_ATOMIC_DETAIL_X86_NO_ASM_AX_DX_PAIRS)
- storage_type old_value;
- __asm__ __volatile__
- (
- "movl %[dest_lo], %%eax\n\t"
- "movl %[dest_hi], %%edx\n\t"
- ".align 16\n\t"
- "1: lock; cmpxchg8b %[dest_lo]\n\t"
- "jne 1b\n\t"
- : "=&A" (old_value), [dest_lo] "+m" (storage), [dest_hi] "+m" (reinterpret_cast< volatile aliasing_uint32_t* >(&storage)[1])
- : "b" ((uint32_t)v), "c" ((uint32_t)(v >> 32))
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
- );
- return old_value;
- #endif // defined(BOOST_ATOMIC_DETAIL_X86_NO_ASM_AX_DX_PAIRS)
- #endif // defined(BOOST_ATOMIC_DETAIL_X86_ASM_PRESERVE_EBX)
- }
- };
- #endif // defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG8B)
- #if defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG16B)
- template< bool Signed >
- struct gcc_dcas_x86_64
- {
- typedef typename make_storage_type< 16u >::type storage_type;
- typedef typename make_storage_type< 16u >::aligned aligned_storage_type;
- typedef uint64_t BOOST_ATOMIC_DETAIL_MAY_ALIAS aliasing_uint64_t;
- static BOOST_CONSTEXPR_OR_CONST bool full_cas_based = true;
- static BOOST_CONSTEXPR_OR_CONST bool is_always_lock_free = true;
- static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
- {
- __asm__ __volatile__
- (
- "movq %[dest_lo], %%rax\n\t"
- "movq %[dest_hi], %%rdx\n\t"
- ".align 16\n\t"
- "1: lock; cmpxchg16b %[dest_lo]\n\t"
- "jne 1b\n\t"
- : [dest_lo] "=m" (storage), [dest_hi] "=m" (reinterpret_cast< volatile aliasing_uint64_t* >(&storage)[1])
- : "b" (reinterpret_cast< const aliasing_uint64_t* >(&v)[0]), "c" (reinterpret_cast< const aliasing_uint64_t* >(&v)[1])
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "rax", "rdx", "memory"
- );
- }
- static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order) BOOST_NOEXCEPT
- {
- // Note that despite const qualification cmpxchg16b below may issue a store to the storage. The storage value
- // will not change, but this prevents the storage to reside in read-only memory.
- #if defined(BOOST_ATOMIC_DETAIL_X86_NO_ASM_AX_DX_PAIRS)
- uint64_t value_bits[2];
- // We don't care for comparison result here; the previous value will be stored into value anyway.
- // Also we don't care for rbx and rcx values, they just have to be equal to rax and rdx before cmpxchg16b.
- __asm__ __volatile__
- (
- "movq %%rbx, %%rax\n\t"
- "movq %%rcx, %%rdx\n\t"
- "lock; cmpxchg16b %[storage]\n\t"
- : "=&a" (value_bits[0]), "=&d" (value_bits[1])
- : [storage] "m" (storage)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
- );
- storage_type value;
- BOOST_ATOMIC_DETAIL_MEMCPY(&value, value_bits, sizeof(value));
- return value;
- #else // defined(BOOST_ATOMIC_DETAIL_X86_NO_ASM_AX_DX_PAIRS)
- storage_type value;
- // We don't care for comparison result here; the previous value will be stored into value anyway.
- // Also we don't care for rbx and rcx values, they just have to be equal to rax and rdx before cmpxchg16b.
- __asm__ __volatile__
- (
- "movq %%rbx, %%rax\n\t"
- "movq %%rcx, %%rdx\n\t"
- "lock; cmpxchg16b %[storage]\n\t"
- : "=&A" (value)
- : [storage] "m" (storage)
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
- );
- return value;
- #endif // defined(BOOST_ATOMIC_DETAIL_X86_NO_ASM_AX_DX_PAIRS)
- }
- static BOOST_FORCEINLINE bool compare_exchange_strong(
- storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order, memory_order) BOOST_NOEXCEPT
- {
- #if defined(__clang__)
- // Clang cannot allocate rax:rdx register pairs but it has sync intrinsics
- storage_type old_expected = expected;
- expected = __sync_val_compare_and_swap(&storage, old_expected, desired);
- return expected == old_expected;
- #elif defined(BOOST_ATOMIC_DETAIL_X86_NO_ASM_AX_DX_PAIRS)
- // Some compilers can't allocate rax:rdx register pair either but also don't support 128-bit __sync_val_compare_and_swap
- bool success;
- __asm__ __volatile__
- (
- "lock; cmpxchg16b %[dest]\n\t"
- "sete %[success]\n\t"
- : [dest] "+m" (storage), "+a" (reinterpret_cast< aliasing_uint64_t* >(&expected)[0]), "+d" (reinterpret_cast< aliasing_uint64_t* >(&expected)[1]), [success] "=q" (success)
- : "b" (reinterpret_cast< const aliasing_uint64_t* >(&desired)[0]), "c" (reinterpret_cast< const aliasing_uint64_t* >(&desired)[1])
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
- );
- return success;
- #else // defined(BOOST_ATOMIC_DETAIL_X86_NO_ASM_AX_DX_PAIRS)
- bool success;
- #if defined(BOOST_ATOMIC_DETAIL_ASM_HAS_FLAG_OUTPUTS)
- __asm__ __volatile__
- (
- "lock; cmpxchg16b %[dest]\n\t"
- : "+A" (expected), [dest] "+m" (storage), "=@ccz" (success)
- : "b" (reinterpret_cast< const aliasing_uint64_t* >(&desired)[0]), "c" (reinterpret_cast< const aliasing_uint64_t* >(&desired)[1])
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
- );
- #else // defined(BOOST_ATOMIC_DETAIL_ASM_HAS_FLAG_OUTPUTS)
- __asm__ __volatile__
- (
- "lock; cmpxchg16b %[dest]\n\t"
- "sete %[success]\n\t"
- : "+A" (expected), [dest] "+m" (storage), [success] "=qm" (success)
- : "b" (reinterpret_cast< const aliasing_uint64_t* >(&desired)[0]), "c" (reinterpret_cast< const aliasing_uint64_t* >(&desired)[1])
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
- );
- #endif // defined(BOOST_ATOMIC_DETAIL_ASM_HAS_FLAG_OUTPUTS)
- return success;
- #endif // defined(BOOST_ATOMIC_DETAIL_X86_NO_ASM_AX_DX_PAIRS)
- }
- 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
- {
- return compare_exchange_strong(storage, expected, desired, success_order, failure_order);
- }
- static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order) BOOST_NOEXCEPT
- {
- #if defined(BOOST_ATOMIC_DETAIL_X86_NO_ASM_AX_DX_PAIRS)
- uint64_t old_bits[2];
- __asm__ __volatile__
- (
- "movq %[dest_lo], %%rax\n\t"
- "movq %[dest_hi], %%rdx\n\t"
- ".align 16\n\t"
- "1: lock; cmpxchg16b %[dest_lo]\n\t"
- "jne 1b\n\t"
- : [dest_lo] "+m" (storage), [dest_hi] "+m" (reinterpret_cast< volatile aliasing_uint64_t* >(&storage)[1]), "=&a" (old_bits[0]), "=&d" (old_bits[1])
- : "b" (reinterpret_cast< const aliasing_uint64_t* >(&v)[0]), "c" (reinterpret_cast< const aliasing_uint64_t* >(&v)[1])
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
- );
- storage_type old_value;
- BOOST_ATOMIC_DETAIL_MEMCPY(&old_value, old_bits, sizeof(old_value));
- return old_value;
- #else // defined(BOOST_ATOMIC_DETAIL_X86_NO_ASM_AX_DX_PAIRS)
- storage_type old_value;
- __asm__ __volatile__
- (
- "movq %[dest_lo], %%rax\n\t"
- "movq %[dest_hi], %%rdx\n\t"
- ".align 16\n\t"
- "1: lock; cmpxchg16b %[dest_lo]\n\t"
- "jne 1b\n\t"
- : "=&A" (old_value), [dest_lo] "+m" (storage), [dest_hi] "+m" (reinterpret_cast< volatile aliasing_uint64_t* >(&storage)[1])
- : "b" (reinterpret_cast< const aliasing_uint64_t* >(&v)[0]), "c" (reinterpret_cast< const aliasing_uint64_t* >(&v)[1])
- : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
- );
- return old_value;
- #endif // defined(BOOST_ATOMIC_DETAIL_X86_NO_ASM_AX_DX_PAIRS)
- }
- };
- #endif // defined(BOOST_ATOMIC_DETAIL_X86_HAS_CMPXCHG16B)
- } // namespace detail
- } // namespace atomics
- } // namespace boost
- #endif // BOOST_ATOMIC_DETAIL_OPS_GCC_X86_DCAS_HPP_INCLUDED_
|