[Squeak-VMdev] [VM] Verification of struct foo differences on x86 needed

[Ned K. <ne...@bi...>]

Hi all,

I noticed the other day that when I build a Linux/x86 VM using the global 
struct foo (which is the default for Unix) that it is faster than when I 
build without using the struct foo. That is, the bytecodes/sec speed is the 
same, and the sends/sec speed is 25-30% slower without struct foo.

This contradicted what Andreas had seen earlier; he's been *not* using the 
struct foo for his Windows builds.

Also, I found (as Andreas has) that gcc 2.95.4 produces a VM that is faster 
than the 3.3 or 3.4 gcc versions that I've tried. I'd earlier thought that a 
3.x version had produced faster code, but I guess I was wrong (or it was on a 
3.x version of GCC that I haven't tested recently).

Anyway, could someone with an x86 machine duplicate my test:

* make a clean build directory
* clean out the generated sources (assuming you're using Unix)
* generate VM sources
* configure and make the VM using GCC 2.95 and the -O3 flag:

 cd build
 CC=gcc-2.95 CFLAGS="-O3" ../platforms/unix/config/configure

run tinyBenchmarks on the resultant VM

then...

* remove the UnixVMMaker>>createCodeGenerator method and repeat the above 
steps.

Here's what I get for various configurations and compilers:

0 tinyBenchmarks

gcc 2.95.4, no struct foo, -O3
  '221837088 bytecodes/sec; 4782330 sends/sec'
  '219178082 bytecodes/sec; 4352264 sends/sec'

gcc 2.95.4, struct foo, -O3
'221453287 bytecodes/sec; 6569575 sends/sec'
 '203013481 bytecodes/sec; 6563459 sends/sec'

gcc 2.95.4, no struct foo
-O3
 '221070811 bytecodes/sec; 4743711 sends/sec'
  '221645021 bytecodes/sec; 4662139 sends/sec'

gcc 2.95.4, struct foo
-O3
 '222996515 bytecodes/sec; 6843839 sends/sec'
 '220309810 bytecodes/sec; 6739153 sends/sec'
-O3 -fno-gcse
 '113980409 bytecodes/sec; 5194661 sends/sec'

gcc 3.3.5, struct foo
-O3
   '200469851 bytecodes/sec; 6857154 sends/sec'
 '198603568 bytecodes/sec; 6778033 sends/sec'
-O2
 '202371541 bytecodes/sec; 6232674 sends/sec'
 '203984063 bytecodes/sec; 6113751 sends/sec'


gcc 3.4.2, struct foo
-O3
 '113374667 bytecodes/sec; 5276313 sends/sec'
 '111888111 bytecodes/sec; 5284224 sends/sec'

gcc 3.4.2, struct foo
-O3 -fnew-ra
 '113374667 bytecodes/sec; 5233225 sends/sec'
 '112379280 bytecodes/sec; 5217731 sends/sec'

gcc 3.4.2, struct foo
-O3 -fno-gcse
 '113174182 bytecodes/sec; 5190835 sends/sec'
 '112478031 bytecodes/sec; 5272367 sends/sec'

Thanks,
-- 
Ned Konz
http://bike-nomad.com