Chris@102: /*
Chris@102:  * Distributed under the Boost Software License, Version 1.0.
Chris@102:  * (See accompanying file LICENSE_1_0.txt or copy at
Chris@102:  * http://www.boost.org/LICENSE_1_0.txt)
Chris@102:  *
Chris@102:  * Copyright (c) 2009 Helge Bahmann
Chris@102:  * Copyright (c) 2013 Tim Blechmann
Chris@102:  * Copyright (c) 2014 Andrey Semashev
Chris@102:  */
Chris@102: /*!
Chris@102:  * \file   atomic/detail/ops_gcc_arm.hpp
Chris@102:  *
Chris@102:  * This header contains implementation of the \c operations template.
Chris@102:  */
Chris@102: 
Chris@102: #ifndef BOOST_ATOMIC_DETAIL_OPS_GCC_ARM_HPP_INCLUDED_
Chris@102: #define BOOST_ATOMIC_DETAIL_OPS_GCC_ARM_HPP_INCLUDED_
Chris@102: 
Chris@102: #include <boost/cstdint.hpp>
Chris@102: #include <boost/memory_order.hpp>
Chris@102: #include <boost/atomic/detail/config.hpp>
Chris@102: #include <boost/atomic/detail/storage_type.hpp>
Chris@102: #include <boost/atomic/detail/operations_fwd.hpp>
Chris@102: #include <boost/atomic/detail/ops_extending_cas_based.hpp>
Chris@102: #include <boost/atomic/capabilities.hpp>
Chris@102: 
Chris@102: #ifdef BOOST_HAS_PRAGMA_ONCE
Chris@102: #pragma once
Chris@102: #endif
Chris@102: 
Chris@102: namespace boost {
Chris@102: namespace atomics {
Chris@102: namespace detail {
Chris@102: 
Chris@102: // From the ARM Architecture Reference Manual for architecture v6:
Chris@102: //
Chris@102: // LDREX{<cond>} <Rd>, [<Rn>]
Chris@102: // <Rd> Specifies the destination register for the memory word addressed by <Rd>
Chris@102: // <Rn> Specifies the register containing the address.
Chris@102: //
Chris@102: // STREX{<cond>} <Rd>, <Rm>, [<Rn>]
Chris@102: // <Rd> Specifies the destination register for the returned status value.
Chris@102: //      0  if the operation updates memory
Chris@102: //      1  if the operation fails to update memory
Chris@102: // <Rm> Specifies the register containing the word to be stored to memory.
Chris@102: // <Rn> Specifies the register containing the address.
Chris@102: // Rd must not be the same register as Rm or Rn.
Chris@102: //
Chris@102: // ARM v7 is like ARM v6 plus:
Chris@102: // There are half-word and byte versions of the LDREX and STREX instructions,
Chris@102: // LDREXH, LDREXB, STREXH and STREXB.
Chris@102: // There are also double-word versions, LDREXD and STREXD.
Chris@102: // (Actually it looks like these are available from version 6k onwards.)
Chris@102: // FIXME these are not yet used; should be mostly a matter of copy-and-paste.
Chris@102: // I think you can supply an immediate offset to the address.
Chris@102: //
Chris@102: // A memory barrier is effected using a "co-processor 15" instruction,
Chris@102: // though a separate assembler mnemonic is available for it in v7.
Chris@102: //
Chris@102: // "Thumb 1" is a subset of the ARM instruction set that uses a 16-bit encoding.  It
Chris@102: // doesn't include all instructions and in particular it doesn't include the co-processor
Chris@102: // instruction used for the memory barrier or the load-locked/store-conditional
Chris@102: // instructions.  So, if we're compiling in "Thumb 1" mode, we need to wrap all of our
Chris@102: // asm blocks with code to temporarily change to ARM mode.
Chris@102: //
Chris@102: // You can only change between ARM and Thumb modes when branching using the bx instruction.
Chris@102: // bx takes an address specified in a register.  The least significant bit of the address
Chris@102: // indicates the mode, so 1 is added to indicate that the destination code is Thumb.
Chris@102: // A temporary register is needed for the address and is passed as an argument to these
Chris@102: // macros.  It must be one of the "low" registers accessible to Thumb code, specified
Chris@102: // using the "l" attribute in the asm statement.
Chris@102: //
Chris@102: // Architecture v7 introduces "Thumb 2", which does include (almost?) all of the ARM
Chris@102: // instruction set.  (Actually, there was an extension of v6 called v6T2 which supported
Chris@102: // "Thumb 2" mode, but its architecture manual is no longer available, referring to v7.)
Chris@102: // So in v7 we don't need to change to ARM mode; we can write "universal
Chris@102: // assembler" which will assemble to Thumb 2 or ARM code as appropriate.  The only thing
Chris@102: // we need to do to make this "universal" assembler mode work is to insert "IT" instructions
Chris@102: // to annotate the conditional instructions.  These are ignored in other modes (e.g. v6),
Chris@102: // so they can always be present.
Chris@102: 
Chris@102: // A note about memory_order_consume. Technically, this architecture allows to avoid
Chris@102: // unnecessary memory barrier after consume load since it supports data dependency ordering.
Chris@102: // However, some compiler optimizations may break a seemingly valid code relying on data
Chris@102: // dependency tracking by injecting bogus branches to aid out of order execution.
Chris@102: // This may happen not only in Boost.Atomic code but also in user's code, which we have no
Chris@102: // control of. See this thread: http://lists.boost.org/Archives/boost/2014/06/213890.php.
Chris@102: // For this reason we promote memory_order_consume to memory_order_acquire.
Chris@102: 
Chris@102: #if defined(__thumb__) && !defined(__thumb2__)
Chris@102: #define BOOST_ATOMIC_DETAIL_ARM_ASM_START(TMPREG) "adr " #TMPREG ", 8f\n" "bx " #TMPREG "\n" ".arm\n" ".align 4\n" "8:\n"
Chris@102: #define BOOST_ATOMIC_DETAIL_ARM_ASM_END(TMPREG)   "adr " #TMPREG ", 9f + 1\n" "bx " #TMPREG "\n" ".thumb\n" ".align 2\n" "9:\n"
Chris@102: #define BOOST_ATOMIC_DETAIL_ARM_ASM_TMPREG_CONSTRAINT(var) "=&l" (var)
Chris@102: #else
Chris@102: // The tmpreg may be wasted in this case, which is non-optimal.
Chris@102: #define BOOST_ATOMIC_DETAIL_ARM_ASM_START(TMPREG)
Chris@102: #define BOOST_ATOMIC_DETAIL_ARM_ASM_END(TMPREG)
Chris@102: #define BOOST_ATOMIC_DETAIL_ARM_ASM_TMPREG_CONSTRAINT(var) "=&r" (var)
Chris@102: #endif
Chris@102: 
Chris@102: struct gcc_arm_operations_base
Chris@102: {
Chris@102:     static BOOST_FORCEINLINE void fence_before(memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         if ((order & memory_order_release) != 0)
Chris@102:             hardware_full_fence();
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE void fence_after(memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         if ((order & (memory_order_consume | memory_order_acquire)) != 0)
Chris@102:             hardware_full_fence();
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE void fence_after_store(memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         if (order == memory_order_seq_cst)
Chris@102:             hardware_full_fence();
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE void hardware_full_fence() BOOST_NOEXCEPT
Chris@102:     {
Chris@102: #if defined(BOOST_ATOMIC_DETAIL_ARM_HAS_DMB)
Chris@102:         // Older binutils (supposedly, older than 2.21.1) didn't support symbolic or numeric arguments of the "dmb" instruction such as "ish" or "#11".
Chris@102:         // As a workaround we have to inject encoded bytes of the instruction. There are two encodings for the instruction: ARM and Thumb. See ARM Architecture Reference Manual, A8.8.43.
Chris@102:         // Since we cannot detect binutils version at compile time, we'll have to always use this hack.
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102: #if defined(__thumb2__)
Chris@102:             ".short 0xF3BF, 0x8F5B\n" // dmb ish
Chris@102: #else
Chris@102:             ".word 0xF57FF05B\n" // dmb ish
Chris@102: #endif
Chris@102:             :
Chris@102:             :
Chris@102:             : "memory"
Chris@102:         );
Chris@102: #else
Chris@102:         int tmp;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%0)
Chris@102:             "mcr\tp15, 0, r0, c7, c10, 5\n"
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%0)
Chris@102:             : "=&l" (tmp)
Chris@102:             :
Chris@102:             : "memory"
Chris@102:         );
Chris@102: #endif
Chris@102:     }
Chris@102: };
Chris@102: 
Chris@102: 
Chris@102: template< bool Signed >
Chris@102: struct operations< 4u, Signed > :
Chris@102:     public gcc_arm_operations_base
Chris@102: {
Chris@102:     typedef typename make_storage_type< 4u, Signed >::type storage_type;
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         fence_before(order);
Chris@102:         storage = v;
Chris@102:         fence_after_store(order);
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         storage_type v = storage;
Chris@102:         fence_after(order);
Chris@102:         return v;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         storage_type original;
Chris@102:         fence_before(order);
Chris@102:         uint32_t tmp;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp])
Chris@102:             "1:\n"
Chris@102:             "ldrex %[original], %[storage]\n"          // load the original value
Chris@102:             "strex %[tmp], %[value], %[storage]\n"     // store the replacement, tmp = store failed
Chris@102:             "teq   %[tmp], #0\n"                       // check if store succeeded
Chris@102:             "bne   1b\n"
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp])
Chris@102:             : [tmp] "=&l" (tmp), [original] "=&r" (original), [storage] "+Q" (storage)
Chris@102:             : [value] "r" (v)
Chris@102:             : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
Chris@102:         );
Chris@102:         fence_after(order);
Chris@102:         return original;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE bool compare_exchange_weak(
Chris@102:         storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         fence_before(success_order);
Chris@102:         uint32_t success;
Chris@102:         uint32_t tmp;
Chris@102:         storage_type original;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp])
Chris@102:             "mov     %[success], #0\n"                      // success = 0
Chris@102:             "ldrex   %[original], %[storage]\n"             // original = *(&storage)
Chris@102:             "cmp     %[original], %[expected]\n"            // flags = original==expected
Chris@102:             "itt     eq\n"                                  // [hint that the following 2 instructions are conditional on flags.equal]
Chris@102:             "strexeq %[success], %[desired], %[storage]\n"  // if (flags.equal) *(&storage) = desired, success = store failed
Chris@102:             "eoreq   %[success], %[success], #1\n"          // if (flags.equal) success ^= 1 (i.e. make it 1 if store succeeded)
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp])
Chris@102:             : [original] "=&r" (original),  // %0
Chris@102:               [success] "=&r" (success),    // %1
Chris@102:               [tmp] "=&l" (tmp),            // %2
Chris@102:               [storage] "+Q" (storage)      // %3
Chris@102:             : [expected] "r" (expected),    // %4
Chris@102:               [desired] "r" (desired)       // %5
Chris@102:             : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
Chris@102:         );
Chris@102:         if (success)
Chris@102:             fence_after(success_order);
Chris@102:         else
Chris@102:             fence_after(failure_order);
Chris@102:         expected = original;
Chris@102:         return !!success;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE bool compare_exchange_strong(
Chris@102:         storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         fence_before(success_order);
Chris@102:         uint32_t success;
Chris@102:         uint32_t tmp;
Chris@102:         storage_type original;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp])
Chris@102:             "mov     %[success], #0\n"                      // success = 0
Chris@102:             "1:\n"
Chris@102:             "ldrex   %[original], %[storage]\n"             // original = *(&storage)
Chris@102:             "cmp     %[original], %[expected]\n"            // flags = original==expected
Chris@102:             "bne     2f\n"                                  // if (!flags.equal) goto end
Chris@102:             "strex   %[success], %[desired], %[storage]\n"  // *(&storage) = desired, success = store failed
Chris@102:             "eors    %[success], %[success], #1\n"          // success ^= 1 (i.e. make it 1 if store succeeded); flags.equal = success == 0
Chris@102:             "beq     1b\n"                                  // if (flags.equal) goto retry
Chris@102:             "2:\n"
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp])
Chris@102:             : [original] "=&r" (original),  // %0
Chris@102:               [success] "=&r" (success),    // %1
Chris@102:               [tmp] "=&l" (tmp),            // %2
Chris@102:               [storage] "+Q" (storage)      // %3
Chris@102:             : [expected] "r" (expected),    // %4
Chris@102:               [desired] "r" (desired)       // %5
Chris@102:             : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
Chris@102:         );
Chris@102:         if (success)
Chris@102:             fence_after(success_order);
Chris@102:         else
Chris@102:             fence_after(failure_order);
Chris@102:         expected = original;
Chris@102:         return !!success;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         fence_before(order);
Chris@102:         uint32_t tmp;
Chris@102:         storage_type original, result;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp])
Chris@102:             "1:\n"
Chris@102:             "ldrex   %[original], %[storage]\n"           // original = *(&storage)
Chris@102:             "add     %[result], %[original], %[value]\n"  // result = original + value
Chris@102:             "strex   %[tmp], %[result], %[storage]\n"     // *(&storage) = result, tmp = store failed
Chris@102:             "teq     %[tmp], #0\n"                        // flags = tmp==0
Chris@102:             "bne     1b\n"                                // if (!flags.equal) goto retry
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp])
Chris@102:             : [original] "=&r" (original),  // %0
Chris@102:               [result] "=&r" (result),      // %1
Chris@102:               [tmp] "=&l" (tmp),            // %2
Chris@102:               [storage] "+Q" (storage)      // %3
Chris@102:             : [value] "r" (v)               // %4
Chris@102:             : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
Chris@102:         );
Chris@102:         fence_after(order);
Chris@102:         return original;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         fence_before(order);
Chris@102:         uint32_t tmp;
Chris@102:         storage_type original, result;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp])
Chris@102:             "1:\n"
Chris@102:             "ldrex   %[original], %[storage]\n"           // original = *(&storage)
Chris@102:             "sub     %[result], %[original], %[value]\n"  // result = original - value
Chris@102:             "strex   %[tmp], %[result], %[storage]\n"     // *(&storage) = result, tmp = store failed
Chris@102:             "teq     %[tmp], #0\n"                        // flags = tmp==0
Chris@102:             "bne     1b\n"                                // if (!flags.equal) goto retry
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp])
Chris@102:             : [original] "=&r" (original),  // %0
Chris@102:               [result] "=&r" (result),      // %1
Chris@102:               [tmp] "=&l" (tmp),            // %2
Chris@102:               [storage] "+Q" (storage)      // %3
Chris@102:             : [value] "r" (v)               // %4
Chris@102:             : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
Chris@102:         );
Chris@102:         fence_after(order);
Chris@102:         return original;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         fence_before(order);
Chris@102:         uint32_t tmp;
Chris@102:         storage_type original, result;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp])
Chris@102:             "1:\n"
Chris@102:             "ldrex   %[original], %[storage]\n"           // original = *(&storage)
Chris@102:             "and     %[result], %[original], %[value]\n"  // result = original & value
Chris@102:             "strex   %[tmp], %[result], %[storage]\n"     // *(&storage) = result, tmp = store failed
Chris@102:             "teq     %[tmp], #0\n"                        // flags = tmp==0
Chris@102:             "bne     1b\n"                                // if (!flags.equal) goto retry
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp])
Chris@102:             : [original] "=&r" (original),  // %0
Chris@102:               [result] "=&r" (result),      // %1
Chris@102:               [tmp] "=&l" (tmp),            // %2
Chris@102:               [storage] "+Q" (storage)      // %3
Chris@102:             : [value] "r" (v)               // %4
Chris@102:             : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
Chris@102:         );
Chris@102:         fence_after(order);
Chris@102:         return original;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         fence_before(order);
Chris@102:         uint32_t tmp;
Chris@102:         storage_type original, result;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp])
Chris@102:             "1:\n"
Chris@102:             "ldrex   %[original], %[storage]\n"           // original = *(&storage)
Chris@102:             "orr     %[result], %[original], %[value]\n"  // result = original | value
Chris@102:             "strex   %[tmp], %[result], %[storage]\n"     // *(&storage) = result, tmp = store failed
Chris@102:             "teq     %[tmp], #0\n"                        // flags = tmp==0
Chris@102:             "bne     1b\n"                                // if (!flags.equal) goto retry
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp])
Chris@102:             : [original] "=&r" (original),  // %0
Chris@102:               [result] "=&r" (result),      // %1
Chris@102:               [tmp] "=&l" (tmp),            // %2
Chris@102:               [storage] "+Q" (storage)      // %3
Chris@102:             : [value] "r" (v)               // %4
Chris@102:             : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
Chris@102:         );
Chris@102:         fence_after(order);
Chris@102:         return original;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         fence_before(order);
Chris@102:         uint32_t tmp;
Chris@102:         storage_type original, result;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp])
Chris@102:             "1:\n"
Chris@102:             "ldrex   %[original], %[storage]\n"           // original = *(&storage)
Chris@102:             "eor     %[result], %[original], %[value]\n"  // result = original ^ value
Chris@102:             "strex   %[tmp], %[result], %[storage]\n"     // *(&storage) = result, tmp = store failed
Chris@102:             "teq     %[tmp], #0\n"                        // flags = tmp==0
Chris@102:             "bne     1b\n"                                // if (!flags.equal) goto retry
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp])
Chris@102:             : [original] "=&r" (original),  // %0
Chris@102:               [result] "=&r" (result),      // %1
Chris@102:               [tmp] "=&l" (tmp),            // %2
Chris@102:               [storage] "+Q" (storage)      // %3
Chris@102:             : [value] "r" (v)               // %4
Chris@102:             : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
Chris@102:         );
Chris@102:         fence_after(order);
Chris@102:         return original;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE bool test_and_set(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return !!exchange(storage, (storage_type)1, order);
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE void clear(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         store(storage, 0, order);
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE bool is_lock_free(storage_type const volatile&) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return true;
Chris@102:     }
Chris@102: };
Chris@102: 
Chris@102: 
Chris@102: template< >
Chris@102: struct operations< 1u, false > :
Chris@102:     public operations< 4u, false >
Chris@102: {
Chris@102:     typedef operations< 4u, false > base_type;
Chris@102:     typedef base_type::storage_type storage_type;
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         fence_before(order);
Chris@102:         uint32_t tmp;
Chris@102:         storage_type original, result;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp])
Chris@102:             "1:\n"
Chris@102:             "ldrex   %[original], %[storage]\n"           // original = *(&storage)
Chris@102:             "add     %[result], %[original], %[value]\n"  // result = original + value
Chris@102:             "uxtb    %[result], %[result]\n"              // zero extend result from 8 to 32 bits
Chris@102:             "strex   %[tmp], %[result], %[storage]\n"     // *(&storage) = result, tmp = store failed
Chris@102:             "teq     %[tmp], #0\n"                        // flags = tmp==0
Chris@102:             "bne     1b\n"                                // if (!flags.equal) goto retry
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp])
Chris@102:             : [original] "=&r" (original),  // %0
Chris@102:               [result] "=&r" (result),      // %1
Chris@102:               [tmp] "=&l" (tmp),            // %2
Chris@102:               [storage] "+Q" (storage)      // %3
Chris@102:             : [value] "r" (v)               // %4
Chris@102:             : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
Chris@102:         );
Chris@102:         fence_after(order);
Chris@102:         return original;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         fence_before(order);
Chris@102:         uint32_t tmp;
Chris@102:         storage_type original, result;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp])
Chris@102:             "1:\n"
Chris@102:             "ldrex   %[original], %[storage]\n"           // original = *(&storage)
Chris@102:             "sub     %[result], %[original], %[value]\n"  // result = original - value
Chris@102:             "uxtb    %[result], %[result]\n"              // zero extend result from 8 to 32 bits
Chris@102:             "strex   %[tmp], %[result], %[storage]\n"     // *(&storage) = result, tmp = store failed
Chris@102:             "teq     %[tmp], #0\n"                        // flags = tmp==0
Chris@102:             "bne     1b\n"                                // if (!flags.equal) goto retry
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp])
Chris@102:             : [original] "=&r" (original),  // %0
Chris@102:               [result] "=&r" (result),      // %1
Chris@102:               [tmp] "=&l" (tmp),            // %2
Chris@102:               [storage] "+Q" (storage)      // %3
Chris@102:             : [value] "r" (v)               // %4
Chris@102:             : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
Chris@102:         );
Chris@102:         fence_after(order);
Chris@102:         return original;
Chris@102:     }
Chris@102: };
Chris@102: 
Chris@102: template< >
Chris@102: struct operations< 1u, true > :
Chris@102:     public operations< 4u, true >
Chris@102: {
Chris@102:     typedef operations< 4u, true > base_type;
Chris@102:     typedef base_type::storage_type storage_type;
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         fence_before(order);
Chris@102:         uint32_t tmp;
Chris@102:         storage_type original, result;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp])
Chris@102:             "1:\n"
Chris@102:             "ldrex   %[original], %[storage]\n"           // original = *(&storage)
Chris@102:             "add     %[result], %[original], %[value]\n"  // result = original + value
Chris@102:             "sxtb    %[result], %[result]\n"              // sign extend result from 8 to 32 bits
Chris@102:             "strex   %[tmp], %[result], %[storage]\n"     // *(&storage) = result, tmp = store failed
Chris@102:             "teq     %[tmp], #0\n"                        // flags = tmp==0
Chris@102:             "bne     1b\n"                                // if (!flags.equal) goto retry
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp])
Chris@102:             : [original] "=&r" (original),  // %0
Chris@102:               [result] "=&r" (result),      // %1
Chris@102:               [tmp] "=&l" (tmp),            // %2
Chris@102:               [storage] "+Q" (storage)      // %3
Chris@102:             : [value] "r" (v)               // %4
Chris@102:             : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
Chris@102:         );
Chris@102:         fence_after(order);
Chris@102:         return original;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         fence_before(order);
Chris@102:         uint32_t tmp;
Chris@102:         storage_type original, result;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp])
Chris@102:             "1:\n"
Chris@102:             "ldrex   %[original], %[storage]\n"           // original = *(&storage)
Chris@102:             "sub     %[result], %[original], %[value]\n"  // result = original - value
Chris@102:             "sxtb    %[result], %[result]\n"              // sign extend result from 8 to 32 bits
Chris@102:             "strex   %[tmp], %[result], %[storage]\n"     // *(&storage) = result, tmp = store failed
Chris@102:             "teq     %[tmp], #0\n"                        // flags = tmp==0
Chris@102:             "bne     1b\n"                                // if (!flags.equal) goto retry
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp])
Chris@102:             : [original] "=&r" (original),  // %0
Chris@102:               [result] "=&r" (result),      // %1
Chris@102:               [tmp] "=&l" (tmp),            // %2
Chris@102:               [storage] "+Q" (storage)      // %3
Chris@102:             : [value] "r" (v)               // %4
Chris@102:             : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
Chris@102:         );
Chris@102:         fence_after(order);
Chris@102:         return original;
Chris@102:     }
Chris@102: };
Chris@102: 
Chris@102: 
Chris@102: template< >
Chris@102: struct operations< 2u, false > :
Chris@102:     public operations< 4u, false >
Chris@102: {
Chris@102:     typedef operations< 4u, false > base_type;
Chris@102:     typedef base_type::storage_type storage_type;
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         fence_before(order);
Chris@102:         uint32_t tmp;
Chris@102:         storage_type original, result;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp])
Chris@102:             "1:\n"
Chris@102:             "ldrex   %[original], %[storage]\n"           // original = *(&storage)
Chris@102:             "add     %[result], %[original], %[value]\n"  // result = original + value
Chris@102:             "uxth    %[result], %[result]\n"              // zero extend result from 16 to 32 bits
Chris@102:             "strex   %[tmp], %[result], %[storage]\n"     // *(&storage) = result, tmp = store failed
Chris@102:             "teq     %[tmp], #0\n"                        // flags = tmp==0
Chris@102:             "bne     1b\n"                                // if (!flags.equal) goto retry
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp])
Chris@102:             : [original] "=&r" (original),  // %0
Chris@102:               [result] "=&r" (result),      // %1
Chris@102:               [tmp] "=&l" (tmp),            // %2
Chris@102:               [storage] "+Q" (storage)      // %3
Chris@102:             : [value] "r" (v)               // %4
Chris@102:             : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
Chris@102:         );
Chris@102:         fence_after(order);
Chris@102:         return original;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         fence_before(order);
Chris@102:         uint32_t tmp;
Chris@102:         storage_type original, result;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp])
Chris@102:             "1:\n"
Chris@102:             "ldrex   %[original], %[storage]\n"           // original = *(&storage)
Chris@102:             "sub     %[result], %[original], %[value]\n"  // result = original - value
Chris@102:             "uxth    %[result], %[result]\n"              // zero extend result from 16 to 32 bits
Chris@102:             "strex   %[tmp], %[result], %[storage]\n"     // *(&storage) = result, tmp = store failed
Chris@102:             "teq     %[tmp], #0\n"                        // flags = tmp==0
Chris@102:             "bne     1b\n"                                // if (!flags.equal) goto retry
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp])
Chris@102:             : [original] "=&r" (original),  // %0
Chris@102:               [result] "=&r" (result),      // %1
Chris@102:               [tmp] "=&l" (tmp),            // %2
Chris@102:               [storage] "+Q" (storage)      // %3
Chris@102:             : [value] "r" (v)               // %4
Chris@102:             : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
Chris@102:         );
Chris@102:         fence_after(order);
Chris@102:         return original;
Chris@102:     }
Chris@102: };
Chris@102: 
Chris@102: template< >
Chris@102: struct operations< 2u, true > :
Chris@102:     public operations< 4u, true >
Chris@102: {
Chris@102:     typedef operations< 4u, true > base_type;
Chris@102:     typedef base_type::storage_type storage_type;
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         fence_before(order);
Chris@102:         uint32_t tmp;
Chris@102:         storage_type original, result;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp])
Chris@102:             "1:\n"
Chris@102:             "ldrex   %[original], %[storage]\n"           // original = *(&storage)
Chris@102:             "add     %[result], %[original], %[value]\n"  // result = original + value
Chris@102:             "sxth    %[result], %[result]\n"              // sign extend result from 16 to 32 bits
Chris@102:             "strex   %[tmp], %[result], %[storage]\n"     // *(&storage) = result, tmp = store failed
Chris@102:             "teq     %[tmp], #0\n"                        // flags = tmp==0
Chris@102:             "bne     1b\n"                                // if (!flags.equal) goto retry
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp])
Chris@102:             : [original] "=&r" (original),  // %0
Chris@102:               [result] "=&r" (result),      // %1
Chris@102:               [tmp] "=&l" (tmp),            // %2
Chris@102:               [storage] "+Q" (storage)      // %3
Chris@102:             : [value] "r" (v)               // %4
Chris@102:             : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
Chris@102:         );
Chris@102:         fence_after(order);
Chris@102:         return original;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         fence_before(order);
Chris@102:         uint32_t tmp;
Chris@102:         storage_type original, result;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%[tmp])
Chris@102:             "1:\n"
Chris@102:             "ldrex   %[original], %[storage]\n"           // original = *(&storage)
Chris@102:             "sub     %[result], %[original], %[value]\n"  // result = original - value
Chris@102:             "sxth    %[result], %[result]\n"              // sign extend result from 16 to 32 bits
Chris@102:             "strex   %[tmp], %[result], %[storage]\n"     // *(&storage) = result, tmp = store failed
Chris@102:             "teq     %[tmp], #0\n"                        // flags = tmp==0
Chris@102:             "bne     1b\n"                                // if (!flags.equal) goto retry
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%[tmp])
Chris@102:             : [original] "=&r" (original),  // %0
Chris@102:               [result] "=&r" (result),      // %1
Chris@102:               [tmp] "=&l" (tmp),            // %2
Chris@102:               [storage] "+Q" (storage)      // %3
Chris@102:             : [value] "r" (v)               // %4
Chris@102:             : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC
Chris@102:         );
Chris@102:         fence_after(order);
Chris@102:         return original;
Chris@102:     }
Chris@102: };
Chris@102: 
Chris@102: 
Chris@102: #if defined(BOOST_ATOMIC_DETAIL_ARM_HAS_LDREXD_STREXD)
Chris@102: 
Chris@102: // Unlike 32-bit operations, for 64-bit loads and stores we must use ldrexd/strexd.
Chris@102: // Any other instructions result in a non-atomic sequence of 32-bit accesses.
Chris@102: // See "ARM Architecture Reference Manual ARMv7-A and ARMv7-R edition",
Chris@102: // Section A3.5.3 "Atomicity in the ARM architecture".
Chris@102: 
Chris@102: // In the asm blocks below we have to use 32-bit register pairs to compose 64-bit values.
Chris@102: // In order to pass the 64-bit operands to/from asm blocks, we use undocumented gcc feature:
Chris@102: // the lower half (Rt) of the operand is accessible normally, via the numbered placeholder (e.g. %0),
Chris@102: // and the upper half (Rt2) - via the same placeholder with an 'H' after the '%' sign (e.g. %H0).
Chris@102: // See: http://hardwarebug.org/2010/07/06/arm-inline-asm-secrets/
Chris@102: 
Chris@102: template< bool Signed >
Chris@102: struct operations< 8u, Signed > :
Chris@102:     public gcc_arm_operations_base
Chris@102: {
Chris@102:     typedef typename make_storage_type< 8u, Signed >::type storage_type;
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE void store(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         exchange(storage, v, order);
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type load(storage_type const volatile& storage, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         storage_type original;
Chris@102:         uint32_t tmp;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%0)
Chris@102:             "ldrexd %1, %H1, [%2]\n"
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%0)
Chris@102:             : BOOST_ATOMIC_DETAIL_ARM_ASM_TMPREG_CONSTRAINT(tmp), // %0
Chris@102:               "=&r" (original)   // %1
Chris@102:             : "r" (&storage)     // %2
Chris@102:         );
Chris@102:         fence_after(order);
Chris@102:         return original;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type exchange(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         storage_type original;
Chris@102:         fence_before(order);
Chris@102:         uint32_t tmp;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%0)
Chris@102:             "1:\n"
Chris@102:             "ldrexd %1, %H1, [%3]\n"        // load the original value
Chris@102:             "strexd %0, %2, %H2, [%3]\n"    // store the replacement, tmp = store failed
Chris@102:             "teq    %0, #0\n"               // check if store succeeded
Chris@102:             "bne    1b\n"
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%0)
Chris@102:             : BOOST_ATOMIC_DETAIL_ARM_ASM_TMPREG_CONSTRAINT(tmp), // %0
Chris@102:               "=&r" (original)   // %1
Chris@102:             : "r" (v),           // %2
Chris@102:               "r" (&storage)     // %3
Chris@102:             : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
Chris@102:         );
Chris@102:         fence_after(order);
Chris@102:         return original;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE bool compare_exchange_weak(
Chris@102:         storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         fence_before(success_order);
Chris@102:         uint32_t tmp;
Chris@102:         storage_type original, old_val = expected;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%0)
Chris@102:             "ldrexd   %1, %H1, [%3]\n"               // original = *(&storage)
Chris@102:             "cmp      %1, %2\n"                      // flags = original.lo==old_val.lo
Chris@102:             "ittt     eq\n"                          // [hint that the following 3 instructions are conditional on flags.equal]
Chris@102:             "cmpeq    %H1, %H2\n"                    // if (flags.equal) flags = original.hi==old_val.hi
Chris@102:             "strexdeq %0, %4, %H4, [%3]\n"           // if (flags.equal) *(&storage) = desired, tmp = store failed
Chris@102:             "teqeq    %0, #0\n"                      // if (flags.equal) flags = tmp==0
Chris@102:             "ite      eq\n"                          // [hint that the following 2 instructions are conditional on flags.equal]
Chris@102:             "moveq    %2, #1\n"                      // if (flags.equal) old_val.lo = 1
Chris@102:             "movne    %2, #0\n"                      // if (!flags.equal) old_val.lo = 0
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%0)
Chris@102:             : BOOST_ATOMIC_DETAIL_ARM_ASM_TMPREG_CONSTRAINT(tmp), // %0
Chris@102:               "=&r" (original),  // %1
Chris@102:               "+r" (old_val)     // %2
Chris@102:             : "r" (&storage),    // %3
Chris@102:               "r" (desired)      // %4
Chris@102:             : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
Chris@102:         );
Chris@102:         const uint32_t success = (uint32_t)old_val;
Chris@102:         if (success)
Chris@102:             fence_after(success_order);
Chris@102:         else
Chris@102:             fence_after(failure_order);
Chris@102:         expected = original;
Chris@102:         return !!success;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE bool compare_exchange_strong(
Chris@102:         storage_type volatile& storage, storage_type& expected, storage_type desired, memory_order success_order, memory_order failure_order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         fence_before(success_order);
Chris@102:         uint32_t tmp;
Chris@102:         storage_type original, old_val = expected;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%0)
Chris@102:             "1:\n"
Chris@102:             "ldrexd  %1, %H1, [%3]\n"               // original = *(&storage)
Chris@102:             "cmp     %1, %2\n"                      // flags = original.lo==old_val.lo
Chris@102:             "it      eq\n"                          // [hint that the following instruction is conditional on flags.equal]
Chris@102:             "cmpeq   %H1, %H2\n"                    // if (flags.equal) flags = original.hi==old_val.hi
Chris@102:             "bne     2f\n"                          // if (!flags.equal) goto end
Chris@102:             "strexd  %0, %4, %H4, [%3]\n"           // *(&storage) = desired, tmp = store failed
Chris@102:             "teq     %0, #0\n"                      // flags.equal = tmp == 0
Chris@102:             "bne     1b\n"                          // if (flags.equal) goto retry
Chris@102:             "2:\n"
Chris@102:             "ite      eq\n"                         // [hint that the following 2 instructions are conditional on flags.equal]
Chris@102:             "moveq    %2, #1\n"                     // if (flags.equal) old_val.lo = 1
Chris@102:             "movne    %2, #0\n"                     // if (!flags.equal) old_val.lo = 0
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%0)
Chris@102:             : BOOST_ATOMIC_DETAIL_ARM_ASM_TMPREG_CONSTRAINT(tmp), // %0
Chris@102:               "=&r" (original),  // %1
Chris@102:               "+r" (old_val)     // %2
Chris@102:             : "r" (&storage),    // %3
Chris@102:               "r" (desired)      // %4
Chris@102:             : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
Chris@102:         );
Chris@102:         const uint32_t success = (uint32_t)old_val;
Chris@102:         if (success)
Chris@102:             fence_after(success_order);
Chris@102:         else
Chris@102:             fence_after(failure_order);
Chris@102:         expected = original;
Chris@102:         return !!success;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_add(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         fence_before(order);
Chris@102:         storage_type original, result;
Chris@102:         uint32_t tmp;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%0)
Chris@102:             "1:\n"
Chris@102:             "ldrexd  %1, %H1, [%3]\n"               // original = *(&storage)
Chris@102:             "adds    %2, %1, %4\n"                  // result = original + value
Chris@102:             "adc     %H2, %H1, %H4\n"
Chris@102:             "strexd  %0, %2, %H2, [%3]\n"           // *(&storage) = result, tmp = store failed
Chris@102:             "teq     %0, #0\n"                      // flags = tmp==0
Chris@102:             "bne     1b\n"                          // if (!flags.equal) goto retry
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%0)
Chris@102:             : BOOST_ATOMIC_DETAIL_ARM_ASM_TMPREG_CONSTRAINT(tmp), // %0
Chris@102:               "=&r" (original),  // %1
Chris@102:               "=&r" (result)     // %2
Chris@102:             : "r" (&storage),    // %3
Chris@102:               "r" (v)            // %4
Chris@102:             : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
Chris@102:         );
Chris@102:         fence_after(order);
Chris@102:         return original;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_sub(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         fence_before(order);
Chris@102:         storage_type original, result;
Chris@102:         uint32_t tmp;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%0)
Chris@102:             "1:\n"
Chris@102:             "ldrexd  %1, %H1, [%3]\n"               // original = *(&storage)
Chris@102:             "subs    %2, %1, %4\n"                  // result = original - value
Chris@102:             "sbc     %H2, %H1, %H4\n"
Chris@102:             "strexd  %0, %2, %H2, [%3]\n"           // *(&storage) = result, tmp = store failed
Chris@102:             "teq     %0, #0\n"                      // flags = tmp==0
Chris@102:             "bne     1b\n"                          // if (!flags.equal) goto retry
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%0)
Chris@102:             : BOOST_ATOMIC_DETAIL_ARM_ASM_TMPREG_CONSTRAINT(tmp), // %0
Chris@102:               "=&r" (original),  // %1
Chris@102:               "=&r" (result)     // %2
Chris@102:             : "r" (&storage),    // %3
Chris@102:               "r" (v)            // %4
Chris@102:             : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
Chris@102:         );
Chris@102:         fence_after(order);
Chris@102:         return original;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_and(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         fence_before(order);
Chris@102:         storage_type original, result;
Chris@102:         uint32_t tmp;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%0)
Chris@102:             "1:\n"
Chris@102:             "ldrexd  %1, %H1, [%3]\n"               // original = *(&storage)
Chris@102:             "and     %2, %1, %4\n"                  // result = original & value
Chris@102:             "and     %H2, %H1, %H4\n"
Chris@102:             "strexd  %0, %2, %H2, [%3]\n"           // *(&storage) = result, tmp = store failed
Chris@102:             "teq     %0, #0\n"                      // flags = tmp==0
Chris@102:             "bne     1b\n"                          // if (!flags.equal) goto retry
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%0)
Chris@102:             : BOOST_ATOMIC_DETAIL_ARM_ASM_TMPREG_CONSTRAINT(tmp), // %0
Chris@102:               "=&r" (original),  // %1
Chris@102:               "=&r" (result)     // %2
Chris@102:             : "r" (&storage),    // %3
Chris@102:               "r" (v)            // %4
Chris@102:             : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
Chris@102:         );
Chris@102:         fence_after(order);
Chris@102:         return original;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_or(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         fence_before(order);
Chris@102:         storage_type original, result;
Chris@102:         uint32_t tmp;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%0)
Chris@102:             "1:\n"
Chris@102:             "ldrexd  %1, %H1, [%3]\n"               // original = *(&storage)
Chris@102:             "orr     %2, %1, %4\n"                  // result = original | value
Chris@102:             "orr     %H2, %H1, %H4\n"
Chris@102:             "strexd  %0, %2, %H2, [%3]\n"           // *(&storage) = result, tmp = store failed
Chris@102:             "teq     %0, #0\n"                      // flags = tmp==0
Chris@102:             "bne     1b\n"                          // if (!flags.equal) goto retry
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%0)
Chris@102:             : BOOST_ATOMIC_DETAIL_ARM_ASM_TMPREG_CONSTRAINT(tmp), // %0
Chris@102:               "=&r" (original),  // %1
Chris@102:               "=&r" (result)     // %2
Chris@102:             : "r" (&storage),    // %3
Chris@102:               "r" (v)            // %4
Chris@102:             : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
Chris@102:         );
Chris@102:         fence_after(order);
Chris@102:         return original;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE storage_type fetch_xor(storage_type volatile& storage, storage_type v, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         fence_before(order);
Chris@102:         storage_type original, result;
Chris@102:         uint32_t tmp;
Chris@102:         __asm__ __volatile__
Chris@102:         (
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_START(%0)
Chris@102:             "1:\n"
Chris@102:             "ldrexd  %1, %H1, [%3]\n"               // original = *(&storage)
Chris@102:             "eor     %2, %1, %4\n"                  // result = original ^ value
Chris@102:             "eor     %H2, %H1, %H4\n"
Chris@102:             "strexd  %0, %2, %H2, [%3]\n"           // *(&storage) = result, tmp = store failed
Chris@102:             "teq     %0, #0\n"                      // flags = tmp==0
Chris@102:             "bne     1b\n"                          // if (!flags.equal) goto retry
Chris@102:             BOOST_ATOMIC_DETAIL_ARM_ASM_END(%0)
Chris@102:             : BOOST_ATOMIC_DETAIL_ARM_ASM_TMPREG_CONSTRAINT(tmp), // %0
Chris@102:               "=&r" (original),  // %1
Chris@102:               "=&r" (result)     // %2
Chris@102:             : "r" (&storage),    // %3
Chris@102:               "r" (v)            // %4
Chris@102:             : BOOST_ATOMIC_DETAIL_ASM_CLOBBER_CC_COMMA "memory"
Chris@102:         );
Chris@102:         fence_after(order);
Chris@102:         return original;
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE bool test_and_set(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return !!exchange(storage, (storage_type)1, order);
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE void clear(storage_type volatile& storage, memory_order order) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         store(storage, 0, order);
Chris@102:     }
Chris@102: 
Chris@102:     static BOOST_FORCEINLINE bool is_lock_free(storage_type const volatile&) BOOST_NOEXCEPT
Chris@102:     {
Chris@102:         return true;
Chris@102:     }
Chris@102: };
Chris@102: 
Chris@102: #endif // defined(BOOST_ATOMIC_DETAIL_ARM_HAS_LDREXD_STREXD)
Chris@102: 
Chris@102: 
Chris@102: BOOST_FORCEINLINE void thread_fence(memory_order order) BOOST_NOEXCEPT
Chris@102: {
Chris@102:     if (order != memory_order_relaxed)
Chris@102:         gcc_arm_operations_base::hardware_full_fence();
Chris@102: }
Chris@102: 
Chris@102: BOOST_FORCEINLINE void signal_fence(memory_order order) BOOST_NOEXCEPT
Chris@102: {
Chris@102:     if (order != memory_order_relaxed)
Chris@102:         __asm__ __volatile__ ("" ::: "memory");
Chris@102: }
Chris@102: 
Chris@102: } // namespace detail
Chris@102: } // namespace atomics
Chris@102: } // namespace boost
Chris@102: 
Chris@102: #endif // BOOST_ATOMIC_DETAIL_OPS_GCC_ARM_HPP_INCLUDED_