Re: Hierarchical States Machines

["Robert I. Eachus" <rieachus@comcast.net>]

tmoran@acm.org wrote:
>>and ran 40% slower. (Register pressure from those state variables push
> 
>   Do you mean the parser state transitions ran 40% slower, or the whole
> compiler, including IO, lexer, symbol tables, code generation, etc
> was slowed by 40%?

Sorry, the first pass of the compiler, which included the scanner 
(lexer) and created the symbol table and AST.  The second pass did the 
semantic analysis and 'decorated' the parse tree, and did all the 
code-generator independent optimizations.

As I remember it, the running times were roughly 1-3-3.  So it would 
have been 1.4-3-3, for a 5 to 6% overall slowdown.

Even so it was pretty huge, and it shocked me.  By the point this 
occurred we had done a lot of optimization of the LALR1 engine and 
structure so that (mostly for error correcting reasons) successive 
reduce operations on the parse stack were deferred, then all called at 
once after two successful succeeding shifts, or a successful shift and 
(pending) reduce.

That sounds complicated, and it was.  But it made for wonderful syntax 
error correction without increasing the parse table size.  (Well we had 
about a half dozen added states to allow "panic mode" if none of the 
single or double token fixes worked.)

We were initially worried about the changes for the error correction 
support slowing down performance, but since it grouped the actions that 
resulted in subroutine calls, it actually sped things up significantly 
by eliminating register spills.  When we looked at the generated code to 
try and explain that 40% number, adding the two flags for the non-goto 
solution increased the number of active variables from 6 to 8 in the key 
loops--and we had six available registers once you eliminated the stack 
pointer and program counter.  Worse the access pattern basically rotated 
through the variables.  We thought about fixing that for both the Ada 
and Multics PL/1 compilers, but a little bit of monitoring indicated 
that this code was about the only thing around that hit this 'misfeature'.

The wonderful thing about being in a compiler group, and using Multics, 
was that we could slip in performance counters like this on the 
production machine, and see whether or not we should actually fix 
anything.  Technically our group supported the DPS-6 line not DPS-8M, 
but we were in Billerica, MA, Multics development was at CISL in 
Cambridge, MA, and GCOS-3/8 support was in Phoenix, AZ.  So I could 
submit a performance counter proposal like that to CISL, and we and they 
would put it into the "development" version of the compiler within a few 
days.

My favorite story about doing that was that we had a debate about using 
static links or a display to manage up-level references in the Ada/SIL 
compiler.  We had a ferocious call-graph analyzer that merged stack 
frames whenever possible, and of course, all variables in library 
packages (and in the main program if it was not called recursively) were 
handled directly.  So we put in a trigger that added the static pointers 
if necessary, and sent e-mail to a list of compiler people.  Almost two 
years later, I finally got one of those messages--for an ACVC test.

We decided we would take that code out and use a dynamic stack walk if 
it was really necessary.  I don't think that change ever happened--it 
was of course, very low priority.

-- 

                                           Robert I. Eachus

"The terrorist enemy holds no territory, defends no population, is 
unconstrained by rules of warfare, and respects no law of morality. Such 
an enemy cannot be deterred, contained, appeased or negotiated with. It 
can only be destroyed--and that, ladies and gentlemen, is the business 
at hand."  -- Dick Cheney