-O2 loop problems

Sun Dec 24 20:16:00 GMT 2000

On 2000年12月23日, Anthony Green wrote:
> I believe the work around for PR's 1159 and 1412 is to compile with 
> -fno-defer-pop.

Thanks! I just built libgcj with -O2 and -fno-defer-pop. It is
measurably faster on my bench application, by about 15% than with the
"standard" standard runtime.
(It's a good sign that this app now runs at all... it contains around 160
classes and interfaces and exercises many parts of the runtime. The
gcj-compiled binary is not yet as fast as the JDK 1.3, but is very, very
close... about 4%. The upcoming allocator improvements could very
well make up the difference!)
> The loops are messed up because we're missing pending stack adjustments 
> somewhere.

Could it be there is a NO_DEFER_POP missing somewhere? Just curious...
Jeff