John Levine wrote:
According to Anton Ertl <anton@mips.complang.tuwien.ac.at>:
Of course, if the performance cost of these features is so high that >>compiler writers prefer to avoid the full-featured call instruction,
the instruction misses the target, too.
The Vax suffered badly from the second system syndrome,
Which is strange because it was not a second system {PDP-1...PDP-11}
in which its designers loaded it up with all the stuff that wouldn't fit on a
PDP-11. Its design also seems to assume a world where main memory is expensive and microcode ROM is much faster,
The handwriting should have been on the wall by the time of VAX design, especially since VAX had caches and TLBs.
hence the dense byte coded instructions that demand to be decoded one operand at a time and the
tiny 512 byte pages.
I also don't think they thought through the costs of implementing
their design. The Vax chapter in the Computer Engineering book says
that they wanted to have a single system wide procedure call, which is
fine, but it should have been evident even at the time that all of the
memory traffic for saving and restoring registers, including ones the
routine doesn't need, would be really slow.
My 66000 separates calling from preservation. Only the caller knows what
it wants saved and whom to call, only the callee knows what it will damage
or return.
Caller has 3 instructions {CALL, CALLX, CALLA} witch transfer control
and provide a return address.
Callee can choose to save nothing if it can do its job entirely within
R0..R15, or it can save as many or as few registers {R16..R31} as it
desires. These choices are expressed through instructions {ENTER, EXIT,
RET}.
I think VAX blew this one rather badly. We tried something similar on
GOULD S.E.L 32/87 and got rid of it before anyone outside the company
saw any of it.
It's perfectly possible to have a standard calling sequence without
baking it all into microcode. All of the languages on S/360 use the
same calling sequence, but it's made from several instructions, and
you can leave out the ones you don't need, e.g., providing a new save
area in leaf routines, or stack management in non-recursive routines.
Yes, YES -- always be in a position where you can LEAVE OUT work that
is not needed in the call/return interface(s).
--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)