Ann: HAllInOne bug fix release
Thu, 31 Jul 2003 11:14:41 +0100
> real 3m14.295s
> user 3m9.879s
> sys 0m2.143s
> real 1m11.380s
> user 0m10.102s
> sys 0m1.711s
> Something very odd is going on. GHC does not generate programs than run
> 3x faster between GHC versions.
This comparison is actually between an nhc98-built program (the first
one) and a ghc-built program (the second one), showing that (in this
specific case) nhc98 is generating code that is three times slower
than that generated by ghc. In my experience, a speed difference of
between 2x and 5x of the (unoptimised) generated code between the two
compilers is fairly typical.
The really odd thing is in the second set of figures. Unless there
has been a cut-n-paste mistake, the user+system times do not add up
even close to the wall-clock time, suggesting a much greater difference
of 18x, which would start to worry me a little if it were true.
However, Hal posted an updated set of figures comparing a program
separately compiled with ghc to all-in-one compiled with ghc, and these
figures seem to have a similar same user+system /= real discrepancy:
> separately compiled with ghc:
> real 0m27.167s
> user 0m9.991s
> sys 0m1.304s
> all-in-one compiled with ghc:
> real 0m31.411s
> user 0m10.007s
> sys 0m1.299s
Is this some kind of known Cygwin timing problem?
Do ghc executables fiddle with the internal timers at runtime? :-)