-fno-strict-aliasing performance

[Kevin Glynn]

duchier at ps.uni-sb.de writes:
 > Konstantin Popov <kost at sics.se> writes:
 > 
 > > .. I forgot to tell that: on i486 the difference is essentially
 > > unmeasurable (except, probably, the 'port' benchmark - no more than
 > > 5%). I switched this flag for all platforms.
 > 
 > I can well understand switching it off for the upcoming release, but
 > this doesn't seem very appealing as a long term solution.  Perhaps we
 > could (later) attempt to fix the problem in a more principled fashion
 > (e.g. using type-puning through a union).
 > 

It would be good to be buildable with all options implied by gcc -O3,
is there also a performance hit? How much pain is involved in fixing
the source properly?


 > Speaking of switching it off on all platforms: if it is not a problem
 > on x86, is it really necessary to switch it off on that platform?
 > platform/emulator/configure.in already has a lot of (platform x
 > compiler) specific logic; the decision could be made there.
 > 

While the mozart source doesn't conform to the constraints for
-fstrict-aliasing then we shouldn't be using it on any platform.

cheers
k