[jira] Commented: (HARMONY-4620) [drlvm][jit] Long return path for floating point values in calling convention

["Alexey Varlamov (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HARMONY-4620?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12575265#action_12575265 ] 

Alexey Varlamov commented on HARMONY-4620:
------------------------------------------

Evgueni, AFAICS you've also fixed HARMONY-5152 in the suggested patch, thanks!

However I do have compatibility concerns, do we really want to cut off completely non-SSE CPUs support? 
The best solution would be defining one more SSE-based managed CC and providing runtime customization to switch between them, accordingly to host architecture. However there is no such customization machinery available in the VM yet, so as a bare minimum we should keep compile-time switch, e.g. ifdef SSE/FPU return path. 
Could you please add such ifdefs to the patch?

> [drlvm][jit] Long return path for floating point values in calling convention
> -----------------------------------------------------------------------------
>
>                 Key: HARMONY-4620
>                 URL: https://issues.apache.org/jira/browse/HARMONY-4620
>             Project: Harmony
>          Issue Type: Improvement
>          Components: DRLVM
>         Environment: appropriate for for Intel architecture
>            Reporter: Naumova Natalya 
>            Assignee: Mikhail Fursov
>         Attachments: return_xmm.patch, return_xmm_2.patch, return_xmm_3.patch
>
>
> DRLVM has too long return path when the return value is floatin point. The reason is FPU usage together with SSE instructions in calling convention: we have "SSE -> mem -> FPU -> (return) mem -> SSE"; return (double) value first is calculated on xmm* registers, then copied to mem, then is put on FPU stack, then extracted from this stack (in calling proc) to memory again, then again calculation is happened in xmm* registers (SSE instructions). This issue overrides the improvement with loop unrolling, overhead from the parameters passing with this calling convention overrides the loop body doubling speed-up. When you increase "arg.optimizer.unroll.medium_loop_unroll_count" option in method where return value is double and it is in loop, then you'll have degradation (example - MonteCarlo benchmark in SciMark).
> Can we avoid using FPU with SSE in this case?

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.