perf regressions on spec2k6 perlbench and gcc

_From [bruen...@google.com](https://code.google.com/u/109494838902877177630/) on September 19, 2011 10:49:28_

perlbench and gcc are MUCH slower (this is -no_delay_frees_stack, measured
as part of issue #205) than back in June 2011:

> spec2k6cmp result/CINT2006.132.ref.txt namedres/defaults-ref/CINT2006.113.ref.txt
400.perlbench 1.27 ( 16429 / 12957)
401.bzip2 1.02 ( 6428 / 6279)
403.gcc 1.65 ( 11054 / 6701)
429.mcf 0.99 ( 1273 / 1283)
445.gobmk 1.03 ( 8545 / 8304)
456.hmmer 1.04 ( 8920 / 8544)
458.sjeng 1.06 ( 10120 / 9540)
462.libquantum 1.01 ( 5104 / 5050)
464.h264ref 1.02 ( 19790 / 19387)
471.omnetpp 1.04 ( 7219 / 6960)
473.astar 1.05 ( 4228 / 4021)
483.xalancbmk 1.01 ( 6782 / 6742)

*\* TODO set up regression tests for perf

the problem is, we'd want an idle machine not a vm on a shared host, and it
takes over 24 hours to run just INT. but running say once a week would be
sufficient.

_Original issue: http://code.google.com/p/drmemory/issues/detail?id=595_