GH-48159 [C++][Gandiva] Projector make is significantly slower after move to OrcJIT #49063
+25
−13
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
|
github-actions
bot
added
Component: C++
Component: Gandiva
awaiting review
Member
|
I think this broke some of our Linux Packaging jobs: Seen this on main: Edit: I've opened: |
Member
|
Oh, sorry... @lriggs Could you take a look at it? |
Contributor
Author
lriggs
added a commit
to lriggs/arrow
that referenced
this pull request
Jan 30, 2026
…after move to OrcJIT (apache#49063) ### Rationale for this change Reduces LLVM TargetMachine object creation from 3 to 1. This object is expensive to create and the extra copies weren't needed. ### What changes are included in this PR? Refactor the Engine class to only create one target machine and pass that to the necessary functions. Before the change (3 TargetMachines created): First TargetMachine: In Engine::Make(), MakeTargetMachineBuilder() is called, then BuildJIT() is called. Inside LLJITBuilder::create(), when prepareForConstruction() runs, if no DataLayout was set, it calls JTMB->getDefaultDataLayoutForTarget() which creates a temporary TargetMachine just to get the DataLayout. Second TargetMachine: Inside BuildJIT(), when setCompileFunctionCreator is used with the lambda, that lambda calls JTMB.createTargetMachine() to create a TargetMachine for the TMOwningSimpleCompiler. Third TargetMachine: Back in Engine::Make(), after BuildJIT() returns, there's an explicit call to jtmb.createTargetMachine() to create target_machine_ for the Engine. After the change (1 TargetMachine created): The key changes are: Create TargetMachine first: The code now creates the TargetMachine explicitly at the start of the Engine in Engine::Make. That machine is passed to BuildJIT. In BuildJiIT that machine's DataLayout is sent to LLJITBuilder which prevents prepareForConstruction() from calling getDefaultDataLayoutForTarget() (which would create a temporary TargetMachine). Use SimpleCompiler instead of TMOwningSimpleCompiler: SimpleCompiler takes a reference to an existing TargetMachine rather than owning one, so no new TargetMachine is created. A shared_ptr is used to ensure that TargetMachine stays around for the lifetime of the LLJIT instance. ### Are these changes tested? Yes, unit and integration. ### Are there any user-facing changes? No. * GitHub Issue: apache#48159 Lead-authored-by: logan.riggs@gmail.com <logan.riggs@gmail.com> Co-authored-by: Logan Riggs <logan.riggs@dremio.com> Signed-off-by: Sutou Kouhei <kou@clear-code.com>
lriggs
added a commit
to dremio/arrow
that referenced
this pull request
Jan 30, 2026
* apacheGH-48159 [C++][Gandiva] Projector make is significantly slower after move to OrcJIT (apache#49063) ### Rationale for this change Reduces LLVM TargetMachine object creation from 3 to 1. This object is expensive to create and the extra copies weren't needed. ### What changes are included in this PR? Refactor the Engine class to only create one target machine and pass that to the necessary functions. Before the change (3 TargetMachines created): First TargetMachine: In Engine::Make(), MakeTargetMachineBuilder() is called, then BuildJIT() is called. Inside LLJITBuilder::create(), when prepareForConstruction() runs, if no DataLayout was set, it calls JTMB->getDefaultDataLayoutForTarget() which creates a temporary TargetMachine just to get the DataLayout. Second TargetMachine: Inside BuildJIT(), when setCompileFunctionCreator is used with the lambda, that lambda calls JTMB.createTargetMachine() to create a TargetMachine for the TMOwningSimpleCompiler. Third TargetMachine: Back in Engine::Make(), after BuildJIT() returns, there's an explicit call to jtmb.createTargetMachine() to create target_machine_ for the Engine. After the change (1 TargetMachine created): The key changes are: Create TargetMachine first: The code now creates the TargetMachine explicitly at the start of the Engine in Engine::Make. That machine is passed to BuildJIT. In BuildJiIT that machine's DataLayout is sent to LLJITBuilder which prevents prepareForConstruction() from calling getDefaultDataLayoutForTarget() (which would create a temporary TargetMachine). Use SimpleCompiler instead of TMOwningSimpleCompiler: SimpleCompiler takes a reference to an existing TargetMachine rather than owning one, so no new TargetMachine is created. A shared_ptr is used to ensure that TargetMachine stays around for the lifetime of the LLJIT instance. ### Are these changes tested? Yes, unit and integration. ### Are there any user-facing changes? No. * GitHub Issue: apache#48159 Lead-authored-by: logan.riggs@gmail.com <logan.riggs@gmail.com> Co-authored-by: Logan Riggs <logan.riggs@dremio.com> Signed-off-by: Sutou Kouhei <kou@clear-code.com> * apacheGH-49034 [C++][Gandiva] Fix binary_string to not trigger error for null strings (apache#49035) ### Rationale for this change The binary_string function will attempt to allocate 0 bytes of memory, which results in a null ptr being returned and the function interprets that as an error. ### What changes are included in this PR? Add kCanReturnErrors to the function definition to match other string functions. Move the check for 0 byte length input earlier in the binary_string function to prevent the 0 allocation. Add a unit test. ### Are these changes tested? Yes, unit and integration testing. ### Are there any user-facing changes? No. * GitHub Issue: apache#49034 Authored-by: Logan Riggs <logan.riggs@dremio.com> Signed-off-by: Sutou Kouhei <kou@clear-code.com> --------- Signed-off-by: Sutou Kouhei <kou@clear-code.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Rationale for this change
Reduces LLVM TargetMachine object creation from 3 to 1. This object is expensive to create and the extra copies weren't needed.
What changes are included in this PR?
Refactor the Engine class to only create one target machine and pass that to the necessary functions.
Before the change (3 TargetMachines created):
First TargetMachine: In Engine::Make(), MakeTargetMachineBuilder() is called, then BuildJIT() is called. Inside LLJITBuilder::create(), when prepareForConstruction() runs, if no DataLayout was set, it calls JTMB->getDefaultDataLayoutForTarget() which creates a temporary TargetMachine just to get the DataLayout.
Second TargetMachine: Inside BuildJIT(), when setCompileFunctionCreator is used with the lambda, that lambda calls JTMB.createTargetMachine() to create a TargetMachine for the TMOwningSimpleCompiler.
Third TargetMachine: Back in Engine::Make(), after BuildJIT() returns, there's an explicit call to jtmb.createTargetMachine() to create target_machine_ for the Engine.
After the change (1 TargetMachine created):
The key changes are:
Create TargetMachine first: The code now creates the TargetMachine explicitly at the start of the Engine in Engine::Make. That machine is passed to BuildJIT. In BuildJiIT that machine's DataLayout is sent to LLJITBuilder which prevents prepareForConstruction() from calling getDefaultDataLayoutForTarget() (which would create a temporary TargetMachine).
Use SimpleCompiler instead of TMOwningSimpleCompiler:
SimpleCompiler takes a reference to an existing TargetMachine rather than owning one, so no new TargetMachine is created.
A shared_ptr is used to ensure that TargetMachine stays around for the lifetime of the LLJIT instance.
Are these changes tested?
Yes, unit and integration.
Are there any user-facing changes?
No.