Re: Easily-Read C++?

[firth@sei.cmu.edu (Robert Firth)]

In article <37bno4$ko4@gnat.cs.nyu.edu> dewar@cs.nyu.edu (Robert Dewar) writes:

>An interesting point about the ++ notation is that it appears as far as
>I can tell to give undefined results or rather implementation dependent
>resultsin the presence of threads, in particular, is it possible for
>++c to store a 4 in memory (where there was a 3 before), and then return
>6 as a result of rereading c, where intermediate updates have occurred.

Unfortunately, as far as I can see this is almost inevitable.  The
construct expands, semantically, to

	valof { c := c+1; resultis c }

Well, one can try to make the increment an atomic operation, giving,
for instance (PE3200 code)

	LIS R1, 1	; load 1 into a temp register
	AM R1, C(RD)	; add it to memory cell C

but then you are forced to do another read

	L R0, C(RD)	; now load new value of C into result reg

and, as Robert Dewar mentioned, who knows what has happened in
the meantime?

The alternative is to make the result computation essentially
atomic:

	L R0, C(RD)	; get value of C into result reg
	AIS R0, 1	; add 1 to it and leave as result
	ST R0, C(RD)	; and store back new value

and, again, anything could have happened to the variable C between
the read and the write.  So the compiler writer, with the best will
in the world (a rare quality, I fear) is still pretty much stuck.

The C code I've seen is absolutely littered with this stuff. I am
continually amazed that *anything* still works when people write
true multitasking code in C.