Peter Gerwinski wrote:

Fixed.

:-)

Fixed.

:-)

...

Another thing I observed when looking at the asm code (see above) was that the i386 backend doesn't seem to know the "push mem" instruction.

[...]

Probably. However if you want to try yourself on that problem, look into the GCC info documentation, and read about "machine descriptions". There is something about how to describe each instruction a processor has, and maybe it's a trivial error and somebody overlooked this instruction.

I looked at i386.h and i386.md, and it seems like there isn't even a way to push anything but registers and constants... ?-(

Or else: Loading the memory reference into a register and pushing that *might* be faster than a direct push (don't know).

That might be a reason. But I don't know, either. I would suppose not -- why should the extra loading a register be faster? I tried some ways to measure the actual time (system clock, ps, /proc/..., even gprof), but they were all terribly inaccurate (+/- 5..10%), so there's no chance to check which is faster. :-( Anyone knows a way to measure runtime presicsely (Linux or DJGPP)?

Jan-Jaap van der Heijden wrote:

unsigned-rel = digit-sequence `.' fractional-part [ `e' scale factor ] | digits-sequence `e' scale factor [...] I see a simular definition in my Turbo pascal Programmer's Guide, so BP should not accept it either.

The BP handbook says (Fig. 2.3, at least in the German version):

unsigned real value (as above)

unsigned numerical value = unsigned integer | unsigned real value

This looks reasonable to me.