diff --git a/hw/xbox/dsp/dsp_cpu.c b/hw/xbox/dsp/dsp_cpu.c
index 05a5ec90b7..beadb4dd97 100644
--- a/hw/xbox/dsp/dsp_cpu.c
+++ b/hw/xbox/dsp/dsp_cpu.c
@@ -67,8 +67,8 @@ static void dsp_compute_ssh_ssl(dsp_core_t* dsp);
 
 /* 56bits arithmetic */
 static uint16_t dsp_abs56(uint32_t *dest);
-static uint16_t dsp_asl56(uint32_t *dest);
-static uint16_t dsp_asr56(uint32_t *dest);
+static uint16_t dsp_asl56(uint32_t *dest, int n);
+static uint16_t dsp_asr56(uint32_t *dest, int n);
 static uint16_t dsp_add56(uint32_t *source, uint32_t *dest);
 static uint16_t dsp_sub56(uint32_t *source, uint32_t *dest);
 static void dsp_mul56(uint32_t source1, uint32_t source2, uint32_t *dest, uint8_t signe);
@@ -164,9 +164,9 @@ static const OpcodeEntry nonparallel_opcodes[] = {
     { "0000000101iiiiii1000d110", "and #xx, D", NULL, NULL },
     { "00000001010000001100d110", "and #xxxx, D", dis_and_long, emu_and_long },
     { "00000000iiiiiiii101110EE", "andi #xx, D", dis_andi, emu_andi },
-    { "0000110000011101SiiiiiiD", "asl #ii, S2, D", NULL, NULL },
+    { "0000110000011101SiiiiiiD", "asl #ii, S2, D", dis_asl_imm, emu_asl_imm },
     { "0000110000011110010SsssD", "asl S1, S2, D", NULL, NULL },
-    { "0000110000011100SiiiiiiD", "asr #ii, S2, D", NULL, NULL },
+    { "0000110000011100SiiiiiiD", "asr #ii, S2, D", dis_asr_imm, emu_asr_imm },
     { "0000110000011110011SsssD", "asr S1, S2, D", NULL, NULL },
     { "00001101000100000100CCCC", "bcc xxxx", dis_bcc_long, emu_bcc_long }, //??
     { "00000101CCCC01aaaa0aaaaa", "bcc xxx", dis_bcc_imm, emu_bcc_imm },
@@ -181,7 +181,7 @@ static const OpcodeEntry nonparallel_opcodes[] = {
     { "0000101010pppppp0S00bbbb", "bclr #n, [X or Y]:pp", dis_bclr_pp, emu_bclr_pp },
     { "0000000100qqqqqq0S00bbbb", "bclr #n, [X or Y]:qq", NULL, NULL },
     { "0000101011DDDDDD010bbbbb", "bclr #n, D", dis_bclr_reg, emu_bclr_reg },
-    { "000011010001000011000000", "bra xxxx", NULL, NULL },
+    { "000011010001000011000000", "bra xxxx", dis_bra_long, emu_bra_long },
     { "00000101000011aaaa0aaaaa", "bra xxx", dis_bra_imm, emu_bra_imm },
     { "0000110100011RRR11000000", "bra Rn", NULL, NULL },
     { "0000110010MMMRRR0S0bbbbb", "brclr #n, [X or Y]:ea, xxxx", NULL, NULL, match_MMMRRR },
@@ -286,17 +286,17 @@ static const OpcodeEntry nonparallel_opcodes[] = {
     { "000011000001111011iiiiiD", "lsr #ii, D", NULL, NULL },
     { "00001100000111100011sssD", "lsr S, D", NULL, NULL },
     { "00000100010MMRRR000ddddd", "lua ea, D", dis_lua, emu_lua },
-    { "0000010000aaaRRRaaaadddd", "lua {Rn + aa}, D", NULL, NULL },
+    { "0000010000aaaRRRaaaadddd", "lua (Rn + aa), D", dis_lua_rel, emu_lua_rel },
     { "00000001000sssss11QQdk10", "mac S, #n, D", NULL, NULL },
     { "000000010100000111qqdk10", "maci #xxxx, S, D", NULL, NULL },
     { "00000001001001101sdkQQQQ", "mac_s_u S1, S2, D", NULL, NULL },
     { "00000001000sssss11QQdk11", "macr S1, S2, D", NULL, NULL },
     { "000000010100000111qqdk11", "macri #xxxx, S, D", NULL, NULL },
     { "00001100000110111000sssD", "merge S, D", NULL, NULL },
-    { "0000101001110RRR1WDDDDDD", "move X:{Rn + xxxx} <-> R", dis_move_x_long, emu_move_x_long },
-    { "0000101101110RRR1WDDDDDD", "move Y:{Rn + xxxx} <-> R", NULL, NULL },
-    { "0000001aaaaaaRRR1a0WDDDD", "move X:{Rn + xxx} <-> R", dis_move_x_imm, emu_move_x_imm },
-    { "0000001aaaaaaRRR1a1WDDDD", "move Y:{Rn + xxx} <-> R", NULL, NULL },
+    { "0000101001110RRR1WDDDDDD", "move X:(Rn + xxxx) <-> R", dis_move_x_long, emu_move_x_long },
+    { "0000101101110RRR1WDDDDDD", "move Y:(Rn + xxxx) <-> R", NULL, NULL },
+    { "0000001aaaaaaRRR1a0WDDDD", "move X:(Rn + xxx) <-> R", dis_move_x_imm, emu_move_x_imm },
+    { "0000001aaaaaaRRR1a1WDDDD", "move Y:(Rn + xxx) <-> R", dis_move_y_imm, emu_move_y_imm },
     { "00000101W1MMMRRR0s1ddddd", "movec [X or Y]:ea <-> R", dis_movec_ea, emu_movec_ea, match_MMMRRR },
     { "00000101W0aaaaaa0s1ddddd", "movec [X or Y]:aa <-> R", dis_movec_aa, emu_movec_aa, match_MMMRRR },
     { "00000100W1eeeeee101ddddd", "movec R1, R2", dis_movec_reg, emu_movec_reg },
@@ -337,7 +337,7 @@ static const OpcodeEntry nonparallel_opcodes[] = {
     { "000000000000000000000100", "rti", NULL, emu_rti },
     { "000000000000000000001100", "rts", NULL, emu_rts },
     { "000000000000000010000111", "stop", NULL, emu_stop },
-    { "0000000101iiiiii1000d100", "sub #xx, D", NULL, NULL },
+    { "0000000101iiiiii1000d100", "sub #xx, D", dis_sub_imm, emu_sub_imm },
     { "00000001010000001100d100", "sub #xxxx, D", NULL, NULL },
     { "00000010CCCC00000JJJd000", "tcc S1, D1", dis_tcc, emu_tcc },
     { "00000011CCCC0ttt0JJJdTTT", "tcc S1,D2 S2,D2", dis_tcc, emu_tcc },
@@ -1227,46 +1227,37 @@ static uint16_t dsp_abs56(uint32_t *dest)
     return newsr;
 }
 
-static uint16_t dsp_asl56(uint32_t *dest)
+static uint16_t dsp_asl56(uint32_t *dest, int n)
 {
-    uint16_t overflow, carry;
+    /* Shift left dest n bits: D<<=n */
 
-    /* Shift left dest 1 bit: D<<=1 */
+    uint64_t dest_v = dest[2] | ((uint64_t)dest[1] << 24) | ((uint64_t)dest[0] << 48);
 
-    carry = (dest[0]>>7) & 1;
+    uint32_t carry = (dest_v >> (56-n)) & 1;
 
-    dest[0] <<= 1;
-    dest[0] |= (dest[1]>>23) & 1;
-    dest[0] &= BITMASK(8);
+    uint64_t dest_s = dest_v << n;
+    dest[2] = dest_s & BITMASK(24);
+    dest[1] = (dest_s >> 24) & BITMASK(24);
+    dest[0] = (dest_s >> 48) & BITMASK(8);
 
-    dest[1] <<= 1;
-    dest[1] |= (dest[2]>>23) & 1;
-    dest[1] &= BITMASK(24);
-    
-    dest[2] <<= 1;
-    dest[2] &= BITMASK(24);
+    uint32_t overflow = (dest_v >> (56-n)) != 0;
+    uint32_t v = ((dest_v >> 55) & 1) != ((dest_s >> 55) & 1);
 
-    overflow = (carry != ((dest[0]>>7) & 1));
-
-    return (overflow<<DSP_SR_L)|(overflow<<DSP_SR_V)|(carry<<DSP_SR_C);
+    return (overflow<<DSP_SR_L)|(v<<DSP_SR_V)|(carry<<DSP_SR_C);
 }
 
-static uint16_t dsp_asr56(uint32_t *dest)
+static uint16_t dsp_asr56(uint32_t *dest, int n)
 {
-    uint16_t carry;
+    /* Shift right dest n bits: D>>=n */
 
-    /* Shift right dest 1 bit: D>>=1 */
+    uint64_t dest_v = dest[2] | ((uint64_t)dest[1] << 24) | ((uint64_t)dest[0] << 48);
 
-    carry = dest[2] & 1;
-
-    dest[2] >>= 1;
-    dest[2] |= (dest[1] & 1)<<23;
-
-    dest[1] >>= 1;
-    dest[1] |= (dest[0] & 1)<<23;
-
-    dest[0] >>= 1;
-    dest[0] |= (dest[0] & (1<<6))<<1;
+    uint16_t carry = (dest_v >> (n-1)) & 1;
+    
+    dest_v >>= n;
+    dest[2] = dest_v & BITMASK(24);
+    dest[1] = (dest_v >> 24) & BITMASK(24);
+    dest[0] = (dest_v >> 48) & BITMASK(8);
 
     return (carry<<DSP_SR_C);
 }
@@ -1373,7 +1364,7 @@ static void dsp_mul56(uint32_t source1, uint32_t source2, uint32_t *dest, uint8_
     }
 
     /* Get rid of extra sign bit */
-    dsp_asl56(dest);
+    dsp_asl56(dest, 1);
 
     if (signe) {
         zerodest[0] = zerodest[1] = zerodest[2] = 0;
diff --git a/hw/xbox/dsp/dsp_dis.inl b/hw/xbox/dsp/dsp_dis.inl
index 801721d07e..50d330294c 100644
--- a/hw/xbox/dsp/dsp_dis.inl
+++ b/hw/xbox/dsp/dsp_dis.inl
@@ -260,6 +260,28 @@ static void dis_andi(dsp_core_t* dsp)
     }
 }
 
+static void dis_asl_imm(dsp_core_t* dsp)
+{
+    uint32_t S = (dsp->disasm_cur_inst >> 7) & 1;
+    uint32_t D = dsp->disasm_cur_inst & 1;
+    uint32_t ii = (dsp->disasm_cur_inst >> 1) & BITMASK(6);
+    sprintf(dsp->disasm_str_instr, "asl #$%02x,%s,%s",
+        ii,
+        registers_name[S ? DSP_REG_B : DSP_REG_A],
+        registers_name[D ? DSP_REG_B : DSP_REG_A]);
+}
+
+static void dis_asr_imm(dsp_core_t* dsp)
+{
+    uint32_t S = (dsp->disasm_cur_inst >> 7) & 1;
+    uint32_t D = dsp->disasm_cur_inst & 1;
+    uint32_t ii = (dsp->disasm_cur_inst >> 1) & BITMASK(6);
+    sprintf(dsp->disasm_str_instr, "asr #$%02x,%s,%s",
+        ii,
+        registers_name[S ? DSP_REG_B : DSP_REG_A],
+        registers_name[D ? DSP_REG_B : DSP_REG_A]);
+}
+
 static void dis_bcc_long(dsp_core_t* dsp) {
     dsp->disasm_cur_inst_len++;
 
@@ -428,6 +450,14 @@ static void dis_bclr_reg(dsp_core_t* dsp)
     sprintf(dsp->disasm_str_instr,"bclr #%d,%s", numbit, registers_name[value]);
 }
 
+static void dis_bra_long(dsp_core_t* dsp)
+{
+    uint32_t xxxx = read_memory_p(dsp, dsp->pc+1);
+    dsp->disasm_cur_inst_len++;
+    sprintf(dsp->disasm_str_instr, "bra p:$%06x",
+        (dsp->pc + xxxx) & BITMASK(24));
+}
+
 static void dis_bra_imm(dsp_core_t* dsp)
 {
     uint32_t xxx = (dsp->disasm_cur_inst & BITMASK(5))
@@ -1256,15 +1286,31 @@ static void dis_jsset_reg(dsp_core_t* dsp)
 
 static void dis_lua(dsp_core_t* dsp)
 {
-    char addr_name[16], numreg;
+    char addr_name[16];
 
     dis_calc_ea(dsp, (dsp->disasm_cur_inst>>8) & BITMASK(5), addr_name);
-    numreg = dsp->disasm_cur_inst & BITMASK(3);
+    uint32_t numreg = dsp->disasm_cur_inst & BITMASK(4);
     
-    if (dsp->disasm_cur_inst & (1<<3))
-        sprintf(dsp->disasm_str_instr,"lua %s,n%d", addr_name, numreg);
-    else
-        sprintf(dsp->disasm_str_instr,"lua %s,r%d", addr_name, numreg);
+    sprintf(dsp->disasm_str_instr,"lua %s,%s", addr_name, registers_name[numreg]);
+}
+
+static void dis_lua_rel(dsp_core_t* dsp)
+{
+    uint32_t aa = (dsp->disasm_cur_inst >> 4) & BITMASK(4)
+                + (((dsp->disasm_cur_inst >> 11) & BITMASK(3)) << 4);
+    uint32_t addrreg = (dsp->disasm_cur_inst >> 8) & BITMASK(3);
+    uint32_t dstreg = (dsp->disasm_cur_inst & BITMASK(3));
+
+    int32_t aa_s = (int32_t)dsp_signextend(7, aa);
+
+    if (dsp->disasm_cur_inst & (1<<3)) {
+        sprintf(dsp->disasm_str_instr,"lua (r%d + %d),n%d",
+            addrreg, aa_s, dstreg);
+    } else {
+        sprintf(dsp->disasm_str_instr,"lua (r%d + %d),r%d",
+            addrreg, aa_s, dstreg);
+    }
+
 }
 
 static void dis_movec_reg(dsp_core_t* dsp)
@@ -1599,31 +1645,46 @@ static void dis_move_x_long(dsp_core_t* dsp) {
     uint32_t xxxx = read_memory_p(dsp, dsp->pc+1);
     uint32_t offreg = DSP_REG_R0 + ((dsp->disasm_cur_inst >> 8) & BITMASK(3));
     uint32_t numreg = dsp->disasm_cur_inst & BITMASK(6);
+
+    int32_t xxxx_s = dsp_signextend(24, xxxx);
     if (W) {
         sprintf(dsp->disasm_str_instr, "move x:(%s + %d), %s",
-            registers_name[offreg], xxxx, registers_name[numreg]);
+            registers_name[offreg], xxxx_s, registers_name[numreg]);
     } else {
         sprintf(dsp->disasm_str_instr, "move %s, x:(%s + %d)",
-            registers_name[numreg], registers_name[offreg], xxxx);
+            registers_name[numreg], registers_name[offreg], xxxx_s);
     }
 }
 
-static void dis_move_x_imm(dsp_core_t* dsp) {
-    // 0000001aaaaaaRRR1a0WDDDD
+static void dis_move_xy_imm(dsp_core_t* dsp, int space)
+{
+    char space_c = space == DSP_SPACE_X ? 'x' : 'y';
+
     int W = (dsp->disasm_cur_inst >> 4) & 1;
     uint32_t xxx = (((dsp->disasm_cur_inst >> 11) & BITMASK(6)) << 1)
              + ((dsp->disasm_cur_inst >> 6) & 1);
     uint32_t offreg = DSP_REG_R0 + ((dsp->disasm_cur_inst >> 8) & BITMASK(3));
     uint32_t numreg = dsp->disasm_cur_inst & BITMASK(4);
+   
+    int32_t xxx_s = dsp_signextend(7, xxx);
     if (W) {
-        sprintf(dsp->disasm_str_instr, "move x:(%s + %d), %s",
-            registers_name[offreg], xxx, registers_name[numreg]);
+        sprintf(dsp->disasm_str_instr, "move %c:(%s + %d), %s",
+            space_c, registers_name[offreg], xxx_s, registers_name[numreg]);
     } else {
-        sprintf(dsp->disasm_str_instr, "move %s, x:(%s + %d)",
-            registers_name[numreg], registers_name[offreg], xxx);
+        sprintf(dsp->disasm_str_instr, "move %s, %c:(%s + %d)",
+            registers_name[numreg], space_c, registers_name[offreg], xxx_s);
     }
 }
 
+static void dis_move_x_imm(dsp_core_t* dsp)
+{
+    dis_move_xy_imm(dsp, DSP_SPACE_X);
+}
+
+static void dis_move_y_imm(dsp_core_t* dsp) {
+    dis_move_xy_imm(dsp, DSP_SPACE_Y);
+}
+
 static void dis_norm(dsp_core_t* dsp)
 {
     uint32_t srcreg, destreg;
@@ -1707,6 +1768,14 @@ static void dis_rep_reg(dsp_core_t* dsp)
     sprintf(dsp->disasm_str_instr,"rep %s", registers_name[(dsp->disasm_cur_inst>>8) & BITMASK(6)]);
 }
 
+static void dis_sub_imm(dsp_core_t* dsp)
+{
+    uint32_t xx = (dsp->disasm_cur_inst >> 8) & BITMASK(6);
+    uint32_t d = (dsp->disasm_cur_inst >> 3) & 1;
+    sprintf(dsp->disasm_str_instr, "sub #$%02x,%s",
+        xx, registers_name[d ? DSP_REG_B : DSP_REG_A]);
+}
+
 static void dis_tcc(dsp_core_t* dsp)
 {
     char ccname[16];
diff --git a/hw/xbox/dsp/dsp_emu.inl b/hw/xbox/dsp/dsp_emu.inl
index 180057f164..e2dfbc19e5 100644
--- a/hw/xbox/dsp/dsp_emu.inl
+++ b/hw/xbox/dsp/dsp_emu.inl
@@ -853,7 +853,7 @@ static void emu_addl_b_a(dsp_core_t* dsp)
     dest[0] = dsp->registers[DSP_REG_A2];
     dest[1] = dsp->registers[DSP_REG_A1];
     dest[2] = dsp->registers[DSP_REG_A0];
-    newsr = dsp_asl56(dest);
+    newsr = dsp_asl56(dest, 1);
 
     source[0] = dsp->registers[DSP_REG_B2];
     source[1] = dsp->registers[DSP_REG_B1];
@@ -878,7 +878,7 @@ static void emu_addl_a_b(dsp_core_t* dsp)
     dest[0] = dsp->registers[DSP_REG_B2];
     dest[1] = dsp->registers[DSP_REG_B1];
     dest[2] = dsp->registers[DSP_REG_B0];
-    newsr = dsp_asl56(dest);
+    newsr = dsp_asl56(dest, 1);
 
     source[0] = dsp->registers[DSP_REG_A2];
     source[1] = dsp->registers[DSP_REG_A1];
@@ -903,7 +903,7 @@ static void emu_addr_b_a(dsp_core_t* dsp)
     dest[0] = dsp->registers[DSP_REG_A2];
     dest[1] = dsp->registers[DSP_REG_A1];
     dest[2] = dsp->registers[DSP_REG_A0];
-    newsr = dsp_asr56(dest);
+    newsr = dsp_asr56(dest, 1);
 
     source[0] = dsp->registers[DSP_REG_B2];
     source[1] = dsp->registers[DSP_REG_B1];
@@ -928,7 +928,7 @@ static void emu_addr_a_b(dsp_core_t* dsp)
     dest[0] = dsp->registers[DSP_REG_B2];
     dest[1] = dsp->registers[DSP_REG_B1];
     dest[2] = dsp->registers[DSP_REG_B0];
-    newsr = dsp_asr56(dest);
+    newsr = dsp_asr56(dest, 1);
 
     source[0] = dsp->registers[DSP_REG_A2];
     source[1] = dsp->registers[DSP_REG_A1];
@@ -1026,7 +1026,7 @@ static void emu_asl_a(dsp_core_t* dsp)
     dest[1] = dsp->registers[DSP_REG_A1];
     dest[2] = dsp->registers[DSP_REG_A0];
 
-    newsr = dsp_asl56(dest);
+    newsr = dsp_asl56(dest, 1);
 
     dsp->registers[DSP_REG_A2] = dest[0];
     dsp->registers[DSP_REG_A1] = dest[1];
@@ -1047,7 +1047,7 @@ static void emu_asl_b(dsp_core_t* dsp)
     dest[1] = dsp->registers[DSP_REG_B1];
     dest[2] = dsp->registers[DSP_REG_B0];
 
-    newsr = dsp_asl56(dest);
+    newsr = dsp_asl56(dest, 1);
 
     dsp->registers[DSP_REG_B2] = dest[0];
     dsp->registers[DSP_REG_B1] = dest[1];
@@ -1068,7 +1068,7 @@ static void emu_asr_a(dsp_core_t* dsp)
     dest[1] = dsp->registers[DSP_REG_A1];
     dest[2] = dsp->registers[DSP_REG_A0];
 
-    newsr = dsp_asr56(dest);
+    newsr = dsp_asr56(dest, 1);
 
     dsp->registers[DSP_REG_A2] = dest[0];
     dsp->registers[DSP_REG_A1] = dest[1];
@@ -1089,7 +1089,7 @@ static void emu_asr_b(dsp_core_t* dsp)
     dest[1] = dsp->registers[DSP_REG_B1];
     dest[2] = dsp->registers[DSP_REG_B0];
 
-    newsr = dsp_asr56(dest);
+    newsr = dsp_asr56(dest, 1);
 
     dsp->registers[DSP_REG_B2] = dest[0];
     dsp->registers[DSP_REG_B1] = dest[1];
@@ -4825,7 +4825,7 @@ static void emu_subl_a(dsp_core_t* dsp)
     dest[0] = dsp->registers[DSP_REG_A2];
     dest[1] = dsp->registers[DSP_REG_A1];
     dest[2] = dsp->registers[DSP_REG_A0];
-    newsr = dsp_asl56(dest);
+    newsr = dsp_asl56(dest, 1);
 
     source[0] = dsp->registers[DSP_REG_B2];
     source[1] = dsp->registers[DSP_REG_B1];
@@ -4850,7 +4850,7 @@ static void emu_subl_b(dsp_core_t* dsp)
     dest[0] = dsp->registers[DSP_REG_B2];
     dest[1] = dsp->registers[DSP_REG_B1];
     dest[2] = dsp->registers[DSP_REG_B0];
-    newsr = dsp_asl56(dest);
+    newsr = dsp_asl56(dest, 1);
 
     source[0] = dsp->registers[DSP_REG_A2];
     source[1] = dsp->registers[DSP_REG_A1];
@@ -4876,7 +4876,7 @@ static void emu_subr_a(dsp_core_t* dsp)
     dest[1] = dsp->registers[DSP_REG_A1];
     dest[2] = dsp->registers[DSP_REG_A0];
     
-    newsr = dsp_asr56(dest);
+    newsr = dsp_asr56(dest, 1);
 
     source[0] = dsp->registers[DSP_REG_B2];
     source[1] = dsp->registers[DSP_REG_B1];
@@ -4903,7 +4903,7 @@ static void emu_subr_b(dsp_core_t* dsp)
     dest[1] = dsp->registers[DSP_REG_B1];
     dest[2] = dsp->registers[DSP_REG_B0];
     
-    newsr = dsp_asr56(dest);
+    newsr = dsp_asr56(dest, 1);
 
     source[0] = dsp->registers[DSP_REG_A2];
     source[1] = dsp->registers[DSP_REG_A1];
@@ -5767,7 +5767,7 @@ static void emu_add_imm(dsp_core_t* dsp)
     if ((dsp->cur_inst >> 3) & 1) {
         dsp->registers[DSP_REG_B2] = dest[0];
         dsp->registers[DSP_REG_B1] = dest[1];
-        dsp->registers[DSP_REG_B0] = dest[2];        
+        dsp->registers[DSP_REG_B0] = dest[2];
     } else {
         dsp->registers[DSP_REG_A2] = dest[0];
         dsp->registers[DSP_REG_A1] = dest[1];
@@ -5860,7 +5860,79 @@ static void emu_andi(dsp_core_t* dsp)
     }
 }
 
-static void emu_bcc_long(dsp_core_t* dsp) {
+static void emu_asl_imm(dsp_core_t* dsp)
+{
+    uint32_t S = ((dsp->cur_inst >> 7) & 1);
+    uint32_t D = dsp->cur_inst & 1;
+    uint32_t ii = (dsp->cur_inst >> 1) & BITMASK(6);
+
+    uint32_t dest[3];
+
+    if (S) {
+        dest[0] = dsp->registers[DSP_REG_B2];
+        dest[1] = dsp->registers[DSP_REG_B1];
+        dest[2] = dsp->registers[DSP_REG_B0];
+    } else {
+        dest[0] = dsp->registers[DSP_REG_A2];
+        dest[1] = dsp->registers[DSP_REG_A1];
+        dest[2] = dsp->registers[DSP_REG_A0];
+    }
+
+    uint16_t newsr = dsp_asl56(dest, ii);
+
+    if (D) {
+        dsp->registers[DSP_REG_B2] = dest[0];
+        dsp->registers[DSP_REG_B1] = dest[1];
+        dsp->registers[DSP_REG_B0] = dest[2];
+    } else {
+        dsp->registers[DSP_REG_A2] = dest[0];
+        dsp->registers[DSP_REG_A1] = dest[1];
+        dsp->registers[DSP_REG_A0] = dest[2];
+    }
+
+    dsp->registers[DSP_REG_SR] &= BITMASK(16)-((1<<DSP_SR_C)|(1<<DSP_SR_V));
+    dsp->registers[DSP_REG_SR] |= newsr;
+
+    emu_ccr_update_e_u_n_z(dsp, dest[0], dest[1], dest[2]);
+}
+
+static void emu_asr_imm(dsp_core_t* dsp)
+{
+    uint32_t S = ((dsp->cur_inst >> 7) & 1);
+    uint32_t D = dsp->cur_inst & 1;
+    uint32_t ii = (dsp->cur_inst >> 1) & BITMASK(6);
+
+    uint32_t dest[3];
+    if (S) {
+        dest[0] = dsp->registers[DSP_REG_B2];
+        dest[1] = dsp->registers[DSP_REG_B1];
+        dest[2] = dsp->registers[DSP_REG_B0];
+    } else {
+        dest[0] = dsp->registers[DSP_REG_A2];
+        dest[1] = dsp->registers[DSP_REG_A1];
+        dest[2] = dsp->registers[DSP_REG_A0];
+    }
+
+    uint16_t newsr = dsp_asr56(dest, ii);
+
+    if (D) {
+        dsp->registers[DSP_REG_B2] = dest[0];
+        dsp->registers[DSP_REG_B1] = dest[1];
+        dsp->registers[DSP_REG_B0] = dest[2];
+    } else {
+        dsp->registers[DSP_REG_A2] = dest[0];
+        dsp->registers[DSP_REG_A1] = dest[1];
+        dsp->registers[DSP_REG_A0] = dest[2];
+    }
+
+    dsp->registers[DSP_REG_SR] &= BITMASK(16)-((1<<DSP_SR_C)|(1<<DSP_SR_V));
+    dsp->registers[DSP_REG_SR] |= newsr;
+
+    emu_ccr_update_e_u_n_z(dsp, dest[0], dest[1], dest[2]);
+}
+
+static void emu_bcc_long(dsp_core_t* dsp)
+{
     uint32_t xxxx = read_memory_p(dsp, dsp->pc+1);
     dsp->cur_inst_len++;
 
@@ -5874,7 +5946,8 @@ static void emu_bcc_long(dsp_core_t* dsp) {
     //TODO: cycles?
 }
 
-static void emu_bcc_imm(dsp_core_t* dsp) {
+static void emu_bcc_imm(dsp_core_t* dsp)
+{
     uint32_t xxx = (dsp->cur_inst & BITMASK(5))
                     + ((dsp->cur_inst & (BITMASK(4) << 6)) >> 1);
 
@@ -6080,7 +6153,18 @@ static void emu_bclr_reg(dsp_core_t* dsp)
     dsp->instr_cycle += 2;
 }
 
-static void emu_bra_imm(dsp_core_t* dsp) {
+static void emu_bra_long(dsp_core_t* dsp)
+{
+    uint32_t xxxx = read_memory_p(dsp, dsp->pc+1);
+    dsp->cur_inst_len++;
+
+    dsp->pc += xxxx;
+    dsp->pc &= BITMASK(24);
+    dsp->cur_inst_len = 0;
+}
+
+static void emu_bra_imm(dsp_core_t* dsp)
+{
     uint32_t xxx = (dsp->cur_inst & BITMASK(5))
                     + ((dsp->cur_inst & (BITMASK(4) << 6)) >> 1);
 
@@ -6089,7 +6173,8 @@ static void emu_bra_imm(dsp_core_t* dsp) {
     dsp->cur_inst_len = 0;
 }
 
-static void emu_brclr_pp(dsp_core_t* dsp) {
+static void emu_brclr_pp(dsp_core_t* dsp)
+{
     uint32_t xxxx = read_memory_p(dsp, dsp->pc+1);
     dsp->cur_inst_len++;
 
@@ -6108,7 +6193,8 @@ static void emu_brclr_pp(dsp_core_t* dsp) {
     }
 }
 
-static void emu_brclr_reg(dsp_core_t* dsp) {
+static void emu_brclr_reg(dsp_core_t* dsp)
+{
     uint32_t xxxx = read_memory_p(dsp, dsp->pc+1);
     dsp->cur_inst_len++;
 
@@ -6131,7 +6217,8 @@ static void emu_brclr_reg(dsp_core_t* dsp) {
     }
 }
 
-static void emu_brset_pp(dsp_core_t* dsp) {
+static void emu_brset_pp(dsp_core_t* dsp)
+{
     uint32_t xxxx = read_memory_p(dsp, dsp->pc+1);
     dsp->cur_inst_len++;
 
@@ -6150,7 +6237,8 @@ static void emu_brset_pp(dsp_core_t* dsp) {
     }
 }
 
-static void emu_brset_reg(dsp_core_t* dsp) {
+static void emu_brset_reg(dsp_core_t* dsp)
+{
     uint32_t xxxx = read_memory_p(dsp, dsp->pc+1);
     dsp->cur_inst_len++;
 
@@ -6375,7 +6463,8 @@ static void emu_btst_reg(dsp_core_t* dsp)
     dsp->instr_cycle += 2;
 }
 
-static void emu_cmp_imm(dsp_core_t* dsp) {
+static void emu_cmp_imm(dsp_core_t* dsp)
+{
     uint32_t xx = (dsp->cur_inst >> 8) & BITMASK(6);
     uint32_t d = (dsp->cur_inst >> 3) & 1;
 
@@ -6403,7 +6492,8 @@ static void emu_cmp_imm(dsp_core_t* dsp) {
     dsp->registers[DSP_REG_SR] |= newsr;
 }
 
-static void emu_cmpu(dsp_core_t* dsp) {
+static void emu_cmpu(dsp_core_t* dsp)
+{
     uint32_t ggg = (dsp->cur_inst >> 1) & BITMASK(3);
     uint32_t d = dsp->cur_inst & 1;
 
@@ -6481,11 +6571,11 @@ static void emu_div(dsp_core_t* dsp)
 
     if (((dest[0]>>7) & 1) ^ ((source[1]>>23) & 1)) {
         /* D += S */
-        newsr = dsp_asl56(dest);
+        newsr = dsp_asl56(dest, 1);
         dsp_add56(source, dest);
     } else {
         /* D -= S */
-        newsr = dsp_asl56(dest);
+        newsr = dsp_asl56(dest, 1);
         dsp_sub56(source, dest);
     }
 
@@ -7144,6 +7234,8 @@ static void emu_lua(dsp_core_t* dsp)
 {
     uint32_t value, srcreg, dstreg, srcsave, srcnew;
 
+    // TODO: I don't think this is right
+
     srcreg = (dsp->cur_inst>>8) & BITMASK(3);
 
     srcsave = dsp->registers[DSP_REG_R0+srcreg];
@@ -7162,6 +7254,25 @@ static void emu_lua(dsp_core_t* dsp)
     dsp->instr_cycle += 2;
 }
 
+static void emu_lua_rel(dsp_core_t* dsp)
+{
+    uint32_t aa = (dsp->cur_inst >> 4) & BITMASK(4)
+                + (((dsp->cur_inst >> 11) & BITMASK(3)) << 4);
+    uint32_t addrreg = (dsp->cur_inst>>8) & BITMASK(3);
+    uint32_t dstreg = dsp->cur_inst & BITMASK(3);
+    
+    uint32_t v = (dsp->registers[DSP_REG_R0+addrreg]
+        + dsp_signextend(7, aa)) & BITMASK(24);
+
+    if (dsp->cur_inst & (1<<3)) {
+        dsp->registers[DSP_REG_N0+dstreg] = v;
+    } else {
+        dsp->registers[DSP_REG_R0+dstreg] = v;
+    }
+
+    dsp->instr_cycle += 2;
+}
+
 static void emu_movec_reg(dsp_core_t* dsp)
 {
     uint32_t numreg1, numreg2, value, dummy;
@@ -7446,7 +7557,8 @@ static void emu_movep_23(dsp_core_t* dsp)
     dsp->instr_cycle += 2;
 }
 
-static void emu_movep_x_qq(dsp_core_t* dsp) {
+static void emu_movep_x_qq(dsp_core_t* dsp)
+{
     // 00000111W1MMMRRR0sqqqqqq
 
     uint32_t x_addr = 0xffff80 + (dsp->cur_inst & BITMASK(6));
@@ -7473,7 +7585,8 @@ static void emu_movep_x_qq(dsp_core_t* dsp) {
     dsp->instr_cycle += 2;
 }
 
-static void emu_move_x_long(dsp_core_t* dsp) {
+static void emu_move_x_long(dsp_core_t* dsp)
+{
     uint32_t xxxx = read_memory_p(dsp, dsp->pc+1);
     dsp->cur_inst_len++;
     
@@ -7497,14 +7610,14 @@ static void emu_move_x_long(dsp_core_t* dsp) {
     // TODO: cycles
 }
 
-static void emu_move_x_imm(dsp_core_t* dsp) {
-    // 0000001aaaaaaRRR1a0WDDDD
+static void emu_move_xy_imm(dsp_core_t* dsp, int space)
+{
     uint32_t xxx = (((dsp->cur_inst >> 11) & BITMASK(6)) << 1)
              + ((dsp->cur_inst >> 6) & 1);
     int W = (dsp->cur_inst >> 4) & 1;
     uint32_t offreg = DSP_REG_R0 + ((dsp->cur_inst >> 8) & BITMASK(3));
     uint32_t numreg = dsp->cur_inst & BITMASK(4);
-    uint32_t x_addr = (dsp->registers[offreg] + dsp_signextend(7, xxx)) & BITMASK(24);
+    uint32_t addr = (dsp->registers[offreg] + dsp_signextend(7, xxx)) & BITMASK(24);
     
     if (!W) {
         uint32_t value;
@@ -7513,14 +7626,24 @@ static void emu_move_x_imm(dsp_core_t* dsp) {
         } else {
             value = dsp->registers[numreg];
         }
-        dsp56k_write_memory(dsp, DSP_SPACE_X, x_addr, value);
+        dsp56k_write_memory(dsp, space, addr, value);
     } else {
-        dsp_write_reg(dsp, numreg, dsp56k_read_memory(dsp, DSP_SPACE_X, x_addr));
+        dsp_write_reg(dsp, numreg, dsp56k_read_memory(dsp, space, addr));
     }
 
     // TODO: cycles
 }
 
+static void emu_move_x_imm(dsp_core_t* dsp)
+{
+    emu_move_xy_imm(dsp, DSP_SPACE_X);
+}
+
+static void emu_move_y_imm(dsp_core_t* dsp)
+{
+    emu_move_xy_imm(dsp, DSP_SPACE_Y);
+}
+
 static void emu_norm(dsp_core_t* dsp)
 {
     uint32_t cursr,cur_e, cur_euz, dest[3], numreg, rreg;
@@ -7540,11 +7663,11 @@ static void emu_norm(dsp_core_t* dsp)
     rreg = DSP_REG_R0+((dsp->cur_inst>>8) & BITMASK(3));
 
     if (cur_euz) {
-        newsr = dsp_asl56(dest);
+        newsr = dsp_asl56(dest, 1);
         --dsp->registers[rreg];
         dsp->registers[rreg] &= BITMASK(16);
     } else if (cur_e) {
-        newsr = dsp_asr56(dest);
+        newsr = dsp_asr56(dest, 1);
         ++dsp->registers[rreg];
         dsp->registers[rreg] &= BITMASK(16);
     } else {
@@ -7718,6 +7841,45 @@ static void emu_swi(dsp_core_t* dsp)
     dsp->instr_cycle += 6;
 }
 
+static void emu_sub_imm(dsp_core_t* dsp)
+{
+    uint32_t xx = (dsp->cur_inst >> 8) & BITMASK(6);
+    uint32_t d = (dsp->cur_inst >> 3) & 1;
+
+    uint32_t source[3], dest[3];
+
+    if (d) {
+        dest[0] = dsp->registers[DSP_REG_B2];
+        dest[1] = dsp->registers[DSP_REG_B1];
+        dest[2] = dsp->registers[DSP_REG_B0];
+    } else {
+        dest[0] = dsp->registers[DSP_REG_A2];
+        dest[1] = dsp->registers[DSP_REG_A1];
+        dest[2] = dsp->registers[DSP_REG_A0];
+    }
+
+    source[2] = 0;
+    source[1] = xx;
+    source[0] = 0;
+
+    uint16_t newsr = dsp_sub56(source, dest);
+
+    if (d) {
+        dsp->registers[DSP_REG_B2] = dest[0];
+        dsp->registers[DSP_REG_B1] = dest[1];
+        dsp->registers[DSP_REG_B0] = dest[2];
+    } else {
+        dsp->registers[DSP_REG_A2] = dest[0];
+        dsp->registers[DSP_REG_A1] = dest[1];
+        dsp->registers[DSP_REG_A0] = dest[2];
+    }
+
+    emu_ccr_update_e_u_n_z(dsp, dest[0], dest[1], dest[2]);
+
+    dsp->registers[DSP_REG_SR] &= BITMASK(16)-((1<<DSP_SR_V)|(1<<DSP_SR_C));
+    dsp->registers[DSP_REG_SR] |= newsr;
+}
+
 static void emu_tcc(dsp_core_t* dsp)
 {
     uint32_t cc_code, regsrc1, regdest1;