Performance impact of GCC 3.4 on ilrun

Norbert and Rhys appear to have made some patches to permit compiling pnet 
under GCC 3.4. Some of the code that affects register allocation appears to 
be problematic under GCC 3.4. I wanted to see if this affects performance. 
Under Pnetmark, the "method" subtype is much, much slower with GCC 3.4, 
although some other tests are faster. Under scimark2, some subtests, like 
Monte Carlo, are dramatically slower under GCC 3.4.

I don't have a pretty spreadsheet view, but here are the numbers:

This is with GCC 3.2.3 

Sieve  ... 6534
Loop   ... 12927
Logic  ... 9289
String ... 1635
Float  ... 768
Method ... 3656

PNetMark rating is 3915

SciMark 2.0a

Composite Score: 12.53690493253059
FFT (1024): 8.347345641893233
SOR (100x100):   19.164091432323104
Monte Carlo : 5.955248784188833
Sparse matmult (N=1000, nz=5000): 15.43118285318456
LU (100x100): 13.786655951063218

*****

This is with GCC 3.4 (with patch to permit compilation under 3.4)

Sieve  ... 6657
Loop   ... 13335
Logic  ... 9628
String ... 1614
Float  ... 777
Method ... 2589

PNetMark rating is 3748


SciMark 2.0a

Composite Score: 11.461565337433854
FFT (1024): 8.716946878282089
SOR (100x100):   18.295035702654964
Monte Carlo : 2.49485293746483
Sparse matmult (N=1000, nz=5000): 15.70739513451045
LU (100x100): 12.093596034256936