diff --git a/src/mem.cpp b/src/mem.cpp
index 9b6e5195..a7434430 100644
--- a/src/mem.cpp
+++ b/src/mem.cpp
@@ -43,7 +43,7 @@ MemBuffer::MemBuffer(unsigned size=0) :
 
 MemBuffer::~MemBuffer()
 {
-    free();
+    this->free();
 }
 
 
@@ -68,8 +68,11 @@ unsigned MemBuffer::getSize() const
 
 void MemBuffer::alloc(unsigned size, unsigned base_offset)
 {
+#if 0
+    // don't automaticlly free a used buffer
+    this->free();
+#endif
     assert(alloc_ptr == NULL);
-    //free();
     assert((int)size > 0);
     size = base_offset + size;
     alloc_ptr = (unsigned char *) malloc(size);
@@ -94,7 +97,7 @@ void MemBuffer::allocForCompression(unsigned uncompressed_size)
     // Idea:
     //   We allocate the buffer at an offset of 4096 so
     //   that we could do an in-place decompression for
-    //   verifying our overlap overhead at the end
+    //   verifying our overlap_overhead at the end
     //   of packing.
     //
     // See Packer::verifyOverlappingDecompression().
@@ -105,7 +108,8 @@ void MemBuffer::allocForCompression(unsigned uncompressed_size)
 
 void MemBuffer::allocForUncompression(unsigned uncompressed_size)
 {
-    alloc(uncompressed_size + 512, 0);  // 512 safety bytes
+    //alloc(uncompressed_size + 512, 0);  // 512 safety bytes
+    alloc(uncompressed_size + 3, 0);  // 3 safety bytes for asm_fast
 }
 
 
diff --git a/src/p_djgpp2.cpp b/src/p_djgpp2.cpp
index f0727c92..580200fd 100644
--- a/src/p_djgpp2.cpp
+++ b/src/p_djgpp2.cpp
@@ -111,7 +111,7 @@ void PackDjgpp2::handleStub(OutputFile *fo)
 }
 
 
-static bool is_dlm(InputFile *fi,long coff_offset)
+static bool is_dlm(InputFile *fi, long coff_offset)
 {
     unsigned char buf[4];
     long off;
@@ -132,7 +132,7 @@ static bool is_dlm(InputFile *fi,long coff_offset)
 }
 
 
-static void handle_allegropak(InputFile *fi,OutputFile *fo)
+static void handle_allegropak(InputFile *fi, OutputFile *fo)
 {
     unsigned char buf[0x4000];
     unsigned pfsize=0, ic;
@@ -285,6 +285,7 @@ void PackDjgpp2::pack(OutputFile *fo)
     ft.verifyUnfilter();
 #else
     // new version using compressWithFilters()
+
     // prepare packheader
     ph.u_len = usize;
     ph.filter = 0;
diff --git a/src/p_exe.cpp b/src/p_exe.cpp
index aba5ff9a..499fd773 100644
--- a/src/p_exe.cpp
+++ b/src/p_exe.cpp
@@ -34,7 +34,7 @@
 static const
 #include "stub/l_exe.h"
 
-#define RSFCRI          4096    // reserved space for compressed relocation info
+#define RSFCRI          4096    // Reserved Space For Compressed Relocation Info
 #define MAXMATCH        0x2000
 #define MAXRELOCS       (0x8000-MAXMATCH)
 
@@ -120,7 +120,7 @@ bool PackExe::canPack()
 static
 unsigned optimize_relocs(upx_byte *b, const unsigned size,
                          const upx_byte *relocs, const unsigned nrelocs,
-                         upx_byte *crel,bool *has_9a)
+                         upx_byte *crel, bool *has_9a)
 {
     upx_byte *crel_save = crel;
     unsigned i;
@@ -270,7 +270,7 @@ void PackExe::pack(OutputFile *fo)
     fi->seek(ih.headsize16*16,SEEK_SET);
     fi->readx(ibuf,imagesize);
 
-    if (find_le32(ibuf,imagesize < 127 ? imagesize : 127,UPX_MAGIC_LE32))
+    if (find_le32(ibuf,imagesize < 127 ? imagesize : 127, UPX_MAGIC_LE32))
         throwAlreadyPacked();
 
     // relocations
@@ -308,7 +308,7 @@ void PackExe::pack(OutputFile *fo)
         relocsize = 0;
     }
 
-    ph.u_len = imagesize+relocsize;
+    ph.u_len = imagesize + relocsize;
     if (!compress(ibuf,obuf,0,MAXMATCH))
         throwNotCompressible();
     const unsigned overlapoh = findOverlapOverhead(obuf,32);
@@ -452,11 +452,13 @@ void PackExe::pack(OutputFile *fo)
     }
 
     putPackHeader(loader,lsize);
-    upx_bytep p = find_le32(loader,lsize,get_le32("IPCS"));
-    if (p == NULL && (flag & USEJUMP))
-        throwBadLoader();
     if (flag & USEJUMP)
+    {
+        upx_bytep p = find_le32(loader,lsize,get_le32("IPCS"));
+        if (p == NULL)
+            throwBadLoader();
         memcpy(p,&ih.ip,4);
+    }
     else
     {
         patch_le16(loader,lsize,"IP",ih.ip);
@@ -495,6 +497,7 @@ void PackExe::pack(OutputFile *fo)
     oh.p512 = (outputlen + 511) >> 9;
 
 //fprintf(stderr,"\ne_len=%x d_len=%x clen=%x oo=%x ulen=%x destp=%x copys=%x images=%x",e_len,d_len,packedsize,overlapoh,ph.u_len,destpara,copysize,imagesize);
+
     // write header + write loader + compressed file
 #ifdef TESTING
     if (opt->debug)
diff --git a/src/stdcxx.cpp b/src/stdcxx.cpp
index 7219db85..b5aeeacc 100644
--- a/src/stdcxx.cpp
+++ b/src/stdcxx.cpp
@@ -33,9 +33,11 @@
 #ifdef WANT_STL
 
 #if defined(__DJGPP__) || defined(__MINGW32__) || defined(__sparc__)
+// provide missing oom_handler
 void (*__malloc_alloc_template<0>::__malloc_alloc_oom_handler)() = 0;
 # if !defined(__USE_MALLOC)
-template class  __default_alloc_template<false, 0>;
+// instantiate default allocator
+template class __default_alloc_template<false, 0>;
 # endif
 #endif
 
diff --git a/src/unupx.h b/src/unupx.h
index a69271d9..c221c4dc 100644
--- a/src/unupx.h
+++ b/src/unupx.h
@@ -71,7 +71,7 @@ typedef unsigned upx_uint32;
 #define UPX_F_DJGPP2_COFF       4
 #define UPX_F_WC_LE             5
 #define UPX_F_VXD_LE            6
-#define UPX_F_DOS_EXEH          7       /* OBSOLETE */
+#define UPX_F_DOS_EXEH          7               /* OBSOLETE */
 #define UPX_F_TMT_ADAM          8
 #define UPX_F_W32_PE            9
 #define UPX_F_LINUX_i386        10