CNv4 JIT compiler for x86-64 and tests

Minimalistic JIT code generator for random math sequence in CryptonightR.

Usage:
- Allocate writable and executable memory
- Call v4_generate_JIT_code with "buf" pointed to memory allocated on the previous step
- Call the generated code instead of "v4_random_math(code, r)", omit the "code" parameter

1 files changed, 99 insertions, 0 deletions

diff --git a/src/crypto/CryptonightR_JIT.c b/src/crypto/CryptonightR_JIT.c
new file mode 100644
index 000000000..0b123b4ee
--- /dev/null
+++ b/src/crypto/CryptonightR_JIT.c
@@ -0,0 +1,99 @@
+#include <assert.h>
+#include <stddef.h>
+#include <stdint.h>
+#include <string.h>
+#include <stdio.h>
+#include <unistd.h>
+
+#include "int-util.h"
+#include "hash-ops.h"
+#include "variant4_random_math.h"
+#include "CryptonightR_JIT.h"
+#include "CryptonightR_template.h"
+
+static const uint8_t prologue[] = {
+	0x4C, 0x8B, 0xD7,	// mov r10, rdi
+	0x53,			// push rbx
+	0x55,			// push rbp
+	0x41, 0x57,		// push r15
+	0x4C, 0x8B, 0xDC,	// mov r11, rsp
+	0x41, 0x8B, 0x1A,	// mov ebx, DWORD PTR [r10]
+	0x41, 0x8B, 0x72, 0x04,	// mov esi, DWORD PTR [r10+4]
+	0x41, 0x8B, 0x7A, 0x08,	// mov edi, DWORD PTR [r10+8]
+	0x41, 0x8B, 0x6A, 0x0C,	// mov ebp, DWORD PTR [r10+12]
+	0x41, 0x8B, 0x62, 0x10,	// mov esp, DWORD PTR [r10+16]
+	0x45, 0x8B, 0x7A, 0x14,	// mov r15d, DWORD PTR [r10+20]
+	0x41, 0x8B, 0x42, 0x18,	// mov eax, DWORD PTR [r10+24]
+	0x41, 0x8B, 0x52, 0x1C,	// mov edx, DWORD PTR [r10+28]
+	0x45, 0x8B, 0x4A, 0x20,	// mov r9d, DWORD PTR [r10+32]
+};
+
+static const uint8_t epilogue[] = {
+	0x49, 0x8B, 0xE3,	// mov rsp, r11
+	0x41, 0x89, 0x1A,	// mov DWORD PTR [r10], ebx
+	0x41, 0x89, 0x72, 0x04,	// mov DWORD PTR [r10+4], esi
+	0x41, 0x89, 0x7A, 0x08,	// mov DWORD PTR [r10+8], edi
+	0x41, 0x89, 0x6A, 0x0C,	// mov DWORD PTR [r10+12], ebp
+	0x41, 0x5F,		// pop r15
+	0x5D,			// pop rbp
+	0x5B,			// pop rbx
+	0xC3,			// ret
+};
+
+#define APPEND_CODE(src, size) \
+	do { \
+		if (JIT_code + (size) > JIT_code_end) \
+			return -1; \
+		memcpy(JIT_code, (src), (size)); \
+		JIT_code += (size); \
+	} while (0)
+
+int v4_generate_JIT_code(const struct V4_Instruction* code, v4_random_math_JIT_func buf, const size_t buf_size)
+{
+	uint8_t* JIT_code = (uint8_t*) buf;
+	const uint8_t* JIT_code_end = JIT_code + buf_size;
+
+	APPEND_CODE(prologue, sizeof(prologue));
+
+	uint32_t prev_rot_src = 0xFFFFFFFFU;
+
+	for (int i = 0;; ++i)
+	{
+		const struct V4_Instruction inst = code[i];
+		if (inst.opcode == RET)
+			break;
+
+		const uint8_t opcode = (inst.opcode == MUL) ? inst.opcode : (inst.opcode + 2);
+
+		const uint32_t a = inst.dst_index;
+		const uint32_t b = inst.src_index;
+		const uint8_t c = opcode | (inst.dst_index << V4_OPCODE_BITS) | (((inst.src_index == 8) ? inst.dst_index : inst.src_index) << (V4_OPCODE_BITS + V4_DST_INDEX_BITS));
+
+		switch (inst.opcode)
+		{
+		case ROR:
+		case ROL:
+			if (b != prev_rot_src)
+			{
+				prev_rot_src = b;
+				const uint8_t* p1 = (const uint8_t*) instructions_mov[c];
+				const uint8_t* p2 = (const uint8_t*) instructions_mov[c + 1];
+				APPEND_CODE(p1, p2 - p1);
+			}
+			break;
+		}
+
+		if (a == prev_rot_src)
+			prev_rot_src = 0xFFFFFFFFU;
+
+		const uint8_t* p1 = (const uint8_t*) instructions[c];
+		const uint8_t* p2 = (const uint8_t*) instructions[c + 1];
+		APPEND_CODE(p1, p2 - p1);
+
+		if (inst.opcode == ADD)
+			*(uint32_t*)(JIT_code - 4) = inst.C;
+	}
+
+	APPEND_CODE(epilogue, sizeof(epilogue));
+	return 0;
+}