About

During 2023 various RISE and non-RISE members worked to improve if-conversion for GCC.  One of the recent discoveries from that work is that it allows if-conversion of a scalar copy of quant_4x4 within the spec2017 benchmark suite.  We have seen this improve the performance of the first (of three) steps in the x264 benchmark by approximately 10%.  The improvement is believed to be primarily due to eliminating a poorly predicted branch in that hot loop.

This work item for LLVM is meant to track a corresponding improvement to LLVM's if-conversion.   While this effort primarily helps the scalar quant_4x4, the belief is that it will help a variety of other codes, just in much smaller ways.

Stakeholders/Partners

Dependencies

Status

Updates

Sep 19, 2025

• Moved to 2H2025

Feb 12, 2025 

• Moved to 1H2025

Jul 11, 2024 

• Moved to 2H2024

Jun 5, 2024 

• Development work is considered done.  Needs to be tested on design to validate improvement though.

Apr 25, 2024 

• Failure to convert test into an equality test continues to be a problem

• Craig thinks that's supposed to already work, so clearly needs deeper dive

• Heuristic to avoid if-converting when some operands are invariant seems reasonable, but implementation seems bogus

• Enablement of early if-conversion

• Misha or another Ventana engineer will pick this up.

Mar 27, 2024 

• Patch under internal review.  Key issues appear to be:

  • Need to enable early if-conversion by default

  • Wire up additional ways to generate czero

  • Still working through decision whether or not to if-convert

Mar 13, 2024 

• Project noted as priority for 20241H