diff --git a/atomic.h b/atomic.h
index b3f7841..aa4d977 100644
--- a/atomic.h
+++ b/atomic.h
@@ -62,6 +62,10 @@ typedef _Atomic long FL2_atomic;
 
 #else  /* No atomics */
 
+#	ifndef FL2_SINGLETHREAD
+#		error  No atomic operations available. Change compiler config or define FL2_SINGLETHREAD for the entire build.
+#	endif
+
 typedef long FL2_atomic;
 #define ATOMIC_INITIAL_VALUE 0
 #define FL2_atomic_increment(n) (n++)
diff --git a/fast-lzma2.h b/fast-lzma2.h
index a0c9b93..44ffd92 100644
--- a/fast-lzma2.h
+++ b/fast-lzma2.h
@@ -55,7 +55,7 @@ Introduction
 /*------   Version   ------*/
 #define FL2_VERSION_MAJOR    0
 #define FL2_VERSION_MINOR    9
-#define FL2_VERSION_RELEASE  1
+#define FL2_VERSION_RELEASE  2
 
 #define FL2_VERSION_NUMBER  (FL2_VERSION_MAJOR *100*100 + FL2_VERSION_MINOR *100 + FL2_VERSION_RELEASE)
 FL2LIB_API unsigned FL2LIB_CALL FL2_versionNumber(void);   /**< useful to check dll version */
@@ -156,7 +156,13 @@ typedef struct {
 
 typedef int (FL2LIB_CALL *FL2_progressFn)(size_t done, void* opaque);
 
+/* Get the size of the overlap section. */
+FL2LIB_API size_t FL2LIB_CALL FL2_blockOverlap(const FL2_CCtx* ctx);
+
+/* Copy the overlap section to the start to prepare for more data */
 FL2LIB_API void FL2LIB_CALL FL2_shiftBlock(const FL2_CCtx* ctx, FL2_blockBuffer *block);
+/* Copy the overlap to a different buffer. This allows a dual-buffer configuration where
+ * data is read into one block while the other is compressed. */
 FL2LIB_API void FL2LIB_CALL FL2_shiftBlock_switch(const FL2_CCtx* ctx, FL2_blockBuffer *block, unsigned char *dst);
 
 FL2LIB_API void FL2LIB_CALL FL2_beginFrame(FL2_CCtx* const cctx);
diff --git a/fl2_compress.c b/fl2_compress.c
index 57fbb94..2825b25 100644
--- a/fl2_compress.c
+++ b/fl2_compress.c
@@ -12,6 +12,7 @@
 #include <string.h>
 #include "fast-lzma2.h"
 #include "fl2_internal.h"
+#include "platform.h"
 #include "mem.h"
 #include "util.h"
 #include "fl2_compress_internal.h"
@@ -22,6 +23,12 @@
 
 #define MIN_BYTES_PER_THREAD 0x10000
 
+#ifdef __64BIT__
+#define ALIGNMENT_MASK ~((size_t)7)
+#else
+#define ALIGNMENT_MASK ~((size_t)3)
+#endif
+
 /*-=====  Pre-defined compression levels  =====-*/
 
 #define FL2_CLEVEL_DEFAULT   9
@@ -443,16 +450,25 @@ FL2LIB_API size_t FL2LIB_CALL FL2_compressCCtx(FL2_CCtx* cctx,
     return dstBuf - (BYTE*)dst;
 }
 
+FL2LIB_API size_t FL2LIB_CALL FL2_blockOverlap(const FL2_CCtx* cctx)
+{
+	return OVERLAP_FROM_DICT_LOG(cctx->params.rParams.dictionary_log, cctx->params.rParams.overlap_fraction);
+}
+
 FL2LIB_API void FL2LIB_CALL FL2_shiftBlock(const FL2_CCtx* cctx, FL2_blockBuffer *block)
 {
     FL2_shiftBlock_switch(cctx, block, NULL);
 }
 
-FL2LIB_API void FL2LIB_CALL FL2_shiftBlock_switch(const FL2_CCtx* cctx, FL2_blockBuffer *block, BYTE *dst)
+FL2LIB_API void FL2LIB_CALL FL2_shiftBlock_switch(const FL2_CCtx* cctx, FL2_blockBuffer *block, unsigned char *dst)
 {
     size_t const block_overlap = OVERLAP_FROM_DICT_LOG(cctx->params.rParams.dictionary_log, cctx->params.rParams.overlap_fraction);
 
-    if (block->end > block_overlap) {
+	if (block_overlap == 0) {
+		block->start = 0;
+		block->end = 0;
+	}
+	else if (block->end > block_overlap) {
         size_t const from = (block->end - block_overlap) & ~(sizeof(size_t) - 1);
         size_t const overlap = block->end - from;
 
@@ -460,7 +476,7 @@ FL2LIB_API void FL2LIB_CALL FL2_shiftBlock_switch(const FL2_CCtx* cctx, FL2_bloc
             DEBUGLOG(5, "Copy overlap data : %u bytes", (U32)overlap);
             memcpy(dst ? dst : block->data, block->data + from, overlap);
         }
-        else {
+		else if (from != 0) {
             DEBUGLOG(5, "Move overlap data : %u bytes", (U32)overlap);
             memmove(block->data, block->data + from, overlap);
         }
diff --git a/radix_mf.c b/radix_mf.c
index 5fa40df..c722c7b 100644
--- a/radix_mf.c
+++ b/radix_mf.c
@@ -153,6 +153,7 @@ FL2_matchTable* RMF_createMatchTable(const RMF_parameters* const p, size_t const
 {
     int isStruct;
     size_t dictionary_size;
+	size_t table_bytes;
     FL2_matchTable* tbl;
     RMF_parameters params = RMF_clampParams(*p);
 
@@ -162,8 +163,10 @@ FL2_matchTable* RMF_createMatchTable(const RMF_parameters* const p, size_t const
 
     DEBUGLOG(3, "RMF_createMatchTable : isStruct %d, dict %u", isStruct, (U32)dictionary_size);
 
+	table_bytes = isStruct ? ((dictionary_size + 3U) / 4U) * sizeof(RMF_unit)
+		: dictionary_size * sizeof(U32);
     tbl = (FL2_matchTable*)malloc(
-        sizeof(FL2_matchTable) + (isStruct ? dictionary_size * sizeof(RMF_unit) : dictionary_size * sizeof(U32)) - sizeof(U32));
+        sizeof(FL2_matchTable) + table_bytes - sizeof(U32));
     if (!tbl) return NULL;
 
     tbl->isStruct = isStruct;