[
https://issues.apache.org/jira/browse/IMPALA-12314?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17748326#comment-17748326
]
Michael Smith commented on IMPALA-12314:
----------------------------------------
I looked at total time spent on code generation for the TPC-H suite on my
workstation, comparing O1 and Os. On average, switching from O1 to Os for
pre-compiled IR reduced code generation time by 24%.
Graphing individual query/fragment pairs shows a few took longer, but most
improved substantially.
!image-2023-07-27-13-55-10-103.png|width=941,height=582!
> Revisit pre-compiled IR at O1
> -----------------------------
>
> Key: IMPALA-12314
> URL: https://issues.apache.org/jira/browse/IMPALA-12314
> Project: IMPALA
> Issue Type: Improvement
> Components: Backend
> Reporter: Michael Smith
> Assignee: Michael Smith
> Priority: Major
> Attachments: image-2023-07-27-13-55-10-103.png
>
>
> From https://github.com/apache/impala/blob/4.2.0/be/CMakeLists.txt#L298-L300
> {quote}
> We enable basic optimizations (-O1) to reduce the IR size and speed up
> runtime JIT. Empirically, the runtime JIT produces slightly better code when
> starting with IR that was optimized at -O1. Higher optimization levels tend
> to bloat the code.
> {quote}
> This evaluation was done with LLVM 3.3. We're now on LLVM 5, and this
> evaluation may no longer hold. Revisit whether this makes sense or whether
> another option (O2, Os, O3) may make more sense.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-all-unsubscr...@impala.apache.org
For additional commands, e-mail: issues-all-h...@impala.apache.org
