🚀 go-pugleaf

In article <7w4813w164w@mantis.co.uk> mathew@mantis.co.uk (mathew) writes:
>eliot@cs.qmw.ac.uk (Eliot Miranda) writes:
>> In article <1991Apr19.132239.9252@daffy.cs.wisc.edu> quale@saavik.cs.wisc.edu
>> >     A compiler for a dynamically typed language can dedicate a register to
>> >     hold a bitmask that will speed type tag operations.  A C program would
>> >     either use an immediate value or a global variable, either of which is
>> >     slower and bulkier on many architectures.
>>
>> I have tried this in my Smalltalk VM. On both 68020 & SPARC there is no
>> significant difference in performance attributable to dedicating a register
>> to tag detection. Add a register here loose it there.
>
>You're saying that you've tried keeping the tags in memory vs. keeping them
>in a register on a 68020, and you've not noticed any difference?  I find this
>a little hard to believe.
>
>Exactly what changes did you make to the code in order to use register-based
>tags rather than memory-based tags, and what sort of program did you run in
>order to test the result?
>
The program is a "dynamic translation to direct threaded code" Smalltalk-80
virtual machine written in C, compiled with GCC & edited into threaded code by
sed-scripts run on the compiler generated assembler.  The system is about
20,000 lines of C.  The system is tested by running the standard Smalltalk-80
macro benchmark suite, which produces a performance figure in percents of a
Dorado, Xerox's fastest D machine.

The format of tagged data is tags in the bottom 2 bits
	01 -> 30 bit signed integer
	10 -> 16 bit unsigned character
	11 -> 30 bit fixed point (16 bit fraction)
	00 -> Ordinary Object Pointer

Using GCC one can declare global register variables, e.g.
register void (**tcip)() asm("a5");
declares the threaded code instruction pointer to be in a5 on 68020s.

To test tags using immediate masks I use:

#define TagMask 3
#define isTagged(o) ((unsigned long)(o)&TagMask)

To test tags using a tag register I use:

#define TagMask 3
register unsigned long tagReg asm("d3");
#define isTagged ((unsigned long)(o)&tagReg)

	...

	tagReg = TagMask

	...

The code generated by the two sequences is either
	movel a3@,d6
	moveq #3,d2
	andl d6,d2
	jeq LABEL
or
	movel a3@,d6
	andl d3,d6
	jeq LABEL

When I compared the two variants they ran the benchmarks to within 0.5% of
each other.  This difference is below the noise and hence not significant.
I suspect that this is because
	a) moveq is very quick
	b) reserving a register for the tag mask slows down other parts of
	   the system, cancelling out.

Baically optimizations are tradeoffs between representations, and each
implementation of a given piece of functionality will have its performance and
its cost.
-- 
Eliot Miranda			email:	eliot@dcs.qmw.ac.uk
Dept of Computer Science	ARPA:	eliot%dcs.qmw.ac.uk@nsf.ac.uk
Queen Mary Westfield College	UUCP:	eliot@qmw-dcs.uucp
Mile End Road			Fax:	081 980 6533 (+44 81 980 6533)
LONDON E1 4NS			Tel:	071 975 5229 (+44 71 975 5229)