diff --git a/intercept_tcmalloc.patch b/intercept_tcmalloc.patch
index a485560..fb4ccc3 100644
--- a/intercept_tcmalloc.patch
+++ b/intercept_tcmalloc.patch
@@ -2,7 +2,7 @@
 ===================================================================
 --- coregrind/m_replacemalloc/vg_replace_malloc.c	(revision 11055)
 +++ coregrind/m_replacemalloc/vg_replace_malloc.c	(working copy)
-@@ -234,6 +234,13 @@
+@@ -234,6 +234,23 @@
  // malloc
  ALLOC_or_NULL(VG_Z_LIBSTDCXX_SONAME, malloc,      malloc);
  ALLOC_or_NULL(VG_Z_LIBC_SONAME,      malloc,      malloc);
@@ -10,154 +10,298 @@
 +// Similar interceptors are added below to handle other libtcmalloc
 +// allocation/deallocation functions.
 +// soname=NONE means that a user's allocation function is intercepted.
++ALLOC_or_NULL(NONE,                  malloc,      malloc);
++// Handle libtcmalloc's malloc() function.
++// Similar interceptors are added below to handle other libtcmalloc
++// allocation/deallocation functions.
++// soname=NONE means that a user's allocation function is intercepted.
 +ALLOC_or_NULL(NONE,                  tc_malloc,      malloc);
++// Bash has sh_malloc() and sh_free() along with standard malloc() and free().
++// Sometimes these functions are called inconsistently (e.g. free() after
++// sh_malloc()). To deal with this we have to intercept the sh_*() functions
++// as well.
++ALLOC_or_NULL(NONE,                  sh_malloc,      malloc);
 +// Handle Python's malloc.
 +ALLOC_or_NULL(NONE,                  PyObject_Malloc,      malloc);
  #if defined(VGP_ppc32_aix5) || defined(VGP_ppc64_aix5)
  ALLOC_or_NULL(VG_Z_LIBC_SONAME,      malloc_common, malloc);
  #elif defined(VGO_darwin)
-@@ -267,6 +274,8 @@
+@@ -246,20 +263,24 @@
+ // operator new(unsigned int), not mangled (for gcc 2.96)
+ ALLOC_or_BOMB(VG_Z_LIBSTDCXX_SONAME,  builtin_new,    __builtin_new);
+ ALLOC_or_BOMB(VG_Z_LIBC_SONAME,       builtin_new,    __builtin_new);
++ALLOC_or_BOMB(NONE,                   builtin_new,    __builtin_new);
+ 
+ ALLOC_or_BOMB(VG_Z_LIBSTDCXX_SONAME,  __builtin_new,  __builtin_new);
+ ALLOC_or_BOMB(VG_Z_LIBC_SONAME,       __builtin_new,  __builtin_new);
++ALLOC_or_BOMB(NONE,                   __builtin_new,  __builtin_new);
+ 
+ // operator new(unsigned int), GNU mangling
+ #if VG_WORDSIZE == 4
+  ALLOC_or_BOMB(VG_Z_LIBSTDCXX_SONAME, _Znwj,          __builtin_new);
+  ALLOC_or_BOMB(VG_Z_LIBC_SONAME,      _Znwj,          __builtin_new);
++ ALLOC_or_BOMB(NONE,                  _Znwj,          __builtin_new);
+ #endif
+ 
+ // operator new(unsigned long), GNU mangling
+ #if VG_WORDSIZE == 8 || defined(VGP_ppc32_aix5) || defined(VGO_darwin)
+  ALLOC_or_BOMB(VG_Z_LIBSTDCXX_SONAME, _Znwm,          __builtin_new);
+  ALLOC_or_BOMB(VG_Z_LIBC_SONAME,      _Znwm,          __builtin_new);
++ ALLOC_or_BOMB(NONE,                  _Znwm,          __builtin_new);
+ #endif
+ 
+ // operator new(unsigned long), ARM/cfront mangling
+@@ -267,19 +288,24 @@
   ALLOC_or_BOMB(VG_Z_LIBC_DOT_A,       __nw__FUl,      __builtin_new);
  #endif
  
 +// libtcmalloc's new
 +ALLOC_or_BOMB(NONE, tc_new, __builtin_new);
  
++
  /*---------------------- new nothrow ----------------------*/
  
-@@ -287,6 +296,8 @@
+ // operator new(unsigned, std::nothrow_t const&), GNU mangling
+ #if VG_WORDSIZE == 4
+  ALLOC_or_NULL(VG_Z_LIBSTDCXX_SONAME, _ZnwjRKSt9nothrow_t,  __builtin_new);
+  ALLOC_or_NULL(VG_Z_LIBC_SONAME,      _ZnwjRKSt9nothrow_t,  __builtin_new);
++ ALLOC_or_NULL(NONE,                  _ZnwjRKSt9nothrow_t,  __builtin_new);
+ #endif
+ 
+ // operator new(unsigned long, std::nothrow_t const&), GNU mangling
+ #if VG_WORDSIZE == 8 || defined(VGP_ppc32_aix5) || defined(VGP_ppc64_aix5) || defined(VGO_darwin)
+  ALLOC_or_NULL(VG_Z_LIBSTDCXX_SONAME, _ZnwmRKSt9nothrow_t,  __builtin_new);
+  ALLOC_or_NULL(VG_Z_LIBC_SONAME,      _ZnwmRKSt9nothrow_t,  __builtin_new);
++ ALLOC_or_NULL(NONE,                  _ZnwmRKSt9nothrow_t,  __builtin_new);
+ #endif
+ 
+ // operator new(unsigned long, std::nothrow_t const&), ARM/cfront mangling
+@@ -287,23 +313,29 @@
   ALLOC_or_NULL(VG_Z_LIBC_DOT_A,    __nw__FUlRCQ2_3std9nothrow_t, __builtin_new);
  #endif
  
 +// libtcmalloc's new nothrow
 +ALLOC_or_NULL(NONE, tc_new_nothrow, __builtin_new);
  
++
  /*---------------------- new [] ----------------------*/
  
-@@ -311,6 +322,8 @@
+ // operator new[](unsigned int), not mangled (for gcc 2.96)
+ ALLOC_or_BOMB(VG_Z_LIBSTDCXX_SONAME,  __builtin_vec_new, __builtin_vec_new );
+ ALLOC_or_BOMB(VG_Z_LIBC_SONAME,       __builtin_vec_new, __builtin_vec_new );
++ALLOC_or_BOMB(NONE,                   __builtin_vec_new, __builtin_vec_new );
+ 
+ // operator new[](unsigned int), GNU mangling
+ #if VG_WORDSIZE == 4
+  ALLOC_or_BOMB(VG_Z_LIBSTDCXX_SONAME, _Znaj,             __builtin_vec_new );
+  ALLOC_or_BOMB(VG_Z_LIBC_SONAME,      _Znaj,             __builtin_vec_new );
++ ALLOC_or_BOMB(NONE,                  _Znaj,             __builtin_vec_new );
+ #endif
+ 
+ // operator new[](unsigned long), GNU mangling
+ #if VG_WORDSIZE == 8 || defined(VGP_ppc32_aix5) || defined(VGP_ppc64_aix5) || defined(VGO_darwin)
+  ALLOC_or_BOMB(VG_Z_LIBSTDCXX_SONAME, _Znam,             __builtin_vec_new );
+  ALLOC_or_BOMB(VG_Z_LIBC_SONAME,      _Znam,             __builtin_vec_new );
++ ALLOC_or_BOMB(NONE,                  _Znam,             __builtin_vec_new );
+ #endif
+ 
+ // operator new[](unsigned long), ARM/cfront mangling
+@@ -311,19 +343,24 @@
   ALLOC_or_BOMB(VG_Z_LIBC_DOT_A,       __vn__FUl,         __builtin_vec_new);
  #endif
  
 +// libtcmalloc's new []
 +ALLOC_or_BOMB(NONE, tc_newarray, __builtin_vec_new);
  
++
  /*---------------------- new [] nothrow ----------------------*/
  
-@@ -331,6 +344,8 @@
+ // operator new[](unsigned, std::nothrow_t const&), GNU mangling
+ #if VG_WORDSIZE == 4
+  ALLOC_or_NULL(VG_Z_LIBSTDCXX_SONAME, _ZnajRKSt9nothrow_t, __builtin_vec_new );
+  ALLOC_or_NULL(VG_Z_LIBC_SONAME,      _ZnajRKSt9nothrow_t, __builtin_vec_new );
++ ALLOC_or_NULL(NONE,                  _ZnajRKSt9nothrow_t, __builtin_vec_new );
+ #endif
+ 
+ // operator new[](unsigned long, std::nothrow_t const&), GNU mangling
+ #if VG_WORDSIZE == 8 || defined(VGP_ppc32_aix5) || defined(VGP_ppc64_aix5) || defined(VGO_darwin)
+  ALLOC_or_NULL(VG_Z_LIBSTDCXX_SONAME, _ZnamRKSt9nothrow_t, __builtin_vec_new );
+  ALLOC_or_NULL(VG_Z_LIBC_SONAME,      _ZnamRKSt9nothrow_t, __builtin_vec_new );
++ ALLOC_or_NULL(NONE,                  _ZnamRKSt9nothrow_t, __builtin_vec_new );
+ #endif
+ 
+ // operator new [](unsigned long, std::nothrow_t const&), ARM/cfront mangling
+@@ -331,7 +368,10 @@
   ALLOC_or_BOMB(VG_Z_LIBC_DOT_A,   __vn__FUlRCQ2_3std9nothrow_t, __builtin_vec_new );
  #endif
  
 +// libtcmalloc's new [] nothrow
 +ALLOC_or_NULL(NONE, tc_newarray_nothrow, __builtin_vec_new);
  
++
  /*---------------------- free ----------------------*/
  
-@@ -364,6 +379,8 @@
+ /* Generate a replacement for 'fnname' in object 'soname', which calls
+@@ -364,6 +404,10 @@
  // free
  FREE(VG_Z_LIBSTDCXX_SONAME,  free,                 free );
  FREE(VG_Z_LIBC_SONAME,       free,                 free );
++FREE(NONE,                   free,                 free );
++FREE(NONE,                   sh_free,              free );
 +FREE(NONE,                   tc_free,                 free );
 +FREE(NONE,                   PyObject_Free,        free );
  #if defined(VGP_ppc32_aix5) || defined(VGP_ppc64_aix5)
  FREE(VG_Z_LIBC_SONAME,       free_common,          free );
  #elif defined(VGO_darwin)
-@@ -376,6 +393,7 @@
+@@ -376,43 +420,57 @@
  // cfree
  FREE(VG_Z_LIBSTDCXX_SONAME,  cfree,                free );
  FREE(VG_Z_LIBC_SONAME,       cfree,                free );
++FREE(NONE,                   cfree,                free );
 +FREE(NONE,                   tc_cfree,                free );
++FREE(NONE,                   sh_cfree,             free );
  
  
  /*---------------------- delete ----------------------*/
-@@ -392,6 +410,8 @@
+ // operator delete(void*), not mangled (for gcc 2.96)
+ FREE(VG_Z_LIBSTDCXX_SONAME,   __builtin_delete,     __builtin_delete );
+ FREE(VG_Z_LIBC_SONAME,        __builtin_delete,     __builtin_delete );
++FREE(NONE,                    __builtin_delete,     __builtin_delete );
+ 
+ // operator delete(void*), GNU mangling
+ FREE(VG_Z_LIBSTDCXX_SONAME,  _ZdlPv,               __builtin_delete );
+ FREE(VG_Z_LIBC_SONAME,       _ZdlPv,               __builtin_delete );
++FREE(NONE,                   _ZdlPv,               __builtin_delete );
+ 
+ // operator delete(void*), ARM/cfront mangling
+ #if defined(VGP_ppc32_aix5) || defined(VGP_ppc64_aix5)
  FREE(VG_Z_LIBC_DOT_A,        __dl__FPv,            __builtin_delete );
  #endif
  
 +// libtcmalloc's delete
 +FREE(NONE, tc_delete, __builtin_delete);
  
++
  /*---------------------- delete nothrow ----------------------*/
  
-@@ -399,6 +419,8 @@
+ // operator delete(void*, std::nothrow_t const&), GNU mangling
  FREE(VG_Z_LIBSTDCXX_SONAME, _ZdlPvRKSt9nothrow_t,  __builtin_delete );
  FREE(VG_Z_LIBC_SONAME,      _ZdlPvRKSt9nothrow_t,  __builtin_delete );
- 
++FREE(NONE,                  _ZdlPvRKSt9nothrow_t,  __builtin_delete );
 +// libtcmalloc's delete nothrow
 +FREE(NONE, tc_delete_nothrow, __builtin_delete);
  
+-
  /*---------------------- delete [] ----------------------*/
  // operator delete[](void*), not mangled (for gcc 2.96)
-@@ -414,6 +436,8 @@
+ FREE(VG_Z_LIBSTDCXX_SONAME,   __builtin_vec_delete, __builtin_vec_delete );
+ FREE(VG_Z_LIBC_SONAME,        __builtin_vec_delete, __builtin_vec_delete );
++FREE(NONE,                    __builtin_vec_delete, __builtin_vec_delete );
+ 
+ // operator delete[](void*), GNU mangling
+ FREE(VG_Z_LIBSTDCXX_SONAME,  _ZdaPv,               __builtin_vec_delete );
+ FREE(VG_Z_LIBC_SONAME,       _ZdaPv,               __builtin_vec_delete );
++FREE(NONE,                   _ZdaPv,               __builtin_vec_delete );
+ 
+ // operator delete[](void*), ARM/cfront mangling
+ #if defined(VGP_ppc32_aix5) || defined(VGP_ppc64_aix5)
  FREE(VG_Z_LIBC_DOT_A,        __vd__FPv,            __builtin_vec_delete );
  #endif
- 
 +// libtcmalloc's delete []
 +FREE(NONE, tc_deletearray, __builtin_vec_delete);
  
- /*---------------------- delete [] nothrow ----------------------*/
  
-@@ -421,6 +445,8 @@
+ /*---------------------- delete [] nothrow ----------------------*/
+@@ -420,6 +478,9 @@
+ // operator delete[](void*, std::nothrow_t const&), GNU mangling
  FREE(VG_Z_LIBSTDCXX_SONAME,  _ZdaPvRKSt9nothrow_t, __builtin_vec_delete );
  FREE(VG_Z_LIBC_SONAME,       _ZdaPvRKSt9nothrow_t, __builtin_vec_delete );
- 
++FREE(NONE,                   _ZdaPvRKSt9nothrow_t, __builtin_vec_delete );
 +// libtcmalloc's delete [] nothrow
 +FREE(NONE, tc_deletearray_nothrow, __builtin_vec_delete);
  
- /*---------------------- calloc ----------------------*/
  
-@@ -465,6 +491,7 @@
+ /*---------------------- calloc ----------------------*/
+@@ -465,6 +526,9 @@
     }
  
  CALLOC(VG_Z_LIBC_SONAME, calloc);
++CALLOC(NONE,             calloc);
 +CALLOC(NONE,             tc_calloc);
++CALLOC(NONE,             sh_calloc);
  #if defined(VGP_ppc32_aix5) || defined(VGP_ppc64_aix5)
  CALLOC(VG_Z_LIBC_SONAME, calloc_common);
  #elif defined(VGO_darwin)
-@@ -523,6 +550,8 @@
+@@ -523,6 +587,10 @@
     }
  
  REALLOC(VG_Z_LIBC_SONAME, realloc);
++REALLOC(NONE,             realloc);
 +REALLOC(NONE,             tc_realloc);
++REALLOC(NONE,             sh_realloc);
 +REALLOC(NONE,             PyObject_Realloc);
  #if defined(VGP_ppc32_aix5) || defined(VGP_ppc64_aix5)
  REALLOC(VG_Z_LIBC_SONAME, realloc_common);
  #elif defined(VGO_darwin)
-@@ -579,6 +608,7 @@
+@@ -579,6 +647,9 @@
     }
  
  MEMALIGN(VG_Z_LIBC_SONAME, memalign);
++MEMALIGN(NONE,             memalign);
 +MEMALIGN(NONE,             tc_memalign);
++MEMALIGN(NONE,          sh_memalign);
  #if defined(VGO_darwin)
  ZONEMEMALIGN(VG_Z_LIBC_SONAME, malloc_zone_memalign);
  #endif
-@@ -621,6 +651,7 @@
+@@ -621,6 +692,9 @@
     }
  
  VALLOC(VG_Z_LIBC_SONAME, valloc);
++VALLOC(NONE,             valloc);
 +VALLOC(NONE,             tc_valloc);
++VALLOC(NONE,          sh_valloc);
  #if defined(VGO_darwin)
  ZONEVALLOC(VG_Z_LIBC_SONAME, malloc_zone_valloc);
  #endif
-@@ -641,6 +672,7 @@
+@@ -641,6 +715,8 @@
     }
  
  MALLOPT(VG_Z_LIBC_SONAME, mallopt);
++MALLOPT(NONE,             mallopt);
 +MALLOPT(NONE,             tc_mallopt);
  
  
  /*---------------------- malloc_trim ----------------------*/
-@@ -707,6 +739,7 @@
+@@ -677,6 +753,7 @@
+    }
+ 
+ MALLOC_TRIM(VG_Z_LIBC_SONAME, malloc_trim);
++MALLOC_TRIM(NONE,             malloc_trim);
+ 
+ 
+ /*---------------------- posix_memalign ----------------------*/
+@@ -707,6 +784,8 @@
     }
  
  POSIX_MEMALIGN(VG_Z_LIBC_SONAME, posix_memalign);
++POSIX_MEMALIGN(NONE,             posix_memalign);
 +POSIX_MEMALIGN(NONE,             tc_posix_memalign);
  #if defined(VGP_ppc32_aix5) || defined(VGP_ppc64_aix5)
  /* 27 Nov 07: it appears that xlc links into executables, a
     posix_memalign, which calls onwards to memalign_common, with the
-@@ -791,6 +824,7 @@
+@@ -737,6 +816,7 @@
+ 
+ MALLOC_USABLE_SIZE(VG_Z_LIBC_SONAME, malloc_usable_size);
+ MALLOC_USABLE_SIZE(VG_Z_LIBC_SONAME, malloc_size);
++MALLOC_USABLE_SIZE(NONE,             malloc_size);
+ 
+ 
+ /*---------------------- (unimplemented) ----------------------*/
+@@ -791,6 +871,8 @@
     }
  
  MALLINFO(VG_Z_LIBC_SONAME, mallinfo);
++MALLINFO(NONE,             mallinfo);
 +MALLINFO(NONE,             tc_mallinfo);