| Issue |
172100
|
| Summary |
[SDAG] Masking vector lanes instead of truncating
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
natanelh-mobileye
|
consider the following pattern:
`(v4i16 (z/sext (foo (trunc v4i16:a)))) -> (mask unused lanes (bitcast (v8i8: foo)))`
masking is potentially cheaper than the shuffling caused by `trunc` especially when there are more than one inputs to `foo`. what should be the limitations for such an optimization?
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
