| Issue |
181708
|
| Summary |
[AMDGPU] Invalid sink of ds_load across workgroup fence
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
AlexAUT
|
In this [example](https://godbolt.org/z/K9d977EeW), `machine-sink` ends up sinking the `ds_load` at `~L200` past an `atomic fence` and a `barrier`. This looks like an error to me because the input IR has a workgroup fence right after that addrspace(3) load, so moving it should be illegal?
To me it looks like machine-sink never looks at instructions in the same block so we are missing this fence, but I am not familiar with this pass?
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
