| Issue |
126107
|
| Summary |
[MachineCP] Assertion `Reg.isPhysical()' failed in MCRegUnitIterator
|
| Labels |
bug,
llvm:regalloc
|
| Assignees |
ostannard
|
| Reporter |
aleks-tmb
|
During our local testing, we encountered the assertion failure `Reg.isPhysical()` in the `MachineCP` pass.
Reduced reproducer:
```
---
name: main
body: |
bb.0.entry:
liveins: $ymm7
renamable $ymm6 = COPY killed renamable $ymm7
CALL64r killed renamable $rax, csr_64_mostregs
renamable $ymm6 = VPADDWZ256rr $ymm6, $ymm6
```
Steps to reproduce:
```
$ bin/llc -run-pass machine-cp test.mir
llc: /root/llvm-project/llvm/include/llvm/MC/MCRegisterInfo.h:649: llvm::MCRegUnitIterator::MCRegUnitIterator(llvm::MCRegister, const llvm::MCRegisterInfo*): Assertion `Reg.isPhysical()' failed
```
Proof: https://godbolt.org/z/PnqdcKbfe
The problematic patch seems to be: https://github.com/llvm/llvm-project/commit/9e436c2daa446da05e9219f0e6a22f932ba8e3cb
Based on my investigation, the issue arises when we remove `renamable $ymm6 = COPY killed renamable $ymm7` due to regmask clobbering. However, during this process, we skip dropping the corresponding RegUnit from the tracking copies, since it's preserved:
```c++
LLVM_DEBUG(dbgs() << "MCP: Removing copy due to regmask clobbering: ";
MaybeDead->dump());
// Invalidate all entries in the copy map which are not preserved by
// this register mask.
for (unsigned RegUnit : TRI->regunits(Reg))
if (!PreservedRegUnits.test(RegUnit))
Tracker.clobberRegUnit(RegUnit, *TRI, *TII, UseCopyInstr);
// erase() will return the next valid iterator pointing to the next
// element after the erased one.
DI = MaybeDeadCopies.erase(DI);
MaybeDead->eraseFromParent();
```
As a result, we retain a pointer to the erased instruction in tracking copies, which leads to a failure when accessing it while processing `renamable $ymm6 = VPADDWZ256rr $ymm6, $ymm6`, clobbering any previous copy definitions.
@ostannard hi, could you please take a look?
![]()
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
