ack/doc/sparc/5

.so init
.nr H1 4
.NH
FUTURE WORK
.NH 2
A critique of EM
.PP
In general, EM fits its purpose quite well. Numerous compilers have been
written using EM as their intermediate language and it has even become a
commercial product. A great deal of its success is probably due to its
simplicity. There are no extravagant instructions but it does have all the
necessary functions to write a decent compiler.
.PP
There are, however, a few functions that come rather close to being
extravagant. The \*(Silar\*(So function for example \(em used
to fetch an element from an array \(em does not make it much easier
to write a frontend, but does make it unnecessary hard to write an
efficient backend. Other instructions for which it is difficult
to generate efficient code for are those that permit
dynamic operators, such as the \*(Silos\*(So. Dynamic operators, however, provide
significant extra possibilities and can therefore not be disposed of.
Note that even though the array operations \*(Silar\*(So and \*(Sisar\*(So
provide dynamic operators, they do not add additional power, since
they can easily be replaced with a sequence using the \*(Silos\*(So or
\*(Sists\*(So instructions.
.PP
EM code to reference arrays generated by the C frontend can be translated
very efficiently for almost any processor. However the same operation
generated by the Modula-2 frontend (which uses the \*(Silar\*(So),
is much less efficient, although the only difference is that the
latter performs range checking whereas the former does not.\(dg
.FS
\(dg Actually this depends on whether or not explicit range checking in enabled.
This clearly shows that the current code generators are not optimal and
often depend on ad-hoc decisions.
.FE
Since range checking can also be expressed explicitly in
EM (\*(Sirck\*(So) there is no need for any of the array operations
(\*(Siaar\*(So, \*(Silar\*(So and \*(Sisar\*(So).
.PP
Besides efficiency of the array-operations themselves, there still is another
major disadvantage of using these array-operations. In sharp contrast to
all other EM instructions except the \*(Silos\*(So and the \*(Sists\*(So,
they allow dynamic operators, so their effect on the stack-pointer can not
always be
determined at compile-time. This means that efficient caching of the
top-of-stack in registers is almost impossible,
so using these array-operations also effects the
efficiency of the surrounding code. Now that processors are produced with
more and more registers it could be very beneficiary to cache the
top-of-stack, so that the memory/register reference ratio decreases
to the benefit of the overall performance.
.PP
As a final critique, we would also like to discuss the semantics of some of
the EM instructions. In
.[ [
Description of a Machine Architecture
.]]
it is said that
all signed instructions such as the \*(Siadi\*(So, should cause an exception
on overflow. The unsigned operations such as \*(Siadu\*(So, however,
should act as modulo operations and therefor not perform overflow checking.
Since it is very expensive to perform overflow checking in EM,
we would suggest that the backend takes care of this. For languages which
do not require overflow checking, a simple message could be generated to
disable overflow checking in backends. This way all backends could be
written to fully comply to the official EM definition without any reduction in
efficiency.\(dd 
.FS
\(dd Currently many backends do not implement error checks because they
are too expensive and almost never needed. Some frontends even have
facilities build in to generate EM-code to force these checks. If this
trend continues we will end up with a de-facto and a de-jure standard
both developed by the same people but nonetheless incompatible.
.FE
When such messages will be added we would like to suggest
that they can enforce overflow checks on unsigned, as well as signed arithmetic.
.PP
As a conclusion we would like to suggest removal of the array operations from
EM, or at least discontinuation of there usage in frontends.
.NH 2
\*(OQWanted: Procedure call information\*(CQ
.PP
The advantage of an intermediate language such as EM is that the backend
no longer has to know about any 'quirks' of the 'input'-language. The major
disadvantage, however, is that the backend no longer knows about any 'quirks'
of the 'input'-language... If the SPARC backend ever has to compete
with Sun's own C-compiler for example, removal of the array-operations
will not be enough. The amount of information that is lost during
the translation to EM is too large to ever generate truly efficient SPARC code.
.PP
To write such an efficient backend one needs to know, for example, whether,
when and what type of parameter is being computed, so the result can be stored
in the proper place and scratch registers can be reused.
(On the SPARC processor, for example, it is very beneficiary
to pass the first six parameters of a procedure call through
registers instead of using the stack.)
One way to express such things in EM is to insert extra messages in
the EM-code. The C statement \*(Sia = f(4, a + b);\*(So for example,
could be translated to the following EM-code:
.DS
.TS
;
l1f6 lf6 l.
lol	-4	! a
lol	-8	! b
mes	x, 2	! next instruction will compute 2nd parameter
adi	4
mes	x, 1	! next instruction will compute 1st parameter
loc	4
cal	_f	! call function f
lfr	4
stl	-4	! store result in a
.TE
.DE
For a code expander it is important that the \*(Simes\*(So pseudo
instructions appear \fIbefore\fR
the EM instruction that computes the parameter, because that way the final
computation (the \*(Siadi\*(So and \*(Siloc\*(So in the previous example)
can be translated to machine code that performs the required computation
and also puts the result in the required place. If it is found to be
too difficult for the frontend to insert these \*(Simes\*(So instructions
at the right place the peep-hole optimizer might swap the \*(Simes\*(So and
the instruction that computes the parameter.
.PP
For some architectures, it is also
possible to generate more efficient code for a procedure when it is a
so-called leaf-procedure: a procedure that doesn't call other procedures.
On the SPARC, for example, it is not necessary to rotate the register
window for a call to a leaf procedure and it is also possible to use
the global registers for register variables in leaf procedures.
It will be a little harder to insert useful messages about leaf procedures,
because just as with register messages, they are only useful to the
backend when they appear immediately
after or before the \*(Sipro\*(So pseudo instruction. The frontend,
however, only knows whether a certain procedure is a leaf-procedure or not
when it has already generated the entire procedure in EM. Just as with the
\*(Sipro ? / end n\*(So-dilemma the peep-hole optimizer
.[ [
Using Peephole Optimization
.]]
might be able to lend a hand
and help us out by delaying EM-code generation until it has reached the
end of the procedure.
.PP
As with most optimizations, the main problem is that they have to be
implemented with the \*(Simes\*(So pseudo instruction.
Because the \*(Simes\*(So instruction can have many different meanings
depending on its argument, 
it is important that all optimizers recognize and respect them. Addition
of even a single message will require careful inspection of, and maybe even
incorporate small changes to each of the optimizers.
.bp