[llvm] [AMDGPU] Stop combining arbitrary offsets into PAL relocs (PR #80034)

[Matt Arsenault via llvm-commits]

================
@@ -7133,6 +7133,17 @@ SDValue SITargetLowering::lowerBUILD_VECTOR(SDValue Op,
 
 bool
 SITargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
+  // Subtargets that use ELF REL relocations (instead of RELA) can only store a
----------------
arsenm wrote:

not a subtarget?

https://github.com/llvm/llvm-project/pull/80034