Index: llvm/lib/Analysis/InstructionSimplify.cpp
===================================================================
--- llvm/lib/Analysis/InstructionSimplify.cpp
+++ llvm/lib/Analysis/InstructionSimplify.cpp
@@ -2329,6 +2329,31 @@
     }
   }
 
+  // A funnel shift (rotate) can be decomposed into simpler shifts. See if we
+  // are mixing in another shift that is redundant with the funnel shift.
+
+  // (fshl X, ?, Y) | (shl X, Y) --> fshl X, ?, Y
+  // (shl X, Y) | (fshl X, ?, Y) --> fshl X, ?, Y
+  if (match(Op0,
+            m_Intrinsic<Intrinsic::fshl>(m_Value(X), m_Value(), m_Value(Y))) &&
+      match(Op1, m_Shl(m_Specific(X), m_Specific(Y))))
+    return Op0;
+  if (match(Op1,
+            m_Intrinsic<Intrinsic::fshl>(m_Value(X), m_Value(), m_Value(Y))) &&
+      match(Op0, m_Shl(m_Specific(X), m_Specific(Y))))
+    return Op1;
+
+  // (fshr ?, X, Y) | (lshr X, Y) --> fshr ?, X, Y
+  // (lshr X, Y) | (fshr ?, X, Y) --> fshr ?, X, Y
+  if (match(Op0,
+            m_Intrinsic<Intrinsic::fshr>(m_Value(), m_Value(X), m_Value(Y))) &&
+      match(Op1, m_LShr(m_Specific(X), m_Specific(Y))))
+    return Op0;
+  if (match(Op1,
+            m_Intrinsic<Intrinsic::fshr>(m_Value(), m_Value(X), m_Value(Y))) &&
+      match(Op0, m_LShr(m_Specific(X), m_Specific(Y))))
+    return Op1;
+
   if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, false))
     return V;
 
Index: llvm/test/Transforms/InstSimplify/or.ll
===================================================================
--- llvm/test/Transforms/InstSimplify/or.ll
+++ llvm/test/Transforms/InstSimplify/or.ll
@@ -1045,10 +1045,8 @@
 
 define i32 @or_shl_fshl(i32 %x, i32 %y, i32 %s) {
 ; CHECK-LABEL: @or_shl_fshl(
-; CHECK-NEXT:    [[SHY:%.*]] = shl i32 [[Y:%.*]], [[S:%.*]]
-; CHECK-NEXT:    [[FUN:%.*]] = call i32 @llvm.fshl.i32(i32 [[Y]], i32 [[X:%.*]], i32 [[S]])
-; CHECK-NEXT:    [[OR:%.*]] = or i32 [[FUN]], [[SHY]]
-; CHECK-NEXT:    ret i32 [[OR]]
+; CHECK-NEXT:    [[FUN:%.*]] = call i32 @llvm.fshl.i32(i32 [[Y:%.*]], i32 [[X:%.*]], i32 [[S:%.*]])
+; CHECK-NEXT:    ret i32 [[FUN]]
 ;
   %shy = shl i32 %y, %s
   %fun = call i32 @llvm.fshl.i32(i32 %y, i32 %x, i32 %s)
@@ -1058,10 +1056,8 @@
 
 define i32 @or_shl_fshl_commute(i32 %x, i32 %y, i32 %s) {
 ; CHECK-LABEL: @or_shl_fshl_commute(
-; CHECK-NEXT:    [[SHY:%.*]] = shl i32 [[Y:%.*]], [[S:%.*]]
-; CHECK-NEXT:    [[FUN:%.*]] = call i32 @llvm.fshl.i32(i32 [[Y]], i32 [[X:%.*]], i32 [[S]])
-; CHECK-NEXT:    [[OR:%.*]] = or i32 [[SHY]], [[FUN]]
-; CHECK-NEXT:    ret i32 [[OR]]
+; CHECK-NEXT:    [[FUN:%.*]] = call i32 @llvm.fshl.i32(i32 [[Y:%.*]], i32 [[X:%.*]], i32 [[S:%.*]])
+; CHECK-NEXT:    ret i32 [[FUN]]
 ;
   %shy = shl i32 %y, %s
   %fun = call i32 @llvm.fshl.i32(i32 %y, i32 %x, i32 %s)
@@ -1069,6 +1065,8 @@
   ret i32 %or
 }
 
+; negative test - fshl operands are not commutative
+
 define i32 @or_shl_fshl_wrong_order(i32 %x, i32 %y, i32 %s) {
 ; CHECK-LABEL: @or_shl_fshl_wrong_order(
 ; CHECK-NEXT:    [[SHY:%.*]] = shl i32 [[Y:%.*]], [[S:%.*]]
@@ -1084,10 +1082,8 @@
 
 define i32 @or_lshr_fshr(i32 %x, i32 %y, i32 %s) {
 ; CHECK-LABEL: @or_lshr_fshr(
-; CHECK-NEXT:    [[SHY:%.*]] = lshr i32 [[Y:%.*]], [[S:%.*]]
-; CHECK-NEXT:    [[FUN:%.*]] = call i32 @llvm.fshr.i32(i32 [[X:%.*]], i32 [[Y]], i32 [[S]])
-; CHECK-NEXT:    [[OR:%.*]] = or i32 [[FUN]], [[SHY]]
-; CHECK-NEXT:    ret i32 [[OR]]
+; CHECK-NEXT:    [[FUN:%.*]] = call i32 @llvm.fshr.i32(i32 [[X:%.*]], i32 [[Y:%.*]], i32 [[S:%.*]])
+; CHECK-NEXT:    ret i32 [[FUN]]
 ;
   %shy = lshr i32 %y, %s
   %fun = call i32 @llvm.fshr.i32(i32 %x, i32 %y, i32 %s)
@@ -1097,10 +1093,8 @@
 
 define i32 @or_lshr_fshr_commute(i32 %x, i32 %y, i32 %s) {
 ; CHECK-LABEL: @or_lshr_fshr_commute(
-; CHECK-NEXT:    [[SHY:%.*]] = lshr i32 [[Y:%.*]], [[S:%.*]]
-; CHECK-NEXT:    [[FUN:%.*]] = call i32 @llvm.fshr.i32(i32 [[X:%.*]], i32 [[Y]], i32 [[S]])
-; CHECK-NEXT:    [[OR:%.*]] = or i32 [[SHY]], [[FUN]]
-; CHECK-NEXT:    ret i32 [[OR]]
+; CHECK-NEXT:    [[FUN:%.*]] = call i32 @llvm.fshr.i32(i32 [[X:%.*]], i32 [[Y:%.*]], i32 [[S:%.*]])
+; CHECK-NEXT:    ret i32 [[FUN]]
 ;
   %shy = lshr i32 %y, %s
   %fun = call i32 @llvm.fshr.i32(i32 %x, i32 %y, i32 %s)
@@ -1108,6 +1102,8 @@
   ret i32 %or
 }
 
+; negative test - fshr operands are not commutative
+
 define i32 @or_lshr_fshr_wrong_order(i32 %x, i32 %y, i32 %s) {
 ; CHECK-LABEL: @or_lshr_fshr_wrong_order(
 ; CHECK-NEXT:    [[SHY:%.*]] = lshr i32 [[Y:%.*]], [[S:%.*]]