Re: RAFTS, compile on EXECUTE and the x86

Alex McDonald <blog@xxxxxxxxxxx> writes:
On 25 Aug, 11:44, an...@xxxxxxxxxxxxxxxxxxxxxxxxxx (Anton Ertl) wrote:
Also, the main point of doing the compilation at the latest
point-in-time possible is to be able to do optimizations that go
beyond singe colon definitions, e.g., inlining or interprocedural
register allocation.

And the point of this paragraph is: If I am collecting code until the
last possible moment in order to see the maximum of code, so that I
can see optimization opportunities, it probably does not make much
sense to compile that big chunk in little pieces.

The code modification seems to upset the timings for several hundreds
if not thousands of iterations if I'm interpreting the output from
Code Analyst correctly.

That's strange. I would expect a big slowdown on the first time the
code is executed after it was changed, from the cache invalidation and
the need to reload the code through the memory hierarchy, but after
that the performance should be pretty stable. Maybe another slowdown
when the next piece of code is compiled, if it shares a cache line
with the current piece of code, but that should be it.

OTOH, who knows what goes on in the complex hierarchy of subsystems on
current CPUs. Or maybe the measurement infrastructure does not report
the numbers correctly.

The MPE benchmark may not be ideal for this
and might need lengthened;

Well, it rather seems like a best-case: Far fewer calls than normal
Forth code, and big basic blocks.

the execution times are measured in sub-
seconds, generally around 100-200ms for most of the tests.

That sounds like plenty. That's more than enough time to load 1M
cache lines, and I very much doubt that your code is nearly as big.

For
example, the first dhrystone iterated 500K times runs at 2.9M
dhrystones/second; the second and subsequent runs at 3.7M/sec.

The first run includes the compilation time, no? What if you try a
1-iteration run, then a 500k-iteration run? How does the 500k-run
fare then?

Compiling on compile is what I don't want to do

Generating native code when the Forth text interpreter compiles is not
what I was suggesting.

- anton
--
M. Anton Ertl http://www.complang.tuwien.ac.at/anton/home.html
comp.lang.forth FAQs: http://www.complang.tuwien.ac.at/forth/faq/toc.html
New standard: http://www.forth200x.org/forth200x.html
EuroForth 2010: http://www.euroforth.org/ef10/
.

Relevant Pages

  • Re: RAFTS, compile on EXECUTE and the x86
    ... point-in-time possible is to be able to do optimizations that go ... sense to compile that big chunk in little pieces. ... when the next piece of code is compiled, if it shares a cache line ... (Benchmarking would not appear to be my strength; ...
    (comp.lang.forth)
  • Re: [OT] Intel Compilers vs. Microsoft Compilers Optimization/Performance on Math Calculations,....
    ... Generally, "optimization" of loops no longer buys as much as it used to, because the x86 ... for cache hits, you will get *substantially* faster computations than if you do not. ... The difference between the Intel and Microsoft compilers might be 10% ... One page fault masks all other optimizations by 5-6 orders of magnitude ...
    (microsoft.public.vc.mfc)
  • Re: Project manager, compile before saving?
    ... Perhaps what you are experiencing is that VFP loves to cache things. ... > modification work, that right?. ... untill I compile it (not only main.prg but all from Project ...
    (microsoft.public.fox.programmer.exchange)
  • Re: The coming death of all RISC chips.
    ... The transforms (i.e. optimizations) are pretty much identical, but the mapping algorithms are quite different. ... If it turned out that the best architecture was something different, there would be a language that would map almost directly to that architecture. ... languages that is completely alien to them. ... I think what you're saying is that compilers don't compile the kinds of programs you're interested in efficiently. ...
    (comp.arch)
  • Re: Inline compilation / HotSpot
    ... from performing these kinds of optimizations? ... JIT compile a method BigBarand detects that a certain method ... SmallFoois worth inlining into BigBar(), ... at any moment of time the JVM can make whatever assumptions ...
    (comp.lang.java.programmer)