On Thursday, 96/11/28, George White wrote to David Noon about "Named
Pipes" as follows:

GW> Borland 3.1 generates duff code if set to allow 386 code
GW> generation (it can't do arithmetic on longs correctly :-( ).

Hi George,

We have found opposite results here, but they are not totally
contradictory.

GW> 100 K or so). If the code generation is set back to 286 everything
GW> works OK (well, I've not yet found any bugs with my C programs that
GW> I can blame on the compiler rather than my code).

The problem I encountered was down to the optimiser in 80286 mode. It
assumed that the initialisation value of NULLHANDLE for a HEV would
still be in place after a call to DosCreateEventSem(). Of course it
wasn't; it had been replaced by the actual semaphore handle. When I
called DosWaitEventSem() it came back with an ERROR_INVALID_HANDLE,
because the 2 PUSH instructions to place the handle onto the stack had
hard-coded zeroes.

When I recompiled in 80386 mode the single PUSH instruction pointed to
the handle location for the value to be pushed onto the stack. I didn't
bother pursuing this until this weekend, since the Watcom compiler was
working.

Ironically, the Watcom 16-bit compiler seems to produce the same
object code whether I specify 80286 or 80386 as the target CPU. [Do any
Watcom experts lurking have any input on this phenomenon?] However, its
object code is always logically correct, provided you stay away from
huge data items.

Regards

Dave


 * KWQ/2 1.2i * She won't last forever........why give her a diamond?

--- Maximus/2 3.01