Re: RFC: Another pass_late_combine after pass_reorder_blocks?

← All discussions

Project / Subsystem

gcc / rfc

Date

2026-04-17

Proposer

Uros Bizjak <[email protected]>

Source type

public_inbox

Consensus

Proposed

Sentiment

8/10

Technical tradeoffs

  • • Adding compiler passes increases compilation time.
  • • The code size improvement may be architecture-specific or depend on the input code.
  • • The benefit of the additional pass needs to be weighed against the cost of maintaining another pass in the compiler.
  • • Consider whether a late combine entry, invoked only when BB reordering duplicates paths, would be sufficient.

All attributes

project
gcc
subsystem
rfc
patch_id
—
discussion_id
CAFULd4a1_CR=NGZUU1X+EfuT0=Ge5d_32R4dNHN8S0mVPBe8dg@mail.gmail.com
source_type
public_inbox
title
Re: RFC: Another pass_late_combine after pass_reorder_blocks?
headline
Consider adding another pass_late_combine after pass_reorder_blocks
tldr
Adding another pass_cprop_hardreg after pass_reorder_blocks shows code size improvements in x86_64 Linux kernel builds.
stakes
Optimizing compiler passes can improve code size and performance, affecting the efficiency of compiled software, especially for large projects like the Linux kernel.
proposer
Uros Bizjak <[email protected]>
consensus
Proposed
outcome
proposed
sentiment_score
8
sentiment_rationale
The email presents empirical data supporting a potential optimization, which is generally viewed positively.
technical_tradeoffs
  • • Adding compiler passes increases compilation time.
  • • The code size improvement may be architecture-specific or depend on the input code.
  • • The benefit of the additional pass needs to be weighed against the cost of maintaining another pass in the compiler.
  • • Consider whether a late combine entry, invoked only when BB reordering duplicates paths, would be sufficient.
series_id
—
series_role
standalone
series_parts
[]
tags
  • • optimization
  • • compiler pass
  • • code size
  • • reordering
  • • x86_64
bugzilla_url
—
date
2026-04-17T00:00:00.000Z

Re: RFC: Another pass_late_combine after pass_reorder_blocks?

Uros Bizjak reports that adding an additional pass_cprop_hardreg after pass_reorder_blocks results in code size reductions for x86_64 Linux kernel builds. The results suggest a benefit from performing constant propagation after basic block reordering.