🚀 go-pugleaf

In article <BRUCE.89Jul18181114@heather.pooh.com> bruce@heather.pooh.com (Bruce Robertson) writes:
>Shouldn't all these optimizations be done in the machine independent
>portion of the compiler?  

Yes, all of the optimizations you mentioned (which I deleted) *should* be
done in the mi part of the compiler.  For compilers which are intended to
support multiple architectures (remember, not all are), it usually is.

>A well designed compiler should have
>all the machine dependencies isolated in one relatively small module
>(relative to the size of the entire compiler). RISCy things like delay
>slot filling are easiest to handle in the assembler, or other post
>processor.

Except that delay-slot filling is not the only type of optimization that
exists.  For some machine (such as a CDC Cyber), doing certain instructions
after another would slow it down, while on other machines, it can speed up.
Specifically:  memory accesses and various register combinations.  Some
microprocessors can be optimized to do certain kinds of memory fetches
faster, while the Cyber does sequential accesses best.  Also, something
like:
	FX6	X5*X3
	FX7	X6/X1

is bad practice on a cyber (put other instructions in between).  On some
other machines (specifically VLIW and the ilk), that could be reduced to a
pseudo-one instruction which did:

	X6 <- X5*X3, X7 <- <multiply unit> / X1

which would execute in parallel as much as possible (I know, I know, that's
really a bad example.  I had some better code than that lying around, but I
forget where I put it).

Then there's the choice of instruction selection.  Sometimes, which
instructions you choose can change some of your "machine independent"
optimizations.  Also, you have to choose registers properly (is it better to
put array addresses into registers, or would it be better to put other
variables in registers and let the hardware compute the array address?  For
something like a '486, this could be a tough choice...).

In other words, life isn't as simple as we would like it to be...

-- 
Sean Eric Fagan  |    "Uhm, excuse me..."
seanf@sco.UUCP   |      -- James T. Kirk (William Shatner), ST V: TFF
(408) 458-1422   | Any opinions expressed are my own, not my employers'.