| Issue |
178867
|
| Summary |
[CodeGen] Crash in SplitKit::addDeadDef when child interval has subrange not in parent
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
kdupontbe
|
Summary
The greedy register allocator crashes in SplitKit::addDeadDef() when splitting a register with partial subranges. The crash occurs when a child interval has a subrange that doesn't exist in the parent interval at the split point.
Environment
LLVM Version: main branch (as of January 2026)
Target: RISC-V 64-bit with vector extensions
Affected File: llvm/lib/CodeGen/SplitKit.cpp
Function: SplitEditor::addDeadDef()
Root Cause
The old code iterates over child subranges and looks them up in the parent:
for (LiveInterval::SubRange &S : LI.subranges()) {
auto &PS = getSubRangeForMask(S.LaneMask, Edit->getParent());
// ↑ Calls llvm_unreachable() if child has lane parent doesn't have
VNInfo *PV = PS.getVNInfoAt(Def);
if (PV != nullptr && PV->def == Def)
S.createDeadDef(Def, LIS.getVNInfoAllocator());
}
This crashes when:
Parent register is split, creating a child interval
Child interval has a subrange (lane mask) that parent doesn't have
getSubRangeForMask() calls llvm_unreachable("SubRange for this mask not found")
I could not find a way to reproduce this with a standalone MIR
Dumping the MIR right before the crash would occur
[splitkit_crash_state.txt](https://github.com/user-attachments/files/24963332/splitkit_crash_state.txt)
SPLITKIT CRASH IMMINENT - MIR DUMP
Function: o
Parent Register: %288 (VRM8)
Child Register: %310 (VRM8)
Def SlotIndex: 2112r
PARENT LiveInterval:
Subranges:
LaneMask: 0000000000000004 (0x0004) ← Parent has only lane 0x0004
CHILD LiveInterval:
Subranges:
LaneMask: 0000000000000002 (0x0002) ← Child has lane 0x0002
*** NOT FOUND in parent - WILL CRASH ***
LaneMask: 0000000000000004 (0x0004)
FOUND in parent
The crash occurs at:
Instruction: undef %288.sub_vrm1_1:vrm8 = INIT_UNDEF
Slot Index: 2112r
Problem: Child has lane 0x0002 which parent doesn't have
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
