/** @file * IPRT - Assembly Functions. */ /* * Copyright (C) 2006-2023 Oracle and/or its affiliates. * * This file is part of VirtualBox base platform packages, as * available from https://www.virtualbox.org. * * This program is free software; you can redistribute it and/or * modify it under the terms of the GNU General Public License * as published by the Free Software Foundation, in version 3 of the * License. * * This program is distributed in the hope that it will be useful, but * WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU * General Public License for more details. * * You should have received a copy of the GNU General Public License * along with this program; if not, see . * * The contents of this file may alternatively be used under the terms * of the Common Development and Distribution License Version 1.0 * (CDDL), a copy of it is provided in the "COPYING.CDDL" file included * in the VirtualBox distribution, in which case the provisions of the * CDDL are applicable instead of those of the GPL. * * You may elect to license modified versions of this file under the * terms and conditions of either the GPL or the CDDL or both. * * SPDX-License-Identifier: GPL-3.0-only OR CDDL-1.0 */ #ifndef IPRT_INCLUDED_asm_h #define IPRT_INCLUDED_asm_h #ifndef RT_WITHOUT_PRAGMA_ONCE # pragma once #endif #include #include #include /** @def RT_INLINE_ASM_USES_INTRIN * Defined as 1 if we're using a _MSC_VER 1400. * Otherwise defined as 0. */ /* Solaris 10 header ugliness */ #ifdef u # undef u #endif #if defined(_MSC_VER) && RT_INLINE_ASM_USES_INTRIN /* Emit the intrinsics at all optimization levels. */ # include # pragma intrinsic(_ReadWriteBarrier) # pragma intrinsic(__cpuid) # pragma intrinsic(__stosd) # pragma intrinsic(__stosw) # pragma intrinsic(__stosb) # pragma intrinsic(_BitScanForward) # pragma intrinsic(_BitScanReverse) # pragma intrinsic(_bittest) # pragma intrinsic(_bittestandset) # pragma intrinsic(_bittestandreset) # pragma intrinsic(_bittestandcomplement) # pragma intrinsic(_byteswap_ushort) # pragma intrinsic(_byteswap_ulong) # pragma intrinsic(_interlockedbittestandset) # pragma intrinsic(_interlockedbittestandreset) # pragma intrinsic(_InterlockedAnd) # pragma intrinsic(_InterlockedOr) # pragma intrinsic(_InterlockedXor) # pragma intrinsic(_InterlockedIncrement) # pragma intrinsic(_InterlockedDecrement) # pragma intrinsic(_InterlockedExchange) # pragma intrinsic(_InterlockedExchangeAdd) # pragma intrinsic(_InterlockedCompareExchange) # pragma intrinsic(_InterlockedCompareExchange8) # pragma intrinsic(_InterlockedCompareExchange16) # pragma intrinsic(_InterlockedCompareExchange64) # pragma intrinsic(_rotl) # pragma intrinsic(_rotr) # pragma intrinsic(_rotl64) # pragma intrinsic(_rotr64) # ifdef RT_ARCH_AMD64 # pragma intrinsic(__stosq) # pragma intrinsic(_byteswap_uint64) # pragma intrinsic(_InterlockedCompareExchange128) # pragma intrinsic(_InterlockedExchange64) # pragma intrinsic(_InterlockedExchangeAdd64) # pragma intrinsic(_InterlockedAnd64) # pragma intrinsic(_InterlockedOr64) # pragma intrinsic(_InterlockedIncrement64) # pragma intrinsic(_InterlockedDecrement64) # endif #endif #if (defined(RT_ARCH_ARM64) && defined(RT_OS_DARWIN)) || defined(DOXYGEN_RUNNING) /** @def RTASM_ARM64_USE_FEAT_LSE * Use instructions from the FEAT_LSE set to implement atomic operations, * assuming that the host CPU always supports these. */ # define RTASM_ARM64_USE_FEAT_LSE 1 /** @def RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB * Set to use DMB w/o barrier in most places and rely on the acquire-release * aspects to do the serializing. The assumption is that the tstRTInline * benchmark may be skewing the results testing an unusual scenario. */ # define RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB 1 #endif /* * Undefine all symbols we have Watcom C/C++ #pragma aux'es for. */ #if defined(__WATCOMC__) && ARCH_BITS == 16 && defined(RT_ARCH_X86) # include "asm-watcom-x86-16.h" #elif defined(__WATCOMC__) && ARCH_BITS == 32 && defined(RT_ARCH_X86) # include "asm-watcom-x86-32.h" #endif /** @defgroup grp_rt_asm ASM - Assembly Routines * @ingroup grp_rt * * @remarks The difference between ordered and unordered atomic operations are * that the former will complete outstanding reads and writes before * continuing while the latter doesn't make any promises about the * order. Ordered operations doesn't, it seems, make any 100% promise * wrt to whether the operation will complete before any subsequent * memory access. (please, correct if wrong.) * * ASMAtomicSomething operations are all ordered, while * ASMAtomicUoSomething are unordered (note the Uo). * * Please note that ordered operations does not necessarily imply a * compiler (memory) barrier. The user has to use the * ASMCompilerBarrier() macro when that is deemed necessary. * * @remarks Some remarks about __volatile__: Without this keyword gcc is allowed * to reorder or even optimize assembler instructions away. For * instance, in the following code the second rdmsr instruction is * optimized away because gcc treats that instruction as deterministic: * * @code * static inline uint64_t rdmsr_low(int idx) * { * uint32_t low; * __asm__ ("rdmsr" : "=a"(low) : "c"(idx) : "edx"); * } * ... * uint32_t msr1 = rdmsr_low(1); * foo(msr1); * msr1 = rdmsr_low(1); * bar(msr1); * @endcode * * The input parameter of rdmsr_low is the same for both calls and * therefore gcc will use the result of the first call as input * parameter for bar() as well. For rdmsr this is not acceptable as * this instruction is _not_ deterministic. This applies to reading * machine status information in general. * * @{ */ /** @def RT_INLINE_ASM_GCC_4_3_X_X86 * Used to work around some 4.3.x register allocation issues in this version of * the compiler. So far this workaround is still required for 4.4 and 4.5 but * definitely not for 5.x */ #if (RT_GNUC_PREREQ(4, 3) && !RT_GNUC_PREREQ(5, 0) && defined(__i386__)) # define RT_INLINE_ASM_GCC_4_3_X_X86 1 #else # define RT_INLINE_ASM_GCC_4_3_X_X86 0 #endif /** @def RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC * i686-apple-darwin9-gcc-4.0.1 (GCC) 4.0.1 (Apple Inc. build 5493) screws up * RTSemRWRequestWrite semsemrw-lockless-generic.cpp in release builds. PIC * mode, x86. * * Some gcc 4.3.x versions may have register allocation issues with cmpxchg8b * when in PIC mode on x86. */ #ifndef RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC # if defined(DOXYGEN_RUNNING) || defined(__WATCOMC__) /* Watcom has trouble with the expression below */ # define RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC 1 # elif defined(_MSC_VER) /* Visual C++ has trouble too, but it'll only tell us when C4688 is enabled. */ # define RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC 0 # elif ( (defined(PIC) || defined(__PIC__)) \ && defined(RT_ARCH_X86) \ && ( RT_INLINE_ASM_GCC_4_3_X_X86 \ || defined(RT_OS_DARWIN)) ) # define RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC 1 # else # define RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC 0 # endif #endif /* * ARM is great fun. */ #if defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) # define RTASM_ARM_NO_BARRIER # ifdef RT_ARCH_ARM64 # define RTASM_ARM_NO_BARRIER_IN_REG # define RTASM_ARM_NO_BARRIER_COMMA_IN_REG # define RTASM_ARM_DSB_SY "dsb sy\n\t" # define RTASM_ARM_DSB_SY_IN_REG # define RTASM_ARM_DSB_SY_COMMA_IN_REG # define RTASM_ARM_DMB_SY "dmb sy\n\t" # define RTASM_ARM_DMB_SY_IN_REG # define RTASM_ARM_DMB_SY_COMMA_IN_REG # define RTASM_ARM_DMB_ST "dmb st\n\t" # define RTASM_ARM_DMB_ST_IN_REG # define RTASM_ARM_DMB_ST_COMMA_IN_REG # define RTASM_ARM_DMB_LD "dmb ld\n\t" # define RTASM_ARM_DMB_LD_IN_REG # define RTASM_ARM_DMB_LD_COMMA_IN_REG # define RTASM_ARM_PICK_6432(expr64, expr32) expr64 # define RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(name, a_pu32Mem, barrier_type, modify64, modify32, in_reg) \ uint32_t rcSpill; \ uint32_t u32NewRet; \ __asm__ __volatile__("Ltry_again_" #name "_%=:\n\t" \ RTASM_ARM_##barrier_type /* before lable? */ \ "ldaxr %w[uNew], %[pMem]\n\t" \ modify64 \ "stlxr %w[rc], %w[uNew], %[pMem]\n\t" \ "cbnz %w[rc], Ltry_again_" #name "_%=\n\t" \ : [pMem] "+Q" (*a_pu32Mem) \ , [uNew] "=&r" (u32NewRet) \ , [rc] "=&r" (rcSpill) \ : in_reg \ : "cc") # define RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(name, a_pu32Mem, barrier_type, modify64, modify32, in_reg) \ uint32_t rcSpill; \ uint32_t u32OldRet; \ uint32_t u32NewSpill; \ __asm__ __volatile__("Ltry_again_" #name "_%=:\n\t" \ RTASM_ARM_##barrier_type /* before lable? */ \ "ldaxr %w[uOld], %[pMem]\n\t" \ modify64 \ "stlxr %w[rc], %w[uNew], %[pMem]\n\t" \ "cbnz %w[rc], Ltry_again_" #name "_%=\n\t" \ : [pMem] "+Q" (*a_pu32Mem) \ , [uOld] "=&r" (u32OldRet) \ , [uNew] "=&r" (u32NewSpill) \ , [rc] "=&r" (rcSpill) \ : in_reg \ : "cc") # define RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_64(name, a_pu64Mem, barrier_type, modify64, modify32, in_reg) \ uint32_t rcSpill; \ uint64_t u64NewRet; \ __asm__ __volatile__("Ltry_again_" #name "_%=:\n\t" \ RTASM_ARM_##barrier_type /* before lable? */ \ "ldaxr %[uNew], %[pMem]\n\t" \ modify64 \ "stlxr %w[rc], %[uNew], %[pMem]\n\t" \ "cbnz %w[rc], Ltry_again_" #name "_%=\n\t" \ : [pMem] "+Q" (*a_pu64Mem) \ , [uNew] "=&r" (u64NewRet) \ , [rc] "=&r" (rcSpill) \ : in_reg \ : "cc") # define RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_64(name, a_pu64Mem, barrier_type, modify64, modify32, in_reg) \ uint32_t rcSpill; \ uint64_t u64OldRet; \ uint64_t u64NewSpill; \ __asm__ __volatile__("Ltry_again_" #name "_%=:\n\t" \ RTASM_ARM_##barrier_type /* before lable? */ \ "ldaxr %[uOld], %[pMem]\n\t" \ modify64 \ "stlxr %w[rc], %[uNew], %[pMem]\n\t" \ "cbnz %w[rc], Ltry_again_" #name "_%=\n\t" \ : [pMem] "+Q" (*a_pu64Mem) \ , [uOld] "=&r" (u64OldRet) \ , [uNew] "=&r" (u64NewSpill) \ , [rc] "=&r" (rcSpill) \ : in_reg \ : "cc") # else /* RT_ARCH_ARM32 */ # define RTASM_ARM_PICK_6432(expr64, expr32) expr32 # if RT_ARCH_ARM32 >= 7 # warning armv7 # define RTASM_ARM_NO_BARRIER_IN_REG # define RTASM_ARM_NO_BARRIER_COMMA_IN_REG # define RTASM_ARM_DSB_SY "dsb sy\n\t" # define RTASM_ARM_DSB_SY_IN_REG "X" (0xfade) # define RTASM_ARM_DMB_SY "dmb sy\n\t" # define RTASM_ARM_DMB_SY_IN_REG "X" (0xfade) # define RTASM_ARM_DMB_ST "dmb st\n\t" # define RTASM_ARM_DMB_ST_IN_REG "X" (0xfade) # define RTASM_ARM_DMB_LD "dmb ld\n\t" # define RTASM_ARM_DMB_LD_IN_REG "X" (0xfade) # elif RT_ARCH_ARM32 >= 6 # warning armv6 # define RTASM_ARM_DSB_SY "mcr p15, 0, %[uZero], c7, c10, 4\n\t" # define RTASM_ARM_DSB_SY_IN_REG [uZero] "r" (0) # define RTASM_ARM_DMB_SY "mcr p15, 0, %[uZero], c7, c10, 5\n\t" # define RTASM_ARM_DMB_SY_IN_REG [uZero] "r" (0) # define RTASM_ARM_DMB_ST RTASM_ARM_DMB_SY # define RTASM_ARM_DMB_ST_IN_REG RTASM_ARM_DMB_SY_IN_REG # define RTASM_ARM_DMB_LD RTASM_ARM_DMB_SY # define RTASM_ARM_DMB_LD_IN_REG RTASM_ARM_DMB_SY_IN_REG # elif RT_ARCH_ARM32 >= 4 # warning armv5 or older # define RTASM_ARM_DSB_SY "mcr p15, 0, %[uZero], c7, c10, 4\n\t" # define RTASM_ARM_DSB_SY_IN_REG [uZero] "r" (0) # define RTASM_ARM_DMB_SY RTASM_ARM_DSB_SY # define RTASM_ARM_DMB_SY_IN_REG RTASM_ARM_DSB_SY_IN_REG # define RTASM_ARM_DMB_ST RTASM_ARM_DSB_SY # define RTASM_ARM_DMB_ST_IN_REG RTASM_ARM_DSB_SY_IN_REG # define RTASM_ARM_DMB_LD RTASM_ARM_DSB_SY # define RTASM_ARM_DMB_LD_IN_REG RTASM_ARM_DSB_SY_IN_REG # else # error "huh? Odd RT_ARCH_ARM32 value!" # endif # define RTASM_ARM_DSB_SY_COMMA_IN_REG , RTASM_ARM_DSB_SY_IN_REG # define RTASM_ARM_DMB_SY_COMMA_IN_REG , RTASM_ARM_DMB_SY_IN_REG # define RTASM_ARM_DMB_ST_COMMA_IN_REG , RTASM_ARM_DMB_ST_IN_REG # define RTASM_ARM_DMB_LD_COMMA_IN_REG , RTASM_ARM_DMB_LD_IN_REG # define RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(name, a_pu32Mem, barrier_type, modify64, modify32, in_reg) \ uint32_t rcSpill; \ uint32_t u32NewRet; \ __asm__ __volatile__("Ltry_again_" #name "_%=:\n\t" \ RT_CONCAT(RTASM_ARM_,barrier_type) /* before lable? */ \ "ldrex %[uNew], %[pMem]\n\t" \ modify32 \ "strex %[rc], %[uNew], %[pMem]\n\t" \ "cmp %[rc], #0\n\t" \ "bne Ltry_again_" #name "_%=\n\t" \ : [pMem] "+m" (*a_pu32Mem) \ , [uNew] "=&r" (u32NewRet) \ , [rc] "=&r" (rcSpill) \ : RT_CONCAT3(RTASM_ARM_,barrier_type,_IN_REG) \ , in_reg \ : "cc") # define RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(name, a_pu32Mem, barrier_type, modify64, modify32, in_reg) \ uint32_t rcSpill; \ uint32_t u32OldRet; \ uint32_t u32NewSpill; \ __asm__ __volatile__("Ltry_again_" #name "_%=:\n\t" \ RT_CONCAT(RTASM_ARM_,barrier_type) /* before lable? */ \ "ldrex %[uOld], %[pMem]\n\t" \ modify32 \ "strex %[rc], %[uNew], %[pMem]\n\t" \ "cmp %[rc], #0\n\t" \ "bne Ltry_again_" #name "_%=\n\t" \ : [pMem] "+m" (*a_pu32Mem) \ , [uOld] "=&r" (u32OldRet) \ , [uNew] "=&r" (u32NewSpill) \ , [rc] "=&r" (rcSpill) \ : RT_CONCAT3(RTASM_ARM_,barrier_type,_IN_REG) \ , in_reg \ : "cc") # define RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_64(name, a_pu64Mem, barrier_type, modify64, modify32, in_reg) \ uint32_t rcSpill; \ uint64_t u64NewRet; \ __asm__ __volatile__("Ltry_again_" #name "_%=:\n\t" \ RT_CONCAT(RTASM_ARM_,barrier_type) /* before lable? */ \ "ldrexd %[uNew], %H[uNew], %[pMem]\n\t" \ modify32 \ "strexd %[rc], %[uNew], %H[uNew], %[pMem]\n\t" \ "cmp %[rc], #0\n\t" \ "bne Ltry_again_" #name "_%=\n\t" \ : [pMem] "+m" (*a_pu64Mem), \ [uNew] "=&r" (u64NewRet), \ [rc] "=&r" (rcSpill) \ : RT_CONCAT3(RTASM_ARM_,barrier_type,_IN_REG) \ , in_reg \ : "cc") # define RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_64(name, a_pu64Mem, barrier_type, modify64, modify32, in_reg) \ uint32_t rcSpill; \ uint64_t u64OldRet; \ uint64_t u64NewSpill; \ __asm__ __volatile__("Ltry_again_" #name "_%=:\n\t" \ RT_CONCAT(RTASM_ARM_,barrier_type) /* before lable? */ \ "ldrexd %[uOld], %H[uOld], %[pMem]\n\t" \ modify32 \ "strexd %[rc], %[uNew], %H[uNew], %[pMem]\n\t" \ "cmp %[rc], #0\n\t" \ "bne Ltry_again_" #name "_%=\n\t" \ : [pMem] "+m" (*a_pu64Mem), \ [uOld] "=&r" (u64OldRet), \ [uNew] "=&r" (u64NewSpill), \ [rc] "=&r" (rcSpill) \ : RT_CONCAT3(RTASM_ARM_,barrier_type,_IN_REG) \ , in_reg \ : "cc") # endif /* RT_ARCH_ARM32 */ #endif /** @def ASMReturnAddress * Gets the return address of the current (or calling if you like) function or method. */ #ifdef _MSC_VER # ifdef __cplusplus extern "C" # endif void * _ReturnAddress(void); # pragma intrinsic(_ReturnAddress) # define ASMReturnAddress() _ReturnAddress() #elif defined(__GNUC__) || defined(DOXYGEN_RUNNING) # define ASMReturnAddress() __builtin_return_address(0) #elif defined(__WATCOMC__) # define ASMReturnAddress() Watcom_does_not_appear_to_have_intrinsic_return_address_function() #else # error "Unsupported compiler." #endif /** * Compiler memory barrier. * * Ensure that the compiler does not use any cached (register/tmp stack) memory * values or any outstanding writes when returning from this function. * * This function must be used if non-volatile data is modified by a * device or the VMM. Typical cases are port access, MMIO access, * trapping instruction, etc. */ #if RT_INLINE_ASM_GNU_STYLE # define ASMCompilerBarrier() do { __asm__ __volatile__("" : : : "memory"); } while (0) #elif RT_INLINE_ASM_USES_INTRIN # define ASMCompilerBarrier() do { _ReadWriteBarrier(); } while (0) #elif defined(__WATCOMC__) void ASMCompilerBarrier(void); #else /* 2003 should have _ReadWriteBarrier() but I guess we're at 2002 level then... */ DECLINLINE(void) ASMCompilerBarrier(void) RT_NOTHROW_DEF { __asm { } } #endif /** @def ASMBreakpoint * Debugger Breakpoint. * @deprecated Use RT_BREAKPOINT instead. * @internal */ #define ASMBreakpoint() RT_BREAKPOINT() /** * Spinloop hint for platforms that have these, empty function on the other * platforms. * * x86 & AMD64: The PAUSE variant of NOP for helping hyperthreaded CPUs detecting * spin locks. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && (defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86)) RT_ASM_DECL_PRAGMA_WATCOM(void) ASMNopPause(void) RT_NOTHROW_PROTO; #else DECLINLINE(void) ASMNopPause(void) RT_NOTHROW_DEF { # if defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__(".byte 0xf3,0x90\n\t"); # else __asm { _emit 0f3h _emit 090h } # endif # elif defined(RT_ARCH_ARM32) || defined(RT_ARCH_ARM64) __asm__ __volatile__("yield\n\t"); /* ARMv6K+ */ # else /* dummy */ # endif } #endif /** * Atomically Exchange an unsigned 8-bit value, ordered. * * @returns Current *pu8 value * @param pu8 Pointer to the 8-bit variable to update. * @param u8 The 8-bit value to assign to *pu8. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM RT_ASM_DECL_PRAGMA_WATCOM(uint8_t) ASMAtomicXchgU8(volatile uint8_t RT_FAR *pu8, uint8_t u8) RT_NOTHROW_PROTO; #else DECLINLINE(uint8_t) ASMAtomicXchgU8(volatile uint8_t RT_FAR *pu8, uint8_t u8) RT_NOTHROW_DEF { # if defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("xchgb %0, %1\n\t" : "=m" (*pu8) , "=q" (u8) /* =r - busted on g++ (GCC) 3.4.4 20050721 (Red Hat 3.4.4-2) */ : "1" (u8) , "m" (*pu8)); # else __asm { # ifdef RT_ARCH_AMD64 mov rdx, [pu8] mov al, [u8] xchg [rdx], al mov [u8], al # else mov edx, [pu8] mov al, [u8] xchg [edx], al mov [u8], al # endif } # endif return u8; # elif defined(RT_ARCH_ARM32) || defined(RT_ARCH_ARM64) uint32_t uOld; # if defined(RTASM_ARM64_USE_FEAT_LSE) /* SWPALB is ~40% more expensive than the non-LSE variant (M1), but since we have the barrier we shouldn't need that, right? Ordering should be taken care of by the DMB. The SWPB is rather cheap (~70% faster). */ __asm__ __volatile__("Lstart_ASMAtomicXchgU8_%=:\n\t" # if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) "swpalb %w[uNew], %w[uOld], %[pMem]\n\t" # else RTASM_ARM_DMB_SY "swpb %w[uNew], %w[uOld], %[pMem]\n\t" # endif : [pMem] "+Q" (*pu8) , [uOld] "=&r" (uOld) : [uNew] "r" ((uint32_t)u8) : ); # else uint32_t rcSpill; __asm__ __volatile__("Ltry_again_ASMAtomicXchgU8_%=:\n\t" RTASM_ARM_DMB_SY # if defined(RT_ARCH_ARM64) "ldaxrb %w[uOld], %[pMem]\n\t" "stlxrb %w[rc], %w[uNew], %[pMem]\n\t" "cbnz %w[rc], Ltry_again_ASMAtomicXchgU8_%=\n\t" # else "ldrexb %[uOld], %[pMem]\n\t" /* ARMv6+ */ "strexb %[rc], %[uNew], %[pMem]\n\t" "cmp %[rc], #0\n\t" "bne Ltry_again_ASMAtomicXchgU8_%=\n\t" # endif : [pMem] "+Q" (*pu8) , [uOld] "=&r" (uOld) , [rc] "=&r" (rcSpill) : [uNew] "r" ((uint32_t)u8) RTASM_ARM_DMB_SY_COMMA_IN_REG : "cc"); # endif return (uint8_t)uOld; # else # error "Port me" # endif } #endif /** * Atomically Exchange a signed 8-bit value, ordered. * * @returns Current *pu8 value * @param pi8 Pointer to the 8-bit variable to update. * @param i8 The 8-bit value to assign to *pi8. */ DECLINLINE(int8_t) ASMAtomicXchgS8(volatile int8_t RT_FAR *pi8, int8_t i8) RT_NOTHROW_DEF { return (int8_t)ASMAtomicXchgU8((volatile uint8_t RT_FAR *)pi8, (uint8_t)i8); } /** * Atomically Exchange a bool value, ordered. * * @returns Current *pf value * @param pf Pointer to the 8-bit variable to update. * @param f The 8-bit value to assign to *pi8. */ DECLINLINE(bool) ASMAtomicXchgBool(volatile bool RT_FAR *pf, bool f) RT_NOTHROW_DEF { #ifdef _MSC_VER return !!ASMAtomicXchgU8((volatile uint8_t RT_FAR *)pf, (uint8_t)f); #else return (bool)ASMAtomicXchgU8((volatile uint8_t RT_FAR *)pf, (uint8_t)f); #endif } /** * Atomically Exchange an unsigned 16-bit value, ordered. * * @returns Current *pu16 value * @param pu16 Pointer to the 16-bit variable to update. * @param u16 The 16-bit value to assign to *pu16. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM RT_ASM_DECL_PRAGMA_WATCOM(uint16_t) ASMAtomicXchgU16(volatile uint16_t RT_FAR *pu16, uint16_t u16) RT_NOTHROW_PROTO; #else DECLINLINE(uint16_t) ASMAtomicXchgU16(volatile uint16_t RT_FAR *pu16, uint16_t u16) RT_NOTHROW_DEF { # if defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("xchgw %0, %1\n\t" : "=m" (*pu16) , "=r" (u16) : "1" (u16) , "m" (*pu16)); # else __asm { # ifdef RT_ARCH_AMD64 mov rdx, [pu16] mov ax, [u16] xchg [rdx], ax mov [u16], ax # else mov edx, [pu16] mov ax, [u16] xchg [edx], ax mov [u16], ax # endif } # endif return u16; # elif defined(RT_ARCH_ARM32) || defined(RT_ARCH_ARM64) uint32_t uOld; # if defined(RTASM_ARM64_USE_FEAT_LSE) /* SWPALH is ~40% more expensive than the non-LSE variant on an M1, 20% slower if we remove the barrier. But since we have the barrier we shouldn't need that, right? Ordering should be taken care of by the DMB. The SWPH is rather cheap (~70% faster). */ __asm__ __volatile__("Lstart_ASMAtomicXchgU16_%=:\n\t" # if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) "swpalh %w[uNew], %w[uOld], %[pMem]\n\t" # else RTASM_ARM_DMB_SY "swph %w[uNew], %w[uOld], %[pMem]\n\t" # endif : [pMem] "+Q" (*pu16) , [uOld] "=&r" (uOld) : [uNew] "r" ((uint32_t)u16) : ); # else uint32_t rcSpill; __asm__ __volatile__("Ltry_again_ASMAtomicXchgU16_%=:\n\t" RTASM_ARM_DMB_SY # if defined(RT_ARCH_ARM64) "ldaxrh %w[uOld], %[pMem]\n\t" "stlxrh %w[rc], %w[uNew], %[pMem]\n\t" "cbnz %w[rc], Ltry_again_ASMAtomicXchgU16_%=\n\t" # else "ldrexh %[uOld], %[pMem]\n\t" /* ARMv6+ */ "strexh %[rc], %[uNew], %[pMem]\n\t" "cmp %[rc], #0\n\t" "bne Ltry_again_ASMAtomicXchgU16_%=\n\t" # endif : [pMem] "+Q" (*pu16) , [uOld] "=&r" (uOld) , [rc] "=&r" (rcSpill) : [uNew] "r" ((uint32_t)u16) RTASM_ARM_DMB_SY_COMMA_IN_REG : "cc"); # endif return (uint16_t)uOld; # else # error "Port me" # endif } #endif /** * Atomically Exchange a signed 16-bit value, ordered. * * @returns Current *pu16 value * @param pi16 Pointer to the 16-bit variable to update. * @param i16 The 16-bit value to assign to *pi16. */ DECLINLINE(int16_t) ASMAtomicXchgS16(volatile int16_t RT_FAR *pi16, int16_t i16) RT_NOTHROW_DEF { return (int16_t)ASMAtomicXchgU16((volatile uint16_t RT_FAR *)pi16, (uint16_t)i16); } /** * Atomically Exchange an unsigned 32-bit value, ordered. * * @returns Current *pu32 value * @param pu32 Pointer to the 32-bit variable to update. * @param u32 The 32-bit value to assign to *pu32. * * @remarks Does not work on 286 and earlier. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMAtomicXchgU32(volatile uint32_t RT_FAR *pu32, uint32_t u32) RT_NOTHROW_PROTO; #else DECLINLINE(uint32_t) ASMAtomicXchgU32(volatile uint32_t RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF { # if defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("xchgl %0, %1\n\t" : "=m" (*pu32) /** @todo r=bird: +m rather than =m here? */ , "=r" (u32) : "1" (u32) , "m" (*pu32)); # elif RT_INLINE_ASM_USES_INTRIN u32 = _InterlockedExchange((long RT_FAR *)pu32, u32); # else __asm { # ifdef RT_ARCH_AMD64 mov rdx, [pu32] mov eax, u32 xchg [rdx], eax mov [u32], eax # else mov edx, [pu32] mov eax, u32 xchg [edx], eax mov [u32], eax # endif } # endif return u32; # elif defined(RT_ARCH_ARM32) || defined(RT_ARCH_ARM64) uint32_t uOld; # if defined(RTASM_ARM64_USE_FEAT_LSE) /* SWPAL is ~40% more expensive than the non-LSE variant on an M1, 20% slower if we remove the barrier. But since we have the barrier we shouldn't need that, right? Ordering should be taken care of by the DMB. The SWP is rather cheap (~70% faster). */ __asm__ __volatile__("Lstart_ASMAtomicXchgU32_%=:\n\t" # if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) "swpal %w[uNew], %w[uOld], %[pMem]\n\t" # else RTASM_ARM_DMB_SY "swp %w[uNew], %w[uOld], %[pMem]\n\t" # endif : [pMem] "+Q" (*pu32) , [uOld] "=&r" (uOld) : [uNew] "r" (u32) : ); # else uint32_t rcSpill; __asm__ __volatile__("Ltry_again_ASMAtomicXchgU32_%=:\n\t" RTASM_ARM_DMB_SY # if defined(RT_ARCH_ARM64) "ldaxr %w[uOld], %[pMem]\n\t" "stlxr %w[rc], %w[uNew], %[pMem]\n\t" "cbnz %w[rc], Ltry_again_ASMAtomicXchgU32_%=\n\t" # else "ldrex %[uOld], %[pMem]\n\t" /* ARMv6+ */ "strex %[rc], %[uNew], %[pMem]\n\t" "cmp %[rc], #0\n\t" "bne Ltry_again_ASMAtomicXchgU32_%=\n\t" # endif : [pMem] "+Q" (*pu32) , [uOld] "=&r" (uOld) , [rc] "=&r" (rcSpill) : [uNew] "r" (u32) RTASM_ARM_DMB_SY_COMMA_IN_REG : "cc"); # endif return uOld; # else # error "Port me" # endif } #endif /** * Atomically Exchange a signed 32-bit value, ordered. * * @returns Current *pu32 value * @param pi32 Pointer to the 32-bit variable to update. * @param i32 The 32-bit value to assign to *pi32. */ DECLINLINE(int32_t) ASMAtomicXchgS32(volatile int32_t RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF { return (int32_t)ASMAtomicXchgU32((volatile uint32_t RT_FAR *)pi32, (uint32_t)i32); } /** * Atomically Exchange an unsigned 64-bit value, ordered. * * @returns Current *pu64 value * @param pu64 Pointer to the 64-bit variable to update. * @param u64 The 64-bit value to assign to *pu64. * * @remarks Works on 32-bit x86 CPUs starting with Pentium. */ #if (RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN) \ || RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC RT_ASM_DECL_PRAGMA_WATCOM(uint64_t) ASMAtomicXchgU64(volatile uint64_t RT_FAR *pu64, uint64_t u64) RT_NOTHROW_PROTO; #else DECLINLINE(uint64_t) ASMAtomicXchgU64(volatile uint64_t RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF { # if defined(RT_ARCH_AMD64) # if RT_INLINE_ASM_USES_INTRIN return _InterlockedExchange64((__int64 *)pu64, u64); # elif RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("xchgq %0, %1\n\t" : "=m" (*pu64) , "=r" (u64) : "1" (u64) , "m" (*pu64)); return u64; # else __asm { mov rdx, [pu64] mov rax, [u64] xchg [rdx], rax mov [u64], rax } return u64; # endif # elif defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE # if defined(PIC) || defined(__PIC__) uint32_t u32EBX = (uint32_t)u64; __asm__ __volatile__(/*"xchgl %%esi, %5\n\t"*/ "xchgl %%ebx, %3\n\t" "1:\n\t" "lock; cmpxchg8b (%5)\n\t" "jnz 1b\n\t" "movl %3, %%ebx\n\t" /*"xchgl %%esi, %5\n\t"*/ : "=A" (u64) , "=m" (*pu64) : "0" (*pu64) , "m" ( u32EBX ) , "c" ( (uint32_t)(u64 >> 32) ) , "S" (pu64) : "cc"); # else /* !PIC */ __asm__ __volatile__("1:\n\t" "lock; cmpxchg8b %1\n\t" "jnz 1b\n\t" : "=A" (u64) , "=m" (*pu64) : "0" (*pu64) , "b" ( (uint32_t)u64 ) , "c" ( (uint32_t)(u64 >> 32) ) : "cc"); # endif # else __asm { mov ebx, dword ptr [u64] mov ecx, dword ptr [u64 + 4] mov edi, pu64 mov eax, dword ptr [edi] mov edx, dword ptr [edi + 4] retry: lock cmpxchg8b [edi] jnz retry mov dword ptr [u64], eax mov dword ptr [u64 + 4], edx } # endif return u64; # elif defined(RT_ARCH_ARM32) || defined(RT_ARCH_ARM64) uint64_t uOld; # if defined(RTASM_ARM64_USE_FEAT_LSE) /* SWPAL is ~40% more expensive than the non-LSE variant on an M1, 20% slower if we remove the barrier. But since we have the barrier we shouldn't need that, right? Ordering should be taken care of by the DMB. The SWP is rather cheap (~70% faster). */ __asm__ __volatile__("Lstart_ASMAtomicXchgU64_%=:\n\t" # if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) "swpal %[uNew], %[uOld], %[pMem]\n\t" # else RTASM_ARM_DMB_SY "swp %[uNew], %[uOld], %[pMem]\n\t" # endif : [pMem] "+Q" (*pu64) , [uOld] "=&r" (uOld) : [uNew] "r" (u64) : ); # else uint32_t rcSpill; __asm__ __volatile__("Ltry_again_ASMAtomicXchgU64_%=:\n\t" RTASM_ARM_DMB_SY # if defined(RT_ARCH_ARM64) "ldaxr %[uOld], %[pMem]\n\t" "stlxr %w[rc], %[uNew], %[pMem]\n\t" "cbnz %w[rc], Ltry_again_ASMAtomicXchgU64_%=\n\t" # else "ldrexd %[uOld], %H[uOld], %[pMem]\n\t" /* ARMv6+ */ "strexd %[rc], %[uNew], %H[uNew], %[pMem]\n\t" "cmp %[rc], #0\n\t" "bne Ltry_again_ASMAtomicXchgU64_%=\n\t" # endif : [pMem] "+Q" (*pu64) , [uOld] "=&r" (uOld) , [rc] "=&r" (rcSpill) : [uNew] "r" (u64) RTASM_ARM_DMB_SY_COMMA_IN_REG : "cc"); # endif return uOld; # else # error "Port me" # endif } #endif /** * Atomically Exchange an signed 64-bit value, ordered. * * @returns Current *pi64 value * @param pi64 Pointer to the 64-bit variable to update. * @param i64 The 64-bit value to assign to *pi64. */ DECLINLINE(int64_t) ASMAtomicXchgS64(volatile int64_t RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF { return (int64_t)ASMAtomicXchgU64((volatile uint64_t RT_FAR *)pi64, (uint64_t)i64); } /** * Atomically Exchange a size_t value, ordered. * * @returns Current *ppv value * @param puDst Pointer to the size_t variable to update. * @param uNew The new value to assign to *puDst. */ DECLINLINE(size_t) ASMAtomicXchgZ(size_t volatile RT_FAR *puDst, const size_t uNew) RT_NOTHROW_DEF { #if ARCH_BITS == 16 AssertCompile(sizeof(size_t) == 2); return ASMAtomicXchgU16((volatile uint16_t RT_FAR *)puDst, uNew); #elif ARCH_BITS == 32 return ASMAtomicXchgU32((volatile uint32_t RT_FAR *)puDst, uNew); #elif ARCH_BITS == 64 return ASMAtomicXchgU64((volatile uint64_t RT_FAR *)puDst, uNew); #else # error "ARCH_BITS is bogus" #endif } /** * Atomically Exchange a pointer value, ordered. * * @returns Current *ppv value * @param ppv Pointer to the pointer variable to update. * @param pv The pointer value to assign to *ppv. */ DECLINLINE(void RT_FAR *) ASMAtomicXchgPtr(void RT_FAR * volatile RT_FAR *ppv, const void RT_FAR *pv) RT_NOTHROW_DEF { #if ARCH_BITS == 32 || ARCH_BITS == 16 return (void RT_FAR *)ASMAtomicXchgU32((volatile uint32_t RT_FAR *)(void RT_FAR *)ppv, (uint32_t)pv); #elif ARCH_BITS == 64 return (void RT_FAR *)ASMAtomicXchgU64((volatile uint64_t RT_FAR *)(void RT_FAR *)ppv, (uint64_t)pv); #else # error "ARCH_BITS is bogus" #endif } /** * Convenience macro for avoiding the annoying casting with ASMAtomicXchgPtr. * * @returns Current *pv value * @param ppv Pointer to the pointer variable to update. * @param pv The pointer value to assign to *ppv. * @param Type The type of *ppv, sans volatile. */ #ifdef __GNUC__ /* 8.2.0 requires -Wno-ignored-qualifiers */ # define ASMAtomicXchgPtrT(ppv, pv, Type) \ __extension__ \ ({\ __typeof__(*(ppv)) volatile * const ppvTypeChecked = (ppv); \ Type const pvTypeChecked = (pv); \ Type pvTypeCheckedRet = (__typeof__(*(ppv))) ASMAtomicXchgPtr((void * volatile *)ppvTypeChecked, (void *)pvTypeChecked); \ pvTypeCheckedRet; \ }) #else # define ASMAtomicXchgPtrT(ppv, pv, Type) \ (Type)ASMAtomicXchgPtr((void RT_FAR * volatile RT_FAR *)(ppv), (void RT_FAR *)(pv)) #endif /** * Atomically Exchange a raw-mode context pointer value, ordered. * * @returns Current *ppv value * @param ppvRC Pointer to the pointer variable to update. * @param pvRC The pointer value to assign to *ppv. */ DECLINLINE(RTRCPTR) ASMAtomicXchgRCPtr(RTRCPTR volatile RT_FAR *ppvRC, RTRCPTR pvRC) RT_NOTHROW_DEF { return (RTRCPTR)ASMAtomicXchgU32((uint32_t volatile RT_FAR *)(void RT_FAR *)ppvRC, (uint32_t)pvRC); } /** * Atomically Exchange a ring-0 pointer value, ordered. * * @returns Current *ppv value * @param ppvR0 Pointer to the pointer variable to update. * @param pvR0 The pointer value to assign to *ppv. */ DECLINLINE(RTR0PTR) ASMAtomicXchgR0Ptr(RTR0PTR volatile RT_FAR *ppvR0, RTR0PTR pvR0) RT_NOTHROW_DEF { #if R0_ARCH_BITS == 32 || ARCH_BITS == 16 return (RTR0PTR)ASMAtomicXchgU32((volatile uint32_t RT_FAR *)(void RT_FAR *)ppvR0, (uint32_t)pvR0); #elif R0_ARCH_BITS == 64 return (RTR0PTR)ASMAtomicXchgU64((volatile uint64_t RT_FAR *)(void RT_FAR *)ppvR0, (uint64_t)pvR0); #else # error "R0_ARCH_BITS is bogus" #endif } /** * Atomically Exchange a ring-3 pointer value, ordered. * * @returns Current *ppv value * @param ppvR3 Pointer to the pointer variable to update. * @param pvR3 The pointer value to assign to *ppv. */ DECLINLINE(RTR3PTR) ASMAtomicXchgR3Ptr(RTR3PTR volatile RT_FAR *ppvR3, RTR3PTR pvR3) RT_NOTHROW_DEF { #if R3_ARCH_BITS == 32 || ARCH_BITS == 16 return (RTR3PTR)ASMAtomicXchgU32((volatile uint32_t RT_FAR *)(void RT_FAR *)ppvR3, (uint32_t)pvR3); #elif R3_ARCH_BITS == 64 return (RTR3PTR)ASMAtomicXchgU64((volatile uint64_t RT_FAR *)(void RT_FAR *)ppvR3, (uint64_t)pvR3); #else # error "R3_ARCH_BITS is bogus" #endif } /** @def ASMAtomicXchgHandle * Atomically Exchange a typical IPRT handle value, ordered. * * @param ph Pointer to the value to update. * @param hNew The new value to assigned to *pu. * @param phRes Where to store the current *ph value. * * @remarks This doesn't currently work for all handles (like RTFILE). */ #if HC_ARCH_BITS == 32 || ARCH_BITS == 16 # define ASMAtomicXchgHandle(ph, hNew, phRes) \ do { \ AssertCompile(sizeof(*(ph)) == sizeof(uint32_t)); \ AssertCompile(sizeof(*(phRes)) == sizeof(uint32_t)); \ *(uint32_t RT_FAR *)(phRes) = ASMAtomicXchgU32((uint32_t volatile RT_FAR *)(ph), (const uint32_t)(hNew)); \ } while (0) #elif HC_ARCH_BITS == 64 # define ASMAtomicXchgHandle(ph, hNew, phRes) \ do { \ AssertCompile(sizeof(*(ph)) == sizeof(uint64_t)); \ AssertCompile(sizeof(*(phRes)) == sizeof(uint64_t)); \ *(uint64_t RT_FAR *)(phRes) = ASMAtomicXchgU64((uint64_t volatile RT_FAR *)(ph), (const uint64_t)(hNew)); \ } while (0) #else # error HC_ARCH_BITS #endif /** * Atomically Exchange a value which size might differ * between platforms or compilers, ordered. * * @param pu Pointer to the variable to update. * @param uNew The value to assign to *pu. * @todo This is busted as its missing the result argument. */ #define ASMAtomicXchgSize(pu, uNew) \ do { \ switch (sizeof(*(pu))) { \ case 1: ASMAtomicXchgU8( (volatile uint8_t RT_FAR *)(void RT_FAR *)(pu), (uint8_t)(uNew)); break; \ case 2: ASMAtomicXchgU16((volatile uint16_t RT_FAR *)(void RT_FAR *)(pu), (uint16_t)(uNew)); break; \ case 4: ASMAtomicXchgU32((volatile uint32_t RT_FAR *)(void RT_FAR *)(pu), (uint32_t)(uNew)); break; \ case 8: ASMAtomicXchgU64((volatile uint64_t RT_FAR *)(void RT_FAR *)(pu), (uint64_t)(uNew)); break; \ default: AssertMsgFailed(("ASMAtomicXchgSize: size %d is not supported\n", sizeof(*(pu)))); \ } \ } while (0) /** * Atomically Exchange a value which size might differ * between platforms or compilers, ordered. * * @param pu Pointer to the variable to update. * @param uNew The value to assign to *pu. * @param puRes Where to store the current *pu value. */ #define ASMAtomicXchgSizeCorrect(pu, uNew, puRes) \ do { \ switch (sizeof(*(pu))) { \ case 1: *(uint8_t RT_FAR *)(puRes) = ASMAtomicXchgU8( (volatile uint8_t RT_FAR *)(void RT_FAR *)(pu), (uint8_t)(uNew)); break; \ case 2: *(uint16_t RT_FAR *)(puRes) = ASMAtomicXchgU16((volatile uint16_t RT_FAR *)(void RT_FAR *)(pu), (uint16_t)(uNew)); break; \ case 4: *(uint32_t RT_FAR *)(puRes) = ASMAtomicXchgU32((volatile uint32_t RT_FAR *)(void RT_FAR *)(pu), (uint32_t)(uNew)); break; \ case 8: *(uint64_t RT_FAR *)(puRes) = ASMAtomicXchgU64((volatile uint64_t RT_FAR *)(void RT_FAR *)(pu), (uint64_t)(uNew)); break; \ default: AssertMsgFailed(("ASMAtomicXchgSize: size %d is not supported\n", sizeof(*(pu)))); \ } \ } while (0) /** * Atomically Compare and Exchange an unsigned 8-bit value, ordered. * * @returns true if xchg was done. * @returns false if xchg wasn't done. * * @param pu8 Pointer to the value to update. * @param u8New The new value to assigned to *pu8. * @param u8Old The old value to *pu8 compare with. * * @remarks x86: Requires a 486 or later. * @todo Rename ASMAtomicCmpWriteU8 */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM || !RT_INLINE_ASM_GNU_STYLE RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicCmpXchgU8(volatile uint8_t RT_FAR *pu8, const uint8_t u8New, const uint8_t u8Old) RT_NOTHROW_PROTO; #else DECLINLINE(bool) ASMAtomicCmpXchgU8(volatile uint8_t RT_FAR *pu8, const uint8_t u8New, uint8_t u8Old) RT_NOTHROW_DEF { # if defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) uint8_t u8Ret; __asm__ __volatile__("lock; cmpxchgb %3, %0\n\t" "setz %1\n\t" : "=m" (*pu8) , "=qm" (u8Ret) , "=a" (u8Old) : "q" (u8New) , "2" (u8Old) , "m" (*pu8) : "cc"); return (bool)u8Ret; # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) union { uint32_t u; bool f; } fXchg; uint32_t u32Spill; # if defined(RTASM_ARM64_USE_FEAT_LSE) __asm__ __volatile__("Lstart_ASMAtomicCmpXchgU8_%=:\n\t" # if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) /* M1 bench: casalb=5625 vs dmb+casb=1597 vs non-lse=5623 (ps/call) */ "casalb %w[uOldActual], %w[uNew], %[pMem]\n\t" # else RTASM_ARM_DMB_SY "casb %w[uOldActual], %w[uNew], %[pMem]\n\t" # endif "cmp %w[uOldActual], %w[uOldOrg]\n\t" "cset %w[fXchg], eq\n\t" : [pMem] "+Q" (*pu8) , [uOldActual] "=&r" (u32Spill) , [fXchg] "=&r" (fXchg.u) : [uNew] "r" ((uint32_t)u8New) , [uOldOrg] "r" ((uint32_t)u8Old) , "[uOldActual]" ((uint32_t)u8Old) : "cc"); # else uint32_t rcSpill; __asm__ __volatile__("Ltry_again_ASMAtomicCmpXchgU8_%=:\n\t" RTASM_ARM_DMB_SY # if defined(RT_ARCH_ARM64) "ldaxrb %w[uOld], %[pMem]\n\t" "cmp %w[uOld], %w[uCmp]\n\t" "bne 1f\n\t" /* stop here if not equal */ "stlxrb %w[rc], %w[uNew], %[pMem]\n\t" "cbnz %w[rc], Ltry_again_ASMAtomicCmpXchgU8_%=\n\t" "mov %w[fXchg], #1\n\t" "1:\n\t" "clrex\n\t" # else "ldrexb %[uOld], %[pMem]\n\t" "teq %[uOld], %[uCmp]\n\t" "strexbeq %[rc], %[uNew], %[pMem]\n\t" "bne 1f\n\t" /* stop here if not equal */ "cmp %[rc], #0\n\t" "bne Ltry_again_ASMAtomicCmpXchgU8_%=\n\t" "mov %[fXchg], #1\n\t" "1:\n\t" /** @todo clrexne on armv7? */ # endif : [pMem] "+Q" (*pu8) , [uOld] "=&r" (u32Spill) , [rc] "=&r" (rcSpill) , [fXchg] "=&r" (fXchg.u) : [uCmp] "r" ((uint32_t)u8Old) , [uNew] "r" ((uint32_t)u8New) , "[fXchg]" (0) RTASM_ARM_DMB_SY_COMMA_IN_REG : "cc"); # endif return fXchg.f; # else # error "Port me" # endif } #endif /** * Atomically Compare and Exchange a signed 8-bit value, ordered. * * @returns true if xchg was done. * @returns false if xchg wasn't done. * * @param pi8 Pointer to the value to update. * @param i8New The new value to assigned to *pi8. * @param i8Old The old value to *pi8 compare with. * * @remarks x86: Requires a 486 or later. * @todo Rename ASMAtomicCmpWriteS8 */ DECLINLINE(bool) ASMAtomicCmpXchgS8(volatile int8_t RT_FAR *pi8, const int8_t i8New, const int8_t i8Old) RT_NOTHROW_DEF { return ASMAtomicCmpXchgU8((volatile uint8_t RT_FAR *)pi8, (uint8_t)i8New, (uint8_t)i8Old); } /** * Atomically Compare and Exchange a bool value, ordered. * * @returns true if xchg was done. * @returns false if xchg wasn't done. * * @param pf Pointer to the value to update. * @param fNew The new value to assigned to *pf. * @param fOld The old value to *pf compare with. * * @remarks x86: Requires a 486 or later. * @todo Rename ASMAtomicCmpWriteBool */ DECLINLINE(bool) ASMAtomicCmpXchgBool(volatile bool RT_FAR *pf, const bool fNew, const bool fOld) RT_NOTHROW_DEF { return ASMAtomicCmpXchgU8((volatile uint8_t RT_FAR *)pf, (uint8_t)fNew, (uint8_t)fOld); } /** * Atomically Compare and Exchange an unsigned 32-bit value, ordered. * * @returns true if xchg was done. * @returns false if xchg wasn't done. * * @param pu32 Pointer to the value to update. * @param u32New The new value to assigned to *pu32. * @param u32Old The old value to *pu32 compare with. * * @remarks x86: Requires a 486 or later. * @todo Rename ASMAtomicCmpWriteU32 */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicCmpXchgU32(volatile uint32_t RT_FAR *pu32, const uint32_t u32New, const uint32_t u32Old) RT_NOTHROW_PROTO; #else DECLINLINE(bool) ASMAtomicCmpXchgU32(volatile uint32_t RT_FAR *pu32, const uint32_t u32New, uint32_t u32Old) RT_NOTHROW_DEF { # if defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE uint8_t u8Ret; __asm__ __volatile__("lock; cmpxchgl %3, %0\n\t" "setz %1\n\t" : "=m" (*pu32) , "=qm" (u8Ret) , "=a" (u32Old) : "r" (u32New) , "2" (u32Old) , "m" (*pu32) : "cc"); return (bool)u8Ret; # elif RT_INLINE_ASM_USES_INTRIN return (uint32_t)_InterlockedCompareExchange((long RT_FAR *)pu32, u32New, u32Old) == u32Old; # else uint32_t u32Ret; __asm { # ifdef RT_ARCH_AMD64 mov rdx, [pu32] # else mov edx, [pu32] # endif mov eax, [u32Old] mov ecx, [u32New] # ifdef RT_ARCH_AMD64 lock cmpxchg [rdx], ecx # else lock cmpxchg [edx], ecx # endif setz al movzx eax, al mov [u32Ret], eax } return !!u32Ret; # endif # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) union { uint32_t u; bool f; } fXchg; uint32_t u32Spill; /* M1 bench: match: casal= 6592 vs dmb+cas= 1562 vs non-lse=5634 (ps/call) mismatch: casal=18794 vs dmb+cas=19697 vs non-lse=2499 (ps/call) */ # if defined(RTASM_ARM64_USE_FEAT_LSE) __asm__ __volatile__("Lstart_ASMAtomicCmpXchgU32_%=:\n\t" # if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) "casal %w[uOldActual], %w[uNew], %[pMem]\n\t" # else RTASM_ARM_DMB_SY "cas %w[uOldActual], %w[uNew], %[pMem]\n\t" # endif "cmp %w[uOldActual], %w[uOldOrg]\n\t" "cset %w[fXchg], eq\n\t" : [pMem] "+Q" (*pu32) , [uOldActual] "=&r" (u32Spill) , [fXchg] "=&r" (fXchg.u) : [uNew] "r" (u32New) , [uOldOrg] "r" (u32Old) , "[uOldActual]" (u32Old) : "cc"); # else uint32_t rcSpill; __asm__ __volatile__("Ltry_again_ASMAtomicCmpXchgU32_%=:\n\t" RTASM_ARM_DMB_SY # if defined(RT_ARCH_ARM64) "ldaxr %w[uOld], %[pMem]\n\t" "cmp %w[uOld], %w[uCmp]\n\t" "bne 1f\n\t" /* stop here if not equal */ "stlxr %w[rc], %w[uNew], %[pMem]\n\t" "cbnz %w[rc], Ltry_again_ASMAtomicCmpXchgU32_%=\n\t" "mov %w[fXchg], #1\n\t" "1:\n\t" "clrex\n\t" # else "ldrex %[uOld], %[pMem]\n\t" "teq %[uOld], %[uCmp]\n\t" "strexeq %[rc], %[uNew], %[pMem]\n\t" "bne 1f\n\t" /* stop here if not equal */ "cmp %[rc], #0\n\t" "bne Ltry_again_ASMAtomicCmpXchgU32_%=\n\t" "mov %[fXchg], #1\n\t" "1:\n\t" /** @todo clrexne on armv7? */ # endif : [pMem] "+Q" (*pu32) , [uOld] "=&r" (u32Spill) , [rc] "=&r" (rcSpill) , [fXchg] "=&r" (fXchg.u) : [uCmp] "r" (u32Old) , [uNew] "r" (u32New) , "[fXchg]" (0) RTASM_ARM_DMB_SY_COMMA_IN_REG : "cc"); # endif return fXchg.f; # else # error "Port me" # endif } #endif /** * Atomically Compare and Exchange a signed 32-bit value, ordered. * * @returns true if xchg was done. * @returns false if xchg wasn't done. * * @param pi32 Pointer to the value to update. * @param i32New The new value to assigned to *pi32. * @param i32Old The old value to *pi32 compare with. * * @remarks x86: Requires a 486 or later. * @todo Rename ASMAtomicCmpWriteS32 */ DECLINLINE(bool) ASMAtomicCmpXchgS32(volatile int32_t RT_FAR *pi32, const int32_t i32New, const int32_t i32Old) RT_NOTHROW_DEF { return ASMAtomicCmpXchgU32((volatile uint32_t RT_FAR *)pi32, (uint32_t)i32New, (uint32_t)i32Old); } /** * Atomically Compare and exchange an unsigned 64-bit value, ordered. * * @returns true if xchg was done. * @returns false if xchg wasn't done. * * @param pu64 Pointer to the 64-bit variable to update. * @param u64New The 64-bit value to assign to *pu64. * @param u64Old The value to compare with. * * @remarks x86: Requires a Pentium or later. * @todo Rename ASMAtomicCmpWriteU64 */ #if (RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN) \ || RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicCmpXchgU64(volatile uint64_t RT_FAR *pu64, const uint64_t u64New, const uint64_t u64Old) RT_NOTHROW_PROTO; #else DECLINLINE(bool) ASMAtomicCmpXchgU64(volatile uint64_t RT_FAR *pu64, uint64_t u64New, uint64_t u64Old) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN return (uint64_t)_InterlockedCompareExchange64((__int64 RT_FAR *)pu64, u64New, u64Old) == u64Old; # elif defined(RT_ARCH_AMD64) # if RT_INLINE_ASM_GNU_STYLE uint8_t u8Ret; __asm__ __volatile__("lock; cmpxchgq %3, %0\n\t" "setz %1\n\t" : "=m" (*pu64) , "=qm" (u8Ret) , "=a" (u64Old) : "r" (u64New) , "2" (u64Old) , "m" (*pu64) : "cc"); return (bool)u8Ret; # else bool fRet; __asm { mov rdx, [pu32] mov rax, [u64Old] mov rcx, [u64New] lock cmpxchg [rdx], rcx setz al mov [fRet], al } return fRet; # endif # elif defined(RT_ARCH_X86) uint32_t u32Ret; # if RT_INLINE_ASM_GNU_STYLE # if defined(PIC) || defined(__PIC__) uint32_t u32EBX = (uint32_t)u64New; uint32_t u32Spill; __asm__ __volatile__("xchgl %%ebx, %4\n\t" "lock; cmpxchg8b (%6)\n\t" "setz %%al\n\t" "movl %4, %%ebx\n\t" "movzbl %%al, %%eax\n\t" : "=a" (u32Ret) , "=d" (u32Spill) # if RT_GNUC_PREREQ(4, 3) , "+m" (*pu64) # else , "=m" (*pu64) # endif : "A" (u64Old) , "m" ( u32EBX ) , "c" ( (uint32_t)(u64New >> 32) ) , "S" (pu64) : "cc"); # else /* !PIC */ uint32_t u32Spill; __asm__ __volatile__("lock; cmpxchg8b %2\n\t" "setz %%al\n\t" "movzbl %%al, %%eax\n\t" : "=a" (u32Ret) , "=d" (u32Spill) , "+m" (*pu64) : "A" (u64Old) , "b" ( (uint32_t)u64New ) , "c" ( (uint32_t)(u64New >> 32) ) : "cc"); # endif return (bool)u32Ret; # else __asm { mov ebx, dword ptr [u64New] mov ecx, dword ptr [u64New + 4] mov edi, [pu64] mov eax, dword ptr [u64Old] mov edx, dword ptr [u64Old + 4] lock cmpxchg8b [edi] setz al movzx eax, al mov dword ptr [u32Ret], eax } return !!u32Ret; # endif # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) union { uint32_t u; bool f; } fXchg; uint64_t u64Spill; /* M1 bench: match: casal= 6599 vs dmb+cas= 1565 vs non-lse=5000 (ps/call) mismatch: casal=18797 vs dmb+cas=19731 vs non-lse=2512 (ps/call) */ # if defined(RTASM_ARM64_USE_FEAT_LSE) __asm__ __volatile__("Lstart_ASMAtomicCmpXchgU75_%=:\n\t" # if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) "casal %[uOldActual], %[uNew], %[pMem]\n\t" # else RTASM_ARM_DMB_SY "cas %[uOldActual], %[uNew], %[pMem]\n\t" # endif "cmp %[uOldActual], %[uOldOrg]\n\t" "cset %w[fXchg], eq\n\t" : [pMem] "+Q" (*pu64) , [uOldActual] "=&r" (u64Spill) , [fXchg] "=&r" (fXchg.u) : [uNew] "r" (u64New) , [uOldOrg] "r" (u64Old) , "[uOldActual]" (u64Old) : "cc"); # else uint32_t rcSpill; __asm__ __volatile__("Ltry_again_ASMAtomicCmpXchgU64_%=:\n\t" RTASM_ARM_DMB_SY # if defined(RT_ARCH_ARM64) "ldaxr %[uOld], %[pMem]\n\t" "cmp %[uOld], %[uCmp]\n\t" "bne 1f\n\t" /* stop here if not equal */ "stlxr %w[rc], %[uNew], %[pMem]\n\t" "cbnz %w[rc], Ltry_again_ASMAtomicCmpXchgU64_%=\n\t" "mov %w[fXchg], #1\n\t" "1:\n\t" "clrex\n\t" # else "ldrexd %[uOld], %H[uOld], %[pMem]\n\t" "teq %[uOld], %[uCmp]\n\t" "teqeq %H[uOld], %H[uCmp]\n\t" "strexdeq %[rc], %[uNew], %H[uNew], %[pMem]\n\t" "bne 1f\n\t" /* stop here if not equal */ "cmp %[rc], #0\n\t" "bne Ltry_again_ASMAtomicCmpXchgU64_%=\n\t" "mov %[fXchg], #1\n\t" "1:\n\t" /** @todo clrexne on armv7? */ # endif : [pMem] "+Q" (*pu64) , [uOld] "=&r" (u64Spill) , [rc] "=&r" (rcSpill) , [fXchg] "=&r" (fXchg.u) : [uCmp] "r" (u64Old) , [uNew] "r" (u64New) , "[fXchg]" (0) RTASM_ARM_DMB_SY_COMMA_IN_REG : "cc"); # endif return fXchg.f; # else # error "Port me" # endif } #endif /** * Atomically Compare and exchange a signed 64-bit value, ordered. * * @returns true if xchg was done. * @returns false if xchg wasn't done. * * @param pi64 Pointer to the 64-bit variable to update. * @param i64 The 64-bit value to assign to *pu64. * @param i64Old The value to compare with. * * @remarks x86: Requires a Pentium or later. * @todo Rename ASMAtomicCmpWriteS64 */ DECLINLINE(bool) ASMAtomicCmpXchgS64(volatile int64_t RT_FAR *pi64, const int64_t i64, const int64_t i64Old) RT_NOTHROW_DEF { return ASMAtomicCmpXchgU64((volatile uint64_t RT_FAR *)pi64, (uint64_t)i64, (uint64_t)i64Old); } #if defined(RT_ARCH_AMD64) || defined(RT_ARCH_ARM64) || defined(DOXYGEN_RUNNING) /** @def RTASM_HAVE_CMP_WRITE_U128 * Indicates that we've got ASMAtomicCmpWriteU128(), ASMAtomicCmpWriteU128v2() * and ASMAtomicCmpWriteExU128() available. */ # define RTASM_HAVE_CMP_WRITE_U128 1 /** * Atomically compare and write an unsigned 128-bit value, ordered. * * @returns true if write was done. * @returns false if write wasn't done. * * @param pu128 Pointer to the 128-bit variable to update. * @param u64NewHi The high 64 bits of the value to assign to *pu128. * @param u64NewLo The low 64 bits of the value to assign to *pu128. * @param u64OldHi The high 64-bit of the value to compare with. * @param u64OldLo The low 64-bit of the value to compare with. * * @remarks AMD64: Not present in the earliest CPUs, so check CPUID. */ # if (RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN) DECLASM(bool) ASMAtomicCmpWriteU128v2(volatile uint128_t *pu128, const uint64_t u64NewHi, const uint64_t u64NewLo, const uint64_t u64OldHi, const uint64_t u64OldLo) RT_NOTHROW_PROTO; # else DECLINLINE(bool) ASMAtomicCmpWriteU128v2(volatile uint128_t *pu128, const uint64_t u64NewHi, const uint64_t u64NewLo, const uint64_t u64OldHi, const uint64_t u64OldLo) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN __int64 ai64Cmp[2]; ai64Cmp[0] = u64OldLo; ai64Cmp[1] = u64OldHi; return _InterlockedCompareExchange128((__int64 volatile *)pu128, u64NewHi, u64NewLo, ai64Cmp) != 0; # elif (defined(__clang_major__) || defined(__GNUC__)) && defined(RT_ARCH_ARM64) return __sync_bool_compare_and_swap(pu128, ((uint128_t)u64OldHi << 64) | u64OldLo, ((uint128_t)u64NewHi << 64) | u64NewLo); # elif defined(RT_ARCH_AMD64) # if RT_INLINE_ASM_GNU_STYLE uint64_t u64Ret; uint64_t u64Spill; __asm__ __volatile__("lock; cmpxchg16b %2\n\t" "setz %%al\n\t" "movzbl %%al, %%eax\n\t" : "=a" (u64Ret) , "=d" (u64Spill) , "+m" (*pu128) : "a" (u64OldLo) , "d" (u64OldHi) , "b" (u64NewLo) , "c" (u64NewHi) : "cc"); return (bool)u64Ret; # else # error "Port me" # endif # else # error "Port me" # endif } # endif /** * Atomically compare and write an unsigned 128-bit value, ordered. * * @returns true if write was done. * @returns false if write wasn't done. * * @param pu128 Pointer to the 128-bit variable to update. * @param u128New The 128-bit value to assign to *pu128. * @param u128Old The value to compare with. * * @remarks AMD64: Not present in the earliest CPUs, so check CPUID. */ DECLINLINE(bool) ASMAtomicCmpWriteU128(volatile uint128_t *pu128, const uint128_t u128New, const uint128_t u128Old) RT_NOTHROW_DEF { # ifdef RT_COMPILER_WITH_128BIT_INT_TYPES # if (defined(__clang_major__) || defined(__GNUC__)) && defined(RT_ARCH_ARM64) return __sync_bool_compare_and_swap(pu128, u128Old, u128New); # else return ASMAtomicCmpWriteU128v2(pu128, (uint64_t)(u128New >> 64), (uint64_t)u128New, (uint64_t)(u128Old >> 64), (uint64_t)u128Old); # endif # else return ASMAtomicCmpWriteU128v2(pu128, u128New.Hi, u128New.Lo, u128Old.Hi, u128Old.Lo); # endif } /** * RTUINT128U wrapper for ASMAtomicCmpWriteU128. */ DECLINLINE(bool) ASMAtomicCmpWriteU128U(volatile RTUINT128U *pu128, const RTUINT128U u128New, const RTUINT128U u128Old) RT_NOTHROW_DEF { # if (defined(__clang_major__) || defined(__GNUC__)) && defined(RT_ARCH_ARM64) return ASMAtomicCmpWriteU128(&pu128->u, u128New.u, u128Old.u); # else return ASMAtomicCmpWriteU128v2(&pu128->u, u128New.s.Hi, u128New.s.Lo, u128Old.s.Hi, u128Old.s.Lo); # endif } #endif /* RT_ARCH_AMD64 || RT_ARCH_ARM64 */ /** * Atomically Compare and Exchange a pointer value, ordered. * * @returns true if xchg was done. * @returns false if xchg wasn't done. * * @param ppv Pointer to the value to update. * @param pvNew The new value to assigned to *ppv. * @param pvOld The old value to *ppv compare with. * * @remarks x86: Requires a 486 or later. * @todo Rename ASMAtomicCmpWritePtrVoid */ DECLINLINE(bool) ASMAtomicCmpXchgPtrVoid(void RT_FAR * volatile RT_FAR *ppv, const void RT_FAR *pvNew, const void RT_FAR *pvOld) RT_NOTHROW_DEF { #if ARCH_BITS == 32 || ARCH_BITS == 16 return ASMAtomicCmpXchgU32((volatile uint32_t RT_FAR *)(void RT_FAR *)ppv, (uint32_t)pvNew, (uint32_t)pvOld); #elif ARCH_BITS == 64 return ASMAtomicCmpXchgU64((volatile uint64_t RT_FAR *)(void RT_FAR *)ppv, (uint64_t)pvNew, (uint64_t)pvOld); #else # error "ARCH_BITS is bogus" #endif } /** * Atomically Compare and Exchange a pointer value, ordered. * * @returns true if xchg was done. * @returns false if xchg wasn't done. * * @param ppv Pointer to the value to update. * @param pvNew The new value to assigned to *ppv. * @param pvOld The old value to *ppv compare with. * * @remarks This is relatively type safe on GCC platforms. * @remarks x86: Requires a 486 or later. * @todo Rename ASMAtomicCmpWritePtr */ #ifdef __GNUC__ # define ASMAtomicCmpXchgPtr(ppv, pvNew, pvOld) \ __extension__ \ ({\ __typeof__(*(ppv)) volatile * const ppvTypeChecked = (ppv); \ __typeof__(*(ppv)) const pvNewTypeChecked = (pvNew); \ __typeof__(*(ppv)) const pvOldTypeChecked = (pvOld); \ bool fMacroRet = ASMAtomicCmpXchgPtrVoid((void * volatile *)ppvTypeChecked, \ (void *)pvNewTypeChecked, (void *)pvOldTypeChecked); \ fMacroRet; \ }) #else # define ASMAtomicCmpXchgPtr(ppv, pvNew, pvOld) \ ASMAtomicCmpXchgPtrVoid((void RT_FAR * volatile RT_FAR *)(ppv), (void RT_FAR *)(pvNew), (void RT_FAR *)(pvOld)) #endif /** @def ASMAtomicCmpXchgHandle * Atomically Compare and Exchange a typical IPRT handle value, ordered. * * @param ph Pointer to the value to update. * @param hNew The new value to assigned to *pu. * @param hOld The old value to *pu compare with. * @param fRc Where to store the result. * * @remarks This doesn't currently work for all handles (like RTFILE). * @remarks x86: Requires a 486 or later. * @todo Rename ASMAtomicCmpWriteHandle */ #if HC_ARCH_BITS == 32 || ARCH_BITS == 16 # define ASMAtomicCmpXchgHandle(ph, hNew, hOld, fRc) \ do { \ AssertCompile(sizeof(*(ph)) == sizeof(uint32_t)); \ (fRc) = ASMAtomicCmpXchgU32((uint32_t volatile RT_FAR *)(ph), (const uint32_t)(hNew), (const uint32_t)(hOld)); \ } while (0) #elif HC_ARCH_BITS == 64 # define ASMAtomicCmpXchgHandle(ph, hNew, hOld, fRc) \ do { \ AssertCompile(sizeof(*(ph)) == sizeof(uint64_t)); \ (fRc) = ASMAtomicCmpXchgU64((uint64_t volatile RT_FAR *)(ph), (const uint64_t)(hNew), (const uint64_t)(hOld)); \ } while (0) #else # error HC_ARCH_BITS #endif /** @def ASMAtomicCmpXchgSize * Atomically Compare and Exchange a value which size might differ * between platforms or compilers, ordered. * * @param pu Pointer to the value to update. * @param uNew The new value to assigned to *pu. * @param uOld The old value to *pu compare with. * @param fRc Where to store the result. * * @remarks x86: Requires a 486 or later. * @todo Rename ASMAtomicCmpWriteSize */ #define ASMAtomicCmpXchgSize(pu, uNew, uOld, fRc) \ do { \ switch (sizeof(*(pu))) { \ case 4: (fRc) = ASMAtomicCmpXchgU32((volatile uint32_t RT_FAR *)(void RT_FAR *)(pu), (uint32_t)(uNew), (uint32_t)(uOld)); \ break; \ case 8: (fRc) = ASMAtomicCmpXchgU64((volatile uint64_t RT_FAR *)(void RT_FAR *)(pu), (uint64_t)(uNew), (uint64_t)(uOld)); \ break; \ default: AssertMsgFailed(("ASMAtomicCmpXchgSize: size %d is not supported\n", sizeof(*(pu)))); \ (fRc) = false; \ break; \ } \ } while (0) /** * Atomically Compare and Exchange an unsigned 8-bit value, additionally passes * back old value, ordered. * * @returns true if xchg was done. * @returns false if xchg wasn't done. * * @param pu8 Pointer to the value to update. * @param u8New The new value to assigned to *pu32. * @param u8Old The old value to *pu8 compare with. * @param pu8Old Pointer store the old value at. * * @remarks x86: Requires a 486 or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicCmpXchgExU8(volatile uint8_t RT_FAR *pu8, const uint8_t u8New, const uint8_t u8Old, uint8_t RT_FAR *pu8Old) RT_NOTHROW_PROTO; #else DECLINLINE(bool) ASMAtomicCmpXchgExU8(volatile uint8_t RT_FAR *pu8, const uint8_t u8New, const uint8_t u8Old, uint8_t RT_FAR *pu8Old) RT_NOTHROW_DEF { # if defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE uint8_t u8Ret; __asm__ __volatile__("lock; cmpxchgb %3, %0\n\t" "setz %1\n\t" : "=m" (*pu8) , "=qm" (u8Ret) , "=a" (*pu8Old) # if defined(RT_ARCH_X86) : "q" (u8New) # else : "r" (u8New) # endif , "a" (u8Old) , "m" (*pu8) : "cc"); return (bool)u8Ret; # elif RT_INLINE_ASM_USES_INTRIN return (*pu8Old = _InterlockedCompareExchange8((char RT_FAR *)pu8, u8New, u8Old)) == u8Old; # else uint8_t u8Ret; __asm { # ifdef RT_ARCH_AMD64 mov rdx, [pu8] # else mov edx, [pu8] # endif mov eax, [u8Old] mov ecx, [u8New] # ifdef RT_ARCH_AMD64 lock cmpxchg [rdx], ecx mov rdx, [pu8Old] mov [rdx], eax # else lock cmpxchg [edx], ecx mov edx, [pu8Old] mov [edx], eax # endif setz al movzx eax, al mov [u8Ret], eax } return !!u8Ret; # endif # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) /* M1 bench: match: casalb= 6594 vs dmb+casb= 1561 vs non-lse=5051 (ps/call) mismatch: casalb=15346 vs dmb+casb=16349 vs non-lse=2505 (ps/call) */ # if defined(RTASM_ARM64_USE_FEAT_LSE) union { uint32_t u; bool f; } fXchg; uint32_t u32Actual; __asm__ __volatile__("Lstart_ASMAtomicCmpXchgExU8_%=:\n\t" # if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) "casalb %w[uOldActual], %w[uNew], %[pMem]\n\t" # else RTASM_ARM_DMB_SY "casb %w[uOldActual], %w[uNew], %[pMem]\n\t" # endif "cmp %w[uOldActual], %w[uOldOrg]\n\t" "cset %w[fXchg], eq\n\t" : [pMem] "+Q" (*pu8) , [uOldActual] "=&r" (u32Actual) , [fXchg] "=&r" (fXchg.u) : [uNew] "r" ((uint32_t)u8New) , [uOldOrg] "r" ((uint32_t)u8Old) , "[uOldActual]" ((uint32_t)u8Old) : "cc"); *pu8Old = (uint8_t)u32Actual; # else union { uint8_t u; bool f; } fXchg; uint8_t u8ActualOld; uint8_t rcSpill; __asm__ __volatile__("Ltry_again_ASMAtomicCmpXchgExU8_%=:\n\t" RTASM_ARM_DMB_SY # if defined(RT_ARCH_ARM64) "ldaxrb %w[uOld], %[pMem]\n\t" "cmp %w[uOld], %w[uCmp]\n\t" "bne 1f\n\t" /* stop here if not equal */ "stlxrb %w[rc], %w[uNew], %[pMem]\n\t" "cbnz %w[rc], Ltry_again_ASMAtomicCmpXchgExU8_%=\n\t" "mov %w[fXchg], #1\n\t" "1:\n\t" "clrex\n\t" # else "ldrexb %[uOld], %[pMem]\n\t" "teq %[uOld], %[uCmp]\n\t" "strexbeq %[rc], %[uNew], %[pMem]\n\t" "bne 1f\n\t" /* stop here if not equal */ "cmp %[rc], #0\n\t" "bne Ltry_again_ASMAtomicCmpXchgExU8_%=\n\t" "mov %[fXchg], #1\n\t" "1:\n\t" /** @todo clrexne on armv7? */ # endif : [pMem] "+Q" (*pu8) , [uOld] "=&r" (u8ActualOld) , [rc] "=&r" (rcSpill) , [fXchg] "=&r" (fXchg.u) : [uCmp] "r" (u8Old) , [uNew] "r" (u8New) , "[fXchg]" (0) RTASM_ARM_DMB_SY_COMMA_IN_REG : "cc"); *pu8Old = u8ActualOld; # endif return fXchg.f; # else # error "Port me" # endif } #endif /** * Atomically Compare and Exchange a signed 8-bit value, additionally * passes back old value, ordered. * * @returns true if xchg was done. * @returns false if xchg wasn't done. * * @param pi8 Pointer to the value to update. * @param i8New The new value to assigned to *pi8. * @param i8Old The old value to *pi8 compare with. * @param pi8Old Pointer store the old value at. * * @remarks x86: Requires a 486 or later. */ DECLINLINE(bool) ASMAtomicCmpXchgExS8(volatile int8_t RT_FAR *pi8, const int8_t i8New, const int8_t i8Old, int8_t RT_FAR *pi8Old) RT_NOTHROW_DEF { return ASMAtomicCmpXchgExU8((volatile uint8_t RT_FAR *)pi8, (uint8_t)i8New, (uint8_t)i8Old, (uint8_t RT_FAR *)pi8Old); } /** * Atomically Compare and Exchange an unsigned 16-bit value, additionally passes * back old value, ordered. * * @returns true if xchg was done. * @returns false if xchg wasn't done. * * @param pu16 Pointer to the value to update. * @param u16New The new value to assigned to *pu16. * @param u16Old The old value to *pu32 compare with. * @param pu16Old Pointer store the old value at. * * @remarks x86: Requires a 486 or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicCmpXchgExU16(volatile uint16_t RT_FAR *pu16, const uint16_t u16New, const uint16_t u16Old, uint16_t RT_FAR *pu16Old) RT_NOTHROW_PROTO; #else DECLINLINE(bool) ASMAtomicCmpXchgExU16(volatile uint16_t RT_FAR *pu16, const uint16_t u16New, const uint16_t u16Old, uint16_t RT_FAR *pu16Old) RT_NOTHROW_DEF { # if defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE uint8_t u8Ret; __asm__ __volatile__("lock; cmpxchgw %3, %0\n\t" "setz %1\n\t" : "=m" (*pu16) , "=qm" (u8Ret) , "=a" (*pu16Old) : "r" (u16New) , "a" (u16Old) , "m" (*pu16) : "cc"); return (bool)u8Ret; # elif RT_INLINE_ASM_USES_INTRIN return (*pu16Old = _InterlockedCompareExchange16((short RT_FAR *)pu16, u16New, u16Old)) == u16Old; # else uint16_t u16Ret; __asm { # ifdef RT_ARCH_AMD64 mov rdx, [pu16] # else mov edx, [pu16] # endif mov eax, [u16Old] mov ecx, [u16New] # ifdef RT_ARCH_AMD64 lock cmpxchg [rdx], ecx mov rdx, [pu16Old] mov [rdx], eax # else lock cmpxchg [edx], ecx mov edx, [pu16Old] mov [edx], eax # endif setz al movzx eax, al mov [u16Ret], eax } return !!u16Ret; # endif # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) /* M1 bench: match: casalh= 6577 vs dmb+cash= 1608 vs non-lse=5078 (ps/call) mismatch: casalh=18791 vs dmb+cash=19721 vs non-lse=2543 (ps/call) */ # if defined(RTASM_ARM64_USE_FEAT_LSE) union { uint32_t u; bool f; } fXchg; uint32_t u32Actual; __asm__ __volatile__("Lstart_ASMAtomicCmpXchgExU16_%=:\n\t" # if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) "casalh %w[uOldActual], %w[uNew], %[pMem]\n\t" # else RTASM_ARM_DMB_SY "cash %w[uOldActual], %w[uNew], %[pMem]\n\t" # endif "cmp %w[uOldActual], %w[uOldOrg]\n\t" "cset %w[fXchg], eq\n\t" : [pMem] "+Q" (*pu16) , [uOldActual] "=&r" (u32Actual) , [fXchg] "=&r" (fXchg.u) : [uNew] "r" ((uint32_t)u16New) , [uOldOrg] "r" ((uint32_t)u16Old) , "[uOldActual]" ((uint32_t)u16Old) : "cc"); *pu16Old = (uint16_t)u32Actual; # else union { uint16_t u; bool f; } fXchg; uint16_t u16ActualOld; uint16_t rcSpill; __asm__ __volatile__("Ltry_again_ASMAtomicCmpXchgExU16_%=:\n\t" RTASM_ARM_DMB_SY # if defined(RT_ARCH_ARM64) "ldaxrh %w[uOld], %[pMem]\n\t" "cmp %w[uOld], %w[uCmp]\n\t" "bne 1f\n\t" /* stop here if not equal */ "stlxrh %w[rc], %w[uNew], %[pMem]\n\t" "cbnz %w[rc], Ltry_again_ASMAtomicCmpXchgExU16_%=\n\t" "mov %w[fXchg], #1\n\t" "1:\n\t" "clrex\n\t" # else "ldrexh %[uOld], %[pMem]\n\t" "teq %[uOld], %[uCmp]\n\t" "strexheq %[rc], %[uNew], %[pMem]\n\t" "bne 1f\n\t" /* stop here if not equal */ "cmp %[rc], #0\n\t" "bne Ltry_again_ASMAtomicCmpXchgExU16_%=\n\t" "mov %[fXchg], #1\n\t" "1:\n\t" /** @todo clrexne on armv7? */ # endif : [pMem] "+Q" (*pu16) , [uOld] "=&r" (u16ActualOld) , [rc] "=&r" (rcSpill) , [fXchg] "=&r" (fXchg.u) : [uCmp] "r" (u16Old) , [uNew] "r" (u16New) , "[fXchg]" (0) RTASM_ARM_DMB_SY_COMMA_IN_REG : "cc"); *pu16Old = u16ActualOld; # endif return fXchg.f; # else # error "Port me" # endif } #endif /** * Atomically Compare and Exchange a signed 16-bit value, additionally * passes back old value, ordered. * * @returns true if xchg was done. * @returns false if xchg wasn't done. * * @param pi16 Pointer to the value to update. * @param i16New The new value to assigned to *pi16. * @param i16Old The old value to *pi16 compare with. * @param pi16Old Pointer store the old value at. * * @remarks x86: Requires a 486 or later. */ DECLINLINE(bool) ASMAtomicCmpXchgExS16(volatile int16_t RT_FAR *pi16, const int16_t i16New, const int16_t i16Old, int16_t RT_FAR *pi16Old) RT_NOTHROW_DEF { return ASMAtomicCmpXchgExU16((volatile uint16_t RT_FAR *)pi16, (uint16_t)i16New, (uint16_t)i16Old, (uint16_t RT_FAR *)pi16Old); } /** * Atomically Compare and Exchange an unsigned 32-bit value, additionally * passes back old value, ordered. * * @returns true if xchg was done. * @returns false if xchg wasn't done. * * @param pu32 Pointer to the value to update. * @param u32New The new value to assigned to *pu32. * @param u32Old The old value to *pu32 compare with. * @param pu32Old Pointer store the old value at. * * @remarks x86: Requires a 486 or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicCmpXchgExU32(volatile uint32_t RT_FAR *pu32, const uint32_t u32New, const uint32_t u32Old, uint32_t RT_FAR *pu32Old) RT_NOTHROW_PROTO; #else DECLINLINE(bool) ASMAtomicCmpXchgExU32(volatile uint32_t RT_FAR *pu32, const uint32_t u32New, const uint32_t u32Old, uint32_t RT_FAR *pu32Old) RT_NOTHROW_DEF { # if defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE uint8_t u8Ret; __asm__ __volatile__("lock; cmpxchgl %3, %0\n\t" "setz %1\n\t" : "=m" (*pu32) , "=qm" (u8Ret) , "=a" (*pu32Old) : "r" (u32New) , "a" (u32Old) , "m" (*pu32) : "cc"); return (bool)u8Ret; # elif RT_INLINE_ASM_USES_INTRIN return (*pu32Old = _InterlockedCompareExchange((long RT_FAR *)pu32, u32New, u32Old)) == u32Old; # else uint32_t u32Ret; __asm { # ifdef RT_ARCH_AMD64 mov rdx, [pu32] # else mov edx, [pu32] # endif mov eax, [u32Old] mov ecx, [u32New] # ifdef RT_ARCH_AMD64 lock cmpxchg [rdx], ecx mov rdx, [pu32Old] mov [rdx], eax # else lock cmpxchg [edx], ecx mov edx, [pu32Old] mov [edx], eax # endif setz al movzx eax, al mov [u32Ret], eax } return !!u32Ret; # endif # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) union { uint32_t u; bool f; } fXchg; /* M1 bench: match: casal= 6590 vs dmb+cas= 1564 vs non-lse=5033 (ps/call) mismatch: casal=18790 vs dmb+cas=19711 vs non-lse=2503 (ps/call) */ # if defined(RTASM_ARM64_USE_FEAT_LSE) __asm__ __volatile__("Lstart_ASMAtomicCmpXchgExU32_%=:\n\t" # if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) "casal %w[uOldActual], %w[uNew], %[pMem]\n\t" # else RTASM_ARM_DMB_SY "cas %w[uOldActual], %w[uNew], %[pMem]\n\t" # endif "cmp %w[uOldActual], %w[uOldOrg]\n\t" "cset %w[fXchg], eq\n\t" : [pMem] "+Q" (*pu32) , [uOldActual] "=&r" (*pu32Old) , [fXchg] "=&r" (fXchg.u) : [uNew] "r" (u32New) , [uOldOrg] "r" (u32Old) , "[uOldActual]" (u32Old) : "cc"); # else uint32_t u32ActualOld; uint32_t rcSpill; __asm__ __volatile__("Ltry_again_ASMAtomicCmpXchgExU32_%=:\n\t" RTASM_ARM_DMB_SY # if defined(RT_ARCH_ARM64) "ldaxr %w[uOld], %[pMem]\n\t" "cmp %w[uOld], %w[uCmp]\n\t" "bne 1f\n\t" /* stop here if not equal */ "stlxr %w[rc], %w[uNew], %[pMem]\n\t" "cbnz %w[rc], Ltry_again_ASMAtomicCmpXchgExU32_%=\n\t" "mov %w[fXchg], #1\n\t" "1:\n\t" "clrex\n\t" # else "ldrex %[uOld], %[pMem]\n\t" "teq %[uOld], %[uCmp]\n\t" "strexeq %[rc], %[uNew], %[pMem]\n\t" "bne 1f\n\t" /* stop here if not equal */ "cmp %[rc], #0\n\t" "bne Ltry_again_ASMAtomicCmpXchgExU32_%=\n\t" "mov %[fXchg], #1\n\t" "1:\n\t" /** @todo clrexne on armv7? */ # endif : [pMem] "+Q" (*pu32) , [uOld] "=&r" (u32ActualOld) , [rc] "=&r" (rcSpill) , [fXchg] "=&r" (fXchg.u) : [uCmp] "r" (u32Old) , [uNew] "r" (u32New) , "[fXchg]" (0) RTASM_ARM_DMB_SY_COMMA_IN_REG : "cc"); *pu32Old = u32ActualOld; # endif return fXchg.f; # else # error "Port me" # endif } #endif /** * Atomically Compare and Exchange a signed 32-bit value, additionally * passes back old value, ordered. * * @returns true if xchg was done. * @returns false if xchg wasn't done. * * @param pi32 Pointer to the value to update. * @param i32New The new value to assigned to *pi32. * @param i32Old The old value to *pi32 compare with. * @param pi32Old Pointer store the old value at. * * @remarks x86: Requires a 486 or later. */ DECLINLINE(bool) ASMAtomicCmpXchgExS32(volatile int32_t RT_FAR *pi32, const int32_t i32New, const int32_t i32Old, int32_t RT_FAR *pi32Old) RT_NOTHROW_DEF { return ASMAtomicCmpXchgExU32((volatile uint32_t RT_FAR *)pi32, (uint32_t)i32New, (uint32_t)i32Old, (uint32_t RT_FAR *)pi32Old); } /** * Atomically Compare and exchange an unsigned 64-bit value, additionally * passing back old value, ordered. * * @returns true if xchg was done. * @returns false if xchg wasn't done. * * @param pu64 Pointer to the 64-bit variable to update. * @param u64New The 64-bit value to assign to *pu64. * @param u64Old The value to compare with. * @param pu64Old Pointer store the old value at. * * @remarks x86: Requires a Pentium or later. */ #if (RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN) \ || RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicCmpXchgExU64(volatile uint64_t RT_FAR *pu64, const uint64_t u64New, const uint64_t u64Old, uint64_t RT_FAR *pu64Old) RT_NOTHROW_PROTO; #else DECLINLINE(bool) ASMAtomicCmpXchgExU64(volatile uint64_t RT_FAR *pu64, const uint64_t u64New, const uint64_t u64Old, uint64_t RT_FAR *pu64Old) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN return (*pu64Old =_InterlockedCompareExchange64((__int64 RT_FAR *)pu64, u64New, u64Old)) == u64Old; # elif defined(RT_ARCH_AMD64) # if RT_INLINE_ASM_GNU_STYLE uint8_t u8Ret; __asm__ __volatile__("lock; cmpxchgq %3, %0\n\t" "setz %1\n\t" : "=m" (*pu64) , "=qm" (u8Ret) , "=a" (*pu64Old) : "r" (u64New) , "a" (u64Old) , "m" (*pu64) : "cc"); return (bool)u8Ret; # else bool fRet; __asm { mov rdx, [pu32] mov rax, [u64Old] mov rcx, [u64New] lock cmpxchg [rdx], rcx mov rdx, [pu64Old] mov [rdx], rax setz al mov [fRet], al } return fRet; # endif # elif defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE uint64_t u64Ret; # if defined(PIC) || defined(__PIC__) /* Note #1: This code uses a memory clobber description, because the clean solution with an output value for *pu64 makes gcc run out of registers. This will cause suboptimal code, and anyone with a better solution is welcome to improve this. Note #2: We must prevent gcc from encoding the memory access, as it may go via the GOT if we're working on a global variable (like in the testcase). Thus we request a register (%3) and dereference it ourselves. */ __asm__ __volatile__("xchgl %%ebx, %1\n\t" "lock; cmpxchg8b (%3)\n\t" "xchgl %%ebx, %1\n\t" : "=A" (u64Ret) : "DS" ((uint32_t)u64New) , "c" ((uint32_t)(u64New >> 32)) , "r" (pu64) /* Do not use "m" here*/ , "0" (u64Old) : "memory" , "cc" ); # else /* !PIC */ __asm__ __volatile__("lock; cmpxchg8b %4\n\t" : "=A" (u64Ret) , "=m" (*pu64) : "b" ((uint32_t)u64New) , "c" ((uint32_t)(u64New >> 32)) , "m" (*pu64) , "0" (u64Old) : "cc"); # endif *pu64Old = u64Ret; return u64Ret == u64Old; # else uint32_t u32Ret; __asm { mov ebx, dword ptr [u64New] mov ecx, dword ptr [u64New + 4] mov edi, [pu64] mov eax, dword ptr [u64Old] mov edx, dword ptr [u64Old + 4] lock cmpxchg8b [edi] mov ebx, [pu64Old] mov [ebx], eax setz al movzx eax, al add ebx, 4 mov [ebx], edx mov dword ptr [u32Ret], eax } return !!u32Ret; # endif # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) union { uint32_t u; bool f; } fXchg; /* M1 bench: match: casal= 6606 vs dmb+cas= 1565 vs non-lse=5006 (ps/call) mismatch: casal=18786 vs dmb+cas=19718 vs non-lse=2503 (ps/call) */ # if defined(RTASM_ARM64_USE_FEAT_LSE) __asm__ __volatile__("Lstart_ASMAtomicCmpXchgExU32_%=:\n\t" # if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) "casal %[uOldActual], %[uNew], %[pMem]\n\t" # else RTASM_ARM_DMB_SY "cas %[uOldActual], %[uNew], %[pMem]\n\t" # endif "cmp %[uOldActual], %[uOldOrg]\n\t" "cset %w[fXchg], eq\n\t" : [pMem] "+Q" (*pu64) , [uOldActual] "=&r" (*pu64Old) , [fXchg] "=&r" (fXchg.u) : [uNew] "r" (u64New) , [uOldOrg] "r" (u64Old) , "[uOldActual]" (u64Old) : "cc"); # else uint64_t u64ActualOld; uint32_t rcSpill; __asm__ __volatile__("Ltry_again_ASMAtomicCmpXchgU64_%=:\n\t" RTASM_ARM_DMB_SY # if defined(RT_ARCH_ARM64) "ldaxr %[uOld], %[pMem]\n\t" "cmp %[uOld], %[uCmp]\n\t" "bne 1f\n\t" /* stop here if not equal */ "stlxr %w[rc], %[uNew], %[pMem]\n\t" "cbnz %w[rc], Ltry_again_ASMAtomicCmpXchgU64_%=\n\t" "mov %w[fXchg], #1\n\t" "1:\n\t" "clrex\n\t" # else "ldrexd %[uOld], %H[uOld], %[pMem]\n\t" "teq %[uOld], %[uCmp]\n\t" "teqeq %H[uOld], %H[uCmp]\n\t" "strexdeq %[rc], %[uNew], %H[uNew], %[pMem]\n\t" "bne 1f\n\t" /* stop here if not equal */ "cmp %[rc], #0\n\t" "bne Ltry_again_ASMAtomicCmpXchgU64_%=\n\t" "mov %[fXchg], #1\n\t" "1:\n\t" /** @todo clrexne on armv7? */ # endif : [pMem] "+Q" (*pu64) , [uOld] "=&r" (u64ActualOld) , [rc] "=&r" (rcSpill) , [fXchg] "=&r" (fXchg.u) : [uCmp] "r" (u64Old) , [uNew] "r" (u64New) , "[fXchg]" (0) RTASM_ARM_DMB_SY_COMMA_IN_REG : "cc"); *pu64Old = u64ActualOld; # endif return fXchg.f; # else # error "Port me" # endif } #endif /** * Atomically Compare and exchange a signed 64-bit value, additionally * passing back old value, ordered. * * @returns true if xchg was done. * @returns false if xchg wasn't done. * * @param pi64 Pointer to the 64-bit variable to update. * @param i64 The 64-bit value to assign to *pu64. * @param i64Old The value to compare with. * @param pi64Old Pointer store the old value at. * * @remarks x86: Requires a Pentium or later. */ DECLINLINE(bool) ASMAtomicCmpXchgExS64(volatile int64_t RT_FAR *pi64, const int64_t i64, const int64_t i64Old, int64_t RT_FAR *pi64Old) RT_NOTHROW_DEF { return ASMAtomicCmpXchgExU64((volatile uint64_t RT_FAR *)pi64, (uint64_t)i64, (uint64_t)i64Old, (uint64_t RT_FAR *)pi64Old); } #if defined(RT_ARCH_AMD64) || defined(RT_ARCH_ARM64) || defined(DOXYGEN_RUNNING) /** @def RTASM_HAVE_CMP_XCHG_U128 * Indicates that we've got ASMAtomicCmpSwapU128(), ASMAtomicCmpSwapU128v2() * and ASMAtomicCmpSwapExU128() available. */ # define RTASM_HAVE_CMP_XCHG_U128 1 /** * Atomically compare and exchange an unsigned 128-bit value, ordered. * * @returns true if exchange was done. * @returns false if exchange wasn't done. * * @param pu128 Pointer to the 128-bit variable to update. * @param u64NewHi The high 64 bits of the value to assign to *pu128. * @param u64NewLo The low 64 bits of the value to assign to *pu128. * @param u64OldHi The high 64-bit of the value to compare with. * @param u64OldLo The low 64-bit of the value to compare with. * @param pu128Old Where to return the old value. * * @remarks AMD64: Not present in the earliest CPUs, so check CPUID. */ # if (RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN) DECLASM(bool) ASMAtomicCmpXchgU128v2(volatile uint128_t *pu128, const uint64_t u64NewHi, const uint64_t u64NewLo, const uint64_t u64OldHi, const uint64_t u64OldLo, uint128_t *pu128Old) RT_NOTHROW_PROTO; # else DECLINLINE(bool) ASMAtomicCmpXchgU128v2(volatile uint128_t *pu128, const uint64_t u64NewHi, const uint64_t u64NewLo, const uint64_t u64OldHi, const uint64_t u64OldLo, uint128_t *pu128Old) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN pu128Old->Hi = u64OldHi; pu128Old->Lo = u64OldLo; AssertCompileMemberOffset(uint128_t, Lo, 0); return _InterlockedCompareExchange128((__int64 volatile *)pu128, u64NewHi, u64NewLo, (__int64 *)&pu128Old->Lo) != 0; # elif (defined(__clang_major__) || defined(__GNUC__)) && defined(RT_ARCH_ARM64) uint128_t const uCmp = ((uint128_t)u64OldHi << 64) | u64OldLo; uint128_t const uOld = __sync_val_compare_and_swap(pu128, uCmp, ((uint128_t)u64NewHi << 64) | u64NewLo); *pu128Old = uOld; return uCmp == uOld; # elif defined(RT_ARCH_AMD64) # if RT_INLINE_ASM_GNU_STYLE uint8_t bRet; uint64_t u64RetHi, u64RetLo; __asm__ __volatile__("lock; cmpxchg16b %3\n\t" "setz %b0\n\t" : "=r" (bRet) , "=a" (u64RetLo) , "=d" (u64RetHi) , "+m" (*pu128) : "a" (u64OldLo) , "d" (u64OldHi) , "b" (u64NewLo) , "c" (u64NewHi) : "cc"); *pu128Old = ((uint128_t)u64RetHi << 64) | u64RetLo; return (bool)bRet; # else # error "Port me" # endif # else # error "Port me" # endif } # endif /** * Atomically compare and exchange an unsigned 128-bit value, ordered. * * @returns true if exchange was done. * @returns false if exchange wasn't done. * * @param pu128 Pointer to the 128-bit variable to update. * @param u128New The 128-bit value to assign to *pu128. * @param u128Old The value to compare with. * @param pu128Old Where to return the old value. * * @remarks AMD64: Not present in the earliest CPUs, so check CPUID. */ DECLINLINE(bool) ASMAtomicCmpXchgU128(volatile uint128_t *pu128, const uint128_t u128New, const uint128_t u128Old, uint128_t *pu128Old) RT_NOTHROW_DEF { # ifdef RT_COMPILER_WITH_128BIT_INT_TYPES # if (defined(__clang_major__) || defined(__GNUC__)) && defined(RT_ARCH_ARM64) uint128_t const uSwapped = __sync_val_compare_and_swap(pu128, u128Old, u128New); *pu128Old = uSwapped; return uSwapped == u128Old; # else return ASMAtomicCmpXchgU128v2(pu128, (uint64_t)(u128New >> 64), (uint64_t)u128New, (uint64_t)(u128Old >> 64), (uint64_t)u128Old, pu128Old); # endif # else return ASMAtomicCmpXchgU128v2(pu128, u128New.Hi, u128New.Lo, u128Old.Hi, u128Old.Lo, pu128Old); # endif } /** * RTUINT128U wrapper for ASMAtomicCmpXchgU128. */ DECLINLINE(bool) ASMAtomicCmpXchgU128U(volatile RTUINT128U *pu128, const RTUINT128U u128New, const RTUINT128U u128Old, PRTUINT128U pu128Old) RT_NOTHROW_DEF { # if (defined(__clang_major__) || defined(__GNUC__)) && defined(RT_ARCH_ARM64) return ASMAtomicCmpXchgU128(&pu128->u, u128New.u, u128Old.u, &pu128Old->u); # else return ASMAtomicCmpXchgU128v2(&pu128->u, u128New.s.Hi, u128New.s.Lo, u128Old.s.Hi, u128Old.s.Lo, &pu128Old->u); # endif } #endif /* RT_ARCH_AMD64 || RT_ARCH_ARM64 */ /** @def ASMAtomicCmpXchgExHandle * Atomically Compare and Exchange a typical IPRT handle value, ordered. * * @param ph Pointer to the value to update. * @param hNew The new value to assigned to *pu. * @param hOld The old value to *pu compare with. * @param fRc Where to store the result. * @param phOldVal Pointer to where to store the old value. * * @remarks This doesn't currently work for all handles (like RTFILE). */ #if HC_ARCH_BITS == 32 || ARCH_BITS == 16 # define ASMAtomicCmpXchgExHandle(ph, hNew, hOld, fRc, phOldVal) \ do { \ AssertCompile(sizeof(*ph) == sizeof(uint32_t)); \ AssertCompile(sizeof(*phOldVal) == sizeof(uint32_t)); \ (fRc) = ASMAtomicCmpXchgExU32((volatile uint32_t RT_FAR *)(ph), (uint32_t)(hNew), (uint32_t)(hOld), (uint32_t RT_FAR *)(phOldVal)); \ } while (0) #elif HC_ARCH_BITS == 64 # define ASMAtomicCmpXchgExHandle(ph, hNew, hOld, fRc, phOldVal) \ do { \ AssertCompile(sizeof(*(ph)) == sizeof(uint64_t)); \ AssertCompile(sizeof(*(phOldVal)) == sizeof(uint64_t)); \ (fRc) = ASMAtomicCmpXchgExU64((volatile uint64_t RT_FAR *)(ph), (uint64_t)(hNew), (uint64_t)(hOld), (uint64_t RT_FAR *)(phOldVal)); \ } while (0) #else # error HC_ARCH_BITS #endif /** @def ASMAtomicCmpXchgExSize * Atomically Compare and Exchange a value which size might differ * between platforms or compilers. Additionally passes back old value. * * @param pu Pointer to the value to update. * @param uNew The new value to assigned to *pu. * @param uOld The old value to *pu compare with. * @param fRc Where to store the result. * @param puOldVal Pointer to where to store the old value. * * @remarks x86: Requires a 486 or later. */ #define ASMAtomicCmpXchgExSize(pu, uNew, uOld, fRc, puOldVal) \ do { \ switch (sizeof(*(pu))) { \ case 4: (fRc) = ASMAtomicCmpXchgExU32((volatile uint32_t RT_FAR *)(void RT_FAR *)(pu), (uint32_t)(uNew), (uint32_t)(uOld), (uint32_t RT_FAR *)(uOldVal)); \ break; \ case 8: (fRc) = ASMAtomicCmpXchgExU64((volatile uint64_t RT_FAR *)(void RT_FAR *)(pu), (uint64_t)(uNew), (uint64_t)(uOld), (uint64_t RT_FAR *)(uOldVal)); \ break; \ default: AssertMsgFailed(("ASMAtomicCmpXchgSize: size %d is not supported\n", sizeof(*(pu)))); \ (fRc) = false; \ (uOldVal) = 0; \ break; \ } \ } while (0) /** * Atomically Compare and Exchange a pointer value, additionally * passing back old value, ordered. * * @returns true if xchg was done. * @returns false if xchg wasn't done. * * @param ppv Pointer to the value to update. * @param pvNew The new value to assigned to *ppv. * @param pvOld The old value to *ppv compare with. * @param ppvOld Pointer store the old value at. * * @remarks x86: Requires a 486 or later. */ DECLINLINE(bool) ASMAtomicCmpXchgExPtrVoid(void RT_FAR * volatile RT_FAR *ppv, const void RT_FAR *pvNew, const void RT_FAR *pvOld, void RT_FAR * RT_FAR *ppvOld) RT_NOTHROW_DEF { #if ARCH_BITS == 32 || ARCH_BITS == 16 return ASMAtomicCmpXchgExU32((volatile uint32_t RT_FAR *)(void RT_FAR *)ppv, (uint32_t)pvNew, (uint32_t)pvOld, (uint32_t RT_FAR *)ppvOld); #elif ARCH_BITS == 64 return ASMAtomicCmpXchgExU64((volatile uint64_t RT_FAR *)(void RT_FAR *)ppv, (uint64_t)pvNew, (uint64_t)pvOld, (uint64_t RT_FAR *)ppvOld); #else # error "ARCH_BITS is bogus" #endif } /** * Atomically Compare and Exchange a pointer value, additionally * passing back old value, ordered. * * @returns true if xchg was done. * @returns false if xchg wasn't done. * * @param ppv Pointer to the value to update. * @param pvNew The new value to assigned to *ppv. * @param pvOld The old value to *ppv compare with. * @param ppvOld Pointer store the old value at. * * @remarks This is relatively type safe on GCC platforms. * @remarks x86: Requires a 486 or later. */ #ifdef __GNUC__ # define ASMAtomicCmpXchgExPtr(ppv, pvNew, pvOld, ppvOld) \ __extension__ \ ({\ __typeof__(*(ppv)) volatile * const ppvTypeChecked = (ppv); \ __typeof__(*(ppv)) const pvNewTypeChecked = (pvNew); \ __typeof__(*(ppv)) const pvOldTypeChecked = (pvOld); \ __typeof__(*(ppv)) * const ppvOldTypeChecked = (ppvOld); \ bool fMacroRet = ASMAtomicCmpXchgExPtrVoid((void * volatile *)ppvTypeChecked, \ (void *)pvNewTypeChecked, (void *)pvOldTypeChecked, \ (void **)ppvOldTypeChecked); \ fMacroRet; \ }) #else # define ASMAtomicCmpXchgExPtr(ppv, pvNew, pvOld, ppvOld) \ ASMAtomicCmpXchgExPtrVoid((void RT_FAR * volatile RT_FAR *)(ppv), (void RT_FAR *)(pvNew), (void RT_FAR *)(pvOld), (void RT_FAR * RT_FAR *)(ppvOld)) #endif /** * Virtualization unfriendly serializing instruction, always exits. */ #if (RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN) || (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86)) RT_ASM_DECL_PRAGMA_WATCOM(void) ASMSerializeInstructionCpuId(void) RT_NOTHROW_PROTO; #else DECLINLINE(void) ASMSerializeInstructionCpuId(void) RT_NOTHROW_DEF { # if RT_INLINE_ASM_GNU_STYLE RTCCUINTREG xAX = 0; # ifdef RT_ARCH_AMD64 __asm__ __volatile__ ("cpuid" : "=a" (xAX) : "0" (xAX) : "rbx", "rcx", "rdx", "memory"); # elif (defined(PIC) || defined(__PIC__)) && defined(__i386__) __asm__ __volatile__ ("push %%ebx\n\t" "cpuid\n\t" "pop %%ebx\n\t" : "=a" (xAX) : "0" (xAX) : "ecx", "edx", "memory"); # else __asm__ __volatile__ ("cpuid" : "=a" (xAX) : "0" (xAX) : "ebx", "ecx", "edx", "memory"); # endif # elif RT_INLINE_ASM_USES_INTRIN int aInfo[4]; _ReadWriteBarrier(); __cpuid(aInfo, 0); # else __asm { push ebx xor eax, eax cpuid pop ebx } # endif } #endif /** * Virtualization friendly serializing instruction, though more expensive. */ #if RT_INLINE_ASM_EXTERNAL || (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86)) RT_ASM_DECL_PRAGMA_WATCOM(void) ASMSerializeInstructionIRet(void) RT_NOTHROW_PROTO; #else DECLINLINE(void) ASMSerializeInstructionIRet(void) RT_NOTHROW_DEF { # if RT_INLINE_ASM_GNU_STYLE # ifdef RT_ARCH_AMD64 __asm__ __volatile__ ("movq %%rsp,%%r10\n\t" "subq $128, %%rsp\n\t" /*redzone*/ "mov %%ss, %%eax\n\t" "pushq %%rax\n\t" "pushq %%r10\n\t" "pushfq\n\t" "movl %%cs, %%eax\n\t" "pushq %%rax\n\t" "leaq 1f(%%rip), %%rax\n\t" "pushq %%rax\n\t" "iretq\n\t" "1:\n\t" ::: "rax", "r10", "memory", "cc"); # else __asm__ __volatile__ ("pushfl\n\t" "pushl %%cs\n\t" "pushl $1f\n\t" "iretl\n\t" "1:\n\t" ::: "memory"); # endif # else __asm { pushfd push cs push la_ret iretd la_ret: } # endif } #endif /** * Virtualization friendlier serializing instruction, may still cause exits. */ #if (RT_INLINE_ASM_EXTERNAL && RT_INLINE_ASM_USES_INTRIN < RT_MSC_VER_VS2008) || (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86)) RT_ASM_DECL_PRAGMA_WATCOM(void) ASMSerializeInstructionRdTscp(void) RT_NOTHROW_PROTO; #else DECLINLINE(void) ASMSerializeInstructionRdTscp(void) RT_NOTHROW_DEF { # if RT_INLINE_ASM_GNU_STYLE /* rdtscp is not supported by ancient linux build VM of course :-( */ # ifdef RT_ARCH_AMD64 /*__asm__ __volatile__("rdtscp\n\t" ::: "rax", "rdx, "rcx"); */ __asm__ __volatile__(".byte 0x0f,0x01,0xf9\n\t" ::: "rax", "rdx", "rcx", "memory"); # else /*__asm__ __volatile__("rdtscp\n\t" ::: "eax", "edx, "ecx"); */ __asm__ __volatile__(".byte 0x0f,0x01,0xf9\n\t" ::: "eax", "edx", "ecx", "memory"); # endif # else # if RT_INLINE_ASM_USES_INTRIN >= RT_MSC_VER_VS2008 uint32_t uIgnore; _ReadWriteBarrier(); (void)__rdtscp(&uIgnore); (void)uIgnore; # else __asm { rdtscp } # endif # endif } #endif /** * Serialize Instruction (both data store and instruction flush). */ #if (defined(RT_ARCH_X86) && ARCH_BITS == 16) || defined(IN_GUEST) # define ASMSerializeInstruction() ASMSerializeInstructionIRet() #elif defined(RT_ARCH_X86) || defined(RT_ARCH_AMD64) # define ASMSerializeInstruction() ASMSerializeInstructionCpuId() #elif defined(RT_ARCH_SPARC64) RTDECL(void) ASMSerializeInstruction(void) RT_NOTHROW_PROTO; #elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) DECLINLINE(void) ASMSerializeInstruction(void) RT_NOTHROW_DEF { __asm__ __volatile__ (RTASM_ARM_DSB_SY :: RTASM_ARM_DSB_SY_IN_REG :); } #else # error "Port me" #endif /** * Memory fence, waits for any pending writes and reads to complete. * @note No implicit compiler barrier (which is probably stupid). */ DECLINLINE(void) ASMMemoryFence(void) RT_NOTHROW_DEF { #if defined(RT_ARCH_AMD64) || (defined(RT_ARCH_X86) && !defined(RT_WITH_OLD_CPU_SUPPORT)) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__ (".byte 0x0f,0xae,0xf0\n\t"); # elif RT_INLINE_ASM_USES_INTRIN _mm_mfence(); # else __asm { _emit 0x0f _emit 0xae _emit 0xf0 } # endif #elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) __asm__ __volatile__ (RTASM_ARM_DMB_SY :: RTASM_ARM_DMB_SY_IN_REG :); #elif ARCH_BITS == 16 uint16_t volatile u16; ASMAtomicXchgU16(&u16, 0); #else uint32_t volatile u32; ASMAtomicXchgU32(&u32, 0); #endif } /** * Write fence, waits for any pending writes to complete. * @note No implicit compiler barrier (which is probably stupid). */ DECLINLINE(void) ASMWriteFence(void) RT_NOTHROW_DEF { #if defined(RT_ARCH_AMD64) || (defined(RT_ARCH_X86) && !defined(RT_WITH_OLD_CPU_SUPPORT)) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__ (".byte 0x0f,0xae,0xf8\n\t"); # elif RT_INLINE_ASM_USES_INTRIN _mm_sfence(); # else __asm { _emit 0x0f _emit 0xae _emit 0xf8 } # endif #elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) __asm__ __volatile__ (RTASM_ARM_DMB_ST :: RTASM_ARM_DMB_ST_IN_REG :); #else ASMMemoryFence(); #endif } /** * Read fence, waits for any pending reads to complete. * @note No implicit compiler barrier (which is probably stupid). */ DECLINLINE(void) ASMReadFence(void) RT_NOTHROW_DEF { #if defined(RT_ARCH_AMD64) || (defined(RT_ARCH_X86) && !defined(RT_WITH_OLD_CPU_SUPPORT)) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__ (".byte 0x0f,0xae,0xe8\n\t"); # elif RT_INLINE_ASM_USES_INTRIN _mm_lfence(); # else __asm { _emit 0x0f _emit 0xae _emit 0xe8 } # endif #elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) __asm__ __volatile__ (RTASM_ARM_DMB_LD :: RTASM_ARM_DMB_LD_IN_REG :); #else ASMMemoryFence(); #endif } /** * Atomically reads an unsigned 8-bit value, ordered. * * @returns Current *pu8 value * @param pu8 Pointer to the 8-bit variable to read. */ DECLINLINE(uint8_t) ASMAtomicReadU8(volatile uint8_t RT_FAR *pu8) RT_NOTHROW_DEF { #if defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) uint32_t u32; # if defined(RTASM_ARM64_USE_FEAT_LSE) && 0 /* very expensive on M1 */ __asm__ __volatile__("Lstart_ASMAtomicReadU8_%=:\n\t" RTASM_ARM_DMB_SY "casab %w[uDst], wzr, %[pMem]\n\t" : [uDst] "=&r" (u32) : [pMem] "Q" (*pu8), "0" (0) RTASM_ARM_DMB_SY_COMMA_IN_REG); # else __asm__ __volatile__("Lstart_ASMAtomicReadU8_%=:\n\t" RTASM_ARM_DMB_SY # if defined(RT_ARCH_ARM64) # if 1 /* shouldn't be any need for more than single-copy atomicity when we've got a proper barrier, just like on x86. */ "ldurb %w[uDst], %[pMem]\n\t" # else "ldxrb %w[uDst], %[pMem]\n\t" "clrex\n\t" # endif # else "ldrexb %[uDst], %[pMem]\n\t" /** @todo clrex */ # endif : [uDst] "=&r" (u32) : [pMem] "Q" (*pu8) RTASM_ARM_DMB_SY_COMMA_IN_REG); # endif return (uint8_t)u32; #else ASMMemoryFence(); return *pu8; /* byte reads are atomic on x86 */ #endif } /** * Atomically reads an unsigned 8-bit value, unordered. * * @returns Current *pu8 value * @param pu8 Pointer to the 8-bit variable to read. */ DECLINLINE(uint8_t) ASMAtomicUoReadU8(volatile uint8_t RT_FAR *pu8) RT_NOTHROW_DEF { #if defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) uint32_t u32; __asm__ __volatile__("Lstart_ASMAtomicUoReadU8_%=:\n\t" # if defined(RT_ARCH_ARM64) "ldurb %w[uDst], %[pMem]\n\t" # else "ldrexb %[uDst], %[pMem]\n\t" /** @todo fix this */ # endif : [uDst] "=&r" (u32) : [pMem] "Q" (*pu8)); return (uint8_t)u32; #else return *pu8; /* byte reads are atomic on x86 */ #endif } /** * Atomically reads a signed 8-bit value, ordered. * * @returns Current *pi8 value * @param pi8 Pointer to the 8-bit variable to read. */ DECLINLINE(int8_t) ASMAtomicReadS8(volatile int8_t RT_FAR *pi8) RT_NOTHROW_DEF { #if defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) return (int8_t)ASMAtomicReadU8((volatile uint8_t RT_FAR *)pi8); #else ASMMemoryFence(); return *pi8; /* byte reads are atomic on x86 */ #endif } /** * Atomically reads a signed 8-bit value, unordered. * * @returns Current *pi8 value * @param pi8 Pointer to the 8-bit variable to read. */ DECLINLINE(int8_t) ASMAtomicUoReadS8(volatile int8_t RT_FAR *pi8) RT_NOTHROW_DEF { #if defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) int32_t i32; __asm__ __volatile__("Lstart_ASMAtomicUoReadS8_%=:\n\t" # if defined(RT_ARCH_ARM64) "ldurb %w[iDst], %[pMem]\n\t" # else "ldrexb %[iDst], %[pMem]\n\t" /** @todo fix this */ # endif : [iDst] "=&r" (i32) : [pMem] "Q" (*pi8)); return (int8_t)i32; #else return *pi8; /* byte reads are atomic on x86 */ #endif } /** * Atomically reads an unsigned 16-bit value, ordered. * * @returns Current *pu16 value * @param pu16 Pointer to the 16-bit variable to read. */ DECLINLINE(uint16_t) ASMAtomicReadU16(volatile uint16_t RT_FAR *pu16) RT_NOTHROW_DEF { Assert(!((uintptr_t)pu16 & 1)); #if defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) uint32_t u32; # if defined(RTASM_ARM64_USE_FEAT_LSE) && 0 /* very expensive on M1, but alignment advantages with LEA2 (M2?). */ __asm__ __volatile__("Lstart_ASMAtomicReadU16_%=:\n\t" RTASM_ARM_DMB_SY "casah %w[uDst], wzr, %[pMem]\n\t" : [uDst] "=&r" (u32) : [pMem] "Q" (*pu16), "0" (0) RTASM_ARM_DMB_SY_COMMA_IN_REG); # else __asm__ __volatile__("Lstart_ASMAtomicReadU16_%=:\n\t" RTASM_ARM_DMB_SY # if defined(RT_ARCH_ARM64) # if 1 /* ASSUMING proper barrier and aligned access, we should be fine with single-copy atomicity, just like on x86. */ "ldurh %w[uDst], %[pMem]\n\t" # else "ldxrh %w[uDst], %[pMem]\n\t" "clrex\n\t" # endif # else "ldrexh %[uDst], %[pMem]\n\t" /** @todo clrex */ # endif : [uDst] "=&r" (u32) : [pMem] "Q" (*pu16) RTASM_ARM_DMB_SY_COMMA_IN_REG); # endif return (uint16_t)u32; #else ASMMemoryFence(); return *pu16; #endif } /** * Atomically reads an unsigned 16-bit value, unordered. * * @returns Current *pu16 value * @param pu16 Pointer to the 16-bit variable to read. */ DECLINLINE(uint16_t) ASMAtomicUoReadU16(volatile uint16_t RT_FAR *pu16) RT_NOTHROW_DEF { Assert(!((uintptr_t)pu16 & 1)); #if defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) uint32_t u32; __asm__ __volatile__("Lstart_ASMAtomicUoReadU16_%=:\n\t" # if defined(RT_ARCH_ARM64) "ldurh %w[uDst], %[pMem]\n\t" # else "ldrexh %[uDst], %[pMem]\n\t" /** @todo fix this */ # endif : [uDst] "=&r" (u32) : [pMem] "Q" (*pu16)); return (uint16_t)u32; #else return *pu16; #endif } /** * Atomically reads a signed 16-bit value, ordered. * * @returns Current *pi16 value * @param pi16 Pointer to the 16-bit variable to read. */ DECLINLINE(int16_t) ASMAtomicReadS16(volatile int16_t RT_FAR *pi16) RT_NOTHROW_DEF { Assert(!((uintptr_t)pi16 & 1)); #if defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) return (int16_t)ASMAtomicReadU16((volatile uint16_t RT_FAR *)pi16); #else ASMMemoryFence(); return *pi16; #endif } /** * Atomically reads a signed 16-bit value, unordered. * * @returns Current *pi16 value * @param pi16 Pointer to the 16-bit variable to read. */ DECLINLINE(int16_t) ASMAtomicUoReadS16(volatile int16_t RT_FAR *pi16) RT_NOTHROW_DEF { Assert(!((uintptr_t)pi16 & 1)); #if defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) int32_t i32; __asm__ __volatile__("Lstart_ASMAtomicUoReadS16_%=:\n\t" # if defined(RT_ARCH_ARM64) "ldurh %w[iDst], %[pMem]\n\t" # else "ldrexh %[iDst], %[pMem]\n\t" /** @todo fix this */ # endif : [iDst] "=&r" (i32) : [pMem] "Q" (*pi16)); return (int16_t)i32; #else return *pi16; #endif } /** * Atomically reads an unsigned 32-bit value, ordered. * * @returns Current *pu32 value * @param pu32 Pointer to the 32-bit variable to read. */ DECLINLINE(uint32_t) ASMAtomicReadU32(volatile uint32_t RT_FAR *pu32) RT_NOTHROW_DEF { Assert(!((uintptr_t)pu32 & 3)); #if defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) uint32_t u32; # if defined(RTASM_ARM64_USE_FEAT_LSE) && 0 /* very expensive on M1, but alignment advantages with LEA2 (M2?). */ __asm__ __volatile__("Lstart_ASMAtomicReadU32_%=:\n\t" RTASM_ARM_DMB_SY "casa %w[uDst], wzr, %[pMem]\n\t" : [uDst] "=&r" (u32) : [pMem] "Q" (*pu32), "0" (0) RTASM_ARM_DMB_SY_COMMA_IN_REG); # else __asm__ __volatile__("Lstart_ASMAtomicReadU32_%=:\n\t" RTASM_ARM_DMB_SY # if defined(RT_ARCH_ARM64) # if 1 /* ASSUMING proper barrier and aligned access, we should be fine with single-copy atomicity, just like on x86. */ "ldur %w[uDst], %[pMem]\n\t" # else "ldxr %w[uDst], %[pMem]\n\t" "clrex\n\t" # endif # else "ldrex %[uDst], %[pMem]\n\t" /** @todo clrex */ # endif : [uDst] "=&r" (u32) : [pMem] "Q" (*pu32) RTASM_ARM_DMB_SY_COMMA_IN_REG); # endif return u32; #else ASMMemoryFence(); # if ARCH_BITS == 16 AssertFailed(); /** @todo 16-bit */ # endif return *pu32; #endif } /** * Atomically reads an unsigned 32-bit value, unordered. * * @returns Current *pu32 value * @param pu32 Pointer to the 32-bit variable to read. */ DECLINLINE(uint32_t) ASMAtomicUoReadU32(volatile uint32_t RT_FAR *pu32) RT_NOTHROW_DEF { Assert(!((uintptr_t)pu32 & 3)); #if defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) uint32_t u32; __asm__ __volatile__("Lstart_ASMAtomicUoReadU32_%=:\n\t" # if defined(RT_ARCH_ARM64) "ldur %w[uDst], %[pMem]\n\t" # else "ldrex %[uDst], %[pMem]\n\t" /** @todo fix this */ # endif : [uDst] "=&r" (u32) : [pMem] "Q" (*pu32)); return u32; #else # if ARCH_BITS == 16 AssertFailed(); /** @todo 16-bit */ # endif return *pu32; #endif } /** * Atomically reads a signed 32-bit value, ordered. * * @returns Current *pi32 value * @param pi32 Pointer to the 32-bit variable to read. */ DECLINLINE(int32_t) ASMAtomicReadS32(volatile int32_t RT_FAR *pi32) RT_NOTHROW_DEF { Assert(!((uintptr_t)pi32 & 3)); #if defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) return (int32_t)ASMAtomicReadU32((volatile uint32_t RT_FAR *)pi32); #else ASMMemoryFence(); # if ARCH_BITS == 16 AssertFailed(); /** @todo 16-bit */ # endif return *pi32; #endif } /** * Atomically reads a signed 32-bit value, unordered. * * @returns Current *pi32 value * @param pi32 Pointer to the 32-bit variable to read. */ DECLINLINE(int32_t) ASMAtomicUoReadS32(volatile int32_t RT_FAR *pi32) RT_NOTHROW_DEF { Assert(!((uintptr_t)pi32 & 3)); #if defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) int32_t i32; __asm__ __volatile__("Lstart_ASMAtomicUoReadS32_%=:\n\t" # if defined(RT_ARCH_ARM64) "ldur %w[iDst], %[pMem]\n\t" # else "ldrex %[iDst], %[pMem]\n\t" /** @todo thix this */ # endif : [iDst] "=&r" (i32) : [pMem] "Q" (*pi32)); return i32; #else # if ARCH_BITS == 16 AssertFailed(); /** @todo 16-bit */ # endif return *pi32; #endif } /** * Atomically reads an unsigned 64-bit value, ordered. * * @returns Current *pu64 value * @param pu64 Pointer to the 64-bit variable to read. * The memory pointed to must be writable. * * @remarks This may fault if the memory is read-only! * @remarks x86: Requires a Pentium or later. */ #if (RT_INLINE_ASM_EXTERNAL_TMP_ARM && !defined(RT_ARCH_AMD64)) \ || RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC RT_ASM_DECL_PRAGMA_WATCOM(uint64_t) ASMAtomicReadU64(volatile uint64_t RT_FAR *pu64) RT_NOTHROW_PROTO; #else DECLINLINE(uint64_t) ASMAtomicReadU64(volatile uint64_t RT_FAR *pu64) RT_NOTHROW_DEF { uint64_t u64; # ifdef RT_ARCH_AMD64 Assert(!((uintptr_t)pu64 & 7)); /*# if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__( "mfence\n\t" "movq %1, %0\n\t" : "=r" (u64) : "m" (*pu64)); # else __asm { mfence mov rdx, [pu64] mov rax, [rdx] mov [u64], rax } # endif*/ ASMMemoryFence(); u64 = *pu64; # elif defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE # if defined(PIC) || defined(__PIC__) uint32_t u32EBX = 0; Assert(!((uintptr_t)pu64 & 7)); __asm__ __volatile__("xchgl %%ebx, %3\n\t" "lock; cmpxchg8b (%5)\n\t" "movl %3, %%ebx\n\t" : "=A" (u64) # if RT_GNUC_PREREQ(4, 3) , "+m" (*pu64) # else , "=m" (*pu64) # endif : "0" (0ULL) , "m" (u32EBX) , "c" (0) , "S" (pu64) : "cc"); # else /* !PIC */ __asm__ __volatile__("lock; cmpxchg8b %1\n\t" : "=A" (u64) , "+m" (*pu64) : "0" (0ULL) , "b" (0) , "c" (0) : "cc"); # endif # else Assert(!((uintptr_t)pu64 & 7)); __asm { xor eax, eax xor edx, edx mov edi, pu64 xor ecx, ecx xor ebx, ebx lock cmpxchg8b [edi] mov dword ptr [u64], eax mov dword ptr [u64 + 4], edx } # endif # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) Assert(!((uintptr_t)pu64 & 7)); # if defined(RTASM_ARM64_USE_FEAT_LSE) && 0 /* very expensive on M1, but alignment advantages with LEA2 (M2?). */ __asm__ __volatile__("Lstart_ASMAtomicReadU64_%=:\n\t" RTASM_ARM_DMB_SY "casa %[uDst], xzr, %[pMem]\n\t" : [uDst] "=&r" (u64) : [pMem] "Q" (*pu64), "0" (0) RTASM_ARM_DMB_SY_COMMA_IN_REG); # else __asm__ __volatile__("Lstart_ASMAtomicReadU64_%=:\n\t" RTASM_ARM_DMB_SY # if defined(RT_ARCH_ARM64) # if 1 /* ASSUMING proper barrier and aligned access, we should be fine with single-copy atomicity, just like on x86. */ "ldur %[uDst], %[pMem]\n\t" # else "ldxr %[uDst], %[pMem]\n\t" "clrex\n\t" # endif # else "ldrexd %[uDst], %H[uDst], %[pMem]\n\t" /** @todo clrex */ # endif : [uDst] "=&r" (u64) : [pMem] "Q" (*pu64) RTASM_ARM_DMB_SY_COMMA_IN_REG); # endif # else # error "Port me" # endif return u64; } #endif /** * Atomically reads an unsigned 64-bit value, unordered. * * @returns Current *pu64 value * @param pu64 Pointer to the 64-bit variable to read. * The memory pointed to must be writable. * * @remarks This may fault if the memory is read-only! * @remarks x86: Requires a Pentium or later. */ #if !defined(RT_ARCH_AMD64) \ && ( (RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN) \ || RT_INLINE_DONT_MIX_CMPXCHG8B_AND_PIC) RT_ASM_DECL_PRAGMA_WATCOM(uint64_t) ASMAtomicUoReadU64(volatile uint64_t RT_FAR *pu64) RT_NOTHROW_PROTO; #else DECLINLINE(uint64_t) ASMAtomicUoReadU64(volatile uint64_t RT_FAR *pu64) RT_NOTHROW_DEF { uint64_t u64; # ifdef RT_ARCH_AMD64 Assert(!((uintptr_t)pu64 & 7)); /*# if RT_INLINE_ASM_GNU_STYLE Assert(!((uintptr_t)pu64 & 7)); __asm__ __volatile__("movq %1, %0\n\t" : "=r" (u64) : "m" (*pu64)); # else __asm { mov rdx, [pu64] mov rax, [rdx] mov [u64], rax } # endif */ u64 = *pu64; # elif defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE # if defined(PIC) || defined(__PIC__) uint32_t u32EBX = 0; uint32_t u32Spill; Assert(!((uintptr_t)pu64 & 7)); __asm__ __volatile__("xor %%eax,%%eax\n\t" "xor %%ecx,%%ecx\n\t" "xor %%edx,%%edx\n\t" "xchgl %%ebx, %3\n\t" "lock; cmpxchg8b (%4)\n\t" "movl %3, %%ebx\n\t" : "=A" (u64) # if RT_GNUC_PREREQ(4, 3) , "+m" (*pu64) # else , "=m" (*pu64) # endif , "=c" (u32Spill) : "m" (u32EBX) , "S" (pu64) : "cc"); # else /* !PIC */ __asm__ __volatile__("lock; cmpxchg8b %1\n\t" : "=A" (u64) , "+m" (*pu64) : "0" (0ULL) , "b" (0) , "c" (0) : "cc"); # endif # else Assert(!((uintptr_t)pu64 & 7)); __asm { xor eax, eax xor edx, edx mov edi, pu64 xor ecx, ecx xor ebx, ebx lock cmpxchg8b [edi] mov dword ptr [u64], eax mov dword ptr [u64 + 4], edx } # endif # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) Assert(!((uintptr_t)pu64 & 7)); __asm__ __volatile__("Lstart_ASMAtomicUoReadU64_%=:\n\t" # if defined(RT_ARCH_ARM64) "ldur %[uDst], %[pMem]\n\t" # else "ldrexd %[uDst], %H[uDst], %[pMem]\n\t" /* this is required for atomic access since it's a pair */ /** @todo clrex? */ # endif : [uDst] "=&r" (u64) : [pMem] "Q" (*pu64)); # else # error "Port me" # endif return u64; } #endif /** * Atomically reads a signed 64-bit value, ordered. * * @returns Current *pi64 value * @param pi64 Pointer to the 64-bit variable to read. * The memory pointed to must be writable. * * @remarks This may fault if the memory is read-only! * @remarks x86: Requires a Pentium or later. */ DECLINLINE(int64_t) ASMAtomicReadS64(volatile int64_t RT_FAR *pi64) RT_NOTHROW_DEF { return (int64_t)ASMAtomicReadU64((volatile uint64_t RT_FAR *)pi64); } /** * Atomically reads a signed 64-bit value, unordered. * * @returns Current *pi64 value * @param pi64 Pointer to the 64-bit variable to read. * The memory pointed to must be writable. * * @remarks This will fault if the memory is read-only! * @remarks x86: Requires a Pentium or later. */ DECLINLINE(int64_t) ASMAtomicUoReadS64(volatile int64_t RT_FAR *pi64) RT_NOTHROW_DEF { return (int64_t)ASMAtomicUoReadU64((volatile uint64_t RT_FAR *)pi64); } /** @def RTASM_HAVE_READ_U128 * Defined in the target architecture supports atomic reading of 128-bit * integers. * * The define value is zero if both ordered and unordered reads are implemented * using ASMAtomicCmpXchgU128v2(). It is 1 if unordered reads are done natively * w/o cmpxchg and 3 if both variants are done natively w/o cmpxchg. * * @note AMD64: Caller must check for cmpxchg16b support before use and make * sure variables are writable (won't be changed). * @sa RTASM_HAVE_CMP_XCHG_U128, RTASM_HAVE_WRITE_U128 */ #if defined(RT_ARCH_ARM64) || defined(DOXYGEN_RUNNING) # define RTASM_HAVE_READ_U128 3 #elif defined(RTASM_HAVE_CMP_XCHG_U128) # define RTASM_HAVE_READ_U128 0 #endif #ifdef RTASM_HAVE_READ_U128 /** * Atomically reads an unsigned 128-bit value, ordered. * * @returns Current *pu128 value * @param pu128 Pointer to the 128-bit variable to read. * The memory pointed to must be writable. * * @remarks AMD64: Requires the memory to be both readable and writable. * @remarks AMD64: Requires support for cmpxchg16b. */ DECLINLINE(uint128_t) ASMAtomicReadU128(volatile uint128_t RT_FAR *pu128) RT_NOTHROW_DEF { RTUINT128U u128Ret; Assert(!((uintptr_t)pu128 & 15)); # if defined(__GNUC__) && defined(RT_ARCH_ARM64) __asm__ __volatile__("Lstart_ASMAtomicReadU128_%=:\n\t" RTASM_ARM_DMB_SY "ldp %[uRetLo], %[uRetHi], %[pMem]\n\t" RTASM_ARM_DMB_SY : [uRetHi] "=r" (u128Ret.s.Hi) , [uRetLo] "=r" (u128Ret.s.Lo) : [pMem] "Q" (*pu128) : ); return u128Ret.u; # else ASMAtomicCmpXchgU128v2(pu128, 0, 0, 0, 0, &u128Ret.u); return u128Ret.u; # endif } /** * Atomically reads an unsigned 128-bit value, ordered. * * @returns Current *pu128 value * @param pu128 Pointer to the 128-bit variable to read. * The memory pointed to must be writable. * * @remarks AMD64: Requires the memory to be both readable and writable. * @remarks AMD64: Requires support for cmpxchg16b. */ DECLINLINE(RTUINT128U) ASMAtomicReadU128U(volatile RTUINT128U RT_FAR *pu128) RT_NOTHROW_DEF { RTUINT128U u128Ret; Assert(!((uintptr_t)pu128 & 15)); # if defined(__GNUC__) && defined(RT_ARCH_ARM64) __asm__ __volatile__("Lstart_ASMAtomicReadU128U_%=:\n\t" RTASM_ARM_DMB_SY "ldp %[uRetLo], %[uRetHi], %[pMem]\n\t" RTASM_ARM_DMB_SY : [uRetHi] "=r" (u128Ret.s.Hi) , [uRetLo] "=r" (u128Ret.s.Lo) : [pMem] "Q" (*pu128) : ); return u128Ret; # else ASMAtomicCmpXchgU128v2(&pu128->u, 0, 0, 0, 0, &u128Ret.u); return u128Ret; # endif } /** * Atomically reads an unsigned 128-bit value, unordered. * * @returns Current *pu128 value * @param pu128 Pointer to the 128-bit variable to read. * The memory pointed to must be writable. * * @remarks AMD64: Requires the memory to be both readable and writable. * @remarks AMD64: Requires support for cmpxchg16b. * @remarks AMD64: Is ordered. */ DECLINLINE(uint128_t) ASMAtomicUoReadU128(volatile uint128_t RT_FAR *pu128) RT_NOTHROW_DEF { Assert(!((uintptr_t)pu128 & 15)); # if defined(__GNUC__) && defined(RT_ARCH_ARM64) RTUINT128U u128Ret; __asm__ __volatile__("Lstart_ASMAtomicUoReadU128_%=:\n\t" "ldp %[uRetLo], %[uRetHi], %[pMem]\n\t" : [uRetHi] "=r" (u128Ret.s.Hi) , [uRetLo] "=r" (u128Ret.s.Lo) : [pMem] "Q" (*pu128) : ); return u128Ret.u; # elif defined(RT_ARCH_AMD64) && 0 /* This doesn't work because __m128i can't be made volatile and we're not able to force MSC (2019) to emit _mm_load_si128 (besides it emits movdqu instead of movdqa). */ __m128i uTmpSse = _mm_load_si128((__m128i volatile *)pu128); __m128i uTmpSseHi = _mm_srli_si128(uTmpSse, 64 / 8); RTUINT128U u128Ret; u128Ret.s.Lo = (uint64_t)_mm_cvtsi128_si64(uTmpSse); u128Ret.s.Hi = (uint64_t)_mm_cvtsi128_si64(uTmpSseHi); return u128Ret.u; # else return ASMAtomicReadU128(pu128); # endif } /** * Atomically reads an unsigned 128-bit value, unordered. * * @returns Current *pu128 value * @param pu128 Pointer to the 128-bit variable to read. * The memory pointed to must be writable. * * @remarks AMD64: Requires the memory to be both readable and writable. * @remarks AMD64: Requires support for cmpxchg16b. * @remarks AMD64: Is ordered. */ DECLINLINE(RTUINT128U) ASMAtomicUoReadU128U(volatile RTUINT128U RT_FAR *pu128) RT_NOTHROW_DEF { Assert(!((uintptr_t)pu128 & 15)); # if defined(__GNUC__) && defined(RT_ARCH_ARM64) RTUINT128U u128Ret; __asm__ __volatile__("Lstart_ASMAtomicUoReadU128U_%=:\n\t" "ldp %[uRetLo], %[uRetHi], %[pMem]\n\t" : [uRetHi] "=r" (u128Ret.s.Hi) , [uRetLo] "=r" (u128Ret.s.Lo) : [pMem] "Q" (*pu128) : ); return u128Ret; # else return ASMAtomicReadU128U(pu128); # endif } #endif /* RTASM_HAVE_READ_U128 */ /** * Atomically reads a size_t value, ordered. * * @returns Current *pcb value * @param pcb Pointer to the size_t variable to read. */ DECLINLINE(size_t) ASMAtomicReadZ(size_t volatile RT_FAR *pcb) RT_NOTHROW_DEF { #if ARCH_BITS == 64 return ASMAtomicReadU64((uint64_t volatile RT_FAR *)pcb); #elif ARCH_BITS == 32 return ASMAtomicReadU32((uint32_t volatile RT_FAR *)pcb); #elif ARCH_BITS == 16 AssertCompileSize(size_t, 2); return ASMAtomicReadU16((uint16_t volatile RT_FAR *)pcb); #else # error "Unsupported ARCH_BITS value" #endif } /** * Atomically reads a size_t value, unordered. * * @returns Current *pcb value * @param pcb Pointer to the size_t variable to read. */ DECLINLINE(size_t) ASMAtomicUoReadZ(size_t volatile RT_FAR *pcb) RT_NOTHROW_DEF { #if ARCH_BITS == 64 || ARCH_BITS == 16 return ASMAtomicUoReadU64((uint64_t volatile RT_FAR *)pcb); #elif ARCH_BITS == 32 return ASMAtomicUoReadU32((uint32_t volatile RT_FAR *)pcb); #elif ARCH_BITS == 16 AssertCompileSize(size_t, 2); return ASMAtomicUoReadU16((uint16_t volatile RT_FAR *)pcb); #else # error "Unsupported ARCH_BITS value" #endif } /** * Atomically reads a pointer value, ordered. * * @returns Current *pv value * @param ppv Pointer to the pointer variable to read. * * @remarks Please use ASMAtomicReadPtrT, it provides better type safety and * requires less typing (no casts). */ DECLINLINE(void RT_FAR *) ASMAtomicReadPtr(void RT_FAR * volatile RT_FAR *ppv) RT_NOTHROW_DEF { #if ARCH_BITS == 32 || ARCH_BITS == 16 return (void RT_FAR *)ASMAtomicReadU32((volatile uint32_t RT_FAR *)(void RT_FAR *)ppv); #elif ARCH_BITS == 64 return (void RT_FAR *)ASMAtomicReadU64((volatile uint64_t RT_FAR *)(void RT_FAR *)ppv); #else # error "ARCH_BITS is bogus" #endif } /** * Convenience macro for avoiding the annoying casting with ASMAtomicReadPtr. * * @returns Current *pv value * @param ppv Pointer to the pointer variable to read. * @param Type The type of *ppv, sans volatile. */ #ifdef __GNUC__ /* 8.2.0 requires -Wno-ignored-qualifiers */ # define ASMAtomicReadPtrT(ppv, Type) \ __extension__ \ ({\ __typeof__(*(ppv)) volatile *ppvTypeChecked = (ppv); \ Type pvTypeChecked = (__typeof__(*(ppv))) ASMAtomicReadPtr((void * volatile *)ppvTypeChecked); \ pvTypeChecked; \ }) #else # define ASMAtomicReadPtrT(ppv, Type) \ (Type)ASMAtomicReadPtr((void RT_FAR * volatile RT_FAR *)(ppv)) #endif /** * Atomically reads a pointer value, unordered. * * @returns Current *pv value * @param ppv Pointer to the pointer variable to read. * * @remarks Please use ASMAtomicUoReadPtrT, it provides better type safety and * requires less typing (no casts). */ DECLINLINE(void RT_FAR *) ASMAtomicUoReadPtr(void RT_FAR * volatile RT_FAR *ppv) RT_NOTHROW_DEF { #if ARCH_BITS == 32 || ARCH_BITS == 16 return (void RT_FAR *)ASMAtomicUoReadU32((volatile uint32_t RT_FAR *)(void RT_FAR *)ppv); #elif ARCH_BITS == 64 return (void RT_FAR *)ASMAtomicUoReadU64((volatile uint64_t RT_FAR *)(void RT_FAR *)ppv); #else # error "ARCH_BITS is bogus" #endif } /** * Convenience macro for avoiding the annoying casting with ASMAtomicUoReadPtr. * * @returns Current *pv value * @param ppv Pointer to the pointer variable to read. * @param Type The type of *ppv, sans volatile. */ #ifdef __GNUC__ /* 8.2.0 requires -Wno-ignored-qualifiers */ # define ASMAtomicUoReadPtrT(ppv, Type) \ __extension__ \ ({\ __typeof__(*(ppv)) volatile * const ppvTypeChecked = (ppv); \ Type pvTypeChecked = (__typeof__(*(ppv))) ASMAtomicUoReadPtr((void * volatile *)ppvTypeChecked); \ pvTypeChecked; \ }) #else # define ASMAtomicUoReadPtrT(ppv, Type) \ (Type)ASMAtomicUoReadPtr((void RT_FAR * volatile RT_FAR *)(ppv)) #endif /** * Atomically reads a boolean value, ordered. * * @returns Current *pf value * @param pf Pointer to the boolean variable to read. */ DECLINLINE(bool) ASMAtomicReadBool(volatile bool RT_FAR *pf) RT_NOTHROW_DEF { ASMMemoryFence(); return *pf; /* byte reads are atomic on x86 */ } /** * Atomically reads a boolean value, unordered. * * @returns Current *pf value * @param pf Pointer to the boolean variable to read. */ DECLINLINE(bool) ASMAtomicUoReadBool(volatile bool RT_FAR *pf) RT_NOTHROW_DEF { return *pf; /* byte reads are atomic on x86 */ } /** * Atomically read a typical IPRT handle value, ordered. * * @param ph Pointer to the handle variable to read. * @param phRes Where to store the result. * * @remarks This doesn't currently work for all handles (like RTFILE). */ #if HC_ARCH_BITS == 32 || ARCH_BITS == 16 # define ASMAtomicReadHandle(ph, phRes) \ do { \ AssertCompile(sizeof(*(ph)) == sizeof(uint32_t)); \ AssertCompile(sizeof(*(phRes)) == sizeof(uint32_t)); \ *(uint32_t RT_FAR *)(phRes) = ASMAtomicReadU32((uint32_t volatile RT_FAR *)(ph)); \ } while (0) #elif HC_ARCH_BITS == 64 # define ASMAtomicReadHandle(ph, phRes) \ do { \ AssertCompile(sizeof(*(ph)) == sizeof(uint64_t)); \ AssertCompile(sizeof(*(phRes)) == sizeof(uint64_t)); \ *(uint64_t RT_FAR *)(phRes) = ASMAtomicReadU64((uint64_t volatile RT_FAR *)(ph)); \ } while (0) #else # error HC_ARCH_BITS #endif /** * Atomically read a typical IPRT handle value, unordered. * * @param ph Pointer to the handle variable to read. * @param phRes Where to store the result. * * @remarks This doesn't currently work for all handles (like RTFILE). */ #if HC_ARCH_BITS == 32 || ARCH_BITS == 16 # define ASMAtomicUoReadHandle(ph, phRes) \ do { \ AssertCompile(sizeof(*(ph)) == sizeof(uint32_t)); \ AssertCompile(sizeof(*(phRes)) == sizeof(uint32_t)); \ *(uint32_t RT_FAR *)(phRes) = ASMAtomicUoReadU32((uint32_t volatile RT_FAR *)(ph)); \ } while (0) #elif HC_ARCH_BITS == 64 # define ASMAtomicUoReadHandle(ph, phRes) \ do { \ AssertCompile(sizeof(*(ph)) == sizeof(uint64_t)); \ AssertCompile(sizeof(*(phRes)) == sizeof(uint64_t)); \ *(uint64_t RT_FAR *)(phRes) = ASMAtomicUoReadU64((uint64_t volatile RT_FAR *)(ph)); \ } while (0) #else # error HC_ARCH_BITS #endif /** * Atomically read a value which size might differ * between platforms or compilers, ordered. * * @param pu Pointer to the variable to read. * @param puRes Where to store the result. */ #define ASMAtomicReadSize(pu, puRes) \ do { \ switch (sizeof(*(pu))) { \ case 1: *(uint8_t RT_FAR *)(puRes) = ASMAtomicReadU8( (volatile uint8_t RT_FAR *)(void RT_FAR *)(pu)); break; \ case 2: *(uint16_t RT_FAR *)(puRes) = ASMAtomicReadU16((volatile uint16_t RT_FAR *)(void RT_FAR *)(pu)); break; \ case 4: *(uint32_t RT_FAR *)(puRes) = ASMAtomicReadU32((volatile uint32_t RT_FAR *)(void RT_FAR *)(pu)); break; \ case 8: *(uint64_t RT_FAR *)(puRes) = ASMAtomicReadU64((volatile uint64_t RT_FAR *)(void RT_FAR *)(pu)); break; \ default: AssertMsgFailed(("ASMAtomicReadSize: size %d is not supported\n", sizeof(*(pu)))); \ } \ } while (0) /** * Atomically read a value which size might differ * between platforms or compilers, unordered. * * @param pu Pointer to the variable to read. * @param puRes Where to store the result. */ #define ASMAtomicUoReadSize(pu, puRes) \ do { \ switch (sizeof(*(pu))) { \ case 1: *(uint8_t RT_FAR *)(puRes) = ASMAtomicUoReadU8( (volatile uint8_t RT_FAR *)(void RT_FAR *)(pu)); break; \ case 2: *(uint16_t RT_FAR *)(puRes) = ASMAtomicUoReadU16((volatile uint16_t RT_FAR *)(void RT_FAR *)(pu)); break; \ case 4: *(uint32_t RT_FAR *)(puRes) = ASMAtomicUoReadU32((volatile uint32_t RT_FAR *)(void RT_FAR *)(pu)); break; \ case 8: *(uint64_t RT_FAR *)(puRes) = ASMAtomicUoReadU64((volatile uint64_t RT_FAR *)(void RT_FAR *)(pu)); break; \ default: AssertMsgFailed(("ASMAtomicReadSize: size %d is not supported\n", sizeof(*(pu)))); \ } \ } while (0) /** * Atomically writes an unsigned 8-bit value, ordered. * * @param pu8 Pointer to the 8-bit variable. * @param u8 The 8-bit value to assign to *pu8. */ DECLINLINE(void) ASMAtomicWriteU8(volatile uint8_t RT_FAR *pu8, uint8_t u8) RT_NOTHROW_DEF { #if defined(RT_ARCH_ARM64) /* The DMB SY will ensure ordering a la x86, the stlrb is probably overkill as all byte accesses are single-copy atomic, which I think suffices here. */ __asm__ __volatile__("Lstart_ASMAtomicWriteU8_%=:\n\t" # if defined(RTASM_ARM64_USE_FEAT_LSE) && 0 /* this is a lot slower and has no alignment benefits with LSE2 */ RTASM_ARM_DMB_SY "swpb %w[uValue], wzr, %[pMem]\n\t" # else RTASM_ARM_DMB_SY "stlrb %w[uValue], %[pMem]\n\t" /* single-copy atomic w/ release semantics. */ # endif : [pMem] "+Q" (*pu8) : [uValue] "r" ((uint32_t)u8) : ); #else ASMAtomicXchgU8(pu8, u8); #endif } /** * Atomically writes an unsigned 8-bit value, unordered. * * @param pu8 Pointer to the 8-bit variable. * @param u8 The 8-bit value to assign to *pu8. */ DECLINLINE(void) ASMAtomicUoWriteU8(volatile uint8_t RT_FAR *pu8, uint8_t u8) RT_NOTHROW_DEF { *pu8 = u8; /* byte writes are atomic on x86 */ } /** * Atomically writes a signed 8-bit value, ordered. * * @param pi8 Pointer to the 8-bit variable to read. * @param i8 The 8-bit value to assign to *pi8. */ DECLINLINE(void) ASMAtomicWriteS8(volatile int8_t RT_FAR *pi8, int8_t i8) RT_NOTHROW_DEF { #if defined(RT_ARCH_ARM64) ASMAtomicWriteU8((volatile uint8_t RT_FAR *)pi8, (uint8_t)i8); #else ASMAtomicXchgS8(pi8, i8); #endif } /** * Atomically writes a signed 8-bit value, unordered. * * @param pi8 Pointer to the 8-bit variable to write. * @param i8 The 8-bit value to assign to *pi8. */ DECLINLINE(void) ASMAtomicUoWriteS8(volatile int8_t RT_FAR *pi8, int8_t i8) RT_NOTHROW_DEF { *pi8 = i8; /* byte writes are atomic on x86 */ } /** * Atomically writes an unsigned 16-bit value, ordered. * * @param pu16 Pointer to the 16-bit variable to write. * @param u16 The 16-bit value to assign to *pu16. */ DECLINLINE(void) ASMAtomicWriteU16(volatile uint16_t RT_FAR *pu16, uint16_t u16) RT_NOTHROW_DEF { #if defined(RT_ARCH_ARM64) __asm__ __volatile__("Lstart_ASMAtomicWriteU16_%=:\n\t" # if defined(RTASM_ARM64_USE_FEAT_LSE) /* slower on M1, but benefits from relaxed LSE2 alignment requirements (M2?). */ RTASM_ARM_DMB_SY "swph %w[uValue], wzr, %[pMem]\n\t" # else RTASM_ARM_DMB_SY "stlrh %w[uValue], %[pMem]\n\t" /* single-copy atomic w/ release semantics. */ # endif : [pMem] "+Q" (*pu16) : [uValue] "r" ((uint32_t)u16) : ); #else ASMAtomicXchgU16(pu16, u16); #endif } /** * Atomically writes an unsigned 16-bit value, unordered. * * @param pu16 Pointer to the 16-bit variable to write. * @param u16 The 16-bit value to assign to *pu16. */ DECLINLINE(void) ASMAtomicUoWriteU16(volatile uint16_t RT_FAR *pu16, uint16_t u16) RT_NOTHROW_DEF { Assert(!((uintptr_t)pu16 & 1)); *pu16 = u16; } /** * Atomically writes a signed 16-bit value, ordered. * * @param pi16 Pointer to the 16-bit variable to write. * @param i16 The 16-bit value to assign to *pi16. */ DECLINLINE(void) ASMAtomicWriteS16(volatile int16_t RT_FAR *pi16, int16_t i16) RT_NOTHROW_DEF { #if defined(RT_ARCH_ARM64) ASMAtomicWriteU16((volatile uint16_t RT_FAR *)pi16, (uint16_t)i16); #else ASMAtomicXchgS16(pi16, i16); #endif } /** * Atomically writes a signed 16-bit value, unordered. * * @param pi16 Pointer to the 16-bit variable to write. * @param i16 The 16-bit value to assign to *pi16. */ DECLINLINE(void) ASMAtomicUoWriteS16(volatile int16_t RT_FAR *pi16, int16_t i16) RT_NOTHROW_DEF { Assert(!((uintptr_t)pi16 & 1)); *pi16 = i16; } /** * Atomically writes an unsigned 32-bit value, ordered. * * @param pu32 Pointer to the 32-bit variable to write. * @param u32 The 32-bit value to assign to *pu32. */ DECLINLINE(void) ASMAtomicWriteU32(volatile uint32_t RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF { #if defined(RT_ARCH_ARM64) __asm__ __volatile__("Lstart_ASMAtomicWriteU32_%=:\n\t" # if defined(RTASM_ARM64_USE_FEAT_LSE) /* slower on M1, but benefits from relaxed LSE2 alignment requirements (M2?). */ RTASM_ARM_DMB_SY "swp %w[uValue], wzr, %[pMem]\n\t" # else RTASM_ARM_DMB_SY "stlr %w[uValue], %[pMem]\n\t" /* single-copy atomic w/ release semantics. */ # endif : [pMem] "+Q" (*pu32) : [uValue] "r" (u32) : "cc"); #else ASMAtomicXchgU32(pu32, u32); #endif } /** * Atomically writes an unsigned 32-bit value, unordered. * * @param pu32 Pointer to the 32-bit variable to write. * @param u32 The 32-bit value to assign to *pu32. */ DECLINLINE(void) ASMAtomicUoWriteU32(volatile uint32_t RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF { Assert(!((uintptr_t)pu32 & 3)); #if ARCH_BITS >= 32 *pu32 = u32; #else ASMAtomicXchgU32(pu32, u32); #endif } /** * Atomically writes a signed 32-bit value, ordered. * * @param pi32 Pointer to the 32-bit variable to write. * @param i32 The 32-bit value to assign to *pi32. */ DECLINLINE(void) ASMAtomicWriteS32(volatile int32_t RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF { #if defined(RT_ARCH_ARM64) ASMAtomicWriteU32((volatile uint32_t RT_FAR *)pi32, (uint32_t)i32); #else ASMAtomicXchgS32(pi32, i32); #endif } /** * Atomically writes a signed 32-bit value, unordered. * * @param pi32 Pointer to the 32-bit variable to write. * @param i32 The 32-bit value to assign to *pi32. */ DECLINLINE(void) ASMAtomicUoWriteS32(volatile int32_t RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF { Assert(!((uintptr_t)pi32 & 3)); #if ARCH_BITS >= 32 *pi32 = i32; #else ASMAtomicXchgS32(pi32, i32); #endif } /** * Atomically writes an unsigned 64-bit value, ordered. * * @param pu64 Pointer to the 64-bit variable to write. * @param u64 The 64-bit value to assign to *pu64. */ DECLINLINE(void) ASMAtomicWriteU64(volatile uint64_t RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF { #if defined(RT_ARCH_ARM64) __asm__ __volatile__("Lstart_ASMAtomicWriteU64_%=:\n\t" # if defined(RTASM_ARM64_USE_FEAT_LSE) /* slower on M1, but benefits from relaxed LSE2 alignment requirements (M2?). */ RTASM_ARM_DMB_SY "swp %[uValue], xzr, %[pMem]\n\t" # else RTASM_ARM_DMB_SY /** @todo necessary? */ "stlr %[uValue], %[pMem]\n\t" # endif : [pMem] "+Q" (*pu64) : [uValue] "r" (u64) : ); #else ASMAtomicXchgU64(pu64, u64); #endif } /** * Atomically writes an unsigned 64-bit value, unordered. * * @param pu64 Pointer to the 64-bit variable to write. * @param u64 The 64-bit value to assign to *pu64. */ DECLINLINE(void) ASMAtomicUoWriteU64(volatile uint64_t RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF { Assert(!((uintptr_t)pu64 & 7)); #if ARCH_BITS == 64 *pu64 = u64; #else ASMAtomicXchgU64(pu64, u64); #endif } /** * Atomically writes a signed 64-bit value, ordered. * * @param pi64 Pointer to the 64-bit variable to write. * @param i64 The 64-bit value to assign to *pi64. */ DECLINLINE(void) ASMAtomicWriteS64(volatile int64_t RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF { #if defined(RT_ARCH_ARM64) ASMAtomicWriteU64((volatile uint64_t RT_FAR *)pi64, (uint64_t)i64); #else ASMAtomicXchgS64(pi64, i64); #endif } /** * Atomically writes a signed 64-bit value, unordered. * * @param pi64 Pointer to the 64-bit variable to write. * @param i64 The 64-bit value to assign to *pi64. */ DECLINLINE(void) ASMAtomicUoWriteS64(volatile int64_t RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF { Assert(!((uintptr_t)pi64 & 7)); #if ARCH_BITS == 64 *pi64 = i64; #else ASMAtomicXchgS64(pi64, i64); #endif } /** @def RTASM_HAVE_WRITE_U128 * Defined in the target architecture supports atomic of 128-bit integers. * * The define value is zero if both ordered and unordered writes are implemented * using ASMAtomicCmpXchgU128v2(). It is 1 if unordered writes are done * natively w/o cmpxchg and 3 if both variants are done natively w/o cmpxchg. * * @note AMD64: Caller must check for cmpxchg16b support before use. * @sa RTASM_HAVE_CMP_XCHG_U128 */ #if defined(RT_ARCH_ARM64) || defined(DOXYGEN_RUNNING) # define RTASM_HAVE_WRITE_U128 3 #elif defined(RTASM_HAVE_CMP_XCHG_U128) # define RTASM_HAVE_WRITE_U128 0 #endif #ifdef RTASM_HAVE_WRITE_U128 /** * Atomically writes an unsigned 128-bit value, ordered. * * @param pu128 Pointer to the variable to overwrite. Must be aligned * on 16 byte boundrary. * @param u64Hi The high 64 bits of the new value. * @param u64Lo The low 64 bits of the new value. */ DECLINLINE(void) ASMAtomicWriteU128v2(volatile uint128_t *pu128, const uint64_t u64Hi, const uint64_t u64Lo) RT_NOTHROW_DEF { # if !defined(__GNUC__) || !defined(RT_ARCH_ARM64) RTUINT128U u128Old; # endif Assert(!((uintptr_t)pu128 & 15)); # if defined(__GNUC__) && defined(RT_ARCH_ARM64) __asm__ __volatile__("Lstart_ASMAtomicWriteU128v2_%=:\n\t" # if 0 && defined(RTASM_ARM64_USE_FEAT_LSE128) /** @todo hw support? test + debug */ RTASM_ARM_DMB_SY "swpp %[uValueLo], %[uValueHi], %[pMem]\n\t" # else RTASM_ARM_DMB_SY "stp %[uValueLo], %[uValueHi], %[pMem]\n\t" "dmb sy\n\t" # endif : [pMem] "+Q" (*pu128) : [uValueHi] "r" (u64Hi) , [uValueLo] "r" (u64Lo) : ); # else # ifdef RT_COMPILER_WITH_128BIT_INT_TYPES u128Old.u = *pu128; # else u128Old.u.Lo = pu128->Lo; u128Old.u.Hi = pu128->Hi; # endif while (!ASMAtomicCmpXchgU128v2(pu128, u64Hi, u64Lo, u128Old.s.Hi, u128Old.s.Lo, &u128Old.u)) { } # endif } /** * Atomically writes an unsigned 128-bit value, ordered. * * @param pu128 Pointer to the variable to overwrite. Must be aligned * on 16 byte boundrary. * @param u64Hi The high 64 bits of the new value. * @param u64Lo The low 64 bits of the new value. * @note This is ordered on AMD64. */ DECLINLINE(void) ASMAtomicUoWriteU128v2(volatile uint128_t *pu128, const uint64_t u64Hi, const uint64_t u64Lo) RT_NOTHROW_DEF { # if !defined(__GNUC__) || !defined(RT_ARCH_ARM64) RTUINT128U u128Old; # endif Assert(!((uintptr_t)pu128 & 15)); # if defined(__GNUC__) && defined(RT_ARCH_ARM64) __asm__ __volatile__("Lstart_ASMAtomicUoWriteU128v2_%=:\n\t" "stp %[uValueLo], %[uValueHi], %[pMem]\n\t" : [pMem] "+Q" (*pu128) : [uValueHi] "r" (u64Hi) , [uValueLo] "r" (u64Lo) : ); # else # ifdef RT_COMPILER_WITH_128BIT_INT_TYPES u128Old.u = *pu128; # else u128Old.u.Lo = pu128->Lo; u128Old.u.Hi = pu128->Hi; # endif while (!ASMAtomicCmpXchgU128v2(pu128, u64Hi, u64Lo, u128Old.s.Hi, u128Old.s.Lo, &u128Old.u)) { } # endif } /** * Atomically writes an unsigned 128-bit value, ordered. * * @param pu128 Pointer to the variable to overwrite. Must be aligned * on 16 byte boundrary. * @param u128 The the new value. */ DECLINLINE(void) ASMAtomicWriteU128(volatile uint128_t *pu128, const uint128_t u128) RT_NOTHROW_DEF { # ifdef RT_COMPILER_WITH_128BIT_INT_TYPES ASMAtomicWriteU128v2(pu128, (uint64_t)(u128 >> 64), (uint64_t)u128); # else ASMAtomicWriteU128v2(pu128, u128.Hi, u128.Lo); # endif } /** * Atomically writes an unsigned 128-bit value, unordered. * * @param pu128 Pointer to the variable to overwrite. Must be aligned * on 16 byte boundrary. * @param u128 The the new value. * @note This is ordered on AMD64. */ DECLINLINE(void) ASMAtomicUoWriteU128(volatile uint128_t *pu128, const uint128_t u128) RT_NOTHROW_DEF { # ifdef RT_COMPILER_WITH_128BIT_INT_TYPES ASMAtomicUoWriteU128v2(pu128, (uint64_t)(u128 >> 64), (uint64_t)u128); # else ASMAtomicUoWriteU128v2(pu128, u128.Hi, u128.Lo); # endif } /** * Atomically writes an unsigned 128-bit value, ordered. * * @param pu128 Pointer to the variable to overwrite. Must be aligned * on 16 byte boundrary. * @param u128 The the new value. */ DECLINLINE(void) ASMAtomicWriteU128U(volatile RTUINT128U *pu128, const RTUINT128U u128) RT_NOTHROW_DEF { ASMAtomicWriteU128v2(&pu128->u, u128.s.Hi, u128.s.Lo); } /** * Atomically writes an unsigned 128-bit value, unordered. * * @param pu128 Pointer to the variable to overwrite. Must be aligned * on 16 byte boundrary. * @param u128 The the new value. * @note This is ordered on AMD64. */ DECLINLINE(void) ASMAtomicUoWriteU128U(volatile RTUINT128U *pu128, const RTUINT128U u128) RT_NOTHROW_DEF { ASMAtomicUoWriteU128v2(&pu128->u, u128.s.Hi, u128.s.Lo); } #endif /* RTASM_HAVE_WRITE_U128 */ /** * Atomically writes a size_t value, ordered. * * @param pcb Pointer to the size_t variable to write. * @param cb The value to assign to *pcb. */ DECLINLINE(void) ASMAtomicWriteZ(volatile size_t RT_FAR *pcb, size_t cb) RT_NOTHROW_DEF { #if ARCH_BITS == 64 ASMAtomicWriteU64((uint64_t volatile *)pcb, cb); #elif ARCH_BITS == 32 ASMAtomicWriteU32((uint32_t volatile *)pcb, cb); #elif ARCH_BITS == 16 AssertCompileSize(size_t, 2); ASMAtomicWriteU16((uint16_t volatile *)pcb, cb); #else # error "Unsupported ARCH_BITS value" #endif } /** * Atomically writes a size_t value, unordered. * * @param pcb Pointer to the size_t variable to write. * @param cb The value to assign to *pcb. */ DECLINLINE(void) ASMAtomicUoWriteZ(volatile size_t RT_FAR *pcb, size_t cb) RT_NOTHROW_DEF { #if ARCH_BITS == 64 ASMAtomicUoWriteU64((uint64_t volatile *)pcb, cb); #elif ARCH_BITS == 32 ASMAtomicUoWriteU32((uint32_t volatile *)pcb, cb); #elif ARCH_BITS == 16 AssertCompileSize(size_t, 2); ASMAtomicUoWriteU16((uint16_t volatile *)pcb, cb); #else # error "Unsupported ARCH_BITS value" #endif } /** * Atomically writes a boolean value, unordered. * * @param pf Pointer to the boolean variable to write. * @param f The boolean value to assign to *pf. */ DECLINLINE(void) ASMAtomicWriteBool(volatile bool RT_FAR *pf, bool f) RT_NOTHROW_DEF { ASMAtomicWriteU8((uint8_t volatile RT_FAR *)pf, f); } /** * Atomically writes a boolean value, unordered. * * @param pf Pointer to the boolean variable to write. * @param f The boolean value to assign to *pf. */ DECLINLINE(void) ASMAtomicUoWriteBool(volatile bool RT_FAR *pf, bool f) RT_NOTHROW_DEF { *pf = f; /* byte writes are atomic on x86 */ } /** * Atomically writes a pointer value, ordered. * * @param ppv Pointer to the pointer variable to write. * @param pv The pointer value to assign to *ppv. */ DECLINLINE(void) ASMAtomicWritePtrVoid(void RT_FAR * volatile RT_FAR *ppv, const void *pv) RT_NOTHROW_DEF { #if ARCH_BITS == 32 || ARCH_BITS == 16 ASMAtomicWriteU32((volatile uint32_t RT_FAR *)(void RT_FAR *)ppv, (uint32_t)pv); #elif ARCH_BITS == 64 ASMAtomicWriteU64((volatile uint64_t RT_FAR *)(void RT_FAR *)ppv, (uint64_t)pv); #else # error "ARCH_BITS is bogus" #endif } /** * Atomically writes a pointer value, unordered. * * @param ppv Pointer to the pointer variable to write. * @param pv The pointer value to assign to *ppv. */ DECLINLINE(void) ASMAtomicUoWritePtrVoid(void RT_FAR * volatile RT_FAR *ppv, const void *pv) RT_NOTHROW_DEF { #if ARCH_BITS == 32 || ARCH_BITS == 16 ASMAtomicUoWriteU32((volatile uint32_t RT_FAR *)(void RT_FAR *)ppv, (uint32_t)pv); #elif ARCH_BITS == 64 ASMAtomicUoWriteU64((volatile uint64_t RT_FAR *)(void RT_FAR *)ppv, (uint64_t)pv); #else # error "ARCH_BITS is bogus" #endif } /** * Atomically writes a pointer value, ordered. * * @param ppv Pointer to the pointer variable to write. * @param pv The pointer value to assign to *ppv. If NULL use * ASMAtomicWriteNullPtr or you'll land in trouble. * * @remarks This is relatively type safe on GCC platforms when @a pv isn't * NULL. */ #ifdef __GNUC__ # define ASMAtomicWritePtr(ppv, pv) \ do \ { \ __typeof__(*(ppv)) volatile RT_FAR * const ppvTypeChecked = (ppv); \ __typeof__(*(ppv)) const pvTypeChecked = (pv); \ \ AssertCompile(sizeof(*ppv) == sizeof(void RT_FAR *)); \ AssertCompile(sizeof(pv) == sizeof(void RT_FAR *)); \ Assert(!( (uintptr_t)ppv & ((ARCH_BITS / 8) - 1) )); \ \ ASMAtomicWritePtrVoid((void RT_FAR * volatile RT_FAR *)(ppvTypeChecked), (void RT_FAR *)(pvTypeChecked)); \ } while (0) #else # define ASMAtomicWritePtr(ppv, pv) \ do \ { \ AssertCompile(sizeof(*ppv) == sizeof(void RT_FAR *)); \ AssertCompile(sizeof(pv) == sizeof(void RT_FAR *)); \ Assert(!( (uintptr_t)ppv & ((ARCH_BITS / 8) - 1) )); \ \ ASMAtomicWritePtrVoid((void RT_FAR * volatile RT_FAR *)(ppv), (void RT_FAR *)(pv)); \ } while (0) #endif /** * Atomically sets a pointer to NULL, ordered. * * @param ppv Pointer to the pointer variable that should be set to NULL. * * @remarks This is relatively type safe on GCC platforms. */ #if RT_GNUC_PREREQ(4, 2) # define ASMAtomicWriteNullPtr(ppv) \ do \ { \ __typeof__(*(ppv)) * const ppvTypeChecked = (ppv); \ AssertCompile(sizeof(*ppv) == sizeof(void RT_FAR *)); \ Assert(!( (uintptr_t)ppv & ((ARCH_BITS / 8) - 1) )); \ ASMAtomicWritePtrVoid((void RT_FAR * volatile RT_FAR *)(ppvTypeChecked), NULL); \ } while (0) #else # define ASMAtomicWriteNullPtr(ppv) \ do \ { \ AssertCompile(sizeof(*ppv) == sizeof(void RT_FAR *)); \ Assert(!( (uintptr_t)ppv & ((ARCH_BITS / 8) - 1) )); \ ASMAtomicWritePtrVoid((void RT_FAR * volatile RT_FAR *)(ppv), NULL); \ } while (0) #endif /** * Atomically writes a pointer value, unordered. * * @returns Current *pv value * @param ppv Pointer to the pointer variable. * @param pv The pointer value to assign to *ppv. If NULL use * ASMAtomicUoWriteNullPtr or you'll land in trouble. * * @remarks This is relatively type safe on GCC platforms when @a pv isn't * NULL. */ #if RT_GNUC_PREREQ(4, 2) # define ASMAtomicUoWritePtr(ppv, pv) \ do \ { \ __typeof__(*(ppv)) volatile * const ppvTypeChecked = (ppv); \ __typeof__(*(ppv)) const pvTypeChecked = (pv); \ \ AssertCompile(sizeof(*ppv) == sizeof(void *)); \ AssertCompile(sizeof(pv) == sizeof(void *)); \ Assert(!( (uintptr_t)ppv & ((ARCH_BITS / 8) - 1) )); \ \ *(ppvTypeChecked) = pvTypeChecked; \ } while (0) #else # define ASMAtomicUoWritePtr(ppv, pv) \ do \ { \ AssertCompile(sizeof(*ppv) == sizeof(void RT_FAR *)); \ AssertCompile(sizeof(pv) == sizeof(void RT_FAR *)); \ Assert(!( (uintptr_t)ppv & ((ARCH_BITS / 8) - 1) )); \ *(ppv) = pv; \ } while (0) #endif /** * Atomically sets a pointer to NULL, unordered. * * @param ppv Pointer to the pointer variable that should be set to NULL. * * @remarks This is relatively type safe on GCC platforms. */ #ifdef __GNUC__ # define ASMAtomicUoWriteNullPtr(ppv) \ do \ { \ __typeof__(*(ppv)) volatile * const ppvTypeChecked = (ppv); \ AssertCompile(sizeof(*ppv) == sizeof(void *)); \ Assert(!( (uintptr_t)ppv & ((ARCH_BITS / 8) - 1) )); \ *(ppvTypeChecked) = NULL; \ } while (0) #else # define ASMAtomicUoWriteNullPtr(ppv) \ do \ { \ AssertCompile(sizeof(*ppv) == sizeof(void RT_FAR *)); \ Assert(!( (uintptr_t)ppv & ((ARCH_BITS / 8) - 1) )); \ *(ppv) = NULL; \ } while (0) #endif /** * Atomically write a typical IPRT handle value, ordered. * * @param ph Pointer to the variable to update. * @param hNew The value to assign to *ph. * * @remarks This doesn't currently work for all handles (like RTFILE). */ #if HC_ARCH_BITS == 32 || ARCH_BITS == 16 # define ASMAtomicWriteHandle(ph, hNew) \ do { \ AssertCompile(sizeof(*(ph)) == sizeof(uint32_t)); \ ASMAtomicWriteU32((uint32_t volatile RT_FAR *)(ph), (const uint32_t)(hNew)); \ } while (0) #elif HC_ARCH_BITS == 64 # define ASMAtomicWriteHandle(ph, hNew) \ do { \ AssertCompile(sizeof(*(ph)) == sizeof(uint64_t)); \ ASMAtomicWriteU64((uint64_t volatile RT_FAR *)(ph), (const uint64_t)(hNew)); \ } while (0) #else # error HC_ARCH_BITS #endif /** * Atomically write a typical IPRT handle value, unordered. * * @param ph Pointer to the variable to update. * @param hNew The value to assign to *ph. * * @remarks This doesn't currently work for all handles (like RTFILE). */ #if HC_ARCH_BITS == 32 || ARCH_BITS == 16 # define ASMAtomicUoWriteHandle(ph, hNew) \ do { \ AssertCompile(sizeof(*(ph)) == sizeof(uint32_t)); \ ASMAtomicUoWriteU32((uint32_t volatile RT_FAR *)(ph), (const uint32_t)hNew); \ } while (0) #elif HC_ARCH_BITS == 64 # define ASMAtomicUoWriteHandle(ph, hNew) \ do { \ AssertCompile(sizeof(*(ph)) == sizeof(uint64_t)); \ ASMAtomicUoWriteU64((uint64_t volatile RT_FAR *)(ph), (const uint64_t)hNew); \ } while (0) #else # error HC_ARCH_BITS #endif /** * Atomically write a value which size might differ * between platforms or compilers, ordered. * * @param pu Pointer to the variable to update. * @param uNew The value to assign to *pu. */ #define ASMAtomicWriteSize(pu, uNew) \ do { \ switch (sizeof(*(pu))) { \ case 1: ASMAtomicWriteU8( (volatile uint8_t RT_FAR *)(void RT_FAR *)(pu), (uint8_t )(uNew)); break; \ case 2: ASMAtomicWriteU16((volatile uint16_t RT_FAR *)(void RT_FAR *)(pu), (uint16_t)(uNew)); break; \ case 4: ASMAtomicWriteU32((volatile uint32_t RT_FAR *)(void RT_FAR *)(pu), (uint32_t)(uNew)); break; \ case 8: ASMAtomicWriteU64((volatile uint64_t RT_FAR *)(void RT_FAR *)(pu), (uint64_t)(uNew)); break; \ default: AssertMsgFailed(("ASMAtomicWriteSize: size %d is not supported\n", sizeof(*(pu)))); \ } \ } while (0) /** * Atomically write a value which size might differ * between platforms or compilers, unordered. * * @param pu Pointer to the variable to update. * @param uNew The value to assign to *pu. */ #define ASMAtomicUoWriteSize(pu, uNew) \ do { \ switch (sizeof(*(pu))) { \ case 1: ASMAtomicUoWriteU8( (volatile uint8_t RT_FAR *)(void RT_FAR *)(pu), (uint8_t )(uNew)); break; \ case 2: ASMAtomicUoWriteU16((volatile uint16_t RT_FAR *)(void RT_FAR *)(pu), (uint16_t)(uNew)); break; \ case 4: ASMAtomicUoWriteU32((volatile uint32_t RT_FAR *)(void RT_FAR *)(pu), (uint32_t)(uNew)); break; \ case 8: ASMAtomicUoWriteU64((volatile uint64_t RT_FAR *)(void RT_FAR *)(pu), (uint64_t)(uNew)); break; \ default: AssertMsgFailed(("ASMAtomicWriteSize: size %d is not supported\n", sizeof(*(pu)))); \ } \ } while (0) /** * Atomically exchanges and adds to a 16-bit value, ordered. * * @returns The old value. * @param pu16 Pointer to the value. * @param u16 Number to add. * * @remarks Currently not implemented, just to make 16-bit code happy. * @remarks x86: Requires a 486 or later. */ RT_ASM_DECL_PRAGMA_WATCOM(uint16_t) ASMAtomicAddU16(uint16_t volatile RT_FAR *pu16, uint32_t u16) RT_NOTHROW_PROTO; /** * Atomically exchanges and adds to a 32-bit value, ordered. * * @returns The old value. * @param pu32 Pointer to the value. * @param u32 Number to add. * * @remarks x86: Requires a 486 or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMAtomicAddU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_PROTO; #else DECLINLINE(uint32_t) ASMAtomicAddU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN u32 = _InterlockedExchangeAdd((long RT_FAR *)pu32, u32); return u32; # elif defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("lock; xaddl %0, %1\n\t" : "=r" (u32) , "=m" (*pu32) : "0" (u32) , "m" (*pu32) : "memory" , "cc"); return u32; # else __asm { mov eax, [u32] # ifdef RT_ARCH_AMD64 mov rdx, [pu32] lock xadd [rdx], eax # else mov edx, [pu32] lock xadd [edx], eax # endif mov [u32], eax } return u32; # endif # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) /* M1 benchmark: ldaddal=6907 vs dmb+ldadd=2114 vs non-lse=6249 (ps/call) */ # if defined(RTASM_ARM64_USE_FEAT_LSE) uint32_t u32OldRet; __asm__ __volatile__("Lstart_ASMAtomicAddU32_%=:\n\t" # if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) "ldaddal %w[uAddend], %w[uOldActual], %[pMem]\n\t" # else RTASM_ARM_DMB_SY "ldadd %w[uAddend], %w[uOldActual], %[pMem]\n\t" # endif : [pMem] "+Q" (*pu32) , [uOldActual] "=&r" (u32OldRet) : [uAddend] "r" (u32) : ); # else RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(ASMAtomicAddU32, pu32, DMB_SY, "add %w[uNew], %w[uOld], %w[uVal]\n\t", "add %[uNew], %[uOld], %[uVal]\n\t", [uVal] "r" (u32)); # endif return u32OldRet; # else # error "Port me" # endif } #endif /** * Atomically exchanges and adds to a signed 32-bit value, ordered. * * @returns The old value. * @param pi32 Pointer to the value. * @param i32 Number to add. * * @remarks x86: Requires a 486 or later. */ DECLINLINE(int32_t) ASMAtomicAddS32(int32_t volatile RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF { return (int32_t)ASMAtomicAddU32((uint32_t volatile RT_FAR *)pi32, (uint32_t)i32); } /** * Atomically exchanges and adds to a 64-bit value, ordered. * * @returns The old value. * @param pu64 Pointer to the value. * @param u64 Number to add. * * @remarks x86: Requires a Pentium or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN DECLASM(uint64_t) ASMAtomicAddU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_PROTO; #else DECLINLINE(uint64_t) ASMAtomicAddU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN && defined(RT_ARCH_AMD64) u64 = _InterlockedExchangeAdd64((__int64 RT_FAR *)pu64, u64); return u64; # elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64) __asm__ __volatile__("lock; xaddq %0, %1\n\t" : "=r" (u64) , "=m" (*pu64) : "0" (u64) , "m" (*pu64) : "memory" , "cc"); return u64; # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) # if defined(RTASM_ARM64_USE_FEAT_LSE) uint64_t u64OldRet; __asm__ __volatile__("Lstart_ASMAtomicAddU64_%=:\n\t" # if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) "ldaddal %[uAddend], %[uOldActual], %[pMem]\n\t" # else RTASM_ARM_DMB_SY "ldadd %[uAddend], %[uOldActual], %[pMem]\n\t" # endif : [pMem] "+Q" (*pu64) , [uOldActual] "=&r" (u64OldRet) : [uAddend] "r" (u64) : ); # else RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_64(ASMAtomicAddU64, pu64, DMB_SY, "add %[uNew], %[uOld], %[uVal]\n\t" , "add %[uNew], %[uOld], %[uVal]\n\t" "adc %H[uNew], %H[uOld], %H[uVal]\n\t", [uVal] "r" (u64)); # endif return u64OldRet; # else uint64_t u64Old; for (;;) { uint64_t u64New; u64Old = ASMAtomicUoReadU64(pu64); u64New = u64Old + u64; if (ASMAtomicCmpXchgU64(pu64, u64New, u64Old)) break; ASMNopPause(); } return u64Old; # endif } #endif /** * Atomically exchanges and adds to a signed 64-bit value, ordered. * * @returns The old value. * @param pi64 Pointer to the value. * @param i64 Number to add. * * @remarks x86: Requires a Pentium or later. */ DECLINLINE(int64_t) ASMAtomicAddS64(int64_t volatile RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF { return (int64_t)ASMAtomicAddU64((uint64_t volatile RT_FAR *)pi64, (uint64_t)i64); } /** * Atomically exchanges and adds to a size_t value, ordered. * * @returns The old value. * @param pcb Pointer to the size_t value. * @param cb Number to add. */ DECLINLINE(size_t) ASMAtomicAddZ(size_t volatile RT_FAR *pcb, size_t cb) RT_NOTHROW_DEF { #if ARCH_BITS == 64 AssertCompileSize(size_t, 8); return ASMAtomicAddU64((uint64_t volatile RT_FAR *)pcb, cb); #elif ARCH_BITS == 32 AssertCompileSize(size_t, 4); return ASMAtomicAddU32((uint32_t volatile RT_FAR *)pcb, cb); #elif ARCH_BITS == 16 AssertCompileSize(size_t, 2); return ASMAtomicAddU16((uint16_t volatile RT_FAR *)pcb, cb); #else # error "Unsupported ARCH_BITS value" #endif } /** * Atomically exchanges and adds a value which size might differ between * platforms or compilers, ordered. * * @param pu Pointer to the variable to update. * @param uNew The value to add to *pu. * @param puOld Where to store the old value. */ #define ASMAtomicAddSize(pu, uNew, puOld) \ do { \ switch (sizeof(*(pu))) { \ case 4: *(uint32_t *)(puOld) = ASMAtomicAddU32((volatile uint32_t RT_FAR *)(void RT_FAR *)(pu), (uint32_t)(uNew)); break; \ case 8: *(uint64_t *)(puOld) = ASMAtomicAddU64((volatile uint64_t RT_FAR *)(void RT_FAR *)(pu), (uint64_t)(uNew)); break; \ default: AssertMsgFailed(("ASMAtomicAddSize: size %d is not supported\n", sizeof(*(pu)))); \ } \ } while (0) /** * Atomically exchanges and subtracts to an unsigned 16-bit value, ordered. * * @returns The old value. * @param pu16 Pointer to the value. * @param u16 Number to subtract. * * @remarks x86: Requires a 486 or later. */ DECLINLINE(uint16_t) ASMAtomicSubU16(uint16_t volatile RT_FAR *pu16, uint32_t u16) RT_NOTHROW_DEF { return ASMAtomicAddU16(pu16, (uint16_t)-(int16_t)u16); } /** * Atomically exchanges and subtracts to a signed 16-bit value, ordered. * * @returns The old value. * @param pi16 Pointer to the value. * @param i16 Number to subtract. * * @remarks x86: Requires a 486 or later. */ DECLINLINE(int16_t) ASMAtomicSubS16(int16_t volatile RT_FAR *pi16, int16_t i16) RT_NOTHROW_DEF { return (int16_t)ASMAtomicAddU16((uint16_t volatile RT_FAR *)pi16, (uint16_t)-i16); } /** * Atomically exchanges and subtracts to an unsigned 32-bit value, ordered. * * @returns The old value. * @param pu32 Pointer to the value. * @param u32 Number to subtract. * * @remarks x86: Requires a 486 or later. */ DECLINLINE(uint32_t) ASMAtomicSubU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF { return ASMAtomicAddU32(pu32, (uint32_t)-(int32_t)u32); } /** * Atomically exchanges and subtracts to a signed 32-bit value, ordered. * * @returns The old value. * @param pi32 Pointer to the value. * @param i32 Number to subtract. * * @remarks x86: Requires a 486 or later. */ DECLINLINE(int32_t) ASMAtomicSubS32(int32_t volatile RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF { return (int32_t)ASMAtomicAddU32((uint32_t volatile RT_FAR *)pi32, (uint32_t)-i32); } /** * Atomically exchanges and subtracts to an unsigned 64-bit value, ordered. * * @returns The old value. * @param pu64 Pointer to the value. * @param u64 Number to subtract. * * @remarks x86: Requires a Pentium or later. */ DECLINLINE(uint64_t) ASMAtomicSubU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF { return ASMAtomicAddU64(pu64, (uint64_t)-(int64_t)u64); } /** * Atomically exchanges and subtracts to a signed 64-bit value, ordered. * * @returns The old value. * @param pi64 Pointer to the value. * @param i64 Number to subtract. * * @remarks x86: Requires a Pentium or later. */ DECLINLINE(int64_t) ASMAtomicSubS64(int64_t volatile RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF { return (int64_t)ASMAtomicAddU64((uint64_t volatile RT_FAR *)pi64, (uint64_t)-i64); } /** * Atomically exchanges and subtracts to a size_t value, ordered. * * @returns The old value. * @param pcb Pointer to the size_t value. * @param cb Number to subtract. * * @remarks x86: Requires a 486 or later. */ DECLINLINE(size_t) ASMAtomicSubZ(size_t volatile RT_FAR *pcb, size_t cb) RT_NOTHROW_DEF { #if ARCH_BITS == 64 return ASMAtomicSubU64((uint64_t volatile RT_FAR *)pcb, cb); #elif ARCH_BITS == 32 return ASMAtomicSubU32((uint32_t volatile RT_FAR *)pcb, cb); #elif ARCH_BITS == 16 AssertCompileSize(size_t, 2); return ASMAtomicSubU16((uint16_t volatile RT_FAR *)pcb, cb); #else # error "Unsupported ARCH_BITS value" #endif } /** * Atomically exchanges and subtracts a value which size might differ between * platforms or compilers, ordered. * * @param pu Pointer to the variable to update. * @param uNew The value to subtract to *pu. * @param puOld Where to store the old value. * * @remarks x86: Requires a 486 or later. */ #define ASMAtomicSubSize(pu, uNew, puOld) \ do { \ switch (sizeof(*(pu))) { \ case 4: *(uint32_t RT_FAR *)(puOld) = ASMAtomicSubU32((volatile uint32_t RT_FAR *)(void RT_FAR *)(pu), (uint32_t)(uNew)); break; \ case 8: *(uint64_t RT_FAR *)(puOld) = ASMAtomicSubU64((volatile uint64_t RT_FAR *)(void RT_FAR *)(pu), (uint64_t)(uNew)); break; \ default: AssertMsgFailed(("ASMAtomicSubSize: size %d is not supported\n", sizeof(*(pu)))); \ } \ } while (0) /** * Atomically increment a 16-bit value, ordered. * * @returns The new value. * @param pu16 Pointer to the value to increment. * @remarks Not implemented. Just to make 16-bit code happy. * * @remarks x86: Requires a 486 or later. */ RT_ASM_DECL_PRAGMA_WATCOM(uint16_t) ASMAtomicIncU16(uint16_t volatile RT_FAR *pu16) RT_NOTHROW_PROTO; /** * Atomically increment a 32-bit value, ordered. * * @returns The new value. * @param pu32 Pointer to the value to increment. * * @remarks x86: Requires a 486 or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMAtomicIncU32(uint32_t volatile RT_FAR *pu32) RT_NOTHROW_PROTO; #else DECLINLINE(uint32_t) ASMAtomicIncU32(uint32_t volatile RT_FAR *pu32) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN return (uint32_t)_InterlockedIncrement((long RT_FAR *)pu32); # elif defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE uint32_t u32; __asm__ __volatile__("lock; xaddl %0, %1\n\t" : "=r" (u32) , "=m" (*pu32) : "0" (1) , "m" (*pu32) : "memory" , "cc"); return u32+1; # else __asm { mov eax, 1 # ifdef RT_ARCH_AMD64 mov rdx, [pu32] lock xadd [rdx], eax # else mov edx, [pu32] lock xadd [edx], eax # endif mov u32, eax } return u32+1; # endif # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) /* M1 benchmark: ldaddal=6887 vs dmb+ldadd=2117 vs non-lse=6247 (ps/call) */ # if defined(RTASM_ARM64_USE_FEAT_LSE) uint32_t u32NewRet; __asm__ __volatile__("Lstart_ASMAtomicIncU32_%=:\n\t" # if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) "ldaddal %w[uAddend], %w[uNewRet], %[pMem]\n\t" # else RTASM_ARM_DMB_SY "ldadd %w[uAddend], %w[uNewRet], %[pMem]\n\t" # endif "add %w[uNewRet], %w[uNewRet], #1\n\t" : [pMem] "+Q" (*pu32) , [uNewRet] "=&r" (u32NewRet) : [uAddend] "r" ((uint32_t)1) : ); # else RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicIncU32, pu32, DMB_SY, "add %w[uNew], %w[uNew], #1\n\t", "add %[uNew], %[uNew], #1\n\t" /* arm6 / thumb2+ */, "X" (0) /* dummy */); # endif return u32NewRet; # else return ASMAtomicAddU32(pu32, 1) + 1; # endif } #endif /** * Atomically increment a signed 32-bit value, ordered. * * @returns The new value. * @param pi32 Pointer to the value to increment. * * @remarks x86: Requires a 486 or later. */ DECLINLINE(int32_t) ASMAtomicIncS32(int32_t volatile RT_FAR *pi32) RT_NOTHROW_DEF { return (int32_t)ASMAtomicIncU32((uint32_t volatile RT_FAR *)pi32); } /** * Atomically increment a 64-bit value, ordered. * * @returns The new value. * @param pu64 Pointer to the value to increment. * * @remarks x86: Requires a Pentium or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN DECLASM(uint64_t) ASMAtomicIncU64(uint64_t volatile RT_FAR *pu64) RT_NOTHROW_PROTO; #else DECLINLINE(uint64_t) ASMAtomicIncU64(uint64_t volatile RT_FAR *pu64) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN && defined(RT_ARCH_AMD64) return (uint64_t)_InterlockedIncrement64((__int64 RT_FAR *)pu64); # elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64) uint64_t u64; __asm__ __volatile__("lock; xaddq %0, %1\n\t" : "=r" (u64) , "=m" (*pu64) : "0" (1) , "m" (*pu64) : "memory" , "cc"); return u64 + 1; # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) # if defined(RTASM_ARM64_USE_FEAT_LSE) uint64_t u64NewRet; __asm__ __volatile__("Lstart_ASMAtomicIncU64_%=:\n\t" # if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) "ldaddal %[uAddend], %[uNewRet], %[pMem]\n\t" # else RTASM_ARM_DMB_SY "ldadd %[uAddend], %[uNewRet], %[pMem]\n\t" # endif "add %[uNewRet], %[uNewRet], #1\n\t" : [pMem] "+Q" (*pu64) , [uNewRet] "=&r" (u64NewRet) : [uAddend] "r" ((uint64_t)1) : ); # else RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_64(ASMAtomicIncU64, pu64, DMB_SY, "add %[uNew], %[uNew], #1\n\t" , "add %[uNew], %[uNew], #1\n\t" /* arm6 / thumb2+ */ "adc %H[uNew], %H[uNew], %[uZeroVal]\n\t", RTASM_ARM_PICK_6432("X" (0) /* dummy */, [uZeroVal] "r" (0)) ); # endif return u64NewRet; # else return ASMAtomicAddU64(pu64, 1) + 1; # endif } #endif /** * Atomically increment a signed 64-bit value, ordered. * * @returns The new value. * @param pi64 Pointer to the value to increment. * * @remarks x86: Requires a Pentium or later. */ DECLINLINE(int64_t) ASMAtomicIncS64(int64_t volatile RT_FAR *pi64) RT_NOTHROW_DEF { return (int64_t)ASMAtomicIncU64((uint64_t volatile RT_FAR *)pi64); } /** * Atomically increment a size_t value, ordered. * * @returns The new value. * @param pcb Pointer to the value to increment. * * @remarks x86: Requires a 486 or later. */ DECLINLINE(size_t) ASMAtomicIncZ(size_t volatile RT_FAR *pcb) RT_NOTHROW_DEF { #if ARCH_BITS == 64 return ASMAtomicIncU64((uint64_t volatile RT_FAR *)pcb); #elif ARCH_BITS == 32 return ASMAtomicIncU32((uint32_t volatile RT_FAR *)pcb); #elif ARCH_BITS == 16 return ASMAtomicIncU16((uint16_t volatile RT_FAR *)pcb); #else # error "Unsupported ARCH_BITS value" #endif } /** * Atomically decrement an unsigned 32-bit value, ordered. * * @returns The new value. * @param pu16 Pointer to the value to decrement. * @remarks Not implemented. Just to make 16-bit code happy. * * @remarks x86: Requires a 486 or later. */ RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMAtomicDecU16(uint16_t volatile RT_FAR *pu16) RT_NOTHROW_PROTO; /** * Atomically decrement an unsigned 32-bit value, ordered. * * @returns The new value. * @param pu32 Pointer to the value to decrement. * * @remarks x86: Requires a 486 or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMAtomicDecU32(uint32_t volatile RT_FAR *pu32) RT_NOTHROW_PROTO; #else DECLINLINE(uint32_t) ASMAtomicDecU32(uint32_t volatile RT_FAR *pu32) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN return (uint32_t)_InterlockedDecrement((long RT_FAR *)pu32); # elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64) # if RT_INLINE_ASM_GNU_STYLE uint32_t u32; __asm__ __volatile__("lock; xaddl %0, %1\n\t" : "=r" (u32) , "=m" (*pu32) : "0" (-1) , "m" (*pu32) : "memory" , "cc"); return u32-1; # else uint32_t u32; __asm { mov eax, -1 # ifdef RT_ARCH_AMD64 mov rdx, [pu32] lock xadd [rdx], eax # else mov edx, [pu32] lock xadd [edx], eax # endif mov u32, eax } return u32-1; # endif # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) /* M1 benchmark: ldaddal=6887 vs dmb+ldadd=2120 vs non-lse=6260 (ps/call) */ # if defined(RTASM_ARM64_USE_FEAT_LSE) uint32_t u32NewRet; __asm__ __volatile__("Lstart_ASMAtomicDecU32_%=:\n\t" # if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) "ldaddal %w[uAddend], %w[uNewRet], %[pMem]\n\t" # else RTASM_ARM_DMB_SY "ldadd %w[uAddend], %w[uNewRet], %[pMem]\n\t" # endif "sub %w[uNewRet], %w[uNewRet], #1\n\t" : [pMem] "+Q" (*pu32) , [uNewRet] "=&r" (u32NewRet) : [uAddend] "r" (~(uint32_t)0) : ); # else RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicDecU32, pu32, DMB_SY, "sub %w[uNew], %w[uNew], #1\n\t", "sub %[uNew], %[uNew], #1\n\t" /* arm6 / thumb2+ */, "X" (0) /* dummy */); # endif return u32NewRet; # else return ASMAtomicSubU32(pu32, 1) - (uint32_t)1; # endif } #endif /** * Atomically decrement a signed 32-bit value, ordered. * * @returns The new value. * @param pi32 Pointer to the value to decrement. * * @remarks x86: Requires a 486 or later. */ DECLINLINE(int32_t) ASMAtomicDecS32(int32_t volatile RT_FAR *pi32) RT_NOTHROW_DEF { return (int32_t)ASMAtomicDecU32((uint32_t volatile RT_FAR *)pi32); } /** * Atomically decrement an unsigned 64-bit value, ordered. * * @returns The new value. * @param pu64 Pointer to the value to decrement. * * @remarks x86: Requires a Pentium or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM(uint64_t) ASMAtomicDecU64(uint64_t volatile RT_FAR *pu64) RT_NOTHROW_PROTO; #else DECLINLINE(uint64_t) ASMAtomicDecU64(uint64_t volatile RT_FAR *pu64) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN && defined(RT_ARCH_AMD64) return (uint64_t)_InterlockedDecrement64((__int64 volatile RT_FAR *)pu64); # elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64) uint64_t u64; __asm__ __volatile__("lock; xaddq %q0, %1\n\t" : "=r" (u64) , "=m" (*pu64) : "0" (~(uint64_t)0) , "m" (*pu64) : "memory" , "cc"); return u64-1; # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) # if defined(RTASM_ARM64_USE_FEAT_LSE) uint64_t u64NewRet; __asm__ __volatile__("Lstart_ASMAtomicDecU64_%=:\n\t" # if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) "ldaddal %[uAddend], %[uNewRet], %[pMem]\n\t" # else RTASM_ARM_DMB_SY "ldadd %[uAddend], %[uNewRet], %[pMem]\n\t" # endif "sub %[uNewRet], %[uNewRet], #1\n\t" : [pMem] "+Q" (*pu64) , [uNewRet] "=&r" (u64NewRet) : [uAddend] "r" (~(uint64_t)0) : ); # else RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_64(ASMAtomicDecU64, pu64, DMB_SY, "sub %[uNew], %[uNew], #1\n\t" , "sub %[uNew], %[uNew], #1\n\t" /* arm6 / thumb2+ */ "sbc %H[uNew], %H[uNew], %[uZeroVal]\n\t", RTASM_ARM_PICK_6432("X" (0) /* dummy */, [uZeroVal] "r" (0)) ); # endif return u64NewRet; # else return ASMAtomicAddU64(pu64, UINT64_MAX) - 1; # endif } #endif /** * Atomically decrement a signed 64-bit value, ordered. * * @returns The new value. * @param pi64 Pointer to the value to decrement. * * @remarks x86: Requires a Pentium or later. */ DECLINLINE(int64_t) ASMAtomicDecS64(int64_t volatile RT_FAR *pi64) RT_NOTHROW_DEF { return (int64_t)ASMAtomicDecU64((uint64_t volatile RT_FAR *)pi64); } /** * Atomically decrement a size_t value, ordered. * * @returns The new value. * @param pcb Pointer to the value to decrement. * * @remarks x86: Requires a 486 or later. */ DECLINLINE(size_t) ASMAtomicDecZ(size_t volatile RT_FAR *pcb) RT_NOTHROW_DEF { #if ARCH_BITS == 64 return ASMAtomicDecU64((uint64_t volatile RT_FAR *)pcb); #elif ARCH_BITS == 32 return ASMAtomicDecU32((uint32_t volatile RT_FAR *)pcb); #elif ARCH_BITS == 16 return ASMAtomicDecU16((uint16_t volatile RT_FAR *)pcb); #else # error "Unsupported ARCH_BITS value" #endif } /** * Atomically Or an unsigned 32-bit value, ordered. * * @param pu32 Pointer to the pointer variable to OR u32 with. * @param u32 The value to OR *pu32 with. * * @remarks x86: Requires a 386 or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicOrU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_PROTO; #else DECLINLINE(void) ASMAtomicOrU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN _InterlockedOr((long volatile RT_FAR *)pu32, (long)u32); # elif defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("lock; orl %1, %0\n\t" : "=m" (*pu32) : "ir" (u32) , "m" (*pu32) : "cc"); # else __asm { mov eax, [u32] # ifdef RT_ARCH_AMD64 mov rdx, [pu32] lock or [rdx], eax # else mov edx, [pu32] lock or [edx], eax # endif } # endif # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) # if defined(RTASM_ARM64_USE_FEAT_LSE) # if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) uint32_t u32Spill; __asm__ __volatile__("Lstart_ASMAtomicOrU32_%=:\n\t" "ldsetal %w[fBitsToSet], %w[uSpill], %[pMem]\n\t" : [pMem] "+Q" (*pu32) , [uSpill] "=&r" (u32Spill) : [fBitsToSet] "r" (u32) : ); # else __asm__ __volatile__("Lstart_ASMAtomicOrU32_%=:\n\t" RTASM_ARM_DMB_SY "stset %w[fBitsToSet], %[pMem]\n\t" : [pMem] "+Q" (*pu32) : [fBitsToSet] "r" (u32) : ); # endif # else /* For more on Orr see https://en.wikipedia.org/wiki/Orr_(Catch-22) ;-) */ RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicOr32, pu32, DMB_SY, "orr %w[uNew], %w[uNew], %w[uVal]\n\t", "orr %[uNew], %[uNew], %[uVal]\n\t", [uVal] "r" (u32)); # endif # else # error "Port me" # endif } #endif /** * Atomically OR an unsigned 32-bit value, ordered, extended version (for bitmap * fallback). * * @returns Old value. * @param pu32 Pointer to the variable to OR @a u32 with. * @param u32 The value to OR @a *pu32 with. */ DECLINLINE(uint32_t) ASMAtomicOrExU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF { #if defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) # if defined(RTASM_ARM64_USE_FEAT_LSE) uint32_t u32OldRet; __asm__ __volatile__("Lstart_ASMAtomicOrExU32_%=:\n\t" # if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) "ldsetal %w[fBitsToSet], %w[uOldRet], %[pMem]\n\t" # else RTASM_ARM_DMB_SY "ldset %w[fBitsToSet], %w[uOldRet], %[pMem]\n\t" # endif : [pMem] "+Q" (*pu32) , [uOldRet] "=&r" (u32OldRet) : [fBitsToSet] "r" (u32) : ); # else RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(ASMAtomicOrEx32, pu32, DMB_SY, "orr %w[uNew], %w[uOld], %w[uVal]\n\t", "orr %[uNew], %[uOld], %[uVal]\n\t", [uVal] "r" (u32)); # endif return u32OldRet; #else uint32_t u32RetOld = ASMAtomicUoReadU32(pu32); uint32_t u32New; do u32New = u32RetOld | u32; while (!ASMAtomicCmpXchgExU32(pu32, u32New, u32RetOld, &u32RetOld)); return u32RetOld; #endif } /** * Atomically Or a signed 32-bit value, ordered. * * @param pi32 Pointer to the pointer variable to OR u32 with. * @param i32 The value to OR *pu32 with. * * @remarks x86: Requires a 386 or later. */ DECLINLINE(void) ASMAtomicOrS32(int32_t volatile RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF { ASMAtomicOrU32((uint32_t volatile RT_FAR *)pi32, (uint32_t)i32); } /** * Atomically Or an unsigned 64-bit value, ordered. * * @param pu64 Pointer to the pointer variable to OR u64 with. * @param u64 The value to OR *pu64 with. * * @remarks x86: Requires a Pentium or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN DECLASM(void) ASMAtomicOrU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_PROTO; #else DECLINLINE(void) ASMAtomicOrU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN && defined(RT_ARCH_AMD64) _InterlockedOr64((__int64 volatile RT_FAR *)pu64, (__int64)u64); # elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64) __asm__ __volatile__("lock; orq %1, %q0\n\t" : "=m" (*pu64) : "r" (u64) , "m" (*pu64) : "cc"); # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) # if defined(RTASM_ARM64_USE_FEAT_LSE) # if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) uint64_t u64Spill; __asm__ __volatile__("Lstart_ASMAtomicOrU64_%=:\n\t" "ldsetal %[fBitsToSet], %[uSpill], %[pMem]\n\t" : [pMem] "+Q" (*pu64) , [uSpill] "=&r" (u64Spill) : [fBitsToSet] "r" (u64) : ); # else __asm__ __volatile__("Lstart_ASMAtomicOrU64_%=:\n\t" RTASM_ARM_DMB_SY "stset %[fBitsToSet], %[pMem]\n\t" : [pMem] "+Q" (*pu64) : [fBitsToSet] "r" (u64) : ); # endif # else RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_64(ASMAtomicOrU64, pu64, DMB_SY, "orr %[uNew], %[uNew], %[uVal]\n\t" , "orr %[uNew], %[uNew], %[uVal]\n\t" "orr %H[uNew], %H[uNew], %H[uVal]\n\t", [uVal] "r" (u64)); # endif # else for (;;) { uint64_t u64Old = ASMAtomicUoReadU64(pu64); uint64_t u64New = u64Old | u64; if (ASMAtomicCmpXchgU64(pu64, u64New, u64Old)) break; ASMNopPause(); } # endif } #endif /** * Atomically Or a signed 64-bit value, ordered. * * @param pi64 Pointer to the pointer variable to OR u64 with. * @param i64 The value to OR *pu64 with. * * @remarks x86: Requires a Pentium or later. */ DECLINLINE(void) ASMAtomicOrS64(int64_t volatile RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF { ASMAtomicOrU64((uint64_t volatile RT_FAR *)pi64, (uint64_t)i64); } /** * Atomically And an unsigned 32-bit value, ordered. * * @param pu32 Pointer to the pointer variable to AND u32 with. * @param u32 The value to AND *pu32 with. * * @remarks x86: Requires a 386 or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicAndU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_PROTO; #else DECLINLINE(void) ASMAtomicAndU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN _InterlockedAnd((long volatile RT_FAR *)pu32, u32); # elif defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("lock; andl %1, %0\n\t" : "=m" (*pu32) : "ir" (u32) , "m" (*pu32) : "cc"); # else __asm { mov eax, [u32] # ifdef RT_ARCH_AMD64 mov rdx, [pu32] lock and [rdx], eax # else mov edx, [pu32] lock and [edx], eax # endif } # endif # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) # if defined(RTASM_ARM64_USE_FEAT_LSE) # if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) uint32_t u32Spill; __asm__ __volatile__("Lstart_ASMAtomicAndU32_%=:\n\t" "ldclral %w[fBitsToClear], %w[uSpill], %[pMem]\n\t" : [pMem] "+Q" (*pu32) , [uSpill] "=&r" (u32Spill) : [fBitsToClear] "r" (~u32) : ); # else __asm__ __volatile__("Lstart_ASMAtomicAndU32_%=:\n\t" RTASM_ARM_DMB_SY "stclr %w[fBitsToClear], %[pMem]\n\t" : [pMem] "+Q" (*pu32) : [fBitsToClear] "r" (~u32) : ); # endif # else RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicAnd32, pu32, DMB_SY, "and %w[uNew], %w[uNew], %w[uVal]\n\t", "and %[uNew], %[uNew], %[uVal]\n\t", [uVal] "r" (u32)); # endif # else # error "Port me" # endif } #endif /** * Atomically AND an unsigned 32-bit value, ordered, extended version. * * @returns Old value. * @param pu32 Pointer to the variable to AND @a u32 with. * @param u32 The value to AND @a *pu32 with. */ DECLINLINE(uint32_t) ASMAtomicAndExU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF { #if defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) # if defined(RTASM_ARM64_USE_FEAT_LSE) uint32_t u32OldRet; __asm__ __volatile__("Lstart_ASMAtomicAndExU32_%=:\n\t" # if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) "ldclral %w[fBitsToClear], %w[uOldRet], %[pMem]\n\t" # else RTASM_ARM_DMB_SY "ldclr %w[fBitsToClear], %w[uOldRet], %[pMem]\n\t" # endif : [pMem] "+Q" (*pu32) , [uOldRet] "=&r" (u32OldRet) : [fBitsToClear] "r" (~u32) : ); # else RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(ASMAtomicAndEx32, pu32, DMB_SY, "and %w[uNew], %w[uOld], %w[uVal]\n\t", "and %[uNew], %[uOld], %[uVal]\n\t", [uVal] "r" (u32)); # endif return u32OldRet; #else uint32_t u32RetOld = ASMAtomicUoReadU32(pu32); uint32_t u32New; do u32New = u32RetOld & u32; while (!ASMAtomicCmpXchgExU32(pu32, u32New, u32RetOld, &u32RetOld)); return u32RetOld; #endif } /** * Atomically And a signed 32-bit value, ordered. * * @param pi32 Pointer to the pointer variable to AND i32 with. * @param i32 The value to AND *pi32 with. * * @remarks x86: Requires a 386 or later. */ DECLINLINE(void) ASMAtomicAndS32(int32_t volatile RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF { ASMAtomicAndU32((uint32_t volatile RT_FAR *)pi32, (uint32_t)i32); } /** * Atomically And an unsigned 64-bit value, ordered. * * @param pu64 Pointer to the pointer variable to AND u64 with. * @param u64 The value to AND *pu64 with. * * @remarks x86: Requires a Pentium or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN DECLASM(void) ASMAtomicAndU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_PROTO; #else DECLINLINE(void) ASMAtomicAndU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN && defined(RT_ARCH_AMD64) _InterlockedAnd64((__int64 volatile RT_FAR *)pu64, u64); # elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64) __asm__ __volatile__("lock; andq %1, %0\n\t" : "=m" (*pu64) : "r" (u64) , "m" (*pu64) : "cc"); # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) # if defined(RTASM_ARM64_USE_FEAT_LSE) # if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) uint64_t u64Spill; __asm__ __volatile__("Lstart_ASMAtomicAndU64_%=:\n\t" "ldclral %[fBitsToClear], %[uSpill], %[pMem]\n\t" : [pMem] "+Q" (*pu64) , [uSpill] "=&r" (u64Spill) : [fBitsToClear] "r" (~u64) : ); # else __asm__ __volatile__("Lstart_ASMAtomicAndU64_%=:\n\t" RTASM_ARM_DMB_SY "stclr %[fBitsToClear], %[pMem]\n\t" : [pMem] "+Q" (*pu64) : [fBitsToClear] "r" (~u64) : ); # endif # else RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_64(ASMAtomicAndU64, pu64, DMB_SY, "and %[uNew], %[uNew], %[uVal]\n\t" , "and %[uNew], %[uNew], %[uVal]\n\t" "and %H[uNew], %H[uNew], %H[uVal]\n\t", [uVal] "r" (u64)); # endif # else for (;;) { uint64_t u64Old = ASMAtomicUoReadU64(pu64); uint64_t u64New = u64Old & u64; if (ASMAtomicCmpXchgU64(pu64, u64New, u64Old)) break; ASMNopPause(); } # endif } #endif /** * Atomically And a signed 64-bit value, ordered. * * @param pi64 Pointer to the pointer variable to AND i64 with. * @param i64 The value to AND *pi64 with. * * @remarks x86: Requires a Pentium or later. */ DECLINLINE(void) ASMAtomicAndS64(int64_t volatile RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF { ASMAtomicAndU64((uint64_t volatile RT_FAR *)pi64, (uint64_t)i64); } /** * Atomically XOR an unsigned 32-bit value and a memory location, ordered. * * @param pu32 Pointer to the variable to XOR @a u32 with. * @param u32 The value to XOR @a *pu32 with. * * @remarks x86: Requires a 386 or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicXorU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_PROTO; #else DECLINLINE(void) ASMAtomicXorU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN _InterlockedXor((long volatile RT_FAR *)pu32, u32); # elif defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("lock; xorl %1, %0\n\t" : "=m" (*pu32) : "ir" (u32) , "m" (*pu32) : "cc"); # else __asm { mov eax, [u32] # ifdef RT_ARCH_AMD64 mov rdx, [pu32] lock xor [rdx], eax # else mov edx, [pu32] lock xor [edx], eax # endif } # endif # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) # if defined(RTASM_ARM64_USE_FEAT_LSE) # if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) uint32_t u32Spill; __asm__ __volatile__("Lstart_ASMAtomicXorU32_%=:\n\t" "ldeoral %w[fBitMask], %w[uSpill], %[pMem]\n\t" : [pMem] "+Q" (*pu32) , [uSpill] "=&r" (u32Spill) : [fBitMask] "r" (u32) : ); # else __asm__ __volatile__("Lstart_ASMAtomicXorU32_%=:\n\t" RTASM_ARM_DMB_SY "steor %w[fBitMask], %[pMem]\n\t" : [pMem] "+Q" (*pu32) : [fBitMask] "r" (u32) : ); # endif # else RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicXor32, pu32, DMB_SY, "eor %w[uNew], %w[uNew], %w[uVal]\n\t", "eor %[uNew], %[uNew], %[uVal]\n\t", [uVal] "r" (u32)); # endif # else # error "Port me" # endif } #endif /** * Atomically XOR an unsigned 32-bit value and a memory location, ordered, * extended version (for bitmaps). * * @returns Old value. * @param pu32 Pointer to the variable to XOR @a u32 with. * @param u32 The value to XOR @a *pu32 with. */ DECLINLINE(uint32_t) ASMAtomicXorExU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF { #if defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) # if defined(RTASM_ARM64_USE_FEAT_LSE) uint32_t u32OldRet; __asm__ __volatile__("Lstart_ASMAtomicXorExU32_%=:\n\t" # if defined(RTASM_ARM64_USE_FEAT_LSE_WITHOUT_DMB) "ldeoral %w[fBitMask], %w[uOldRet], %[pMem]\n\t" # else RTASM_ARM_DMB_SY "ldeor %w[fBitMask], %w[uOldRet], %[pMem]\n\t" # endif : [pMem] "+Q" (*pu32) , [uOldRet] "=&r" (u32OldRet) : [fBitMask] "r" (u32) : ); # else RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(ASMAtomicXorEx32, pu32, DMB_SY, "eor %w[uNew], %w[uOld], %w[uVal]\n\t", "eor %[uNew], %[uOld], %[uVal]\n\t", [uVal] "r" (u32)); # endif return u32OldRet; #else uint32_t u32RetOld = ASMAtomicUoReadU32(pu32); uint32_t u32New; do u32New = u32RetOld ^ u32; while (!ASMAtomicCmpXchgExU32(pu32, u32New, u32RetOld, &u32RetOld)); return u32RetOld; #endif } /** * Atomically XOR a signed 32-bit value, ordered. * * @param pi32 Pointer to the variable to XOR i32 with. * @param i32 The value to XOR *pi32 with. * * @remarks x86: Requires a 386 or later. */ DECLINLINE(void) ASMAtomicXorS32(int32_t volatile RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF { ASMAtomicXorU32((uint32_t volatile RT_FAR *)pi32, (uint32_t)i32); } /** * Atomically OR an unsigned 32-bit value, unordered but interrupt safe. * * @param pu32 Pointer to the pointer variable to OR u32 with. * @param u32 The value to OR *pu32 with. * * @remarks x86: Requires a 386 or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicUoOrU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_PROTO; #else DECLINLINE(void) ASMAtomicUoOrU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF { # if defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("orl %1, %0\n\t" : "=m" (*pu32) : "ir" (u32) , "m" (*pu32) : "cc"); # else __asm { mov eax, [u32] # ifdef RT_ARCH_AMD64 mov rdx, [pu32] or [rdx], eax # else mov edx, [pu32] or [edx], eax # endif } # endif # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) /* M1 benchmark: stset=1974 vs non-lse=6271 */ # if defined(RTASM_ARM64_USE_FEAT_LSE) __asm__ __volatile__("Lstart_ASMAtomicUoOrU32_%=:\n\t" "stset %w[fBitsToSet], %[pMem]\n\t" : [pMem] "+Q" (*pu32) : [fBitsToSet] "r" (u32) : ); # else RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicUoOrU32, pu32, NO_BARRIER, "orr %w[uNew], %w[uNew], %w[uVal]\n\t", "orr %[uNew], %[uNew], %[uVal]\n\t", [uVal] "r" (u32)); # endif # else # error "Port me" # endif } #endif /** * Atomically OR an unsigned 32-bit value, unordered but interrupt safe, * extended version (for bitmap fallback). * * @returns Old value. * @param pu32 Pointer to the variable to OR @a u32 with. * @param u32 The value to OR @a *pu32 with. */ DECLINLINE(uint32_t) ASMAtomicUoOrExU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF { #if defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) # if defined(RTASM_ARM64_USE_FEAT_LSE) uint32_t u32OldRet; __asm__ __volatile__("Lstart_ASMAtomicOrExU32_%=:\n\t" "ldset %w[fBitsToSet], %w[uOldRet], %[pMem]\n\t" : [pMem] "+Q" (*pu32) , [uOldRet] "=&r" (u32OldRet) : [fBitsToSet] "r" (u32) : ); # else RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(ASMAtomicUoOrExU32, pu32, NO_BARRIER, "orr %w[uNew], %w[uOld], %w[uVal]\n\t", "orr %[uNew], %[uOld], %[uVal]\n\t", [uVal] "r" (u32)); # endif return u32OldRet; #else return ASMAtomicOrExU32(pu32, u32); /* (we have no unordered cmpxchg primitive atm.) */ #endif } /** * Atomically OR a signed 32-bit value, unordered. * * @param pi32 Pointer to the pointer variable to OR u32 with. * @param i32 The value to OR *pu32 with. * * @remarks x86: Requires a 386 or later. */ DECLINLINE(void) ASMAtomicUoOrS32(int32_t volatile RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF { ASMAtomicUoOrU32((uint32_t volatile RT_FAR *)pi32, (uint32_t)i32); } /** * Atomically OR an unsigned 64-bit value, unordered. * * @param pu64 Pointer to the pointer variable to OR u64 with. * @param u64 The value to OR *pu64 with. * * @remarks x86: Requires a Pentium or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM DECLASM(void) ASMAtomicUoOrU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_PROTO; #else DECLINLINE(void) ASMAtomicUoOrU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF { # if RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64) __asm__ __volatile__("orq %1, %q0\n\t" : "=m" (*pu64) : "r" (u64) , "m" (*pu64) : "cc"); # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) # if defined(RTASM_ARM64_USE_FEAT_LSE) __asm__ __volatile__("Lstart_ASMAtomicUoOrU64_%=:\n\t" "stset %[fBitsToSet], %[pMem]\n\t" : [pMem] "+Q" (*pu64) : [fBitsToSet] "r" (u64) : ); # else RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_64(ASMAtomicUoOrU64, pu64, NO_BARRIER, "orr %[uNew], %[uNew], %[uVal]\n\t" , "orr %[uNew], %[uNew], %[uVal]\n\t" "orr %H[uNew], %H[uNew], %H[uVal]\n\t", [uVal] "r" (u64)); # endif # else for (;;) { uint64_t u64Old = ASMAtomicUoReadU64(pu64); uint64_t u64New = u64Old | u64; if (ASMAtomicCmpXchgU64(pu64, u64New, u64Old)) break; ASMNopPause(); } # endif } #endif /** * Atomically Or a signed 64-bit value, unordered. * * @param pi64 Pointer to the pointer variable to OR u64 with. * @param i64 The value to OR *pu64 with. * * @remarks x86: Requires a Pentium or later. */ DECLINLINE(void) ASMAtomicUoOrS64(int64_t volatile RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF { ASMAtomicUoOrU64((uint64_t volatile RT_FAR *)pi64, (uint64_t)i64); } /** * Atomically And an unsigned 32-bit value, unordered. * * @param pu32 Pointer to the pointer variable to AND u32 with. * @param u32 The value to AND *pu32 with. * * @remarks x86: Requires a 386 or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicUoAndU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_PROTO; #else DECLINLINE(void) ASMAtomicUoAndU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF { # if defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("andl %1, %0\n\t" : "=m" (*pu32) : "ir" (u32) , "m" (*pu32) : "cc"); # else __asm { mov eax, [u32] # ifdef RT_ARCH_AMD64 mov rdx, [pu32] and [rdx], eax # else mov edx, [pu32] and [edx], eax # endif } # endif # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) /* M1 benchmark: stclr=1884 vs non-lse=6299 (ps/call) */ # if defined(RTASM_ARM64_USE_FEAT_LSE) __asm__ __volatile__("Lstart_ASMAtomicUoAndU32_%=:\n\t" "stclr %w[fBitsToClear], %[pMem]\n\t" : [pMem] "+Q" (*pu32) : [fBitsToClear] "r" (~u32) : ); # else RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicUoAnd32, pu32, NO_BARRIER, "and %w[uNew], %w[uNew], %w[uVal]\n\t", "and %[uNew], %[uNew], %[uVal]\n\t", [uVal] "r" (u32)); # endif # else # error "Port me" # endif } #endif /** * Atomically AND an unsigned 32-bit value, unordered, extended version (for * bitmap fallback). * * @returns Old value. * @param pu32 Pointer to the pointer to AND @a u32 with. * @param u32 The value to AND @a *pu32 with. */ DECLINLINE(uint32_t) ASMAtomicUoAndExU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF { #if defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) # if defined(RTASM_ARM64_USE_FEAT_LSE) uint32_t u32OldRet; __asm__ __volatile__("Lstart_ASMAtomicAndExU32_%=:\n\t" "ldclr %w[fBitsToClear], %w[uOldRet], %[pMem]\n\t" : [pMem] "+Q" (*pu32) , [uOldRet] "=&r" (u32OldRet) : [fBitsToClear] "r" (~u32) : ); # else RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(ASMAtomicUoAndEx32, pu32, NO_BARRIER, "and %w[uNew], %w[uOld], %w[uVal]\n\t", "and %[uNew], %[uOld], %[uVal]\n\t", [uVal] "r" (u32)); # endif return u32OldRet; #else return ASMAtomicAndExU32(pu32, u32); /* (we have no unordered cmpxchg primitive atm.) */ #endif } /** * Atomically And a signed 32-bit value, unordered. * * @param pi32 Pointer to the pointer variable to AND i32 with. * @param i32 The value to AND *pi32 with. * * @remarks x86: Requires a 386 or later. */ DECLINLINE(void) ASMAtomicUoAndS32(int32_t volatile RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF { ASMAtomicUoAndU32((uint32_t volatile RT_FAR *)pi32, (uint32_t)i32); } /** * Atomically And an unsigned 64-bit value, unordered. * * @param pu64 Pointer to the pointer variable to AND u64 with. * @param u64 The value to AND *pu64 with. * * @remarks x86: Requires a Pentium or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM DECLASM(void) ASMAtomicUoAndU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_PROTO; #else DECLINLINE(void) ASMAtomicUoAndU64(uint64_t volatile RT_FAR *pu64, uint64_t u64) RT_NOTHROW_DEF { # if RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64) __asm__ __volatile__("andq %1, %0\n\t" : "=m" (*pu64) : "r" (u64) , "m" (*pu64) : "cc"); # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) # if defined(RTASM_ARM64_USE_FEAT_LSE) __asm__ __volatile__("Lstart_ASMAtomicUoAndU64_%=:\n\t" "stclr %[fBitsToClear], %[pMem]\n\t" : [pMem] "+Q" (*pu64) : [fBitsToClear] "r" (~u64) : ); # else RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_64(ASMAtomicUoAndU64, pu64, NO_BARRIER, "and %[uNew], %[uNew], %[uVal]\n\t" , "and %[uNew], %[uNew], %[uVal]\n\t" "and %H[uNew], %H[uNew], %H[uVal]\n\t", [uVal] "r" (u64)); # endif # else for (;;) { uint64_t u64Old = ASMAtomicUoReadU64(pu64); uint64_t u64New = u64Old & u64; if (ASMAtomicCmpXchgU64(pu64, u64New, u64Old)) break; ASMNopPause(); } # endif } #endif /** * Atomically And a signed 64-bit value, unordered. * * @param pi64 Pointer to the pointer variable to AND i64 with. * @param i64 The value to AND *pi64 with. * * @remarks x86: Requires a Pentium or later. */ DECLINLINE(void) ASMAtomicUoAndS64(int64_t volatile RT_FAR *pi64, int64_t i64) RT_NOTHROW_DEF { ASMAtomicUoAndU64((uint64_t volatile RT_FAR *)pi64, (uint64_t)i64); } /** * Atomically XOR an unsigned 32-bit value, unordered but interrupt safe. * * @param pu32 Pointer to the variable to XOR @a u32 with. * @param u32 The value to OR @a *pu32 with. * * @remarks x86: Requires a 386 or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicUoXorU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_PROTO; #else DECLINLINE(void) ASMAtomicUoXorU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF { # if defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("xorl %1, %0\n\t" : "=m" (*pu32) : "ir" (u32) , "m" (*pu32) : "cc"); # else __asm { mov eax, [u32] # ifdef RT_ARCH_AMD64 mov rdx, [pu32] xor [rdx], eax # else mov edx, [pu32] xor [edx], eax # endif } # endif # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) # if defined(RTASM_ARM64_USE_FEAT_LSE) __asm__ __volatile__("Lstart_ASMAtomicUoXorU32_%=:\n\t" "steor %w[fBitMask], %[pMem]\n\t" : [pMem] "+Q" (*pu32) : [fBitMask] "r" (u32) : ); # else RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicUoXorU32, pu32, NO_BARRIER, "eor %w[uNew], %w[uNew], %w[uVal]\n\t", "eor %[uNew], %[uNew], %[uVal]\n\t", [uVal] "r" (u32)); # endif # else # error "Port me" # endif } #endif /** * Atomically XOR an unsigned 32-bit value, unordered but interrupt safe, * extended version (for bitmap fallback). * * @returns Old value. * @param pu32 Pointer to the variable to XOR @a u32 with. * @param u32 The value to OR @a *pu32 with. */ DECLINLINE(uint32_t) ASMAtomicUoXorExU32(uint32_t volatile RT_FAR *pu32, uint32_t u32) RT_NOTHROW_DEF { #if defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) # if defined(RTASM_ARM64_USE_FEAT_LSE) uint32_t u32OldRet; __asm__ __volatile__("Lstart_ASMAtomicUoXorExU32_%=:\n\t" "ldeor %w[fBitMask], %w[uOldRet], %[pMem]\n\t" : [pMem] "+Q" (*pu32) , [uOldRet] "=&r" (u32OldRet) : [fBitMask] "r" (u32) : ); # else RTASM_ARM_LOAD_MODIFY_STORE_RET_OLD_32(ASMAtomicUoXorExU32, pu32, NO_BARRIER, "eor %w[uNew], %w[uOld], %w[uVal]\n\t", "eor %[uNew], %[uOld], %[uVal]\n\t", [uVal] "r" (u32)); # endif return u32OldRet; #else return ASMAtomicXorExU32(pu32, u32); /* (we have no unordered cmpxchg primitive atm.) */ #endif } /** * Atomically XOR a signed 32-bit value, unordered. * * @param pi32 Pointer to the variable to XOR @a u32 with. * @param i32 The value to XOR @a *pu32 with. * * @remarks x86: Requires a 386 or later. */ DECLINLINE(void) ASMAtomicUoXorS32(int32_t volatile RT_FAR *pi32, int32_t i32) RT_NOTHROW_DEF { ASMAtomicUoXorU32((uint32_t volatile RT_FAR *)pi32, (uint32_t)i32); } /** * Atomically increment an unsigned 32-bit value, unordered. * * @returns the new value. * @param pu32 Pointer to the variable to increment. * * @remarks x86: Requires a 486 or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMAtomicUoIncU32(uint32_t volatile RT_FAR *pu32) RT_NOTHROW_PROTO; #else DECLINLINE(uint32_t) ASMAtomicUoIncU32(uint32_t volatile RT_FAR *pu32) RT_NOTHROW_DEF { # if defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) uint32_t u32; # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("xaddl %0, %1\n\t" : "=r" (u32) , "=m" (*pu32) : "0" (1) , "m" (*pu32) : "memory" /** @todo why 'memory'? */ , "cc"); return u32 + 1; # else __asm { mov eax, 1 # ifdef RT_ARCH_AMD64 mov rdx, [pu32] xadd [rdx], eax # else mov edx, [pu32] xadd [edx], eax # endif mov u32, eax } return u32 + 1; # endif # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) /* M1 benchmark: ldadd=2031 vs non-lse=6301 (ps/call) */ # if defined(RTASM_ARM64_USE_FEAT_LSE) uint32_t u32NewRet; __asm__ __volatile__("Lstart_ASMAtomicUoIncU32_%=:\n\t" "ldadd %w[uAddend], %w[uNewRet], %[pMem]\n\t" "add %w[uNewRet], %w[uNewRet], #1\n\t" : [pMem] "+Q" (*pu32) , [uNewRet] "=&r" (u32NewRet) : [uAddend] "r" ((uint32_t)1) : ); # else RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicUoIncU32, pu32, NO_BARRIER, "add %w[uNew], %w[uNew], #1\n\t", "add %[uNew], %[uNew], #1\n\t" /* arm6 / thumb2+ */, "X" (0) /* dummy */); # endif return u32NewRet; # else # error "Port me" # endif } #endif /** * Atomically decrement an unsigned 32-bit value, unordered. * * @returns the new value. * @param pu32 Pointer to the variable to decrement. * * @remarks x86: Requires a 486 or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMAtomicUoDecU32(uint32_t volatile RT_FAR *pu32) RT_NOTHROW_PROTO; #else DECLINLINE(uint32_t) ASMAtomicUoDecU32(uint32_t volatile RT_FAR *pu32) RT_NOTHROW_DEF { # if defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) uint32_t u32; # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("lock; xaddl %0, %1\n\t" : "=r" (u32) , "=m" (*pu32) : "0" (-1) , "m" (*pu32) : "memory" , "cc"); return u32 - 1; # else __asm { mov eax, -1 # ifdef RT_ARCH_AMD64 mov rdx, [pu32] xadd [rdx], eax # else mov edx, [pu32] xadd [edx], eax # endif mov u32, eax } return u32 - 1; # endif # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) /* M1 benchmark: ldadd=2101 vs non-lse=6268 (ps/call) */ # if defined(RTASM_ARM64_USE_FEAT_LSE) uint32_t u32NewRet; __asm__ __volatile__("Lstart_ASMAtomicUoDecU32_%=:\n\t" "ldadd %w[uAddend], %w[uNewRet], %[pMem]\n\t" "sub %w[uNewRet], %w[uNewRet], #1\n\t" : [pMem] "+Q" (*pu32) , [uNewRet] "=&r" (u32NewRet) : [uAddend] "r" (~(uint32_t)0) : ); # else RTASM_ARM_LOAD_MODIFY_STORE_RET_NEW_32(ASMAtomicUoDecU32, pu32, NO_BARRIER, "sub %w[uNew], %w[uNew], #1\n\t", "sub %[uNew], %[uNew], #1\n\t" /* arm6 / thumb2+ */, "X" (0) /* dummy */); # endif return u32NewRet; # else # error "Port me" # endif } #endif /** @todo Move ASMByteSwapU16, ASMByteSwapU32 and ASMByteSwapU64 in their own * header as it's a common reason for including asm.h. */ /** * Reverse the byte order of the given 16-bit integer. * * @returns Revert * @param u16 16-bit integer value. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM(uint16_t) ASMByteSwapU16(uint16_t u16) RT_NOTHROW_PROTO; #else DECLINLINE(uint16_t) ASMByteSwapU16(uint16_t u16) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN return _byteswap_ushort(u16); # elif defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ ("rorw $8, %0" : "=r" (u16) : "0" (u16) : "cc"); # else _asm { mov ax, [u16] ror ax, 8 mov [u16], ax } # endif return u16; # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) uint32_t u32Ret; __asm__ __volatile__( # if defined(RT_ARCH_ARM64) "rev16 %w[uRet], %w[uVal]\n\t" # else "rev16 %[uRet], %[uVal]\n\t" # endif : [uRet] "=r" (u32Ret) : [uVal] "r" (u16)); return (uint16_t)u32Ret; # else # error "Port me" # endif } #endif /** * Reverse the byte order of the given 32-bit integer. * * @returns Revert * @param u32 32-bit integer value. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMByteSwapU32(uint32_t u32) RT_NOTHROW_PROTO; #else DECLINLINE(uint32_t) ASMByteSwapU32(uint32_t u32) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN return _byteswap_ulong(u32); # elif defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ ("bswapl %0" : "=r" (u32) : "0" (u32)); # else _asm { mov eax, [u32] bswap eax mov [u32], eax } # endif return u32; # elif defined(RT_ARCH_ARM64) uint64_t u64Ret; __asm__ __volatile__("rev32 %[uRet], %[uVal]\n\t" : [uRet] "=r" (u64Ret) : [uVal] "r" ((uint64_t)u32)); return (uint32_t)u64Ret; # elif defined(RT_ARCH_ARM32) __asm__ __volatile__("rev %[uRet], %[uVal]\n\t" : [uRet] "=r" (u32) : [uVal] "[uRet]" (u32)); return u32; # else # error "Port me" # endif } #endif /** * Reverse the byte order of the given 64-bit integer. * * @returns Revert * @param u64 64-bit integer value. */ DECLINLINE(uint64_t) ASMByteSwapU64(uint64_t u64) RT_NOTHROW_DEF { #if defined(RT_ARCH_AMD64) && RT_INLINE_ASM_USES_INTRIN return _byteswap_uint64(u64); # elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64) __asm__ ("bswapq %0" : "=r" (u64) : "0" (u64)); return u64; # elif defined(RT_ARCH_ARM64) __asm__ __volatile__("rev %[uRet], %[uVal]\n\t" : [uRet] "=r" (u64) : [uVal] "[uRet]" (u64)); return u64; #else return (uint64_t)ASMByteSwapU32((uint32_t)u64) << 32 | (uint64_t)ASMByteSwapU32((uint32_t)(u64 >> 32)); #endif } /** @defgroup grp_inline_bits Bitmap Operations * * @todo Move these into a separate header, with standard IPRT prefix * (RTBitmapXxx). Move the more complex (searched) stuff into C source * files. * * @{ */ /** * Sets a bit in a bitmap. * * @param pvBitmap Pointer to the bitmap (little endian). This should be * 32-bit aligned. * @param iBit The bit to set. * * @remarks The 32-bit aligning of pvBitmap is not a strict requirement. * However, doing so will yield better performance as well as avoiding * traps accessing the last bits in the bitmap. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM(void) ASMBitSet(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO; #else DECLINLINE(void) ASMBitSet(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN _bittestandset((long RT_FAR *)pvBitmap, iBit); # elif defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("btsl %1, %0" : "=m" (*(volatile long RT_FAR *)pvBitmap) : "Ir" (iBit) , "m" (*(volatile long RT_FAR *)pvBitmap) : "memory" , "cc"); # else __asm { # ifdef RT_ARCH_AMD64 mov rax, [pvBitmap] mov edx, [iBit] bts [rax], edx # else mov eax, [pvBitmap] mov edx, [iBit] bts [eax], edx # endif } # endif # else int32_t offBitmap = iBit / 32; AssertStmt(!((uintptr_t)pvBitmap & 3), offBitmap += (uintptr_t)pvBitmap & 3; iBit += ((uintptr_t)pvBitmap & 3) * 8); ASMAtomicUoOrU32(&((uint32_t volatile *)pvBitmap)[offBitmap], RT_H2LE_U32(RT_BIT_32(iBit & 31))); # endif } #endif /** * Atomically sets a bit in a bitmap, ordered. * * @param pvBitmap Pointer to the bitmap (little endian). Must be 32-bit * aligned, otherwise the memory access isn't atomic! * @param iBit The bit to set. * * @remarks x86: Requires a 386 or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicBitSet(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO; #else DECLINLINE(void) ASMAtomicBitSet(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF { AssertMsg(!((uintptr_t)pvBitmap & 3), ("address %p not 32-bit aligned", pvBitmap)); # if RT_INLINE_ASM_USES_INTRIN _interlockedbittestandset((long RT_FAR *)pvBitmap, iBit); # elif defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("lock; btsl %1, %0" : "=m" (*(volatile long *)pvBitmap) : "Ir" (iBit) , "m" (*(volatile long *)pvBitmap) : "memory" , "cc"); # else __asm { # ifdef RT_ARCH_AMD64 mov rax, [pvBitmap] mov edx, [iBit] lock bts [rax], edx # else mov eax, [pvBitmap] mov edx, [iBit] lock bts [eax], edx # endif } # endif # else ASMAtomicOrU32(&((uint32_t volatile *)pvBitmap)[iBit / 32], RT_H2LE_U32(RT_BIT_32(iBit & 31))); # endif } #endif /** * Clears a bit in a bitmap. * * @param pvBitmap Pointer to the bitmap (little endian). * @param iBit The bit to clear. * * @remarks The 32-bit aligning of pvBitmap is not a strict requirement. * However, doing so will yield better performance as well as avoiding * traps accessing the last bits in the bitmap. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM(void) ASMBitClear(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO; #else DECLINLINE(void) ASMBitClear(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN _bittestandreset((long RT_FAR *)pvBitmap, iBit); # elif defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("btrl %1, %0" : "=m" (*(volatile long RT_FAR *)pvBitmap) : "Ir" (iBit) , "m" (*(volatile long RT_FAR *)pvBitmap) : "memory" , "cc"); # else __asm { # ifdef RT_ARCH_AMD64 mov rax, [pvBitmap] mov edx, [iBit] btr [rax], edx # else mov eax, [pvBitmap] mov edx, [iBit] btr [eax], edx # endif } # endif # else int32_t offBitmap = iBit / 32; AssertStmt(!((uintptr_t)pvBitmap & 3), offBitmap += (uintptr_t)pvBitmap & 3; iBit += ((uintptr_t)pvBitmap & 3) * 8); ASMAtomicUoAndU32(&((uint32_t volatile *)pvBitmap)[offBitmap], RT_H2LE_U32(~RT_BIT_32(iBit & 31))); # endif } #endif /** * Atomically clears a bit in a bitmap, ordered. * * @param pvBitmap Pointer to the bitmap (little endian). Must be 32-bit * aligned, otherwise the memory access isn't atomic! * @param iBit The bit to toggle set. * * @remarks No memory barrier, take care on smp. * @remarks x86: Requires a 386 or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicBitClear(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO; #else DECLINLINE(void) ASMAtomicBitClear(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF { AssertMsg(!((uintptr_t)pvBitmap & 3), ("address %p not 32-bit aligned", pvBitmap)); # if defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("lock; btrl %1, %0" : "=m" (*(volatile long RT_FAR *)pvBitmap) : "Ir" (iBit) , "m" (*(volatile long RT_FAR *)pvBitmap) : "memory" , "cc"); # else __asm { # ifdef RT_ARCH_AMD64 mov rax, [pvBitmap] mov edx, [iBit] lock btr [rax], edx # else mov eax, [pvBitmap] mov edx, [iBit] lock btr [eax], edx # endif } # endif # else ASMAtomicAndU32(&((uint32_t volatile *)pvBitmap)[iBit / 32], RT_H2LE_U32(~RT_BIT_32(iBit & 31))); # endif } #endif /** * Toggles a bit in a bitmap. * * @param pvBitmap Pointer to the bitmap (little endian). * @param iBit The bit to toggle. * * @remarks The 32-bit aligning of pvBitmap is not a strict requirement. * However, doing so will yield better performance as well as avoiding * traps accessing the last bits in the bitmap. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM(void) ASMBitToggle(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO; #else DECLINLINE(void) ASMBitToggle(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN _bittestandcomplement((long RT_FAR *)pvBitmap, iBit); # elif defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("btcl %1, %0" : "=m" (*(volatile long *)pvBitmap) : "Ir" (iBit) , "m" (*(volatile long *)pvBitmap) : "memory" , "cc"); # else __asm { # ifdef RT_ARCH_AMD64 mov rax, [pvBitmap] mov edx, [iBit] btc [rax], edx # else mov eax, [pvBitmap] mov edx, [iBit] btc [eax], edx # endif } # endif # else int32_t offBitmap = iBit / 32; AssertStmt(!((uintptr_t)pvBitmap & 3), offBitmap += (uintptr_t)pvBitmap & 3; iBit += ((uintptr_t)pvBitmap & 3) * 8); ASMAtomicUoXorU32(&((uint32_t volatile *)pvBitmap)[offBitmap], RT_H2LE_U32(RT_BIT_32(iBit & 31))); # endif } #endif /** * Atomically toggles a bit in a bitmap, ordered. * * @param pvBitmap Pointer to the bitmap (little endian). Must be 32-bit * aligned, otherwise the memory access isn't atomic! * @param iBit The bit to test and set. * * @remarks x86: Requires a 386 or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM RT_ASM_DECL_PRAGMA_WATCOM(void) ASMAtomicBitToggle(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO; #else DECLINLINE(void) ASMAtomicBitToggle(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF { AssertMsg(!((uintptr_t)pvBitmap & 3), ("address %p not 32-bit aligned", pvBitmap)); # if defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("lock; btcl %1, %0" : "=m" (*(volatile long RT_FAR *)pvBitmap) : "Ir" (iBit) , "m" (*(volatile long RT_FAR *)pvBitmap) : "memory" , "cc"); # else __asm { # ifdef RT_ARCH_AMD64 mov rax, [pvBitmap] mov edx, [iBit] lock btc [rax], edx # else mov eax, [pvBitmap] mov edx, [iBit] lock btc [eax], edx # endif } # endif # else ASMAtomicXorU32(&((uint32_t volatile *)pvBitmap)[iBit / 32], RT_H2LE_U32(RT_BIT_32(iBit & 31))); # endif } #endif /** * Tests and sets a bit in a bitmap. * * @returns true if the bit was set. * @returns false if the bit was clear. * * @param pvBitmap Pointer to the bitmap (little endian). * @param iBit The bit to test and set. * * @remarks The 32-bit aligning of pvBitmap is not a strict requirement. * However, doing so will yield better performance as well as avoiding * traps accessing the last bits in the bitmap. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMBitTestAndSet(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO; #else DECLINLINE(bool) ASMBitTestAndSet(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF { union { bool f; uint32_t u32; uint8_t u8; } rc; # if RT_INLINE_ASM_USES_INTRIN rc.u8 = _bittestandset((long RT_FAR *)pvBitmap, iBit); # elif defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("btsl %2, %1\n\t" "setc %b0\n\t" "andl $1, %0\n\t" : "=q" (rc.u32) , "=m" (*(volatile long RT_FAR *)pvBitmap) : "Ir" (iBit) , "m" (*(volatile long RT_FAR *)pvBitmap) : "memory" , "cc"); # else __asm { mov edx, [iBit] # ifdef RT_ARCH_AMD64 mov rax, [pvBitmap] bts [rax], edx # else mov eax, [pvBitmap] bts [eax], edx # endif setc al and eax, 1 mov [rc.u32], eax } # endif # else int32_t offBitmap = iBit / 32; AssertStmt(!((uintptr_t)pvBitmap & 3), offBitmap += (uintptr_t)pvBitmap & 3; iBit += ((uintptr_t)pvBitmap & 3) * 8); rc.u32 = RT_LE2H_U32(ASMAtomicUoOrExU32(&((uint32_t volatile *)pvBitmap)[offBitmap], RT_H2LE_U32(RT_BIT_32(iBit & 31)))) >> (iBit & 31); rc.u32 &= 1; # endif return rc.f; } #endif /** * Atomically tests and sets a bit in a bitmap, ordered. * * @returns true if the bit was set. * @returns false if the bit was clear. * * @param pvBitmap Pointer to the bitmap (little endian). Must be 32-bit * aligned, otherwise the memory access isn't atomic! * @param iBit The bit to set. * * @remarks x86: Requires a 386 or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicBitTestAndSet(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO; #else DECLINLINE(bool) ASMAtomicBitTestAndSet(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF { union { bool f; uint32_t u32; uint8_t u8; } rc; AssertMsg(!((uintptr_t)pvBitmap & 3), ("address %p not 32-bit aligned", pvBitmap)); # if RT_INLINE_ASM_USES_INTRIN rc.u8 = _interlockedbittestandset((long RT_FAR *)pvBitmap, iBit); # elif defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("lock; btsl %2, %1\n\t" "setc %b0\n\t" "andl $1, %0\n\t" : "=q" (rc.u32) , "=m" (*(volatile long RT_FAR *)pvBitmap) : "Ir" (iBit) , "m" (*(volatile long RT_FAR *)pvBitmap) : "memory" , "cc"); # else __asm { mov edx, [iBit] # ifdef RT_ARCH_AMD64 mov rax, [pvBitmap] lock bts [rax], edx # else mov eax, [pvBitmap] lock bts [eax], edx # endif setc al and eax, 1 mov [rc.u32], eax } # endif # else rc.u32 = RT_LE2H_U32(ASMAtomicOrExU32(&((uint32_t volatile *)pvBitmap)[iBit / 32], RT_H2LE_U32(RT_BIT_32(iBit & 31)))) >> (iBit & 31); rc.u32 &= 1; # endif return rc.f; } #endif /** * Tests and clears a bit in a bitmap. * * @returns true if the bit was set. * @returns false if the bit was clear. * * @param pvBitmap Pointer to the bitmap (little endian). * @param iBit The bit to test and clear. * * @remarks The 32-bit aligning of pvBitmap is not a strict requirement. * However, doing so will yield better performance as well as avoiding * traps accessing the last bits in the bitmap. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMBitTestAndClear(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO; #else DECLINLINE(bool) ASMBitTestAndClear(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF { union { bool f; uint32_t u32; uint8_t u8; } rc; # if RT_INLINE_ASM_USES_INTRIN rc.u8 = _bittestandreset((long RT_FAR *)pvBitmap, iBit); # elif defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("btrl %2, %1\n\t" "setc %b0\n\t" "andl $1, %0\n\t" : "=q" (rc.u32) , "=m" (*(volatile long RT_FAR *)pvBitmap) : "Ir" (iBit) , "m" (*(volatile long RT_FAR *)pvBitmap) : "memory" , "cc"); # else __asm { mov edx, [iBit] # ifdef RT_ARCH_AMD64 mov rax, [pvBitmap] btr [rax], edx # else mov eax, [pvBitmap] btr [eax], edx # endif setc al and eax, 1 mov [rc.u32], eax } # endif # else int32_t offBitmap = iBit / 32; AssertStmt(!((uintptr_t)pvBitmap & 3), offBitmap += (uintptr_t)pvBitmap & 3; iBit += ((uintptr_t)pvBitmap & 3) * 8); rc.u32 = RT_LE2H_U32(ASMAtomicUoAndExU32(&((uint32_t volatile *)pvBitmap)[offBitmap], RT_H2LE_U32(~RT_BIT_32(iBit & 31)))) >> (iBit & 31); rc.u32 &= 1; # endif return rc.f; } #endif /** * Atomically tests and clears a bit in a bitmap, ordered. * * @returns true if the bit was set. * @returns false if the bit was clear. * * @param pvBitmap Pointer to the bitmap (little endian). Must be 32-bit * aligned, otherwise the memory access isn't atomic! * @param iBit The bit to test and clear. * * @remarks No memory barrier, take care on smp. * @remarks x86: Requires a 386 or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicBitTestAndClear(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO; #else DECLINLINE(bool) ASMAtomicBitTestAndClear(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF { union { bool f; uint32_t u32; uint8_t u8; } rc; AssertMsg(!((uintptr_t)pvBitmap & 3), ("address %p not 32-bit aligned", pvBitmap)); # if RT_INLINE_ASM_USES_INTRIN rc.u8 = _interlockedbittestandreset((long RT_FAR *)pvBitmap, iBit); # elif defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("lock; btrl %2, %1\n\t" "setc %b0\n\t" "andl $1, %0\n\t" : "=q" (rc.u32) , "=m" (*(volatile long RT_FAR *)pvBitmap) : "Ir" (iBit) , "m" (*(volatile long RT_FAR *)pvBitmap) : "memory" , "cc"); # else __asm { mov edx, [iBit] # ifdef RT_ARCH_AMD64 mov rax, [pvBitmap] lock btr [rax], edx # else mov eax, [pvBitmap] lock btr [eax], edx # endif setc al and eax, 1 mov [rc.u32], eax } # endif # else rc.u32 = RT_LE2H_U32(ASMAtomicAndExU32(&((uint32_t volatile *)pvBitmap)[iBit / 32], RT_H2LE_U32(~RT_BIT_32(iBit & 31)))) >> (iBit & 31); rc.u32 &= 1; # endif return rc.f; } #endif /** * Tests and toggles a bit in a bitmap. * * @returns true if the bit was set. * @returns false if the bit was clear. * * @param pvBitmap Pointer to the bitmap (little endian). * @param iBit The bit to test and toggle. * * @remarks The 32-bit aligning of pvBitmap is not a strict requirement. * However, doing so will yield better performance as well as avoiding * traps accessing the last bits in the bitmap. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMBitTestAndToggle(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO; #else DECLINLINE(bool) ASMBitTestAndToggle(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF { union { bool f; uint32_t u32; uint8_t u8; } rc; # if RT_INLINE_ASM_USES_INTRIN rc.u8 = _bittestandcomplement((long RT_FAR *)pvBitmap, iBit); # elif defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("btcl %2, %1\n\t" "setc %b0\n\t" "andl $1, %0\n\t" : "=q" (rc.u32) , "=m" (*(volatile long RT_FAR *)pvBitmap) : "Ir" (iBit) , "m" (*(volatile long RT_FAR *)pvBitmap) : "memory" , "cc"); # else __asm { mov edx, [iBit] # ifdef RT_ARCH_AMD64 mov rax, [pvBitmap] btc [rax], edx # else mov eax, [pvBitmap] btc [eax], edx # endif setc al and eax, 1 mov [rc.u32], eax } # endif # else int32_t offBitmap = iBit / 32; AssertStmt(!((uintptr_t)pvBitmap & 3), offBitmap += (uintptr_t)pvBitmap & 3; iBit += ((uintptr_t)pvBitmap & 3) * 8); rc.u32 = RT_LE2H_U32(ASMAtomicUoXorExU32(&((uint32_t volatile *)pvBitmap)[offBitmap], RT_H2LE_U32(RT_BIT_32(iBit & 31)))) >> (iBit & 31); rc.u32 &= 1; # endif return rc.f; } #endif /** * Atomically tests and toggles a bit in a bitmap, ordered. * * @returns true if the bit was set. * @returns false if the bit was clear. * * @param pvBitmap Pointer to the bitmap (little endian). Must be 32-bit * aligned, otherwise the memory access isn't atomic! * @param iBit The bit to test and toggle. * * @remarks x86: Requires a 386 or later. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMAtomicBitTestAndToggle(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO; #else DECLINLINE(bool) ASMAtomicBitTestAndToggle(volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF { union { bool f; uint32_t u32; uint8_t u8; } rc; AssertMsg(!((uintptr_t)pvBitmap & 3), ("address %p not 32-bit aligned", pvBitmap)); # if defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("lock; btcl %2, %1\n\t" "setc %b0\n\t" "andl $1, %0\n\t" : "=q" (rc.u32) , "=m" (*(volatile long RT_FAR *)pvBitmap) : "Ir" (iBit) , "m" (*(volatile long RT_FAR *)pvBitmap) : "memory" , "cc"); # else __asm { mov edx, [iBit] # ifdef RT_ARCH_AMD64 mov rax, [pvBitmap] lock btc [rax], edx # else mov eax, [pvBitmap] lock btc [eax], edx # endif setc al and eax, 1 mov [rc.u32], eax } # endif # else rc.u32 = RT_H2LE_U32(ASMAtomicXorExU32(&((uint32_t volatile *)pvBitmap)[iBit / 32], RT_LE2H_U32(RT_BIT_32(iBit & 31)))) >> (iBit & 31); rc.u32 &= 1; # endif return rc.f; } #endif /** * Tests if a bit in a bitmap is set. * * @returns true if the bit is set. * @returns false if the bit is clear. * * @param pvBitmap Pointer to the bitmap (little endian). * @param iBit The bit to test. * * @remarks The 32-bit aligning of pvBitmap is not a strict requirement. * However, doing so will yield better performance as well as avoiding * traps accessing the last bits in the bitmap. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM(bool) ASMBitTest(const volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_PROTO; #else DECLINLINE(bool) ASMBitTest(const volatile void RT_FAR *pvBitmap, int32_t iBit) RT_NOTHROW_DEF { union { bool f; uint32_t u32; uint8_t u8; } rc; # if RT_INLINE_ASM_USES_INTRIN rc.u32 = _bittest((long *)pvBitmap, iBit); # elif defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("btl %2, %1\n\t" "setc %b0\n\t" "andl $1, %0\n\t" : "=q" (rc.u32) : "m" (*(const volatile long RT_FAR *)pvBitmap) , "Ir" (iBit) : "memory" , "cc"); # else __asm { mov edx, [iBit] # ifdef RT_ARCH_AMD64 mov rax, [pvBitmap] bt [rax], edx # else mov eax, [pvBitmap] bt [eax], edx # endif setc al and eax, 1 mov [rc.u32], eax } # endif # else int32_t offBitmap = iBit / 32; AssertRelease(!((uintptr_t)pvBitmap & (sizeof(uint32_t) - 1))); rc.u32 = RT_LE2H_U32(ASMAtomicUoReadU32(&((uint32_t volatile *)pvBitmap)[offBitmap])) >> (iBit & 31); rc.u32 &= 1; # endif return rc.f; } #endif #ifdef IPRT_INCLUDED_asm_mem_h /** * Clears a bit range within a bitmap. * * @param pvBitmap Pointer to the bitmap (little endian). * @param iBitStart The First bit to clear. * @param iBitEnd The first bit not to clear. */ DECLINLINE(void) ASMBitClearRange(volatile void RT_FAR *pvBitmap, size_t iBitStart, size_t iBitEnd) RT_NOTHROW_DEF { if (iBitStart < iBitEnd) { uint32_t volatile RT_FAR *pu32 = (volatile uint32_t RT_FAR *)pvBitmap + (iBitStart >> 5); size_t iStart = iBitStart & ~(size_t)31; size_t iEnd = iBitEnd & ~(size_t)31; if (iStart == iEnd) *pu32 &= RT_H2LE_U32(((UINT32_C(1) << (iBitStart & 31)) - 1) | ~((UINT32_C(1) << (iBitEnd & 31)) - 1)); else { /* bits in first dword. */ if (iBitStart & 31) { *pu32 &= RT_H2LE_U32((UINT32_C(1) << (iBitStart & 31)) - 1); pu32++; iBitStart = iStart + 32; } /* whole dwords. */ if (iBitStart != iEnd) ASMMemZero32(pu32, (iEnd - iBitStart) >> 3); /* bits in last dword. */ if (iBitEnd & 31) { pu32 = (volatile uint32_t RT_FAR *)pvBitmap + (iBitEnd >> 5); *pu32 &= RT_H2LE_U32(~((UINT32_C(1) << (iBitEnd & 31)) - 1)); } } } } /** * Sets a bit range within a bitmap. * * @param pvBitmap Pointer to the bitmap (little endian). * @param iBitStart The First bit to set. * @param iBitEnd The first bit not to set. */ DECLINLINE(void) ASMBitSetRange(volatile void RT_FAR *pvBitmap, size_t iBitStart, size_t iBitEnd) RT_NOTHROW_DEF { if (iBitStart < iBitEnd) { uint32_t volatile RT_FAR *pu32 = (volatile uint32_t RT_FAR *)pvBitmap + (iBitStart >> 5); size_t iStart = iBitStart & ~(size_t)31; size_t iEnd = iBitEnd & ~(size_t)31; if (iStart == iEnd) *pu32 |= RT_H2LE_U32(((UINT32_C(1) << (iBitEnd - iBitStart)) - 1) << (iBitStart & 31)); else { /* bits in first dword. */ if (iBitStart & 31) { *pu32 |= RT_H2LE_U32(~((UINT32_C(1) << (iBitStart & 31)) - 1)); pu32++; iBitStart = iStart + 32; } /* whole dword. */ if (iBitStart != iEnd) ASMMemFill32(pu32, (iEnd - iBitStart) >> 3, ~UINT32_C(0)); /* bits in last dword. */ if (iBitEnd & 31) { pu32 = (volatile uint32_t RT_FAR *)pvBitmap + (iBitEnd >> 5); *pu32 |= RT_H2LE_U32((UINT32_C(1) << (iBitEnd & 31)) - 1); } } } } #endif /* IPRT_INCLUDED_asm_mem_h */ /** * Finds the first clear bit in a bitmap. * * @returns Index of the first zero bit. * @returns -1 if no clear bit was found. * @param pvBitmap Pointer to the bitmap (little endian). * @param cBits The number of bits in the bitmap. Multiple of 32. */ #if RT_INLINE_ASM_EXTERNAL || (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86)) DECLASM(int32_t) ASMBitFirstClear(const volatile void RT_FAR *pvBitmap, uint32_t cBits) RT_NOTHROW_PROTO; #else DECLINLINE(int32_t) ASMBitFirstClear(const volatile void RT_FAR *pvBitmap, uint32_t cBits) RT_NOTHROW_DEF { if (cBits) { int32_t iBit; # if RT_INLINE_ASM_GNU_STYLE RTCCUINTREG uEAX, uECX, uEDI; cBits = RT_ALIGN_32(cBits, 32); __asm__ __volatile__("repe; scasl\n\t" "je 1f\n\t" # ifdef RT_ARCH_AMD64 "lea -4(%%rdi), %%rdi\n\t" "xorl (%%rdi), %%eax\n\t" "subq %5, %%rdi\n\t" # else "lea -4(%%edi), %%edi\n\t" "xorl (%%edi), %%eax\n\t" "subl %5, %%edi\n\t" # endif "shll $3, %%edi\n\t" "bsfl %%eax, %%edx\n\t" "addl %%edi, %%edx\n\t" "1:\t\n" : "=d" (iBit) , "=&c" (uECX) , "=&D" (uEDI) , "=&a" (uEAX) : "0" (0xffffffff) , "mr" (pvBitmap) , "1" (cBits >> 5) , "2" (pvBitmap) , "3" (0xffffffff) : "cc"); # else cBits = RT_ALIGN_32(cBits, 32); __asm { # ifdef RT_ARCH_AMD64 mov rdi, [pvBitmap] mov rbx, rdi # else mov edi, [pvBitmap] mov ebx, edi # endif mov edx, 0ffffffffh mov eax, edx mov ecx, [cBits] shr ecx, 5 repe scasd je done # ifdef RT_ARCH_AMD64 lea rdi, [rdi - 4] xor eax, [rdi] sub rdi, rbx # else lea edi, [edi - 4] xor eax, [edi] sub edi, ebx # endif shl edi, 3 bsf edx, eax add edx, edi done: mov [iBit], edx } # endif return iBit; } return -1; } #endif /** * Finds the next clear bit in a bitmap. * * @returns Index of the first zero bit. * @returns -1 if no clear bit was found. * @param pvBitmap Pointer to the bitmap (little endian). * @param cBits The number of bits in the bitmap. Multiple of 32. * @param iBitPrev The bit returned from the last search. * The search will start at iBitPrev + 1. */ #if RT_INLINE_ASM_EXTERNAL || (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86)) DECLASM(int) ASMBitNextClear(const volatile void RT_FAR *pvBitmap, uint32_t cBits, uint32_t iBitPrev) RT_NOTHROW_PROTO; #else DECLINLINE(int) ASMBitNextClear(const volatile void RT_FAR *pvBitmap, uint32_t cBits, uint32_t iBitPrev) RT_NOTHROW_DEF { const volatile uint32_t RT_FAR *pau32Bitmap = (const volatile uint32_t RT_FAR *)pvBitmap; int iBit = ++iBitPrev & 31; if (iBit) { /* * Inspect the 32-bit word containing the unaligned bit. */ uint32_t u32 = ~pau32Bitmap[iBitPrev / 32] >> iBit; # if RT_INLINE_ASM_USES_INTRIN unsigned long ulBit = 0; if (_BitScanForward(&ulBit, u32)) return ulBit + iBitPrev; # else # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("bsf %1, %0\n\t" "jnz 1f\n\t" "movl $-1, %0\n\t" /** @todo use conditional move for 64-bit? */ "1:\n\t" : "=r" (iBit) : "r" (u32) : "cc"); # else __asm { mov edx, [u32] bsf eax, edx jnz done mov eax, 0ffffffffh done: mov [iBit], eax } # endif if (iBit >= 0) return iBit + (int)iBitPrev; # endif /* * Skip ahead and see if there is anything left to search. */ iBitPrev |= 31; iBitPrev++; if (cBits <= (uint32_t)iBitPrev) return -1; } /* * 32-bit aligned search, let ASMBitFirstClear do the dirty work. */ iBit = ASMBitFirstClear(&pau32Bitmap[iBitPrev / 32], cBits - iBitPrev); if (iBit >= 0) iBit += iBitPrev; return iBit; } #endif /** * Finds the first set bit in a bitmap. * * @returns Index of the first set bit. * @returns -1 if no clear bit was found. * @param pvBitmap Pointer to the bitmap (little endian). * @param cBits The number of bits in the bitmap. Multiple of 32. */ #if RT_INLINE_ASM_EXTERNAL || (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86)) DECLASM(int32_t) ASMBitFirstSet(const volatile void RT_FAR *pvBitmap, uint32_t cBits) RT_NOTHROW_PROTO; #else DECLINLINE(int32_t) ASMBitFirstSet(const volatile void RT_FAR *pvBitmap, uint32_t cBits) RT_NOTHROW_DEF { if (cBits) { int32_t iBit; # if RT_INLINE_ASM_GNU_STYLE RTCCUINTREG uEAX, uECX, uEDI; cBits = RT_ALIGN_32(cBits, 32); __asm__ __volatile__("repe; scasl\n\t" "je 1f\n\t" # ifdef RT_ARCH_AMD64 "lea -4(%%rdi), %%rdi\n\t" "movl (%%rdi), %%eax\n\t" "subq %5, %%rdi\n\t" # else "lea -4(%%edi), %%edi\n\t" "movl (%%edi), %%eax\n\t" "subl %5, %%edi\n\t" # endif "shll $3, %%edi\n\t" "bsfl %%eax, %%edx\n\t" "addl %%edi, %%edx\n\t" "1:\t\n" : "=d" (iBit) , "=&c" (uECX) , "=&D" (uEDI) , "=&a" (uEAX) : "0" (0xffffffff) , "mr" (pvBitmap) , "1" (cBits >> 5) , "2" (pvBitmap) , "3" (0) : "cc"); # else cBits = RT_ALIGN_32(cBits, 32); __asm { # ifdef RT_ARCH_AMD64 mov rdi, [pvBitmap] mov rbx, rdi # else mov edi, [pvBitmap] mov ebx, edi # endif mov edx, 0ffffffffh xor eax, eax mov ecx, [cBits] shr ecx, 5 repe scasd je done # ifdef RT_ARCH_AMD64 lea rdi, [rdi - 4] mov eax, [rdi] sub rdi, rbx # else lea edi, [edi - 4] mov eax, [edi] sub edi, ebx # endif shl edi, 3 bsf edx, eax add edx, edi done: mov [iBit], edx } # endif return iBit; } return -1; } #endif /** * Finds the next set bit in a bitmap. * * @returns Index of the next set bit. * @returns -1 if no set bit was found. * @param pvBitmap Pointer to the bitmap (little endian). * @param cBits The number of bits in the bitmap. Multiple of 32. * @param iBitPrev The bit returned from the last search. * The search will start at iBitPrev + 1. */ #if RT_INLINE_ASM_EXTERNAL || (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86)) DECLASM(int) ASMBitNextSet(const volatile void RT_FAR *pvBitmap, uint32_t cBits, uint32_t iBitPrev) RT_NOTHROW_PROTO; #else DECLINLINE(int) ASMBitNextSet(const volatile void RT_FAR *pvBitmap, uint32_t cBits, uint32_t iBitPrev) RT_NOTHROW_DEF { const volatile uint32_t RT_FAR *pau32Bitmap = (const volatile uint32_t RT_FAR *)pvBitmap; int iBit = ++iBitPrev & 31; if (iBit) { /* * Inspect the 32-bit word containing the unaligned bit. */ uint32_t u32 = pau32Bitmap[iBitPrev / 32] >> iBit; # if RT_INLINE_ASM_USES_INTRIN unsigned long ulBit = 0; if (_BitScanForward(&ulBit, u32)) return ulBit + iBitPrev; # else # if RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("bsf %1, %0\n\t" "jnz 1f\n\t" /** @todo use conditional move for 64-bit? */ "movl $-1, %0\n\t" "1:\n\t" : "=r" (iBit) : "r" (u32) : "cc"); # else __asm { mov edx, [u32] bsf eax, edx jnz done mov eax, 0ffffffffh done: mov [iBit], eax } # endif if (iBit >= 0) return iBit + (int)iBitPrev; # endif /* * Skip ahead and see if there is anything left to search. */ iBitPrev |= 31; iBitPrev++; if (cBits <= (uint32_t)iBitPrev) return -1; } /* * 32-bit aligned search, let ASMBitFirstClear do the dirty work. */ iBit = ASMBitFirstSet(&pau32Bitmap[iBitPrev / 32], cBits - iBitPrev); if (iBit >= 0) iBit += iBitPrev; return iBit; } #endif /** @} */ /** @defgroup grp_inline_bits Bitmap Operations * @{ */ /** * Finds the first bit which is set in the given 32-bit integer. * Bits are numbered from 1 (least significant) to 32. * * @returns index [1..32] of the first set bit. * @returns 0 if all bits are cleared. * @param u32 Integer to search for set bits. * @remarks Similar to ffs() in BSD. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMBitFirstSetU32(uint32_t u32) RT_NOTHROW_PROTO; #else DECLINLINE(unsigned) ASMBitFirstSetU32(uint32_t u32) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN unsigned long iBit; if (_BitScanForward(&iBit, u32)) iBit++; else iBit = 0; # elif defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE uint32_t iBit; __asm__ __volatile__("bsf %1, %0\n\t" "jnz 1f\n\t" "xorl %0, %0\n\t" "jmp 2f\n" "1:\n\t" "incl %0\n" "2:\n\t" : "=r" (iBit) : "rm" (u32) : "cc"); # else uint32_t iBit; _asm { bsf eax, [u32] jnz found xor eax, eax jmp done found: inc eax done: mov [iBit], eax } # endif # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) /* * Using the "count leading zeros (clz)" instruction here because there * is no dedicated instruction to get the first set bit. * Need to reverse the bits in the value with "rbit" first because * "clz" starts counting from the most significant bit. */ uint32_t iBit; __asm__ __volatile__( # if defined(RT_ARCH_ARM64) "rbit %w[uVal], %w[uVal]\n\t" "clz %w[iBit], %w[uVal]\n\t" # else "rbit %[uVal], %[uVal]\n\t" "clz %[iBit], %[uVal]\n\t" # endif : [uVal] "=r" (u32) , [iBit] "=r" (iBit) : "[uVal]" (u32)); if (iBit != 32) iBit++; else iBit = 0; /* No bit set. */ # else # error "Port me" # endif return iBit; } #endif /** * Finds the first bit which is set in the given 32-bit integer. * Bits are numbered from 1 (least significant) to 32. * * @returns index [1..32] of the first set bit. * @returns 0 if all bits are cleared. * @param i32 Integer to search for set bits. * @remark Similar to ffs() in BSD. */ DECLINLINE(unsigned) ASMBitFirstSetS32(int32_t i32) RT_NOTHROW_DEF { return ASMBitFirstSetU32((uint32_t)i32); } /** * Finds the first bit which is set in the given 64-bit integer. * * Bits are numbered from 1 (least significant) to 64. * * @returns index [1..64] of the first set bit. * @returns 0 if all bits are cleared. * @param u64 Integer to search for set bits. * @remarks Similar to ffs() in BSD. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMBitFirstSetU64(uint64_t u64) RT_NOTHROW_PROTO; #else DECLINLINE(unsigned) ASMBitFirstSetU64(uint64_t u64) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN unsigned long iBit; # if ARCH_BITS == 64 if (_BitScanForward64(&iBit, u64)) iBit++; else iBit = 0; # else if (_BitScanForward(&iBit, (uint32_t)u64)) iBit++; else if (_BitScanForward(&iBit, (uint32_t)(u64 >> 32))) iBit += 33; else iBit = 0; # endif # elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64) uint64_t iBit; __asm__ __volatile__("bsfq %1, %0\n\t" "jnz 1f\n\t" "xorl %k0, %k0\n\t" "jmp 2f\n" "1:\n\t" "incl %k0\n" "2:\n\t" : "=r" (iBit) : "rm" (u64) : "cc"); # elif defined(RT_ARCH_ARM64) uint64_t iBit; __asm__ __volatile__("rbit %[uVal], %[uVal]\n\t" "clz %[iBit], %[uVal]\n\t" : [uVal] "=r" (u64) , [iBit] "=r" (iBit) : "[uVal]" (u64)); if (iBit != 64) iBit++; else iBit = 0; /* No bit set. */ # else unsigned iBit = ASMBitFirstSetU32((uint32_t)u64); if (!iBit) { iBit = ASMBitFirstSetU32((uint32_t)(u64 >> 32)); if (iBit) iBit += 32; } # endif return (unsigned)iBit; } #endif /** * Finds the first bit which is set in the given 16-bit integer. * * Bits are numbered from 1 (least significant) to 16. * * @returns index [1..16] of the first set bit. * @returns 0 if all bits are cleared. * @param u16 Integer to search for set bits. * @remarks For 16-bit bs3kit code. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMBitFirstSetU16(uint16_t u16) RT_NOTHROW_PROTO; #else DECLINLINE(unsigned) ASMBitFirstSetU16(uint16_t u16) RT_NOTHROW_DEF { return ASMBitFirstSetU32((uint32_t)u16); } #endif /** * Finds the last bit which is set in the given 32-bit integer. * Bits are numbered from 1 (least significant) to 32. * * @returns index [1..32] of the last set bit. * @returns 0 if all bits are cleared. * @param u32 Integer to search for set bits. * @remark Similar to fls() in BSD. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMBitLastSetU32(uint32_t u32) RT_NOTHROW_PROTO; #else DECLINLINE(unsigned) ASMBitLastSetU32(uint32_t u32) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN unsigned long iBit; if (_BitScanReverse(&iBit, u32)) iBit++; else iBit = 0; # elif defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) # if RT_INLINE_ASM_GNU_STYLE uint32_t iBit; __asm__ __volatile__("bsrl %1, %0\n\t" "jnz 1f\n\t" "xorl %0, %0\n\t" "jmp 2f\n" "1:\n\t" "incl %0\n" "2:\n\t" : "=r" (iBit) : "rm" (u32) : "cc"); # else uint32_t iBit; _asm { bsr eax, [u32] jnz found xor eax, eax jmp done found: inc eax done: mov [iBit], eax } # endif # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) uint32_t iBit; __asm__ __volatile__( # if defined(RT_ARCH_ARM64) "clz %w[iBit], %w[uVal]\n\t" # else "clz %[iBit], %[uVal]\n\t" # endif : [iBit] "=r" (iBit) : [uVal] "r" (u32)); iBit = 32 - iBit; # else # error "Port me" # endif return iBit; } #endif /** * Finds the last bit which is set in the given 32-bit integer. * Bits are numbered from 1 (least significant) to 32. * * @returns index [1..32] of the last set bit. * @returns 0 if all bits are cleared. * @param i32 Integer to search for set bits. * @remark Similar to fls() in BSD. */ DECLINLINE(unsigned) ASMBitLastSetS32(int32_t i32) RT_NOTHROW_DEF { return ASMBitLastSetU32((uint32_t)i32); } /** * Finds the last bit which is set in the given 64-bit integer. * * Bits are numbered from 1 (least significant) to 64. * * @returns index [1..64] of the last set bit. * @returns 0 if all bits are cleared. * @param u64 Integer to search for set bits. * @remark Similar to fls() in BSD. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMBitLastSetU64(uint64_t u64) RT_NOTHROW_PROTO; #else DECLINLINE(unsigned) ASMBitLastSetU64(uint64_t u64) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN unsigned long iBit; # if ARCH_BITS == 64 if (_BitScanReverse64(&iBit, u64)) iBit++; else iBit = 0; # else if (_BitScanReverse(&iBit, (uint32_t)(u64 >> 32))) iBit += 33; else if (_BitScanReverse(&iBit, (uint32_t)u64)) iBit++; else iBit = 0; # endif # elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64) uint64_t iBit; __asm__ __volatile__("bsrq %1, %0\n\t" "jnz 1f\n\t" "xorl %k0, %k0\n\t" "jmp 2f\n" "1:\n\t" "incl %k0\n" "2:\n\t" : "=r" (iBit) : "rm" (u64) : "cc"); # elif defined(RT_ARCH_ARM64) uint64_t iBit; __asm__ __volatile__("clz %[iBit], %[uVal]\n\t" : [iBit] "=r" (iBit) : [uVal] "r" (u64)); iBit = 64 - iBit; # else unsigned iBit = ASMBitLastSetU32((uint32_t)(u64 >> 32)); if (iBit) iBit += 32; else iBit = ASMBitLastSetU32((uint32_t)u64); # endif return (unsigned)iBit; } #endif /** * Finds the last bit which is set in the given 16-bit integer. * * Bits are numbered from 1 (least significant) to 16. * * @returns index [1..16] of the last set bit. * @returns 0 if all bits are cleared. * @param u16 Integer to search for set bits. * @remarks For 16-bit bs3kit code. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMBitLastSetU16(uint16_t u16) RT_NOTHROW_PROTO; #else DECLINLINE(unsigned) ASMBitLastSetU16(uint16_t u16) RT_NOTHROW_DEF { return ASMBitLastSetU32((uint32_t)u16); } #endif /** * Count the number of leading zero bits in the given 32-bit integer. * * The counting starts with the most significate bit. * * @returns Number of most significant zero bits. * @returns 32 if all bits are cleared. * @param u32 Integer to consider. * @remarks Similar to __builtin_clz() in gcc, except defined zero input result. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMCountLeadingZerosU32(uint32_t u32) RT_NOTHROW_PROTO; #else DECLINLINE(unsigned) ASMCountLeadingZerosU32(uint32_t u32) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN unsigned long iBit; if (!_BitScanReverse(&iBit, u32)) return 32; return 31 - (unsigned)iBit; # elif defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) uint32_t iBit; # if RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64) && 0 /* significantly slower on 10980xe; 929 vs 237 ps/call */ __asm__ __volatile__("bsrl %1, %0\n\t" "cmovzl %2, %0\n\t" : "=&r" (iBit) : "rm" (u32) , "rm" ((int32_t)-1) : "cc"); # elif RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("bsr %1, %0\n\t" "jnz 1f\n\t" "mov $-1, %0\n\t" "1:\n\t" : "=r" (iBit) : "rm" (u32) : "cc"); # else _asm { bsr eax, [u32] jnz found mov eax, -1 found: mov [iBit], eax } # endif return 31 - iBit; # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) uint32_t iBit; __asm__ __volatile__( # if defined(RT_ARCH_ARM64) "clz %w[iBit], %w[uVal]\n\t" # else "clz %[iBit], %[uVal]\n\t" # endif : [uVal] "=r" (u32) , [iBit] "=r" (iBit) : "[uVal]" (u32)); return iBit; # elif defined(__GNUC__) AssertCompile(sizeof(u32) == sizeof(unsigned int)); return u32 ? __builtin_clz(u32) : 32; # else # error "Port me" # endif } #endif /** * Count the number of leading zero bits in the given 64-bit integer. * * The counting starts with the most significate bit. * * @returns Number of most significant zero bits. * @returns 64 if all bits are cleared. * @param u64 Integer to consider. * @remarks Similar to __builtin_clzl() in gcc, except defined zero input * result. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMCountLeadingZerosU64(uint64_t u64) RT_NOTHROW_PROTO; #else DECLINLINE(unsigned) ASMCountLeadingZerosU64(uint64_t u64) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN unsigned long iBit; # if ARCH_BITS == 64 if (_BitScanReverse64(&iBit, u64)) return 63 - (unsigned)iBit; # else if (_BitScanReverse(&iBit, (uint32_t)(u64 >> 32))) return 31 - (unsigned)iBit; if (_BitScanReverse(&iBit, (uint32_t)u64)) return 63 - (unsigned)iBit; # endif return 64; # elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64) uint64_t iBit; # if 0 /* 10980xe benchmark: 932 ps/call - the slower variant */ __asm__ __volatile__("bsrq %1, %0\n\t" "cmovzq %2, %0\n\t" : "=&r" (iBit) : "rm" (u64) , "rm" ((int64_t)-1) : "cc"); # else /* 10980xe benchmark: 262 ps/call */ __asm__ __volatile__("bsrq %1, %0\n\t" "jnz 1f\n\t" "mov $-1, %0\n\t" "1:\n\t" : "=&r" (iBit) : "rm" (u64) : "cc"); # endif return 63 - (unsigned)iBit; # elif defined(RT_ARCH_ARM64) uint64_t iBit; __asm__ __volatile__("clz %[iBit], %[uVal]\n\t" : [uVal] "=r" (u64) , [iBit] "=r" (iBit) : "[uVal]" (u64)); return (unsigned)iBit; # elif defined(__GNUC__) && ARCH_BITS == 64 AssertCompile(sizeof(u64) == sizeof(unsigned long)); return u64 ? __builtin_clzl(u64) : 64; # else unsigned iBit = ASMCountLeadingZerosU32((uint32_t)(u64 >> 32)); if (iBit == 32) iBit = ASMCountLeadingZerosU32((uint32_t)u64) + 32; return iBit; # endif } #endif /** * Count the number of leading zero bits in the given 16-bit integer. * * The counting starts with the most significate bit. * * @returns Number of most significant zero bits. * @returns 16 if all bits are cleared. * @param u16 Integer to consider. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMCountLeadingZerosU16(uint16_t u16) RT_NOTHROW_PROTO; #else DECLINLINE(unsigned) ASMCountLeadingZerosU16(uint16_t u16) RT_NOTHROW_DEF { # if RT_INLINE_ASM_GNU_STYLE && (defined(RT_ARCH_X86) || defined(RT_ARCH_AMD64)) && 0 /* slower (10980xe: 987 vs 292 ps/call) */ uint16_t iBit; __asm__ __volatile__("bsrw %1, %0\n\t" "jnz 1f\n\t" "mov $-1, %0\n\t" "1:\n\t" : "=r" (iBit) : "rm" (u16) : "cc"); return 15 - (int16_t)iBit; # else return ASMCountLeadingZerosU32((uint32_t)u16) - 16; # endif } #endif /** * Count the number of trailing zero bits in the given 32-bit integer. * * The counting starts with the least significate bit, i.e. the zero bit. * * @returns Number of least significant zero bits. * @returns 32 if all bits are cleared. * @param u32 Integer to consider. * @remarks Similar to __builtin_ctz() in gcc, except defined zero input result. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMCountTrailingZerosU32(uint32_t u32) RT_NOTHROW_PROTO; #else DECLINLINE(unsigned) ASMCountTrailingZerosU32(uint32_t u32) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN unsigned long iBit; if (!_BitScanForward(&iBit, u32)) return 32; return (unsigned)iBit; # elif defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86) uint32_t iBit; # if RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64) && 0 /* significantly slower on 10980xe; 932 vs 240 ps/call */ __asm__ __volatile__("bsfl %1, %0\n\t" "cmovzl %2, %0\n\t" : "=&r" (iBit) : "rm" (u32) , "rm" ((int32_t)32) : "cc"); # elif RT_INLINE_ASM_GNU_STYLE __asm__ __volatile__("bsfl %1, %0\n\t" "jnz 1f\n\t" "mov $32, %0\n\t" "1:\n\t" : "=r" (iBit) : "rm" (u32) : "cc"); # else _asm { bsf eax, [u32] jnz found mov eax, 32 found: mov [iBit], eax } # endif return iBit; # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) /* Invert the bits and use clz. */ uint32_t iBit; __asm__ __volatile__( # if defined(RT_ARCH_ARM64) "rbit %w[uVal], %w[uVal]\n\t" "clz %w[iBit], %w[uVal]\n\t" # else "rbit %[uVal], %[uVal]\n\t" "clz %[iBit], %[uVal]\n\t" # endif : [uVal] "=r" (u32) , [iBit] "=r" (iBit) : "[uVal]" (u32)); return iBit; # elif defined(__GNUC__) AssertCompile(sizeof(u32) == sizeof(unsigned int)); return u32 ? __builtin_ctz(u32) : 32; # else # error "Port me" # endif } #endif /** * Count the number of trailing zero bits in the given 64-bit integer. * * The counting starts with the least significate bit. * * @returns Number of least significant zero bits. * @returns 64 if all bits are cleared. * @param u64 Integer to consider. * @remarks Similar to __builtin_ctzl() in gcc, except defined zero input * result. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMCountTrailingZerosU64(uint64_t u64) RT_NOTHROW_PROTO; #else DECLINLINE(unsigned) ASMCountTrailingZerosU64(uint64_t u64) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN unsigned long iBit; # if ARCH_BITS == 64 if (_BitScanForward64(&iBit, u64)) return (unsigned)iBit; # else if (_BitScanForward(&iBit, (uint32_t)u64)) return (unsigned)iBit; if (_BitScanForward(&iBit, (uint32_t)(u64 >> 32))) return (unsigned)iBit + 32; # endif return 64; # elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64) uint64_t iBit; # if 0 /* 10980xe benchmark: 932 ps/call - the slower variant */ __asm__ __volatile__("bsfq %1, %0\n\t" "cmovzq %2, %0\n\t" : "=&r" (iBit) : "rm" (u64) , "rm" ((int64_t)64) : "cc"); # else /* 10980xe benchmark: 262 ps/call */ __asm__ __volatile__("bsfq %1, %0\n\t" "jnz 1f\n\t" "mov $64, %0\n\t" "1:\n\t" : "=&r" (iBit) : "rm" (u64) : "cc"); # endif return (unsigned)iBit; # elif defined(RT_ARCH_ARM64) /* Invert the bits and use clz. */ uint64_t iBit; __asm__ __volatile__("rbit %[uVal], %[uVal]\n\t" "clz %[iBit], %[uVal]\n\t" : [uVal] "=r" (u64) , [iBit] "=r" (iBit) : "[uVal]" (u64)); return (unsigned)iBit; # elif defined(__GNUC__) && ARCH_BITS == 64 AssertCompile(sizeof(u64) == sizeof(unsigned long)); return u64 ? __builtin_ctzl(u64) : 64; # else unsigned iBit = ASMCountTrailingZerosU32((uint32_t)u64); if (iBit == 32) iBit = ASMCountTrailingZerosU32((uint32_t)(u64 >> 32)) + 32; return iBit; # endif } #endif /** * Count the number of trailing zero bits in the given 16-bit integer. * * The counting starts with the most significate bit. * * @returns Number of most significant zero bits. * @returns 16 if all bits are cleared. * @param u16 Integer to consider. */ #if RT_INLINE_ASM_EXTERNAL_TMP_ARM && !RT_INLINE_ASM_USES_INTRIN RT_ASM_DECL_PRAGMA_WATCOM_386(unsigned) ASMCountTrailingZerosU16(uint16_t u16) RT_NOTHROW_PROTO; #else DECLINLINE(unsigned) ASMCountTrailingZerosU16(uint16_t u16) RT_NOTHROW_DEF { # if RT_INLINE_ASM_GNU_STYLE && (defined(RT_ARCH_X86) || defined(RT_ARCH_AMD64)) && 0 /* slower (10980xe: 992 vs 349 ps/call) */ uint16_t iBit; __asm__ __volatile__("bsfw %1, %0\n\t" "jnz 1f\n\t" "mov $16, %0\n\t" "1:\n\t" : "=r" (iBit) : "rm" (u16) : "cc"); return iBit; # else return ASMCountTrailingZerosU32((uint32_t)u16 | UINT32_C(0x10000)); #endif } #endif /** * Rotate 32-bit unsigned value to the left by @a cShift. * * @returns Rotated value. * @param u32 The value to rotate. * @param cShift How many bits to rotate by. */ #ifdef __WATCOMC__ RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMRotateLeftU32(uint32_t u32, unsigned cShift) RT_NOTHROW_PROTO; #else DECLINLINE(uint32_t) ASMRotateLeftU32(uint32_t u32, uint32_t cShift) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN return _rotl(u32, cShift); # elif RT_INLINE_ASM_GNU_STYLE && (defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86)) __asm__ __volatile__("roll %b1, %0" : "=g" (u32) : "Ic" (cShift), "0" (u32) : "cc"); return u32; # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) __asm__ __volatile__( # if defined(RT_ARCH_ARM64) "ror %w[uRet], %w[uVal], %w[cShift]\n\t" # else "ror %[uRet], %[uVal], %[cShift]\n\t" # endif : [uRet] "=r" (u32) : [uVal] "[uRet]" (u32) , [cShift] "r" (32 - (cShift & 31))); /** @todo there is an immediate form here */ return u32; # else cShift &= 31; return (u32 << cShift) | (u32 >> (32 - cShift)); # endif } #endif /** * Rotate 32-bit unsigned value to the right by @a cShift. * * @returns Rotated value. * @param u32 The value to rotate. * @param cShift How many bits to rotate by. */ #ifdef __WATCOMC__ RT_ASM_DECL_PRAGMA_WATCOM(uint32_t) ASMRotateRightU32(uint32_t u32, unsigned cShift) RT_NOTHROW_PROTO; #else DECLINLINE(uint32_t) ASMRotateRightU32(uint32_t u32, uint32_t cShift) RT_NOTHROW_DEF { # if RT_INLINE_ASM_USES_INTRIN return _rotr(u32, cShift); # elif RT_INLINE_ASM_GNU_STYLE && (defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86)) __asm__ __volatile__("rorl %b1, %0" : "=g" (u32) : "Ic" (cShift), "0" (u32) : "cc"); return u32; # elif defined(RT_ARCH_ARM64) || defined(RT_ARCH_ARM32) __asm__ __volatile__( # if defined(RT_ARCH_ARM64) "ror %w[uRet], %w[uVal], %w[cShift]\n\t" # else "ror %[uRet], %[uVal], %[cShift]\n\t" # endif : [uRet] "=r" (u32) : [uVal] "[uRet]" (u32) , [cShift] "r" (cShift & 31)); /** @todo there is an immediate form here */ return u32; # else cShift &= 31; return (u32 >> cShift) | (u32 << (32 - cShift)); # endif } #endif /** * Rotate 64-bit unsigned value to the left by @a cShift. * * @returns Rotated value. * @param u64 The value to rotate. * @param cShift How many bits to rotate by. */ DECLINLINE(uint64_t) ASMRotateLeftU64(uint64_t u64, uint32_t cShift) RT_NOTHROW_DEF { #if RT_INLINE_ASM_USES_INTRIN return _rotl64(u64, cShift); #elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64) __asm__ __volatile__("rolq %b1, %0" : "=g" (u64) : "Jc" (cShift), "0" (u64) : "cc"); return u64; #elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_X86) uint32_t uSpill; __asm__ __volatile__("testb $0x20, %%cl\n\t" /* if (cShift >= 0x20) { swap(u64.hi, u64lo); cShift -= 0x20; } */ "jz 1f\n\t" "xchgl %%eax, %%edx\n\t" "1:\n\t" "andb $0x1f, %%cl\n\t" /* if (cShift & 0x1f) { */ "jz 2f\n\t" "movl %%edx, %2\n\t" /* save the hi value in %3. */ "shldl %%cl,%%eax,%%edx\n\t" /* shift the hi value left, feeding MSBits from the low value. */ "shldl %%cl,%2,%%eax\n\t" /* shift the lo value left, feeding MSBits from the saved hi value. */ "2:\n\t" /* } */ : "=A" (u64) , "=c" (cShift) , "=r" (uSpill) : "0" (u64) , "1" (cShift) : "cc"); return u64; # elif defined(RT_ARCH_ARM64) __asm__ __volatile__("ror %[uRet], %[uVal], %[cShift]\n\t" : [uRet] "=r" (u64) : [uVal] "[uRet]" (u64) , [cShift] "r" ((uint64_t)(64 - (cShift & 63)))); /** @todo there is an immediate form here */ return u64; #else cShift &= 63; return (u64 << cShift) | (u64 >> (64 - cShift)); #endif } /** * Rotate 64-bit unsigned value to the right by @a cShift. * * @returns Rotated value. * @param u64 The value to rotate. * @param cShift How many bits to rotate by. */ DECLINLINE(uint64_t) ASMRotateRightU64(uint64_t u64, uint32_t cShift) RT_NOTHROW_DEF { #if RT_INLINE_ASM_USES_INTRIN return _rotr64(u64, cShift); #elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_AMD64) __asm__ __volatile__("rorq %b1, %0" : "=g" (u64) : "Jc" (cShift), "0" (u64) : "cc"); return u64; #elif RT_INLINE_ASM_GNU_STYLE && defined(RT_ARCH_X86) uint32_t uSpill; __asm__ __volatile__("testb $0x20, %%cl\n\t" /* if (cShift >= 0x20) { swap(u64.hi, u64lo); cShift -= 0x20; } */ "jz 1f\n\t" "xchgl %%eax, %%edx\n\t" "1:\n\t" "andb $0x1f, %%cl\n\t" /* if (cShift & 0x1f) { */ "jz 2f\n\t" "movl %%edx, %2\n\t" /* save the hi value in %3. */ "shrdl %%cl,%%eax,%%edx\n\t" /* shift the hi value right, feeding LSBits from the low value. */ "shrdl %%cl,%2,%%eax\n\t" /* shift the lo value right, feeding LSBits from the saved hi value. */ "2:\n\t" /* } */ : "=A" (u64) , "=c" (cShift) , "=r" (uSpill) : "0" (u64) , "1" (cShift) : "cc"); return u64; # elif defined(RT_ARCH_ARM64) __asm__ __volatile__("ror %[uRet], %[uVal], %[cShift]\n\t" : [uRet] "=r" (u64) : [uVal] "[uRet]" (u64) , [cShift] "r" ((uint64_t)(cShift & 63))); /** @todo there is an immediate form here */ return u64; #else cShift &= 63; return (u64 >> cShift) | (u64 << (64 - cShift)); #endif } /** @} */ /** @} */ /* * Include #pragma aux definitions for Watcom C/C++. */ #if defined(__WATCOMC__) && ARCH_BITS == 16 && defined(RT_ARCH_X86) # define IPRT_ASM_WATCOM_X86_16_WITH_PRAGMAS # undef IPRT_INCLUDED_asm_watcom_x86_16_h # include "asm-watcom-x86-16.h" #elif defined(__WATCOMC__) && ARCH_BITS == 32 && defined(RT_ARCH_X86) # define IPRT_ASM_WATCOM_X86_32_WITH_PRAGMAS # undef IPRT_INCLUDED_asm_watcom_x86_32_h # include "asm-watcom-x86-32.h" #endif #endif /* !IPRT_INCLUDED_asm_h */