| Issue |
107702
|
| Summary |
MachineCopyPropagation remove copy instruction due to no live-out succ, but the def is indeed used.
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
zzk0
|
Suppose we have the following scenario:
```
Before MachineCopyPropagation Pass:
bb.1: (no successors)
B = COPY C
early-clobber A = OP B (B is a sub-register of A, and B is not renamable.)
After MachineCopyPropagation Pass:
bb.1: (no successors)
early-clobber A = OP B (B is a sub-register of A, and B is not renamable.)
```
Is this reasonable?
The following might be somewhat difficult to understand. If you feel confused, please let me know. According to the code, we encounter the following situation:
1. `B = COPY C` is added to MaybeDeadCopies.
2. The def register in the instruction `early-clobber A = OP B` is checked, and Tracker.clobberRegister is called to remove A's RegUnit (including B) from the CopyTracker map.
3. When reading the operand `early-clobber A = OP B`, ReadRegister is called for B. However, at this point, B has already been removed from the CopyTracker, so the copy instruction `B = COPY C` is not removed from MaybeDeadCopies.
As a result, B = COPY C is mistakenly deleted when there are no successors.
Any suggestions would be appreciated.
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
