Index: lib/CodeGen/SelectionDAG/DAGCombiner.cpp
===================================================================
--- lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -8273,6 +8273,12 @@
     }
   }
 
+  // For reciprocal optimization of a division node, check the division
+  // node's optimization flags as well as the function-wide fast-math setting.
+  bool AllowRecip = false;
+  if (const auto *NodeWithFlags = dyn_cast<BinaryWithFlagsSDNode>(N))
+    AllowRecip = NodeWithFlags->Flags.hasAllowReciprocal();
+  
   // Combine multiple FDIVs with the same divisor into multiple FMULs by the
   // reciprocal.
   // E.g., (a / D; b / D;) -> (recip = 1.0 / D; a * recip; b * recip)
@@ -8280,7 +8286,7 @@
   // may have different costs for FDIV and FMUL, so sometimes the cost of two
   // FDIVs may be lower than the cost of one FDIV and two FMULs. Another reason
   // is the critical path is increased from "one FDIV" to "one FDIV + one FMUL".
-  if (Options.UnsafeFPMath) {
+  if (Options.UnsafeFPMath || AllowRecip) {
     // Skip if current node is a reciprocal.
     if (N0CFP && N0CFP->isExactlyValue(1.0))
       return SDValue();
@@ -8292,7 +8298,9 @@
          UI != UE; ++UI) {
       SDNode *User = UI.getUse().getUser();
       if (User->getOpcode() == ISD::FDIV && User->getOperand(1) == N1)
-        Users.push_back(User);
+        if (const auto *NodeWithFlags = dyn_cast<BinaryWithFlagsSDNode>(User))
+          if (Options.UnsafeFPMath || NodeWithFlags->Flags.hasAllowReciprocal())
+            Users.push_back(User);
     }
 
     if (TLI.combineRepeatedFPDivisors(Users.size())) {
Index: test/CodeGen/X86/fdiv-combine.ll
===================================================================
--- test/CodeGen/X86/fdiv-combine.ll
+++ test/CodeGen/X86/fdiv-combine.ll
@@ -1,9 +1,11 @@
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown | FileCheck %s
 
-; Anything more than one division using a single divisor operand
+; More than one 'arcp' division using a single divisor operand
 ; should be converted into a reciprocal and multiplication.
 
-define float @div1_arcp(float %x, float %y, float %z) #0 {
+; Don't do anything for just one division.
+;
+define float @div1_arcp(float %x, float %y, float %z) {
 ; CHECK-LABEL: div1_arcp:
 ; CHECK:       # BB#0:
 ; CHECK-NEXT:    divss %xmm1, %xmm0
@@ -12,8 +14,10 @@
   ret float %div1
 }
 
-define float @div2_arcp(float %x, float %y, float %z) #0 {
-; CHECK-LABEL: div2_arcp:
+; All math instructions are 'arcp', so optimize.
+;
+define float @div2_arcp_all(float %x, float %y, float %z) {
+; CHECK-LABEL: div2_arcp_all:
 ; CHECK:       # BB#0:
 ; CHECK-NEXT:    movss {{.*#+}} xmm3 = mem[0],zero,zero,zero
 ; CHECK-NEXT:    divss %xmm2, %xmm3
@@ -27,5 +31,50 @@
   ret float %div2
 }
 
-; FIXME: If the backend understands 'arcp', then this attribute is unnecessary.
-attributes #0 = { "unsafe-fp-math"="true" }
+; The first division is not 'arcp', so do not optimize.
+;
+define float @div2_arcp_partial1(float %x, float %y, float %z) {
+; CHECK-LABEL: div2_arcp_partial1:
+; CHECK:       # BB#0:
+; CHECK-NEXT:    divss %xmm2, %xmm0
+; CHECK-NEXT:    mulss %xmm1, %xmm0
+; CHECK-NEXT:    divss %xmm2, %xmm0
+; CHECK-NEXT:    retq
+  %div1 = fdiv float %x, %z
+  %mul = fmul arcp float %div1, %y
+  %div2 = fdiv arcp float %mul, %z
+  ret float %div2
+}
+
+; The second division is not 'arcp', so do not optimize.
+;
+define float @div2_arcp_partial2(float %x, float %y, float %z) {
+; CHECK-LABEL: div2_arcp_partial2:
+; CHECK:       # BB#0:
+; CHECK-NEXT:    divss %xmm2, %xmm0
+; CHECK-NEXT:    mulss %xmm1, %xmm0
+; CHECK-NEXT:    divss %xmm2, %xmm0
+; CHECK-NEXT:    retq
+  %div1 = fdiv arcp float %x, %z
+  %mul = fmul arcp float %div1, %y
+  %div2 = fdiv float %mul, %z
+  ret float %div2
+}
+
+; The multiply is not 'arcp', but that does not prevent optimizing the divisions.
+;
+define float @div2_arcp_partial3(float %x, float %y, float %z) {
+; CHECK-LABEL: div2_arcp_partial3:
+; CHECK:       # BB#0:
+; CHECK-NEXT:    movss {{.*#+}} xmm3 = mem[0],zero,zero,zero
+; CHECK-NEXT:    divss %xmm2, %xmm3
+; CHECK-NEXT:    mulss %xmm3, %xmm0
+; CHECK-NEXT:    mulss %xmm1, %xmm0
+; CHECK-NEXT:    mulss %xmm3, %xmm0
+; CHECK-NEXT:    retq
+  %div1 = fdiv arcp float %x, %z
+  %mul = fmul float %div1, %y
+  %div2 = fdiv arcp float %mul, %z
+  ret float %div2
+}
+