mirrors/xmrig
1
0
Fork 1
mirror of https://github.com/xmrig/xmrig.git synced 2026-06-18 10:22:39 -04:00
Code Issues Projects Releases Wiki Activity

Merge pull request #3816 from SChernykh/dev

Browse Source 
ARM64 RandomX optimizations
This commit is contained in:
xmrig
2026-05-18 02:12:24 +07:00
committed by GitHub
parent 27f116e2da 720325c40f
commit 7eadfdc9c6
2 changed files with 33 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
src/crypto/common/VirtualMemory_unix.cpp
View File

@@ -74,6 +74,11 @@
#endif #endif
#ifndef MADV_COLLAPSE
# define MADV_COLLAPSE 25
#endif
#if defined(XMRIG_OS_LINUX) || (!defined(XMRIG_OS_APPLE) && !defined(XMRIG_OS_FREEBSD)) #if defined(XMRIG_OS_LINUX) || (!defined(XMRIG_OS_APPLE) && !defined(XMRIG_OS_FREEBSD))
static inline int hugePagesFlag(size_t size) static inline int hugePagesFlag(size_t size)
{ {
@@ -278,8 +283,9 @@ bool xmrig::VirtualMemory::allocateOneGbPagesMemory()
bool xmrig::VirtualMemory::adviseLargePages(void *p, size_t size) bool xmrig::VirtualMemory::adviseLargePages(void *p, size_t size)
{ {
# ifdef XMRIG_OS_LINUX # if defined(XMRIG_OS_ANDROID) || defined(XMRIG_OS_LINUX)
return (madvise(p, size, MADV_HUGEPAGE) == 0); // MADV_COLLAPSE works even if /sys/kernel/mm/transparent_hugepage/enabled is set to "never", but only on Linux 6.1+
return (madvise(p, size, MADV_COLLAPSE) == 0) || (madvise(p, size, MADV_HUGEPAGE) == 0);
# else # else
return false; return false;
# endif # endif

33
src/crypto/randomx/jit_compiler_a64.cpp
View File

@@ -141,7 +141,7 @@ void JitCompilerA64::generateProgram(Program& program, ProgramConfiguration& con
codePos = PrologueSize; codePos = PrologueSize;
literalPos = ImulRcpLiteralsEnd; literalPos = ImulRcpLiteralsEnd;
num32bitLiterals = 0; num32bitLiterals = 64; // effectively disabled because it's slower than plain movn/movz+movk
for (uint32_t i = 0; i < RegistersCount; ++i) for (uint32_t i = 0; i < RegistersCount; ++i)
reg_changed_offset[i] = codePos; reg_changed_offset[i] = codePos;
@@ -237,7 +237,7 @@ void JitCompilerA64::generateProgramLight(Program& program, ProgramConfiguration
codePos = PrologueSize; codePos = PrologueSize;
literalPos = ImulRcpLiteralsEnd; literalPos = ImulRcpLiteralsEnd;
num32bitLiterals = 0; num32bitLiterals = 64; // effectively disabled because it's slower than plain movn/movz+movk
for (uint32_t i = 0; i < RegistersCount; ++i) for (uint32_t i = 0; i < RegistersCount; ++i)
reg_changed_offset[i] = codePos; reg_changed_offset[i] = codePos;
@@ -488,13 +488,31 @@ void JitCompilerA64::emitMovImmediate(uint32_t dst, uint32_t imm, uint8_t* code,
{ {
uint32_t k = codePos; uint32_t k = codePos;
// 196606 different values can be encoded with a single instruction, the rest requires smov/umov load, or movn/movz+movk pair
if (imm < (1 << 16)) if (imm < (1 << 16))
{ {
// Sign-extended 64-bit value: 0x000000000000xxxx
// movz tmp_reg, imm32 (16 low bits) // movz tmp_reg, imm32 (16 low bits)
emit32(ARMV8A::MOVZ | dst | (imm << 5), code, k); emit32(ARMV8A::MOVZ | dst | (imm << 5), code, k);
} }
else if ((imm >> 16) == 0xFFFF) {
// Sign-extended 64-bit value: 0xFFFFFFFFFFFFxxxx
// movn tmp_reg, ~imm32 (16 low bits)
emit32(ARMV8A::MOVN | dst | ((~imm & 0xFFFF) << 5), code, k);
}
else if (((imm & 0xFFFF) == 0xFFFF) && (static_cast<int32_t>(imm) < 0)) {
// Sign-extended 64-bit value: 0xFFFFFFFFxxxxFFFF
// movn tmp_reg, ~imm32 (16 high bits)
emit32(ARMV8A::MOVN | dst | (1 << 21) | ((~imm >> 16) << 5), code, k);
}
else if (((imm & 0xFFFF) == 0) && (static_cast<int32_t>(imm) >= 0)) {
// Sign-extended 64-bit value: 0x00000000xxxx0000
// movz tmp_reg, imm32 (16 high bits)
emit32(ARMV8A::MOVZ | dst | (1 << 21) | ((imm >> 16) << 5), code, k);
}
else else
{ {
// Full sign-extended 64-bit value: 0x00000000xxxxxxxx or 0xFFFFFFFFxxxxxxxx
if (num32bitLiterals < 64) if (num32bitLiterals < 64)
{ {
if (static_cast<int32_t>(imm) < 0) if (static_cast<int32_t>(imm) < 0)
@@ -611,18 +629,17 @@ void JitCompilerA64::emitMemLoad(uint32_t dst, uint32_t src, Instruction& instr,
else else
{ {
imm = (imm & ScratchpadL3Mask) >> 3; imm = (imm & ScratchpadL3Mask) >> 3;
if (imm) if (imm < 4096) {
// ldr tmp_reg, [x2, #imm*8]
emit32(0xf9400040 | tmp_reg | (imm << 10), code, k);
}
else
{ {
emitMovImmediate(tmp_reg, imm, code, k); emitMovImmediate(tmp_reg, imm, code, k);
// ldr tmp_reg, [x2, tmp_reg, lsl 3] // ldr tmp_reg, [x2, tmp_reg, lsl 3]
emit32(0xf8607840 | tmp_reg | (tmp_reg << 16), code, k); emit32(0xf8607840 | tmp_reg | (tmp_reg << 16), code, k);
} }
else
{
// ldr tmp_reg, [x2]
emit32(0xf9400040 | tmp_reg, code, k);
}
} }
codePos = k; codePos = k;