Regression in Scimark benchmark on amd64

Monte Carlo performance regression in b90
is due to Chuck's putback 20060619160059.rasbold.c2_baseline:
6417150: Performance regression in NCAR microbenchmark from Jre 1.5.0_06

Chucks's fix brought about 8% Sparse improvement for b90 over b89.
But it introduced an additional spill on stack in Monte Carlo's loop
which causes 20% regression:

Composite Score: 551.3180154521758
FFT (1024): 440.6873836219782
SOR (100x100): 715.8079123854665
Monte Carlo : 161.7081092756569
Sparse matmult (N=1000, nz=5000): 375.45687027106914
LU (100x100): 1062.929801706708

b90 which includes Ross's fix for the unrolling and
Chuck's fix for the code blocks scheduling and few
other not related fixes:

Composite Score: 553.1541457196678
FFT (1024): 440.7987446128855
SOR (100x100): 715.5421284974674
Monte Carlo : 131.6505402060507
Sparse matmult (N=1000, nz=5000): 413.6067966129366
LU (100x100): 1064.1725186689985