diff --git a/src-md/crt0.s b/src-md/crt0.s
index 0027d30c3..f855ee95c 100644
--- a/src-md/crt0.s
+++ b/src-md/crt0.s
@@ -2597,19 +2597,25 @@ bump_fm:
 13:
         /* EOF reached */
         move.l  fm_loop,d0
+        andi.w  #0x01FF,d0          /* 1 read buffer */
         add.w   d3,d0
         andi.w  #0x0FFF,d0          /* 4KB buffer */
-        move.l  d0,fm_loop2         /* offset to z80 loop start */
+        move.w  d0,fm_loop2         /* offset to z80 loop start */
 
         /* reset */
         jsr     vgm_reset           /* restart at start of compressed data */
-        move.w  #0,offs68k
-        move.w  d3,offsz80
-        move.w  #7,preread_cnt      /* refill buffer if room */
-        bra.w   7f
+        move.l  fm_loop,d2
+        andi.l  #0xFFFFFE00,d2
+        beq.b   14f
+        move.l  d2,-(sp)
+        jsr     vgm_read2
+        addq.l  #4,sp
+        move.w  d0,d2
+        andi.w  #0x7FFF,d2          /* amount data pre-read (with wrap around) */
 14:
         move.w  d2,offs68k
         move.w  d3,offsz80
+        move.w  #7,preread_cnt      /* refill buffer if room */
         bra.w   7f
 
 5:
@@ -2625,9 +2631,13 @@ bump_fm:
 
         /* music ended, check for loop */
         tst.w   fm_rep
-        beq.w   7f                  /* no repeat, leave FM_IDX as 0 */
+        bne.b   50f                 /* repeat, loop vgm */
+        /* no repeat, reset FM */
+        jsr     rst_ym2612
+        bra.w   7f
 
-        move.l  fm_loop2,d0
+50:
+        move.w  fm_loop2,d0
         z80wr   FM_START,d0
         lsr.w   #8,d0
         z80wr   FM_START+1,d0       /* music start = loop offset */
@@ -3154,9 +3164,6 @@ fm_start:
         dc.l    0
         .global    fm_loop
 fm_loop:
-        dc.l    0
-        .global    fm_loop2
-fm_loop2:
         dc.l    0
         .global pcm_baseoffs
 pcm_baseoffs:
@@ -3180,6 +3187,8 @@ fm_stream:
         dc.w    0
 fm_stream_freq:
         dc.w    0
+fm_loop2:
+        dc.w    0
 
 dac_freq:
         .global    dac_freq
diff --git a/src-md/scd_fs.c b/src-md/scd_fs.c
index 34cb78ac0..3df369d84 100644
--- a/src-md/scd_fs.c
+++ b/src-md/scd_fs.c
@@ -11,8 +11,9 @@
 #define CHUNK_BLOCKS (1<<CHUNK_SHIFT) 
 #define CHUNK_SIZE (CHUNK_BLOCKS*BLOCK_SIZE)
 
-#define MCD_DISC_BUFFER (void*)(MCD_WORDRAM + 0x20000 - CHUNK_SIZE)
-#define MD_DISC_BUFFER (void*)(MD_WORDRAM + 0x20000 - CHUNK_SIZE)
+#define DISC_BUFFER_WORDRAM_OFS (0x20000 - CHUNK_SIZE)
+#define MCD_DISC_BUFFER (void*)(MCD_WORDRAM + DISC_BUFFER_WORDRAM_OFS)
+#define MD_DISC_BUFFER (void*)(MD_WORDRAM + DISC_BUFFER_WORDRAM_OFS)
 
 extern int64_t scd_open_file(const char *name);
 extern void scd_read_sectors(void *ptr, int lba, int len, void (*wait)(void));
diff --git a/src-md/vgm.c b/src-md/vgm.c
index f9501ce4c..1d68b1f6e 100644
--- a/src-md/vgm.c
+++ b/src-md/vgm.c
@@ -12,7 +12,7 @@
 #define VGM_WORDRAM_OFS     0x3000
 #define VGM_READAHEAD       0x200
 #define VGM_LZSS_BUF_SIZE   0x8000
-#define VGM_MAX_SIZE        0x18000
+#define VGM_MAX_SIZE        0x18800
 
 #define VGM_USE_PWM_FOR_DAC
 
@@ -48,6 +48,7 @@ void lzss_reset(lzss_state_t* lzss) __attribute__((section(".data"), aligned(16)
 
 int vgm_setup(void* fm_ptr) __attribute__((section(".data"), aligned(16)));
 int vgm_read(void) __attribute__((section(".data"), aligned(16)));
+int vgm_read2(int length) __attribute__((section(".data"), aligned(16)));
 
 extern int64_t scd_open_file(const char *name);
 extern void scd_read_sectors(void *ptr, int lba, int len, void (*wait)(void));
@@ -161,6 +162,11 @@ int vgm_read(void)
     return lzss_read(&vgm_lzss, VGM_READAHEAD);
 }
 
+int vgm_read2(int length)
+{
+    return lzss_read(&vgm_lzss, length);
+}
+
 void *vgm_cache_scd(const char *name, int offset, int length)
 {
     int64_t lo;