Request for reviews (L): 7009309: JSR 292: compiler/6991596/Test6991596.java crashes on fastdebug JDK7/b122

[Vladimir Kozlov]

I think this looks good. Thank you for cleaning it.

Vladimir

On 1/24/11 4:42 AM, Christian Thalinger wrote:
> HEADS-UP: Please take a look at this one (or review it) if you are using or supporting older pre-SSE2 systems or use UseSSE=0 or UseSSE=1!  Thank you!
>
> http://cr.openjdk.java.net/~twisti/7009309/webrev.01/
>
> 7009309: JSR 292: compiler/6991596/Test6991596.java crashes on fastdebug JDK7/b122
> Reviewed-by:
>
> The reason the test fails is because of the extra cleanup code in the
> I2C adapter that's executed in C2 when SSE2 is not available (or
> UseSSE=1 is used).
>
> The suggested fix is to remove that extra logic completely and always
> do the FPU cleaning when SSE2 is not available.  This would also be a
> nice cleanup of the code in question.
>
> I did a refworkload run of specjvm98_client with 5 iterations and
> -Xint on a dual P3 system and it's slightly slower:
>
> ============================================================================
> logs.baseline-Xint
>   Benchmark         Samples        Mean     Stdev             Geomean Weight
>   specjvm98_client        5       11.78      0.10
> ============================================================================
> logs.fix-Xint
>   Benchmark         Samples        Mean     Stdev   %Diff    P   Significant
>   specjvm98_client        5       11.63      0.09   -1.27 0.035            *
> ============================================================================
>
> This slowdown would only apply to very old systems and given the fact
> that most x86 chips today are 64-bit and/or have SSE2, I think this
> should not be a big deal.  And for smaller systems which use C1 and
> rely on startup performance this does not apply anyway.
>
> However, in mixed mode the slowdown should completely disappear as the
> extra code is only executed when returning from compiled code to the
> interpreter.  A quick mixed mode run on that system even shows a
> speedup:
>
> ============================================================================
> logs.baseline
>   Benchmark         Samples        Mean     Stdev             Geomean Weight
>   specjvm98_client        5       66.81      2.31
> ============================================================================
> logs.fix
>   Benchmark         Samples        Mean     Stdev   %Diff    P   Significant
>   specjvm98_client        5       68.20      2.04    2.08 0.343            *
> ============================================================================
>
>