[GreedyRA] Assert triggered: "Non-empty but used interval"

[serguei-katkov]

The test below triggers an assert in greedy register allocator.

At the moment I see the following reason of an assert:

1. During last chance recoloring some live interval is unassigned from physical register
2. Then it is split by region split and completely replaces by others live intervals corresponding to new registers. On this stage instructions are updated.
3. If re-coloring failed, the rollback is started. On this stage the original live interval corresponding to original register which does not have any uses now is assigned back to original physical register (it looks very strange).
4. Later on handling another live interval , our live interval is evicted.
5. Finally RegAllocBase in allocatePhysRegs complains with assert that Live interval has no nodbg uses but interval is not empty.

How to reproduce:

$ llc -x mir -mattr="+reserve-x28" --start-before=greedy 5.mir
...RegAllocBase.cpp:152: void llvm::RegAllocBase::allocatePhysRegs(): Assertion `SplitVirtReg->empty() && "Non-empty but used interval"' failed

The reproducer is attached as 5.txt in the next comment.

[serguei-katkov]

The reproducer:
5.txt

[BlaM]

Why do I get email notifications for this issue? I don't even know the project?

[arsenm]

Could it be picking up "@BlaM" from the attached testcase?

[serguei-katkov]

Could it be picking up "@BlaM" from the attached testcase?

What does it mean?

[kito-cheng]

Could it be picking up "@BlaM" from the attached testcase?
What does it mean?

I guess he mean there is a function called @blam in attached testcase(5.txt) and that might be the reason why @BlaM got mail :P

[serguei-katkov]

Fixed with
commit 095bf6b (HEAD -> ZVM-23622-2, origin/main, origin/HEAD)
Author: Serguei Katkov serguei.katkov@azul.com
Date: Wed Jun 8 16:21:01 2022 +0700

[Greedy RegAlloc] Fix the handling of split register in last chance re-coloring.

This is a fix for https://github.com/llvm/llvm-project/issues/55827.

When register we are trying to re-color is split the original register (we tried to recover)
has no uses after the split. However in rollback actions we assign back physical register to it.
Later it causes different assertions. One of them is in attached test.

This CL fixes this by avoiding assigning physical register back to register which has no usage
or its live interval now is empty.

Reviewed By: arsenm, qcolombet
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D127281

[serguei-katkov]

Fixed.