Add non-null check before accessing pointer

[MartinWehking]

Add a check if RC is not null to ensure that a consecutive access is safe.

A static analyzer flagged this issue since hasVectorRegisters potentially dereferences RC.

[llvmbot]

@llvm/pr-subscribers-backend-amdgpu

Author: Martin Wehking (MartinWehking)

Changes

Add a check if RC is not null to ensure that a consecutive access is safe.

A static analyzer flagged this issue since hasVectorRegisters potentially dereferences RC.

---

Full diff: https://github.com/llvm/llvm-project/pull/83459.diff

1 Files Affected:

• (modified) llvm/lib/Target/AMDGPU/SIFoldOperands.cpp (+1-1)

diff --git a/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp b/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
index 634b4aeb30a730..1cef4f1fa81e3a 100644
--- a/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
+++ b/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
@@ -1054,7 +1054,7 @@ void SIFoldOperands::foldOperand(
       // Don't fold if OpToFold doesn't hold an aligned register.
       const TargetRegisterClass *RC =
           TRI->getRegClassForReg(*MRI, OpToFold.getReg());
-      if (TRI->hasVectorRegisters(RC) && OpToFold.getSubReg()) {
+      if (RC && TRI->hasVectorRegisters(RC) && OpToFold.getSubReg()) {
         unsigned SubReg = OpToFold.getSubReg();
         if (const TargetRegisterClass *SubRC =
                 TRI->getSubRegisterClass(RC, SubReg))

[arsenm]

There's no context that can reach here where this can fail

[MartinWehking]

Is getRegClassForReg always returning a non-null pointer for every possible context?
When looking at the body of the function:
return Reg.isVirtual() ? MRI.getRegClass(Reg) : getPhysRegBaseClass(Reg);
I spotted that getPhysRegBaseClass(Reg) might return a nullptr.

[arsenm]

This would fail for unallocatable physical registers which shouldn't have been copy folding candidates to begin with

[MartinWehking]

Ok, then we can close this PR again

[arsenm]

Will adding an assert work?

[MartinWehking]

Yes an assert should also silence the issue:1af318e