[AMDGPU] Convert mul24 sdwa detection to tablegen predicates
Open
backend:AMDGPU
- Dominant language
- LLVM
- Stars
- 40.5k
- Forks
- 18.7k
- PR merge metrics
- PR metrics pending
Description
#193033 allowed detection of operands to a mul24 node that would probably fold using SDWA patterns to prevent folding to mad24 (which can't use SDWA) - but its SelectionDAG-only.
This ticket tracks conversion to a purely-tablegen predicate to allow us to remove the AMDGPUDAGToDAGISel::isSDWAOperand hook and support globalisel as well (with suitable additional test coverage).
Contributor guide
Assessment
This issue has not been assessed yet.