Landing: 39e7936c2ba2
Project / Subsystem
gcc / x86
Date
2026-09-28
Author
Richard Biener
Commit
39e7936c2ba2dea099bca58d12d9d0721d4eae55
Source
github
Perf win
No
Breaking
No
All attributes
- project
- gcc
- subsystem
- x86
- patch_id
- —
- commit_hash
- 39e7936c2ba2dea099bca58d12d9d0721d4eae55
- source_type
- github
- headline
- Fix load combining peepholes for adjacent stack loads
- tldr
- Corrected x86 load combining to prevent reordering of loads around stores.
- author
- Richard Biener
- outcome
- committed
- performance_win
- false
- breaking_change
- false
- series_id
- —
- series_parts
- []
- tags
-
- • x86
- • optimization
- • peephole
- • compiler
- discussion_id_link
- —
- bugzilla_pr
- —
- date
- 2026-09-28T00:00:00.000Z
This commit fixes an issue in x86 load combining peepholes that could lead to incorrect instruction scheduling. When combining two adjacent stack loads into a wider V2DFmode load using adjust_address, the MEM_EXPR and MEM_ALIAS_SET were not properly updated. This allowed later scheduling passes to move the widened load before a preceding store, causing incorrect program behavior. The fix replaces adjust_address with widen_memory_access, which correctly updates MEM_EXPR and resets MEM_ALIAS_SET, while also adjusting operand checking to avoid merging loads from different parameters.