RFR(M): 7133857: exp() and pow() should use the x87 ISA
vladimir.kozlov at oracle.com
Fri May 11 13:51:39 PDT 2012
Deneau, Tom wrote:
> Vladimir --
> On a related topic, newer processors may tend to downplay the
> performance of x87, so for better performance it would be good
> to have equivalent SSE routines for those things which are currently
> only in x87.
I agree, it would be nice and it would simplify/improve generated code.
> -- Tom
> Vladimir Kozlov wrote:
>> May be I misunderstand your question but we can't "retire x87". Some
>> instructions we use are present only in x87, like in these changes.
>> Nils Eliasson wrote:
>>> Just curious, have there ever been any discussion about retiring
>>> x87 support? For server, there can't be many non-SSE2 user (JRockit
>>> retired x87 in the R28.0 release in 2010), for client some P3s and
>>> Athlon XPs are probably still out there. For embedded, I have no clue
>>> at all, except that there are some Geodes without SSE2.
>>> Do we have any download count for different x86/x64 hardware?
>>> If we agree that server and client can do without, then we should
>>> start with removing x87 from c2.
>>> Roland Westrelin skrev 2012-04-02 10:56:
>>>> Implements Math.exp() and Math.pow() using the x87 ISA. It is based on an existing implementation in C2 that was disabled because it was found to be broken. This change fixes some corner cases of the previous implementation and uses the same code in the interpreter, c1 and c2 so that consistent results are obtained from exp() and pow().
More information about the hotspot-compiler-dev