30% faster JIT compiler

src/asm/randomx_reciprocal.inc

@@ -0,0 +1,7 @@
+	mov edx, 1
+	mov r8, rcx
+	xor eax, eax
+	bsr rcx, rcx
+	shl rdx, cl
+	div r8
+	ret

src/jit_compiler_x86.cpp

@@ -616,7 +616,7 @@ namespace randomx {
 		if (!isPowerOf2(divisor)) {
 			registerUsage[instr.dst].lastUsed = i;
 			emit(MOV_RAX_I);
-			emit64(randomx_reciprocal(divisor));
+			emit64(randomx_reciprocal_fast(divisor));
 			emit(REX_IMUL_RM);
 			emitByte(0xc0 + 8 * instr.dst);
 		}

src/jit_compiler_x86.hpp

@@ -78,23 +78,13 @@ namespace randomx {
 		}
 
 		void emit32(uint32_t val) {
-			code[codePos + 0] = val;
-			code[codePos + 1] = val >> 8;
-			code[codePos + 2] = val >> 16;
-			code[codePos + 3] = val >> 24;
-			codePos += 4;
+			memcpy(code + codePos, &val, sizeof val);
+			codePos += sizeof val;
 		}
 
 		void emit64(uint64_t val) {
-			code[codePos + 0] = val;
-			code[codePos + 1] = val >> 8;
-			code[codePos + 2] = val >> 16;
-			code[codePos + 3] = val >> 24;
-			code[codePos + 4] = val >> 32;
-			code[codePos + 5] = val >> 40;
-			code[codePos + 6] = val >> 48;
-			code[codePos + 7] = val >> 56;
-			codePos += 8;
+			memcpy(code + codePos, &val, sizeof val);
+			codePos += sizeof val;
 		}
 
 		template<size_t N>

src/jit_compiler_x86_static.S

@@ -42,6 +42,7 @@
 .global DECL(randomx_sshash_end)
 .global DECL(randomx_sshash_init)
 .global DECL(randomx_program_end)
+.global DECL(randomx_reciprocal_fast)
 
 #define db .byte
 
@@ -158,3 +159,7 @@ DECL(randomx_sshash_init):
 .balign 64
 DECL(randomx_program_end):
 	nop
+
+DECL(randomx_reciprocal_fast):
+	mov rcx, rdi
+	#include "asm/randomx_reciprocal.inc"

src/jit_compiler_x86_static.asm

@@ -35,6 +35,7 @@ PUBLIC randomx_sshash_prefetch
 PUBLIC randomx_sshash_end
 PUBLIC randomx_sshash_init
 PUBLIC randomx_program_end
+PUBLIC randomx_reciprocal_fast
 
 ALIGN 64
 randomx_program_prologue PROC
@@ -169,6 +170,10 @@ randomx_program_end PROC
 	nop
 randomx_program_end ENDP
 
+randomx_reciprocal_fast PROC
+	include asm/randomx_reciprocal.inc
+randomx_reciprocal_fast ENDP
+
 _RANDOMX_JITX86_STATIC ENDS
 
 ENDIF

src/reciprocal.c

@@ -56,4 +56,4 @@ uint64_t randomx_reciprocal(uint64_t divisor) {
 	}
 
 	return quotient;
-}
+}

src/reciprocal.h

@@ -26,6 +26,7 @@ extern "C" {
 #endif
 
 uint64_t randomx_reciprocal(uint64_t);
+uint64_t randomx_reciprocal_fast(uint64_t);
 
 #if defined(__cplusplus)
 }

src/tests/jit-performance.cpp

@@ -0,0 +1,44 @@
+#include "../aes_hash.hpp"
+#include "../jit_compiler_x86.hpp"
+#include "../program.hpp"
+#include "utility.hpp"
+#include "stopwatch.hpp"
+#include "../blake2/blake2.h"
+#include "../reciprocal.h"
+
+int main(int argc, char** argv) {
+	int count;
+	readInt(argc, argv, count, 1000000);
+
+	const char seed[] = "JIT performance test seed";
+	uint8_t hash[64];
+
+	blake2b(&hash, sizeof hash, &seed, sizeof seed, nullptr, 0);
+
+	randomx::ProgramConfiguration config;
+
+	randomx::Program program;
+	randomx::JitCompilerX86 jit;
+
+	std::cout << "Compiling " << count << " programs..." << std::endl;
+
+	Stopwatch sw(true);
+
+	for (int i = 0; i < count; ++i) {
+		fillAes1Rx4<false>(hash, sizeof(program), &program);
+		auto addressRegisters = program.getEntropy(12);
+		config.readReg0 = 0 + (addressRegisters & 1);
+		addressRegisters >>= 1;
+		config.readReg1 = 2 + (addressRegisters & 1);
+		addressRegisters >>= 1;
+		config.readReg2 = 4 + (addressRegisters & 1);
+		addressRegisters >>= 1;
+		config.readReg3 = 6 + (addressRegisters & 1);
+		jit.generateProgram(program, config);
+	}
+
+	std::cout << "Elapsed: " << sw.getElapsed() << " s" << std::endl;
+
+	dump((const char*)jit.getProgramFunc(), randomx::CodeSize, "program.bin");
+	return 0;
+}