apache / apache/arrow

[C++][SIMD] Avoid one-definition-rule violation of `arrow::internal::BitmapWriter` without depending on `-O2`

Open
#36,902 1 comment 0 reactions 0 assignees View on GitHub
Component: C++ Type: enhancement
Dominant language
C++
Stars
17.1k
Forks
4.3k
Avg merge
3d 13h
Merged PRs (30d)
88

Description

### Describe the enhancement requested

Contexts:

* https://github.com/apache/arrow/issues/31132#issuecomment-1378082076
* GH-14342
* GH-36685

We have runtime SIMD dispatch feature. For example, it's used in `cpp/src/parquet/level_conversion.cc` to run BMI2 based `DefLevelsToBitmap` implementation if available.

`DefLevelsToBitmap` (`DefLevelsToBitmapSimd`) is implemented in `cpp/src/parquet/level_conversion_inc.h` and it's built with/without BMI2. `DefLevelsToBitmapSimd` uses `arrow::internal::BitmapWriter`. If `arrow::internal::BitmapWriter` in `DefLevelsToBitmapSimd` is not inlined, one-definition-rule violation is happen. If it's happen, `libparquet.dylib` has two `BitmapWriter` definitions for with/without BMI2. If BMI2 version `BitmapWriter` is used on BMI2 unavailable machine, "illegal opcode" is happen.

GH-14342 resolved this by forcing to use `-O2` to enforce inlining. But it doesn't work with Homebrew. Because Homebrew overrides `-O*` specified by us.

If we can avoid one-definition-rule violation of `arrow::internal::BitmapWriter` without depending on `-O2`, we can provide a Homebrew bottle that can use the BMI2 based implementation on machines that support BMI2.

### Component(s)

C++

Contributor guide

Open the contributing guide

Assessment

This issue has not been assessed yet.

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.