Path: utzoo!attcan!uunet!cs.utexas.edu!sdd.hp.com!ucsd!ucbvax!MITCH.ENG.SUN.COM!wmb
From: wmb@MITCH.ENG.SUN.COM
Newsgroups: comp.lang.forth
Subject: setjmp/longjmp vs. CATCH/THROW
Message-ID: <9007062100.AA05671@ucbvax.Berkeley.EDU>
Date: 6 Jul 90 14:30:43 GMT
Sender: daemon@ucbvax.BERKELEY.EDU
Reply-To: wmb%MITCH.ENG.SUN.COM@SCFVM.GSFC.NASA.GOV
Organization: The Internet
Lines: 36

> You can push the jmp_buf on the return stack.

You can do so if you happen to know how many return stack cells are
necessary to hold the information in jmp_buf.  Since the size of
jmp_buf is not necessarily the same for all systems, the portable code
for doing so looks like:

    #JMP_BUF  BEGIN  JMP_BUF OVER CELLS+ @ >R  1- DUP  0<  UNTIL

and the popping code which is required at the end is of similar ilk.
And you can't build this code inside a colon definition because a
standard program cannot make detailed assumptions about exactly
what happens to the return stack when a colon definition is called.
In particular,  : FOO  R>  {push stuff on the return stack}  >R  ;
isn't guaranteed to be portable.

Note that you can't use a DO .. LOOP for the above code because you can't
shove stuff on the return stack inside a DO .. LOOP .  Which brings up the
other obvious problem with using the return stack in this fashion; it
"partitions" the return stack in the middle of a definition, so that
anything that is placed on the return stack before the jmp_buf stuff
is put there becomes inaccessible while the jmp_buf stuff is there.

CATCH/THROW avoids this partitioning effect by making the return stack
effects occur at a procedure boundary, where the return stack details
are already hidden from a standard program.

Specifying that the JMP_BUF information "must be" e.g. 2 cells worth of
data would work in some systems, and would simplify the jmp_buf save
code to
	JMP_BUF 2@ 2>R
but the restriction to 2 cells is impractical, because many systems will
need to save additional information, such as a floating point stack pointer
or a local variable frame pointer.

Mitch