diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index fc0f1d91f80059..3ad0a66ace92a7 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -7366,6 +7366,14 @@ SDValue SITargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op,
     DAG.setNodeMemRefs(NewNode, {MemRef});
     return SDValue(NewNode, 0);
   }
+  case Intrinsic::amdgcn_waterfall_readfirstlane: {
+    if (!Op->getOperand(3)->isDivergent()) {
+      // If waterfall_readfirstlane is uniform, it can be removed
+      DAG.ReplaceAllUsesWith(Op.getNode(), Op->getOperand(3).getNode());
+      return SDValue();
+    }
+    return Op;
+  }
   default:
     if (const AMDGPU::ImageDimIntrinsicInfo *ImageDimIntr =
             AMDGPU::getImageDimIntrinsicInfo(IntrID))
@@ -7690,7 +7698,6 @@ SDValue SITargetLowering::LowerINTRINSIC_VOID(SDValue Op,
   case Intrinsic::amdgcn_end_cf:
     return SDValue(DAG.getMachineNode(AMDGPU::SI_END_CF, DL, MVT::Other,
                                       Op->getOperand(2), Chain), 0);
-
   default: {
     if (const AMDGPU::ImageDimIntrinsicInfo *ImageDimIntr =
             AMDGPU::getImageDimIntrinsicInfo(IntrinsicID))
diff --git a/llvm/lib/Target/AMDGPU/SIInsertWaterfall.cpp b/llvm/lib/Target/AMDGPU/SIInsertWaterfall.cpp
index 58ec0c292fecbc..072d52f2faefeb 100644
--- a/llvm/lib/Target/AMDGPU/SIInsertWaterfall.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInsertWaterfall.cpp
@@ -370,6 +370,9 @@ bool SIInsertWaterfall::removeRedundantWaterfall(WaterfallWorkitem &Item) {
   // If all the readfirstlane intrinsics are actually for uniform values and 
   // the token used in the begin/end isn't used in anything else the waterfall
   // can be removed.
+  // Alternatively, prior passes may have removed the readfirstlane intrinsics
+  // altogether, in this case the begin/end intrinsics are now redundant and can
+  // also be removed.
   // The readfirstlane intrinsics are replaced with the uniform source value,
   // the loop is removed and the defs in the end intrinsics are just replaced with
   // the input operands
@@ -399,6 +402,8 @@ bool SIInsertWaterfall::removeRedundantWaterfall(WaterfallWorkitem &Item) {
     }
   }
 
+  // Note: this test also returns true when there are NO RFL intrinsics, the
+  // case where a prior pass has removed all of them and the loop is now redundant
   if (Removed == Item.RFLList.size()) {
     // Removed all of the RFLs
     // We can remove the waterfall loop entirely
@@ -468,16 +473,16 @@ bool SIInsertWaterfall::processWaterfall(MachineBasicBlock &MBB) {
                  "Linked WATERFALL pseudo ops found in different BBs");
         });
 
-    assert(Item.RFLList.size() &&
-           (Item.EndList.size() || Item.LastUseList.size()) &&
-           "SI_WATERFALL* pseudo instruction group must have at least 1 of "
-           "each type");
-
     if (removeRedundantWaterfall(Item)) {
       Changed = true;
       continue;
     }
 
+    assert(Item.RFLList.size() &&
+           (Item.EndList.size() || Item.LastUseList.size()) &&
+           "SI_WATERFALL* pseudo instruction group must have at least 1 of "
+           "each type");
+
     // Insert the waterfall loop code around the identified region of
     // instructions
     // Loop starts at the SI_WATERFALL_BEGIN