diff --git a/docs/pdds/draft/pdd32_m0.pod b/docs/pdds/draft/pdd32_m0.pod
index 1b993d7b04..921040b2e8 100644
--- a/docs/pdds/draft/pdd32_m0.pod
+++ b/docs/pdds/draft/pdd32_m0.pod
@@ -297,10 +297,29 @@ Store C<*$2> ^ C<*$3> into C<*$1>.
 
 =back
 
-=head3 Memory ops (4 ops)
+=head3 Memory ops (7 ops)
 
 =over
 
+=item * gc_alloc - allocate memory from the GC
+
+Allocate C<*$2> bytes, store the pointer C<*$1>.  C<*$3> contains flags
+indicating any special properties required of the returned chunk of memory.  A
+pointer returned by C<gc_alloc> does not need to be freed.
+
+=item * sys_alloc - allocate memory using malloc or its equivalent
+
+Allocate C<*$2> bytes, store the pointer C<*$1>.  C<$3> is ignored.  The
+pointer returned by C<sys_alloc> must be released with C<sys_free> to avoid
+memory leaks.  This is a direct interface to the underlying implementation's
+C<malloc()> implementation (or equivalent) and does not go through ffi.
+
+=item * sys_free - free memory using free() or its equivalent
+
+Free the region of memory at C<*$1>.  C<$2> and C<$3> are ignored.  This is a
+direct interface to the underlying implementation's C<free()> implementation
+(or equivalent) and does not go through ffi.
+
 =item * set - sets a register to a value
 
 Set C<*$1> to whatever is in C<*$2>.  C<$3> is ignored.
@@ -369,7 +388,7 @@ its arguments have been set up correctly.  C<$2> and C<$3> are ignored.
 
 =back
 
-=head3 Temporary ops (6 ops)
+=head3 Temporary ops (3 ops)
 
 =over
 
@@ -385,14 +404,6 @@ Print the integer at C<*$2> to the filehandle C<*$1>.
 
 Print the floating-point number at C<*$2> to the filehandle C<*$1>.
 
-=item * alloc - allocate memory (go through ffi)
-
-Allocate C<*$2> bytes, store them in C<*$1>.  C<$3> is ignored.
-
-=item * free - free memory (go through ffi)
-
-Free the region of memory at C<*$1>.  C<$2> and C<$3> are ignored.
-
 =item * exit - exit with the given status (go through ffi)
 
 Call C's C<exit> function with C<*$1>.  C<$2> and C<$3> are ignored.
@@ -723,20 +734,23 @@ until we're reasonably confident that we have a final set of ops.
     0x13 and
     0x14 or
     0x15 xor
-    0x16 set
-    0x17 set_mem
-    0x18 get_mem
-    0x19 set_var
-    0x1A csym
-    0x1B ccall_arg
-    0x1C ccall_ret
-    0x1D ccall
-    0x1E print_s
-    0x1F print_i
-    0x20 print_n
-    0x21 alloc
-    0x22 free
-    0x23 exit
+         gc_alloc
+         sys_alloc
+         sys_free
+         set
+         set_mem
+         get_mem
+         set_var
+         csym
+         ccall_arg
+         ccall_ret
+         ccall
+         print_s
+         print_i
+         print_n
+         alloc
+         free
+         exit
 
 =head2 Execution
 
@@ -861,6 +875,10 @@ machine code, to jitting and to compilation to static C.  Where possible, M0's
 design should not impose any constraints that preclude efficient implementation
 of any of type of interpreter.
 
+M0 must be implemented such that only programs which make direct use of ffi
+require it to be presence.  An M0 program which does not make use of FFI should
+run without modification on an implementation that does not support ffi.
+
 =head3 Implementation
 
 M0's FFI will implement similar functionality to dlfunc and dlvar in the form