Improve sljit_emit_op2_shift on x86 (#342)

Author: zherczeg

sljit_src/sljitLir.c

@@ -1986,15 +1986,16 @@ static SLJIT_INLINE CHECK_RETURN_TYPE check_sljit_emit_op2_shift(struct sljit_co
 {
 	SLJIT_UNUSED_ARG(shift_arg);
 #if (defined SLJIT_ARGUMENT_CHECKS && SLJIT_ARGUMENT_CHECKS)
-	CHECK_ARGUMENT(op == (SLJIT_ADD | SLJIT_SHL_IMM));
+	CHECK_ARGUMENT((op & ~SLJIT_SRC2_UNDEFINED) == (SLJIT_ADD | SLJIT_SHL_IMM));
 	FUNCTION_CHECK_DST(dst, dstw);
 	FUNCTION_CHECK_SRC(src1, src1w);
 	FUNCTION_CHECK_SRC(src2, src2w);
 	compiler->last_flags = 0;
 #endif /* SLJIT_ARGUMENT_CHECKS */
 #if (defined SLJIT_VERBOSE && SLJIT_VERBOSE)
 	if (SLJIT_UNLIKELY(!!compiler->verbose)) {
-		fprintf(compiler->verbose, "  %s.shl_imm ", op2_names[GET_OPCODE(op) - SLJIT_OP2_BASE]);
+		fprintf(compiler->verbose, "  %s.shl_imm%s ", op2_names[GET_OPCODE(op) - SLJIT_OP2_BASE],
+			(op & SLJIT_SRC2_UNDEFINED) ? ".src2und" : "");
 
 		sljit_verbose_param(compiler, dst, dstw);
 		fprintf(compiler->verbose, ", ");

sljit_src/sljitLir.h

@@ -1468,7 +1468,10 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_shift_into(struct sljit_compiler *
 
 /* The following options are used by sljit_emit_op2_shift. */
 
+/* The src2 argument is shifted left by an immedate value. */
 #define SLJIT_SHL_IMM			(1 << 9)
+/* When src2 argument is a register, its value is undefined after the operation. */
+#define SLJIT_SRC2_UNDEFINED		(1 << 10)
 
 /* Emits an addition operation, where the second argument is shifted by a value.
 

sljit_src/sljitNativeX86_common.c

@@ -3094,11 +3094,16 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op2_shift(struct sljit_compiler *c
 		return SLJIT_SUCCESS;
 	}
 
-	dst_r = FAST_IS_REG(dst) && (dst != src1) ? dst : TMP_REG1;
+	if ((op & SLJIT_SRC2_UNDEFINED) != 0 && FAST_IS_REG(src2) && src1 != src2)
+		dst_r = src2;
+	else {
+		dst_r = FAST_IS_REG(dst) && (dst != src1) ? dst : TMP_REG1;
 
-	if (src2 != dst_r) {
-		EMIT_MOV(compiler, dst_r, 0, src2, src2w);
+		if (src2 != dst_r) {
+			EMIT_MOV(compiler, dst_r, 0, src2, src2w);
+		}
 	}
+
 	inst = emit_x86_instruction(compiler, 1 | EX86_SHIFT_INS, SLJIT_IMM, shift_arg, dst_r, 0);
 	FAIL_IF(!inst);
 	inst[1] |= SHL;
@@ -3110,6 +3115,13 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op2_shift(struct sljit_compiler *c
 		return SLJIT_SUCCESS;
 	}
 
+	if (FAST_IS_REG(dst) && FAST_IS_REG(src1)) {
+		inst = emit_x86_instruction(compiler, 1, dst, 0, SLJIT_MEM2(src1, dst_r), 0);
+		FAIL_IF(!inst);
+		*inst = LEA_r_m;
+		return SLJIT_SUCCESS;
+	}
+
 	if (src1 == SLJIT_IMM) {
 		BINARY_IMM(ADD, ADD_rm_r, src1w, dst_r, 0);
 	} else {

test_src/sljitTest.c

@@ -9860,12 +9860,12 @@ static void test80(void)
 	executable_code code;
 	struct sljit_compiler *compiler = sljit_create_compiler(NULL);
 	sljit_s32 i;
-	sljit_sw buf[20];
+	sljit_sw buf[23];
 
 	if (verbose)
 		printf("Run test80\n");
 
-	for (i = 0; i < 20; i++)
+	for (i = 0; i < 23; i++)
 		buf[i] = -1;
 
 	FAILED(!compiler, "cannot create compiler\n");
@@ -9986,6 +9986,23 @@ static void test80(void)
 	/* buf[19] */
 	sljit_emit_op1(compiler, SLJIT_MOV, SLJIT_MEM1(SLJIT_S0), 19 * sizeof(sljit_sw), SLJIT_TMP_DEST_REG, 0);
 
+	sljit_emit_op1(compiler, SLJIT_MOV, SLJIT_R0, 0, SLJIT_IMM, WCONST(0x8e2c4e79b79500a8, 0x8e2c4e79));
+	sljit_emit_op1(compiler, SLJIT_MOV, SLJIT_R1, 0, SLJIT_IMM, WCONST(0x5ce8fa2c60c7c10c, 0x5ce8fa2c));
+	/* buf[20] */
+	sljit_emit_op2_shift(compiler, SLJIT_ADD | SLJIT_SHL_IMM | SLJIT_SRC2_UNDEFINED, SLJIT_MEM1(SLJIT_S0), 20 * sizeof(sljit_sw), SLJIT_R0, 0, SLJIT_R1, 0, 7);
+
+	sljit_emit_op1(compiler, SLJIT_MOV, SLJIT_R1, 0, SLJIT_IMM, WCONST(0xbdfe33c03babe728, 0xbdfe33c0));
+	sljit_emit_op1(compiler, SLJIT_MOV, SLJIT_R2, 0, SLJIT_IMM, WCONST(0x6dda8d28a705b87a, 0x6dda8d28));
+	sljit_emit_op2_shift(compiler, SLJIT_ADD | SLJIT_SHL_IMM | SLJIT_SRC2_UNDEFINED, SLJIT_R2, 0, SLJIT_R1, 0, SLJIT_R2, 0, 9);
+	/* buf[21] */
+	sljit_emit_op1(compiler, SLJIT_MOV, SLJIT_MEM1(SLJIT_S0), 21 * sizeof(sljit_sw), SLJIT_R2, 0);
+
+	sljit_emit_op1(compiler, SLJIT_MOV, SLJIT_R1, 0, SLJIT_IMM, WCONST(0xadb9530fe12dca23, 0xadb9530f));
+	sljit_emit_op1(compiler, SLJIT_MOV, SLJIT_R2, 0, SLJIT_IMM, WCONST(0x678fe74fab8bcc2b, 0x678fe74f));
+	sljit_emit_op2_shift(compiler, SLJIT_ADD | SLJIT_SHL_IMM | SLJIT_SRC2_UNDEFINED, SLJIT_R0, 0, SLJIT_R1, 0, SLJIT_R2, 0, 11);
+	/* buf[22] */
+	sljit_emit_op1(compiler, SLJIT_MOV, SLJIT_MEM1(SLJIT_S0), 22 * sizeof(sljit_sw), SLJIT_R0, 0);
+
 	sljit_emit_return_void(compiler);
 
 	code.code = sljit_generate_code(compiler, 0, NULL);
@@ -10015,6 +10032,9 @@ static void test80(void)
 	FAILED(buf[17] != WCONST(0xe7204ac2adf575cd, 0xe72075cd), "test80 case 18 failed\n");
 	FAILED(buf[18] != WCONST(0x90e766522922d467, 0x90e76550), "test80 case 19 failed\n");
 	FAILED(buf[19] != WCONST(0xbf01437a347858bf, 0xbf014378), "test80 case 20 failed\n");
+	FAILED(buf[20] != WCONST(0x2a964aa1b7586a8, 0x2a96479), "test80 case 21 failed\n");
+	FAILED(buf[21] != WCONST(0x7318850e471cdb28, 0x731883c0), "test80 case 22 failed\n");
+	FAILED(buf[22] != WCONST(0x2cf3d06c3f8f2223, 0x2cf3cb0f), "test80 case 23 failed\n");
 
 	sljit_free_code(code.code, NULL);
 	successful_tests++;