Landing: 9394c71d3698
Project / Subsystem
gcc / rtl-optimization
Date
2026-08-25
Author
Kyrylo Tkachov
Commit
9394c71d36980a0abf1502be26608d6055893ed6
Source
github
Perf win
Yes
Breaking
No
All attributes
- project
- gcc
- subsystem
- rtl-optimization
- patch_id
- —
- commit_hash
- 9394c71d36980a0abf1502be26608d6055893ed6
- source_type
- github
- headline
- lra: Canonicalize whole-address reloads
- tldr
- Ensures correct instruction selection for addresses involving multiplications in the AArch64 reloader.
- author
- Kyrylo Tkachov
- outcome
- committed
- performance_win
- true
- breaking_change
- false
- series_id
- —
- series_parts
- []
- tags
-
- • rtl-optimization
- • aarch64
- • lra
- discussion_id_link
- —
- bugzilla_pr
- —
- date
- 2026-08-25T00:00:00.000Z
GCC’s instruction reloader (LRA) on AArch64 had issues recognizing addresses that involved a multiplication by a power of two, especially when this multiplication was part of a larger address calculation. The reloader would incorrectly generate a SET instruction for this part of the address, which didn’t match any available instruction, causing LRA to fail. This commit introduces a call to canonicalize_reload_addr to convert such multiplications into a canonical left-shift form (AShIFT), which LRA can then correctly recognize and emit the appropriate instruction. This fixes PR127056.