Merge f0f6d1666c into fee51b20fa

ARM64 JIT: don't use `x18` register

CMakeLists.txt

@@ -32,6 +32,7 @@ option(WITH_VAES            "Enable VAES instructions for Cryptonight" ON)
 option(WITH_BENCHMARK       "Enable builtin RandomX benchmark and stress test" ON)
 option(WITH_SECURE_JIT      "Enable secure access to JIT memory" OFF)
 option(WITH_DMI             "Enable DMI/SMBIOS reader" ON)
+option(WITH_BUNDLED_FMT     "Force use of bundled fmt library" ON)
 
 option(BUILD_STATIC         "Build static binary" OFF)
 option(ARM_V8               "Force ARMv8 (64 bit) architecture, use with caution if automatic detection fails, but you sure it may work" OFF)
@@ -198,6 +199,7 @@ add_definitions(-D__STDC_FORMAT_MACROS -DUNICODE -D_FILE_OFFSET_BITS=64)
 find_package(UV REQUIRED)
 
 include(cmake/flags.cmake)
+include(cmake/fmt.cmake)
 include(cmake/randomx.cmake)
 include(cmake/argon2.cmake)
 include(cmake/kawpow.cmake)
@@ -229,7 +231,7 @@ include(src/hw/api/api.cmake)
 include(src/hw/dmi/dmi.cmake)
 
 include_directories(src)
-include_directories(src/3rdparty)
+include_directories(src/3rdparty/CL)
 include_directories(${UV_INCLUDE_DIR})
 
 if (WITH_DEBUG_LOG)
@@ -237,7 +239,7 @@ if (WITH_DEBUG_LOG)
 endif()
 
 add_executable(${CMAKE_PROJECT_NAME} ${HEADERS} ${SOURCES} ${SOURCES_OS} ${HEADERS_CRYPTO} ${SOURCES_CRYPTO} ${SOURCES_SYSLOG} ${TLS_SOURCES} ${XMRIG_ASM_SOURCES})
-target_link_libraries(${CMAKE_PROJECT_NAME} ${XMRIG_ASM_LIBRARY} ${OPENSSL_LIBRARIES} ${UV_LIBRARIES} ${EXTRA_LIBS} ${CPUID_LIB} ${ARGON2_LIBRARY} ${ETHASH_LIBRARY} ${GHOSTRIDER_LIBRARY})
+target_link_libraries(${CMAKE_PROJECT_NAME} ${XMRIG_ASM_LIBRARY} ${OPENSSL_LIBRARIES} ${UV_LIBRARIES} ${EXTRA_LIBS} ${CPUID_LIB} ${ARGON2_LIBRARY} ${ETHASH_LIBRARY} ${GHOSTRIDER_LIBRARY} ${FMT_LIBRARY})
 
 if (WIN32)
     add_custom_command(TARGET ${CMAKE_PROJECT_NAME} POST_BUILD COMMAND ${CMAKE_COMMAND} -E copy_if_different "${CMAKE_SOURCE_DIR}/bin/WinRing0/WinRing0x64.sys" $<TARGET_FILE_DIR:${CMAKE_PROJECT_NAME}>)

cmake/fmt.cmake

@@ -0,0 +1,13 @@
+# SPDX-FileCopyrightText: © 2023 Jean-Pierre De Jesus DIAZ <me@jeandudey.tech>
+# SPDX-License-Identifier: GPL-3.0-or-later
+
+
+if(WITH_BUNDLED_FMT)
+    add_library(fmt INTERFACE)
+    target_sources(fmt INTERFACE ${CMAKE_CURRENT_SOURCE_DIR}/src/3rdparty/fmt/format.cc)
+    target_include_directories(fmt INTERFACE ${CMAKE_CURRENT_SOURCE_DIR}/src/3rdparty/fmt)
+else()
+    set(FMT_LIBRARY fmt)
+    find_package(fmt REQUIRED)
+    set(FMT_LIBRARY fmt::fmt)
+endif()

src/3rdparty/cl.h

@@ -29,7 +29,7 @@
 #if defined(__APPLE__)
 #   include <OpenCL/cl.h>
 #else
-#   include "3rdparty/CL/cl.h"
+#   include <CL/cl.h>
 #endif
 
 

src/3rdparty/fmt/fmt/format.cc

@@ -5,7 +5,7 @@
 //
 // For the license information refer to format.h.
 
-#include "3rdparty/fmt/format-inl.h"
+#include "fmt/format-inl.h"
 
 FMT_BEGIN_NAMESPACE
 namespace detail {

src/backend/cpu/platform/lscpu_arm.cpp

@@ -18,11 +18,11 @@
  */
 
 #include "base/tools/String.h"
-#include "3rdparty/fmt/core.h"
 
 
 #include <cstdio>
 #include <cctype>
+#include <fmt/core.h>
 
 
 namespace xmrig {

src/backend/opencl/wrappers/AdlLib_linux.cpp

@@ -20,7 +20,6 @@
 
 
 #include "backend/opencl/wrappers/AdlLib.h"
-#include "3rdparty/fmt/core.h"
 #include "backend/opencl/wrappers/OclDevice.h"
 
 
@@ -31,6 +30,7 @@
 #include <string>
 #include <sys/stat.h>
 #include <sys/types.h>
+#include <fmt/core.h>
 
 
 namespace xmrig {

src/base/base.cmake

@@ -88,7 +88,6 @@ set(HEADERS_BASE
    )
 
 set(SOURCES_BASE
-    src/3rdparty/fmt/format.cc
     src/base/crypto/Algorithm.cpp
     src/base/crypto/Coin.cpp
     src/base/crypto/keccak.cpp

src/base/kernel/Process.cpp

@@ -22,10 +22,10 @@
 
 
 #include "base/kernel/Process.h"
-#include "3rdparty/fmt/core.h"
 #include "base/tools/Chrono.h"
 #include "version.h"
 
+#include <fmt/core.h>
 
 #ifdef XMRIG_OS_WIN
 #   ifdef _MSC_VER

src/base/net/stratum/BaseClient.cpp

@@ -17,7 +17,6 @@
  */
 
 #include "base/net/stratum/BaseClient.h"
-#include "3rdparty/fmt/core.h"
 #include "3rdparty/rapidjson/document.h"
 #include "base/io/Env.h"
 #include "base/io/log/Log.h"
@@ -25,6 +24,8 @@
 #include "base/kernel/interfaces/IClientListener.h"
 #include "base/net/stratum/SubmitResult.h"
 
+#include <fmt/core.h>
+
 
 namespace xmrig {
 

src/base/net/stratum/benchmark/BenchClient.cpp

@@ -17,7 +17,6 @@
  */
 
 #include "base/net/stratum/benchmark/BenchClient.h"
-#include "3rdparty/fmt/core.h"
 #include "3rdparty/rapidjson/document.h"
 #include "backend/common/benchmark/BenchState.h"
 #include "backend/common/interfaces/IBackend.h"
@@ -39,6 +38,8 @@
 #   include "hw/dmi/DmiReader.h"
 #endif
 
+#include <fmt/core.h>
+
 
 xmrig::BenchClient::BenchClient(const std::shared_ptr<BenchConfig> &benchmark, IClientListener* listener) :
     m_listener(listener),

src/base/net/stratum/benchmark/BenchConfig.cpp

@@ -17,12 +17,12 @@
  */
 
 #include "base/net/stratum/benchmark/BenchConfig.h"
-#include "3rdparty/fmt/core.h"
 #include "3rdparty/rapidjson/document.h"
 #include "base/io/json/Json.h"
 
 
 #include <string>
+#include <fmt/core.h>
 
 
 #ifdef _MSC_VER

src/crypto/common/LinuxMemory.cpp

@@ -17,12 +17,12 @@
  */
 
 #include "crypto/common/LinuxMemory.h"
-#include "3rdparty/fmt/core.h"
 #include "crypto/common/VirtualMemory.h"
 
 
 #include <algorithm>
 #include <fstream>
+#include <fmt/core.h>
 #include <mutex>
 #include <string>
 

src/crypto/randomx/jit_compiler_a64.cpp

@@ -131,8 +131,8 @@ void JitCompilerA64::generateProgram(Program& program, ProgramConfiguration& con
 	// and w16, w10, ScratchpadL3Mask64
 	emit32(0x121A0000 | 16 | (10 << 5) | ((RandomX_CurrentConfig.Log2_ScratchpadL3 - 7) << 10), code, codePos);
 
-	// and w17, w18, ScratchpadL3Mask64
+	// and w17, w20, ScratchpadL3Mask64
-	emit32(0x121A0000 | 17 | (18 << 5) | ((RandomX_CurrentConfig.Log2_ScratchpadL3 - 7) << 10), code, codePos);
+	emit32(0x121A0000 | 17 | (20 << 5) | ((RandomX_CurrentConfig.Log2_ScratchpadL3 - 7) << 10), code, codePos);
 
 	codePos = PrologueSize;
 	literalPos = ImulRcpLiteralsEnd;
@@ -148,16 +148,16 @@ void JitCompilerA64::generateProgram(Program& program, ProgramConfiguration& con
 	}
 
 	// Update spMix2
-	// eor w18, config.readReg2, config.readReg3
+	// eor w20, config.readReg2, config.readReg3
-	emit32(ARMV8A::EOR32 | 18 | (IntRegMap[config.readReg2] << 5) | (IntRegMap[config.readReg3] << 16), code, codePos);
+	emit32(ARMV8A::EOR32 | 20 | (IntRegMap[config.readReg2] << 5) | (IntRegMap[config.readReg3] << 16), code, codePos);
 
 	// Jump back to the main loop
 	const uint32_t offset = (((uint8_t*)randomx_program_aarch64_vm_instructions_end) - ((uint8_t*)randomx_program_aarch64)) - codePos;
 	emit32(ARMV8A::B | (offset / 4), code, codePos);
 
-	// and w18, w18, CacheLineAlignMask
+	// and w20, w20, CacheLineAlignMask
 	codePos = (((uint8_t*)randomx_program_aarch64_cacheline_align_mask1) - ((uint8_t*)randomx_program_aarch64));
-	emit32(0x121A0000 | 18 | (18 << 5) | ((RandomX_CurrentConfig.Log2_DatasetBaseSize - 7) << 10), code, codePos);
+	emit32(0x121A0000 | 20 | (20 << 5) | ((RandomX_CurrentConfig.Log2_DatasetBaseSize - 7) << 10), code, codePos);
 
 	// and w10, w10, CacheLineAlignMask
 	codePos = (((uint8_t*)randomx_program_aarch64_cacheline_align_mask2) - ((uint8_t*)randomx_program_aarch64));
@@ -189,8 +189,8 @@ void JitCompilerA64::generateProgramLight(Program& program, ProgramConfiguration
 	// and w16, w10, ScratchpadL3Mask64
 	emit32(0x121A0000 | 16 | (10 << 5) | ((RandomX_CurrentConfig.Log2_ScratchpadL3 - 7) << 10), code, codePos);
 
-	// and w17, w18, ScratchpadL3Mask64
+	// and w17, w20, ScratchpadL3Mask64
-	emit32(0x121A0000 | 17 | (18 << 5) | ((RandomX_CurrentConfig.Log2_ScratchpadL3 - 7) << 10), code, codePos);
+	emit32(0x121A0000 | 17 | (20 << 5) | ((RandomX_CurrentConfig.Log2_ScratchpadL3 - 7) << 10), code, codePos);
 
 	codePos = PrologueSize;
 	literalPos = ImulRcpLiteralsEnd;
@@ -206,8 +206,8 @@ void JitCompilerA64::generateProgramLight(Program& program, ProgramConfiguration
 	}
 
 	// Update spMix2
-	// eor w18, config.readReg2, config.readReg3
+	// eor w20, config.readReg2, config.readReg3
-	emit32(ARMV8A::EOR32 | 18 | (IntRegMap[config.readReg2] << 5) | (IntRegMap[config.readReg3] << 16), code, codePos);
+	emit32(ARMV8A::EOR32 | 20 | (IntRegMap[config.readReg2] << 5) | (IntRegMap[config.readReg3] << 16), code, codePos);
 
 	// Jump back to the main loop
 	const uint32_t offset = (((uint8_t*)randomx_program_aarch64_vm_instructions_end_light) - ((uint8_t*)randomx_program_aarch64)) - codePos;
@@ -477,7 +477,7 @@ void JitCompilerA64::emitAddImmediate(uint32_t dst, uint32_t src, uint32_t imm,
 	}
 	else
 	{
-		constexpr uint32_t tmp_reg = 18;
+		constexpr uint32_t tmp_reg = 20;
 		emitMovImmediate(tmp_reg, imm, code, k);
 
 		// add dst, src, tmp_reg
@@ -526,7 +526,7 @@ void JitCompilerA64::emitMemLoadFP(uint32_t src, Instruction& instr, uint8_t* co
 	uint32_t k = codePos;
 
 	uint32_t imm = instr.getImm32();
-	constexpr uint32_t tmp_reg = 18;
+	constexpr uint32_t tmp_reg = 19;
 
 	imm &= instr.getModMem() ? (RandomX_CurrentConfig.ScratchpadL1_Size - 1) : (RandomX_CurrentConfig.ScratchpadL2_Size - 1);
 	emitAddImmediate(tmp_reg, src, imm, code, k);
@@ -580,7 +580,7 @@ void JitCompilerA64::h_IADD_M(Instruction& instr, uint32_t& codePos)
 	const uint32_t src = IntRegMap[instr.src];
 	const uint32_t dst = IntRegMap[instr.dst];
 
-	constexpr uint32_t tmp_reg = 18;
+	constexpr uint32_t tmp_reg = 20;
 	emitMemLoad<tmp_reg>(dst, src, instr, code, k);
 
 	// add dst, dst, tmp_reg
@@ -618,7 +618,7 @@ void JitCompilerA64::h_ISUB_M(Instruction& instr, uint32_t& codePos)
 	const uint32_t src = IntRegMap[instr.src];
 	const uint32_t dst = IntRegMap[instr.dst];
 
-	constexpr uint32_t tmp_reg = 18;
+	constexpr uint32_t tmp_reg = 20;
 	emitMemLoad<tmp_reg>(dst, src, instr, code, k);
 
 	// sub dst, dst, tmp_reg
@@ -637,7 +637,7 @@ void JitCompilerA64::h_IMUL_R(Instruction& instr, uint32_t& codePos)
 
 	if (src == dst)
 	{
-		src = 18;
+		src = 20;
 		emitMovImmediate(src, instr.getImm32(), code, k);
 	}
 
@@ -655,7 +655,7 @@ void JitCompilerA64::h_IMUL_M(Instruction& instr, uint32_t& codePos)
 	const uint32_t src = IntRegMap[instr.src];
 	const uint32_t dst = IntRegMap[instr.dst];
 
-	constexpr uint32_t tmp_reg = 18;
+	constexpr uint32_t tmp_reg = 20;
 	emitMemLoad<tmp_reg>(dst, src, instr, code, k);
 
 	// sub dst, dst, tmp_reg
@@ -686,7 +686,7 @@ void JitCompilerA64::h_IMULH_M(Instruction& instr, uint32_t& codePos)
 	const uint32_t src = IntRegMap[instr.src];
 	const uint32_t dst = IntRegMap[instr.dst];
 
-	constexpr uint32_t tmp_reg = 18;
+	constexpr uint32_t tmp_reg = 20;
 	emitMemLoad<tmp_reg>(dst, src, instr, code, k);
 
 	// umulh dst, dst, tmp_reg
@@ -717,7 +717,7 @@ void JitCompilerA64::h_ISMULH_M(Instruction& instr, uint32_t& codePos)
 	const uint32_t src = IntRegMap[instr.src];
 	const uint32_t dst = IntRegMap[instr.dst];
 
-	constexpr uint32_t tmp_reg = 18;
+	constexpr uint32_t tmp_reg = 20;
 	emitMemLoad<tmp_reg>(dst, src, instr, code, k);
 
 	// smulh dst, dst, tmp_reg
@@ -735,7 +735,7 @@ void JitCompilerA64::h_IMUL_RCP(Instruction& instr, uint32_t& codePos)
 
 	uint32_t k = codePos;
 
-	constexpr uint32_t tmp_reg = 18;
+	constexpr uint32_t tmp_reg = 20;
 	const uint32_t dst = IntRegMap[instr.dst];
 
 	constexpr uint64_t N = 1ULL << 63;
@@ -754,9 +754,9 @@ void JitCompilerA64::h_IMUL_RCP(Instruction& instr, uint32_t& codePos)
 	literalPos -= sizeof(uint64_t);
 	*(uint64_t*)(code + literalPos) = (q << shift) + ((r << shift) / divisor);
 
-	if (literal_id < 13)
+	if (literal_id < 12)
 	{
-		static constexpr uint32_t literal_regs[13] = { 30 << 16, 29 << 16, 28 << 16, 27 << 16, 26 << 16, 25 << 16, 24 << 16, 23 << 16, 22 << 16, 21 << 16, 20 << 16, 11 << 16, 0 };
+		static constexpr uint32_t literal_regs[12] = { 30 << 16, 29 << 16, 28 << 16, 27 << 16, 26 << 16, 25 << 16, 24 << 16, 23 << 16, 22 << 16, 21 << 16, 11 << 16, 0 };
 
 		// mul dst, dst, literal_reg
 		emit32(ARMV8A::MUL | dst | (dst << 5) | literal_regs[literal_id], code, k);
@@ -794,7 +794,7 @@ void JitCompilerA64::h_IXOR_R(Instruction& instr, uint32_t& codePos)
 
 	if (src == dst)
 	{
-		src = 18;
+		src = 20;
 		emitMovImmediate(src, instr.getImm32(), code, k);
 	}
 
@@ -812,7 +812,7 @@ void JitCompilerA64::h_IXOR_M(Instruction& instr, uint32_t& codePos)
 	const uint32_t src = IntRegMap[instr.src];
 	const uint32_t dst = IntRegMap[instr.dst];
 
-	constexpr uint32_t tmp_reg = 18;
+	constexpr uint32_t tmp_reg = 20;
 	emitMemLoad<tmp_reg>(dst, src, instr, code, k);
 
 	// eor dst, dst, tmp_reg
@@ -850,7 +850,7 @@ void JitCompilerA64::h_IROL_R(Instruction& instr, uint32_t& codePos)
 
 	if (src != dst)
 	{
-		constexpr uint32_t tmp_reg = 18;
+		constexpr uint32_t tmp_reg = 20;
 
 		// sub tmp_reg, xzr, src
 		emit32(ARMV8A::SUB | tmp_reg | (31 << 5) | (src << 16), code, k);
@@ -878,7 +878,7 @@ void JitCompilerA64::h_ISWAP_R(Instruction& instr, uint32_t& codePos)
 
 	uint32_t k = codePos;
 
-	constexpr uint32_t tmp_reg = 18;
+	constexpr uint32_t tmp_reg = 20;
 	emit32(ARMV8A::MOV_REG | tmp_reg | (dst << 16), code, k);
 	emit32(ARMV8A::MOV_REG | dst | (src << 16), code, k);
 	emit32(ARMV8A::MOV_REG | src | (tmp_reg << 16), code, k);
@@ -1026,7 +1026,7 @@ void JitCompilerA64::h_CFROUND(Instruction& instr, uint32_t& codePos)
 
 	const uint32_t src = IntRegMap[instr.src];
 
-	constexpr uint32_t tmp_reg = 18;
+	constexpr uint32_t tmp_reg = 20;
 	constexpr uint32_t fpcr_tmp_reg = 8;
 
 	// ror tmp_reg, src, imm
@@ -1050,7 +1050,7 @@ void JitCompilerA64::h_ISTORE(Instruction& instr, uint32_t& codePos)
 
 	const uint32_t src = IntRegMap[instr.src];
 	const uint32_t dst = IntRegMap[instr.dst];
-	constexpr uint32_t tmp_reg = 18;
+	constexpr uint32_t tmp_reg = 20;
 
 	uint32_t imm = instr.getImm32();
 

src/crypto/randomx/jit_compiler_a64_static.S

@@ -72,9 +72,9 @@
 # x15 -> "r7"
 # x16 -> spAddr0
 # x17 -> spAddr1
-# x18 -> temporary
+# x18 -> unused (platform register, don't touch it)
 # x19 -> temporary
-# x20 -> literal for IMUL_RCP
+# x20 -> temporary
 # x21 -> literal for IMUL_RCP
 # x22 -> literal for IMUL_RCP
 # x23 -> literal for IMUL_RCP
@@ -109,7 +109,7 @@ DECL(randomx_program_aarch64):
 	# Save callee-saved registers
 	sub	sp, sp, 192
 	stp	x16, x17, [sp]
-	stp	x18, x19, [sp, 16]
+	str	x19, [sp, 16]
 	stp	x20, x21, [sp, 32]
 	stp	x22, x23, [sp, 48]
 	stp	x24, x25, [sp, 64]
@@ -164,7 +164,6 @@ DECL(randomx_program_aarch64):
 	# Read literals
 	ldr	x0, literal_x0
 	ldr	x11, literal_x11
-	ldr	x20, literal_x20
 	ldr	x21, literal_x21
 	ldr	x22, literal_x22
 	ldr	x23, literal_x23
@@ -196,11 +195,11 @@ DECL(randomx_program_aarch64):
 DECL(randomx_program_aarch64_main_loop):
 	# spAddr0 = spMix1 & ScratchpadL3Mask64;
 	# spAddr1 = (spMix1 >> 32) & ScratchpadL3Mask64;
-	lsr	x18, x10, 32
+	lsr	x20, x10, 32
 
 	# Actual mask will be inserted by JIT compiler
 	and	w16, w10, 1
-	and	w17, w18, 1
+	and	w17, w20, 1
 
 	# x16 = scratchpad + spAddr0
 	# x17 = scratchpad + spAddr1
@@ -208,31 +207,31 @@ DECL(randomx_program_aarch64_main_loop):
 	add	x17, x17, x2
 
 	# xor integer registers with scratchpad data (spAddr0)
-	ldp	x18, x19, [x16]
+	ldp	x20, x19, [x16]
-	eor	x4, x4, x18
+	eor	x4, x4, x20
 	eor	x5, x5, x19
-	ldp	x18, x19, [x16, 16]
+	ldp	x20, x19, [x16, 16]
-	eor	x6, x6, x18
+	eor	x6, x6, x20
 	eor	x7, x7, x19
-	ldp	x18, x19, [x16, 32]
+	ldp	x20, x19, [x16, 32]
-	eor	x12, x12, x18
+	eor	x12, x12, x20
 	eor	x13, x13, x19
-	ldp	x18, x19, [x16, 48]
+	ldp	x20, x19, [x16, 48]
-	eor	x14, x14, x18
+	eor	x14, x14, x20
 	eor	x15, x15, x19
 
 	# Load group F registers (spAddr1)
-	ldpsw	x18, x19, [x17]
+	ldpsw	x20, x19, [x17]
-	ins	v16.d[0], x18
+	ins	v16.d[0], x20
 	ins	v16.d[1], x19
-	ldpsw	x18, x19, [x17, 8]
+	ldpsw	x20, x19, [x17, 8]
-	ins	v17.d[0], x18
+	ins	v17.d[0], x20
 	ins	v17.d[1], x19
-	ldpsw	x18, x19, [x17, 16]
+	ldpsw	x20, x19, [x17, 16]
-	ins	v18.d[0], x18
+	ins	v18.d[0], x20
 	ins	v18.d[1], x19
-	ldpsw	x18, x19, [x17, 24]
+	ldpsw	x20, x19, [x17, 24]
-	ins	v19.d[0], x18
+	ins	v19.d[0], x20
 	ins	v19.d[1], x19
 	scvtf	v16.2d, v16.2d
 	scvtf	v17.2d, v17.2d
@@ -240,17 +239,17 @@ DECL(randomx_program_aarch64_main_loop):
 	scvtf	v19.2d, v19.2d
 
 	# Load group E registers (spAddr1)
-	ldpsw	x18, x19, [x17, 32]
+	ldpsw	x20, x19, [x17, 32]
-	ins	v20.d[0], x18
+	ins	v20.d[0], x20
 	ins	v20.d[1], x19
-	ldpsw	x18, x19, [x17, 40]
+	ldpsw	x20, x19, [x17, 40]
-	ins	v21.d[0], x18
+	ins	v21.d[0], x20
 	ins	v21.d[1], x19
-	ldpsw	x18, x19, [x17, 48]
+	ldpsw	x20, x19, [x17, 48]
-	ins	v22.d[0], x18
+	ins	v22.d[0], x20
 	ins	v22.d[1], x19
-	ldpsw	x18, x19, [x17, 56]
+	ldpsw	x20, x19, [x17, 56]
-	ins	v23.d[0], x18
+	ins	v23.d[0], x20
 	ins	v23.d[1], x19
 	scvtf	v20.2d, v20.2d
 	scvtf	v21.2d, v21.2d
@@ -273,7 +272,6 @@ DECL(randomx_program_aarch64_vm_instructions):
 
 literal_x0:  .fill 1,8,0
 literal_x11: .fill 1,8,0
-literal_x20: .fill 1,8,0
 literal_x21: .fill 1,8,0
 literal_x22: .fill 1,8,0
 literal_x23: .fill 1,8,0
@@ -309,17 +307,17 @@ DECL(randomx_program_aarch64_vm_instructions_end):
 	lsr	x10, x9, 32
 
 	# mx ^= r[readReg2] ^ r[readReg3];
-	eor	x9, x9, x18
+	eor	x9, x9, x20
 
 	# Calculate dataset pointer for dataset prefetch
-	mov	w18, w9
+	mov	w20, w9
 DECL(randomx_program_aarch64_cacheline_align_mask1):
 	# Actual mask will be inserted by JIT compiler
-	and	x18, x18, 1
+	and	x20, x20, 1
-	add	x18, x18, x1
+	add	x20, x20, x1
 
 	# Prefetch dataset data
-	prfm	pldl2strm, [x18]
+	prfm	pldl2strm, [x20]
 
 	# mx <-> ma
 	ror	x9, x9, 32
@@ -331,17 +329,17 @@ DECL(randomx_program_aarch64_cacheline_align_mask2):
 
 DECL(randomx_program_aarch64_xor_with_dataset_line):
 	# xor integer registers with dataset data
-	ldp	x18, x19, [x10]
+	ldp	x20, x19, [x10]
-	eor	x4, x4, x18
+	eor	x4, x4, x20
 	eor	x5, x5, x19
-	ldp	x18, x19, [x10, 16]
+	ldp	x20, x19, [x10, 16]
-	eor	x6, x6, x18
+	eor	x6, x6, x20
 	eor	x7, x7, x19
-	ldp	x18, x19, [x10, 32]
+	ldp	x20, x19, [x10, 32]
-	eor	x12, x12, x18
+	eor	x12, x12, x20
 	eor	x13, x13, x19
-	ldp	x18, x19, [x10, 48]
+	ldp	x20, x19, [x10, 48]
-	eor	x14, x14, x18
+	eor	x14, x14, x20
 	eor	x15, x15, x19
 
 DECL(randomx_program_aarch64_update_spMix1):
@@ -384,7 +382,7 @@ DECL(randomx_program_aarch64_update_spMix1):
 
 	# Restore callee-saved registers
 	ldp	x16, x17, [sp]
-	ldp	x18, x19, [sp, 16]
+	ldr	x19, [sp, 16]
 	ldp	x20, x21, [sp, 32]
 	ldp	x22, x23, [sp, 48]
 	ldp	x24, x25, [sp, 64]
@@ -405,7 +403,7 @@ DECL(randomx_program_aarch64_vm_instructions_end_light):
 	stp	x2, x30, [sp, 80]
 
 	# mx ^= r[readReg2] ^ r[readReg3];
-	eor	x9, x9, x18
+	eor	x9, x9, x20
 
 	# mx <-> ma
 	ror	x9, x9, 32
@@ -447,8 +445,8 @@ DECL(randomx_program_aarch64_light_dataset_offset):
 # x3 -> end item
 
 DECL(randomx_init_dataset_aarch64):
-	# Save x30 (return address)
+	# Save x20 (used as temporary, but must be saved to not break ABI) and x30 (return address)
-	str	x30, [sp, -16]!
+	stp	x20, x30, [sp, -16]!
 
 	# Load pointer to cache memory
 	ldr	x0, [x0]
@@ -460,8 +458,8 @@ DECL(randomx_init_dataset_aarch64_main_loop):
 	cmp	x2, x3
 	bne	DECL(randomx_init_dataset_aarch64_main_loop)
 
-	# Restore x30 (return address)
+	# Restore x20 and x30
-	ldr	x30, [sp], 16
+	ldp	x20, x30, [sp], 16
 
 	ret
 

src/hw/dmi/DmiMemory.cpp

@@ -20,7 +20,6 @@
 
 
 #include "hw/dmi/DmiMemory.h"
-#include "3rdparty/fmt/format.h"
 #include "3rdparty/rapidjson/document.h"
 #include "hw/dmi/DmiTools.h"
 
@@ -28,6 +27,7 @@
 #include <algorithm>
 #include <array>
 #include <regex>
+#include <fmt/format.h>
 
 
 namespace xmrig {

src/hw/dmi/DmiReader.cpp

@@ -19,10 +19,11 @@
  */
 
 #include "hw/dmi/DmiReader.h"
-#include "3rdparty/fmt/core.h"
 #include "3rdparty/rapidjson/document.h"
 #include "hw/dmi/DmiTools.h"
 
+#include <fmt/core.h>
+
 
 namespace xmrig {
 

src/hw/msr/Msr_linux.cpp

@@ -18,7 +18,6 @@
 
 
 #include "hw/msr/Msr.h"
-#include "3rdparty/fmt/core.h"
 #include "backend/cpu/Cpu.h"
 #include "base/io/log/Log.h"
 
@@ -29,6 +28,7 @@
 #include <cstdio>
 #include <dirent.h>
 #include <fcntl.h>
+#include <fmt/core.h>
 #include <fstream>
 #include <sys/stat.h>
 #include <sys/types.h>