AMDGPU: Add disjoint to or produced from lowering vector ops

[arsenm]

No description provided.

[arsenm]

• AMDGPU: Add disjoint to or produced from lowering vector ops #122424 👈 (View in Graphite)
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[llvmbot]

@llvm/pr-subscribers-backend-amdgpu

Author: Matt Arsenault (arsenm)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/122424.diff

1 Files Affected:

• (modified) llvm/lib/Target/AMDGPU/SIISelLowering.cpp (+4-2)

diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 992f7ed99d3bb7..5a7ff81e08064e 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -7429,7 +7429,8 @@ SDValue SITargetLowering::lowerINSERT_VECTOR_ELT(SDValue Op,
       DAG.getNode(ISD::AND, SL, IntVT, DAG.getNOT(SL, BFM, IntVT), BCVec);
 
   // 4. Get (2) and (3) ORed into the target vector.
-  SDValue BFI = DAG.getNode(ISD::OR, SL, IntVT, LHS, RHS);
+  SDValue BFI =
+      DAG.getNode(ISD::OR, SL, IntVT, LHS, RHS, SDNodeFlags::Disjoint);
 
   return DAG.getNode(ISD::BITCAST, SL, VecVT, BFI);
 }
@@ -7637,7 +7638,8 @@ SDValue SITargetLowering::lowerBUILD_VECTOR(SDValue Op,
     Lo = DAG.getNode(ISD::BITCAST, SL, MVT::i16, Lo);
     Lo = DAG.getNode(ISD::ZERO_EXTEND, SL, MVT::i32, Lo);
 
-    SDValue Or = DAG.getNode(ISD::OR, SL, MVT::i32, Lo, ShlHi);
+    SDValue Or =
+        DAG.getNode(ISD::OR, SL, MVT::i32, Lo, ShlHi, SDNodeFlags::Disjoint);
     return DAG.getNode(ISD::BITCAST, SL, VT, Or);
   }
 

[jayfoad]

Makes sense to me.

[shiltian]

surprised that no test change is needed

[arsenm]

surprised that no test change is needed

It's not surprising, this rarely does anything. The only times it's really useful is when the or appears in an addressing expression, which won't really appear as a user of a vector expansion