Index: lib/Target/R600/AMDGPUISelLowering.h
===================================================================
--- lib/Target/R600/AMDGPUISelLowering.h
+++ lib/Target/R600/AMDGPUISelLowering.h
@@ -146,6 +146,10 @@
   SDValue CombineMinMax(SDNode *N, SelectionDAG &DAG) const;
   const char* getTargetNodeName(unsigned Opcode) const override;
 
+  SDValue getRsqrtEstimate(SDValue Operand,
+                           DAGCombinerInfo &DCI,
+                           unsigned &RefinementSteps) const override;
+
   virtual SDNode *PostISelFolding(MachineSDNode *N,
                                   SelectionDAG &DAG) const {
     return N;
Index: lib/Target/R600/AMDGPUISelLowering.cpp
===================================================================
--- lib/Target/R600/AMDGPUISelLowering.cpp
+++ lib/Target/R600/AMDGPUISelLowering.cpp
@@ -2325,6 +2325,23 @@
   }
 }
 
+SDValue AMDGPUTargetLowering::getRsqrtEstimate(SDValue Operand,
+                                               DAGCombinerInfo &DCI,
+                                               unsigned &RefinementSteps) const {
+  SelectionDAG &DAG = DCI.DAG;
+  EVT VT = Operand.getValueType();
+
+  if (VT == MVT::f32) {
+    RefinementSteps = 0;
+    return DAG.getNode(AMDGPUISD::RSQ, SDLoc(Operand), VT, Operand);
+  }
+
+  // TODO: There is also f64 rsq instruction, but the documentation is less
+  // clear on its precision.
+
+  return SDValue();
+}
+
 static void computeKnownBitsForMinMax(const SDValue Op0,
                                       const SDValue Op1,
                                       APInt &KnownZero,
Index: test/CodeGen/R600/fsqrt.ll
===================================================================
--- test/CodeGen/R600/fsqrt.ll
+++ test/CodeGen/R600/fsqrt.ll
@@ -1,4 +1,7 @@
-; RUN: llc < %s -march=r600 -mcpu=tahiti -verify-machineinstrs | FileCheck %s
+; RUN: llc -march=r600 -mcpu=tahiti -verify-machineinstrs < %s | FileCheck %s
+; RUN: llc -march=r600 -mcpu=tahiti -verify-machineinstrs -enable-unsafe-fp-math < %s | FileCheck %s
+
+; Run with unsafe-fp-math to make sure nothing tries to turn this into 1 / rsqrt(x)
 
 ; CHECK: {{^}}fsqrt_f32:
 ; CHECK: V_SQRT_F32_e32 {{v[0-9]+, v[0-9]+}}
Index: test/CodeGen/R600/rsq.ll
===================================================================
--- test/CodeGen/R600/rsq.ll
+++ test/CodeGen/R600/rsq.ll
@@ -1,6 +1,7 @@
 ; RUN: llc -march=r600 -mcpu=SI -mattr=-fp32-denormals -verify-machineinstrs -enable-unsafe-fp-math < %s | FileCheck -check-prefix=SI-UNSAFE -check-prefix=SI %s
 ; RUN: llc -march=r600 -mcpu=SI -mattr=-fp32-denormals -verify-machineinstrs < %s | FileCheck -check-prefix=SI-SAFE -check-prefix=SI %s
 
+declare i32 @llvm.r600.read.tidig.x() nounwind readnone
 declare float @llvm.sqrt.f32(float) nounwind readnone
 declare double @llvm.sqrt.f64(double) nounwind readnone
 
@@ -36,3 +37,38 @@
   store float %div, float addrspace(1)* %out, align 4
   ret void
 }
+
+; Recognize that this is rsqrt(a) * rcp(b) * c,
+; not 1 / ( 1 / sqrt(a)) * rcp(b) * c.
+
+; SI-LABEL: @rsqrt_fmul
+; SI-DAG: BUFFER_LOAD_DWORD [[A:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: BUFFER_LOAD_DWORD [[B:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:0x4
+; SI-DAG: BUFFER_LOAD_DWORD [[C:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:0x8
+
+; SI-UNSAFE-DAG: V_RSQ_F32_e32 [[RSQA:v[0-9]+]], [[A]]
+; SI-UNSAFE-DAG: V_RCP_F32_e32 [[RCPB:v[0-9]+]], [[B]]
+; SI-UNSAFE-DAG: V_MUL_F32_e32 [[TMP:v[0-9]+]], [[RCPB]], [[RSQA]]
+; SI-UNSAFE: V_MUL_F32_e32 [[RESULT:v[0-9]+]], [[TMP]], [[C]]
+; SI-UNSAFE: BUFFER_STORE_DWORD [[RESULT]]
+
+; SI-SAFE-NOT: V_RSQ_F32
+
+; SI: S_ENDPGM
+define void @rsqrt_fmul(float addrspace(1)* %out, float addrspace(1)* %in) {
+  %tid = call i32 @llvm.r600.read.tidig.x() nounwind readnone
+  %out.gep = getelementptr float addrspace(1)* %out, i32 %tid
+  %gep.0 = getelementptr float addrspace(1)* %in, i32 %tid
+  %gep.1 = getelementptr float addrspace(1)* %gep.0, i32 1
+  %gep.2 = getelementptr float addrspace(1)* %gep.0, i32 2
+
+  %a = load float addrspace(1)* %gep.0
+  %b = load float addrspace(1)* %gep.1
+  %c = load float addrspace(1)* %gep.2
+
+  %x = call float @llvm.sqrt.f32(float %a)
+  %y = fmul float %x, %b
+  %z = fdiv float %c, %y
+  store float %z, float addrspace(1)* %out.gep
+  ret void
+}