From 720325c40f812739e964a127a640039b15ee0db8 Mon Sep 17 00:00:00 2001
From: SChernykh <15806605+SChernykh@users.noreply.github.com>
Date: Sun, 17 May 2026 17:35:40 +0200
Subject: [PATCH] RandomX optimizations: - ARM64: optimized
 emitMovImmediate/emitMemLoad - ARM64: disabled 32-bit literal preloading (it
 was slower) - Android and Linux: added MADV_COLLAPSE support to memory
 allocation

---
 src/crypto/common/VirtualMemory_unix.cpp | 10 +++++--
 src/crypto/randomx/jit_compiler_a64.cpp  | 33 ++++++++++++++++++------
 2 files changed, 33 insertions(+), 10 deletions(-)

diff --git a/src/crypto/common/VirtualMemory_unix.cpp b/src/crypto/common/VirtualMemory_unix.cpp
index 471c9cf07..dedf4603a 100644
--- a/src/crypto/common/VirtualMemory_unix.cpp
+++ b/src/crypto/common/VirtualMemory_unix.cpp
@@ -74,6 +74,11 @@
 #endif
 
 
+#ifndef MADV_COLLAPSE
+#   define MADV_COLLAPSE 25
+#endif
+
+
 #if defined(XMRIG_OS_LINUX) || (!defined(XMRIG_OS_APPLE) && !defined(XMRIG_OS_FREEBSD))
 static inline int hugePagesFlag(size_t size)
 {
@@ -278,8 +283,9 @@ bool xmrig::VirtualMemory::allocateOneGbPagesMemory()
 
 bool xmrig::VirtualMemory::adviseLargePages(void *p, size_t size)
 {
-#   ifdef XMRIG_OS_LINUX
-    return (madvise(p, size, MADV_HUGEPAGE) == 0);
+#   if defined(XMRIG_OS_ANDROID) || defined(XMRIG_OS_LINUX)
+    // 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
     return false;
 #   endif
diff --git a/src/crypto/randomx/jit_compiler_a64.cpp b/src/crypto/randomx/jit_compiler_a64.cpp
index 0d624dfa9..bae695cf1 100644
--- a/src/crypto/randomx/jit_compiler_a64.cpp
+++ b/src/crypto/randomx/jit_compiler_a64.cpp
@@ -141,7 +141,7 @@ void JitCompilerA64::generateProgram(Program& program, ProgramConfiguration& con
 
 	codePos = PrologueSize;
 	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)
 		reg_changed_offset[i] = codePos;
@@ -237,7 +237,7 @@ void JitCompilerA64::generateProgramLight(Program& program, ProgramConfiguration
 
 	codePos = PrologueSize;
 	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)
 		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;
 
+	// 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))
 	{
+		// Sign-extended 64-bit value: 0x000000000000xxxx
 		// movz tmp_reg, imm32 (16 low bits)
 		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
 	{
+		// Full sign-extended 64-bit value: 0x00000000xxxxxxxx or 0xFFFFFFFFxxxxxxxx
 		if (num32bitLiterals < 64)
 		{
 			if (static_cast<int32_t>(imm) < 0)
@@ -611,18 +629,17 @@ void JitCompilerA64::emitMemLoad(uint32_t dst, uint32_t src, Instruction& instr,
 	else
 	{
 		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);
 
 			// ldr tmp_reg, [x2, tmp_reg, lsl 3]
 			emit32(0xf8607840 | tmp_reg | (tmp_reg << 16), code, k);
 		}
-		else
-		{
-			// ldr tmp_reg, [x2]
-			emit32(0xf9400040 | tmp_reg, code, k);
-		}
 	}
 
 	codePos = k;