Request for review (XS): 7027232: JSR 292: wrong numeric value returned by MH on solaris-sparc
christian.thalinger at oracle.com
Fri Mar 18 09:35:06 PDT 2011
On Mar 17, 2011, at 5:56 PM, Vladimir Kozlov wrote:
> Sorry for arguing about few lines but I think without renaming the code will be better (6 new lines instead of 7):
> + #ifdef _LP64
> + __ ldsw(arg_lsw, O2_scratch); // Load LSW
> + #else
> + __ ldsw(arg_lsw, O3_scratch); // Load LSW
> + __ srlx(O3_scratch, BitsPerInt, O2_scratch)); // Move MSW value to lower 32-bits for STD
> + #endif
> __ st_long(O2_scratch, arg_msw); // Uses O2/O3 on !_LP64
Alright. I changed my code. Okay to push?
> On 3/17/11 2:24 AM, Christian Thalinger wrote:
>> On Mar 15, 2011, at 6:22 PM, Vladimir Kozlov wrote:
>>> We don't support little-endian SPARC, why add a code which will never be executed?
>> I thought to be on the safe side if anyone ever wants to support it. Anyway, I removed it and updated the webrev.
>> -- Christian
>>> On side note (RFE): we should officially drop v8 support (C2 does not support already, only C1). There are already several places where we use stx/ldx without check that it is v9 (yes, we still have an assert which is fine). Also std/ldd is emulated on modern SPARC and so is very slow.
>>> Christian Thalinger wrote:
>>>> 7027232: JSR 292: wrong numeric value returned by MH on solaris-sparc
>>>> After 7018378 one of the tests of java/dyn/MethodHandlesTest fails on
>>>> 32-bit SPARC. st_long uses STD on 32-bit SPARC and the LSW and MSW
>>>> bits are not moved to the proper registers resulting in a wrong
>>>> The fix is to use the correct registers for 32-bit big-endian and
>>>> little-endian (just in case) SPARC.
More information about the hotspot-compiler-dev