tree-ssa performance

[Steven Bosscher]

Diego Novillo wrote:

>On Tue, 2003-06-10 at 12:18, Andrew MacLeod wrote:
>
>  
>
>>FWIW, The very limited amount of compile time experience Ive got (an
>>early cut at reducing compile time for out-of-SSA) indicated that
>>tree-ssa requires more memory. If you have a smaller memory machine, its
>>compile times shoot up quite a bit due to swapping. On machines where it
>>can compile without swapping, compiles times are not so bad.
>>
>>    
>>
>Perhaps.  The machines that run SPEC are not small.  The SPEC95 machine
>has 256Mb of RAM.  The SPEC2k machine has 1Gb of RAM.
>
>We get comparable results in both machines.  From what I've seen,
>anything below 128Mb would cause the machine to swap itself silly while
>bootstrapping.
>

Now _that_ is something I can confirm!  Until a couple of weeks ago, I 
had a P-II 500Mhz, 192 MB RAM, and a tree-ssa bootstrap with C, C++, and 
G95, checking enabled, took about 7 hours while I was working on that 
machine.  I now have a P-II 450 MHz with only 96 MB ram, and I cannot 
even complete a C only bootstrap overnight!  :-)

1GB is probably not what the avarage GCC user has, and even 512MB is 
still a lot.  I wonder if you ever garbage-collect at all with a machine 
like that!!!  Maybe you should try and run the SPEC tester with --param 
gcc-min-{heapsize,expand} to some smaller values that are typical for 
the avarage GCC power-user, see how it compares to mainline then?

Gr.
Steven