fine edged compiler (and language?) problem -- order of evaluation

[Jeffrey E. F. Friedl]

Hi...

I'm having problems with the actions of a compiler I'm using
("cc" on the CalComp 7500), and am wondering if it's me or
the compiler that's acting up.

This is strictly a pre-X3J11 matter, but post K&R-I
(i.e. C with "common extensions").

I have a situation where I need to multiply doubles with
unsigned chars.  There is a definite compiler bug in that
I get different results between:
    double_result = double_var * ((int)uchar_var);
and
    double_result = double_var * (temp_int_var = uchar_var, temp_int_var);

The difference is that the first one "sign-extends" and the
2nd doesn't (which is what I desire).  As was discussed several
weeks ago, the problem is not the matter of sign extending or not,
but simply that the assign to a temp acts differently than a cast.

So, because of this bug, I was forced to change my code (which
has several such multiplies in one expression) from something
natural like (where D == Double, UC == unsigned char):

    result = D1 * UC1   +    D2 * UC2   +    D3 * UC3;

to the kludge like:

    result = D1 * (TEMP_INT = UC1, TEMP_INT) +
	     D2 * (TEMP_INT = UC2, TEMP_INT) +
	     D3 * (TEMP_INT = UC3, TEMP_INT); 

The problem is this: the compiler generates assembler as if
the latter had been written:

    TEMP_INT = UC1;	/* side effects notwithstanding, worthless */
    TEMP_INT = UC2;	/* side effects notwithstanding, worthless */
    TEMP_INT = UC3;	/* this value gets used throughout...(agh!) */

    result = D1 * TEMP_INT +
	     D2 * TEMP_INT +
	     D3 * TEMP_INT;

Not quite what I intended.  As a general case,
I've checked that it generates code for:
    x = (ExpA_1, ..., ExpA_N-1, ExpA_N) + (ExpB_1, ..., ExpB_N-1, ExpB_N);
as:
    ExpA_1; ...; ExpA_N-1; ExpB_1; ...; ExpB_N-1;
    x = ExpA_N + ExpB_N;

Is this correct?  It most certainly is not to my intuition.

However, K&R-I says that the expressions are to be executed
sequentially; does that (or other quotes -- I've been to
perturbed to look this one up carefully) absolutely imply
that they must be done together (i.e. not interleaved as
this compiler does)?  Could the compiler's actions be considered
correct by any stretch of the rules?

One one of us is brain damaged -- which?.
(what friedl!vsi says [(-:] notwithstanding, I think it's the compiler).
Any thoughts out there on this?  How about in a context of 
compilation to parallel code -- could this be a way of specifying
parallelism? (I hope not).


As an aside: in checking the assembly listing, I noticed a *lot* of
comments in the listing.  Not the type that I would have loved
to see -- "this register has that variable" -- but those cryptic
kind (such as I have in my compiler under construction) obviously
there for the compiler writer's benefit.  So shocked at the number
of such comments (in a released compiler) that I stripped all
comments away to see just how much was there.

The size of the file shrank from 63K to 34K.  47% of the "code" was
tied up in comments.  This is a *slow* system as it is .. they most
certainly don't need the 47% overhead in I/O.   Is this a normal and
acceptable thing (by the tone of my question, my feeling should be
quite obvious)?

To end with one nice word about CalComp:  the strawberries are
impressive (if you know what I'm talking about, then you.. uh...
well, *know* what I'm talking about [(-:]).


	*jeff*


-------------------------------------------------------------------------------
Jeffrey Eric Francis Friedl, Box 2146 Babcock House, Durham New Hampshire 03824
..!{uunet,decvax}!unh!jeff   BITNET%"j_friedl at unhh"  ..!ucbvax!kentvax!jfriedl

I hope I'm not around Jan 18, 2038 at 10:14:08PM