Date: Fri, 18 Sep 92 19:03:24 CDT
From: "George Jetson" <pynq AT midway DOT uchicago DOT edu>
To: djgpp AT sun DOT soe DOT clarkson DOT edu
Subject: Problems with multiplication inside a loop inside a function, ...

I have found what I believe to be a compiler bug, relating to doing
multiplication inside a loop inside a function.  The code fragment given
below may or may not be helpful to anyone.  I shall probably end up
spending some time trying to isolate the problem later on.

Anyway, the code frag below is inside a function.  The function is nested
in my main() function. ('i' is declared inside the function, at the top)

-----
int i;
for (i=0; i<nkeys; i++) {
    int ln = keys[i].line;
    int off = keys[i].offset;
    char *s1 = t1 -> textlen[ln] > off ? t1 -> text[ln] + off : "";
    char *s2 = t2 -> textlen[ln] > off ? t2 -> text[ln] + off : "";
    int w = keys[i].width;

    if (res = keys[i].direction *
        (keys[i].lexical ? (Strnicmp(s1,s2,w) ? : strncmp(s1,s2,w))
         strncmp(s1,s2,w)))
                return (res);
    }
-----

nkeys is the number of keys and if it is one, then this works fine.
But if nkeys > 1, then I get an Exception 13.

Tracing it through with DEBUG32 reveals that the start of the for loop gets
implemented as something like:

	mov     ebx,[eax-16]		(Watch this space)
	cmp     [ebp-8],ebx
	jge     _big_cmp.13+384 (0x184c )
	mov     edx,[eax-44]

I think that eax-16 is the address of 'i', and ebp-8 is nkeys.

Anyway, what happens is that the first time through, this works fine,
but in the second iteration, eax is 0, and, obviously, [eax-16] is a
bogus address.  Stepping the loop through all the way shows that eax
gets trashed by an imul, and does not get restored.

Also, if I make a slight source code modification, this ends up using
edi instead of eax, but edi is also trashed by the multiplication routine.

Also, a couple of related questions about DEBUG32:
	1) How do I make the source code available to the debugger?
	   (DEBUG32.DOC implies that it is possible - I assume it is
	   some option on the compilation)  Yes, I know I am being lazy
	   and could certainly look it up...
	2) In the above frag, when I do a g to the "mov ebx,[eax-16]",
	   I should be able to do a "d %eax-16" and get the result that
	   is about to be loaded into ebx, right?  Unfortunately, I
	   can't get this to work.  "d %eax-16" gives me an unrelated
	   value.  (I.e., when I single step it, ebx gets the correct
	   value)