Index: lib/Target/R600/AMDGPUISelLowering.cpp
===================================================================
--- lib/Target/R600/AMDGPUISelLowering.cpp
+++ lib/Target/R600/AMDGPUISelLowering.cpp
@@ -137,6 +137,9 @@
   if (!Subtarget->hasFP32Denormals())
     setOperationAction(ISD::FMAD, MVT::f32, Legal);
 
+  // Expand to fneg + fadd.
+  setOperationAction(ISD::FSUB, MVT::f64, Expand);
+
   // Lower floating point store/load to integer store/load to reduce the number
   // of patterns in tablegen.
   setOperationAction(ISD::STORE, MVT::f32, Promote);
Index: lib/Target/R600/SIISelLowering.cpp
===================================================================
--- lib/Target/R600/SIISelLowering.cpp
+++ lib/Target/R600/SIISelLowering.cpp
@@ -605,19 +605,8 @@
   switch (MI->getOpcode()) {
   default:
     return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
-  case AMDGPU::BRANCH: return BB;
-  case AMDGPU::V_SUB_F64: {
-    unsigned DestReg = MI->getOperand(0).getReg();
-    BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_ADD_F64), DestReg)
-      .addImm(0)  // SRC0 modifiers
-      .addReg(MI->getOperand(1).getReg())
-      .addImm(1)  // SRC1 modifiers
-      .addReg(MI->getOperand(2).getReg())
-      .addImm(0)  // CLAMP
-      .addImm(0); // OMOD
-    MI->eraseFromParent();
-    break;
-  }
+  case AMDGPU::BRANCH:
+    return BB;
   case AMDGPU::SI_RegisterStorePseudo: {
     MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
     unsigned Reg = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
Index: lib/Target/R600/SIInstructions.td
===================================================================
--- lib/Target/R600/SIInstructions.td
+++ lib/Target/R600/SIInstructions.td
@@ -1947,17 +1947,6 @@
 
 } // Uses = [EXEC,VCC,M0], Defs = [EXEC,VCC,M0]
 
-let usesCustomInserter = 1 in {
-
-def V_SUB_F64 : InstSI <
-  (outs VReg_64:$dst),
-  (ins VReg_64:$src0, VReg_64:$src1),
-  "v_sub_f64 $dst, $src0, $src1",
-  [(set f64:$dst, (fsub f64:$src0, f64:$src1))]
->;
-
-} // end usesCustomInserter
-
 multiclass SI_SPILL_SGPR <RegisterClass sgpr_class> {
 
   let UseNamedOperandTable = 1 in {
Index: test/CodeGen/R600/fneg-fabs.f64.ll
===================================================================
--- test/CodeGen/R600/fneg-fabs.f64.ll
+++ test/CodeGen/R600/fneg-fabs.f64.ll
@@ -5,9 +5,7 @@
 ; into 2 modifiers, although theoretically that should work.
 
 ; FUNC-LABEL: {{^}}fneg_fabs_fadd_f64:
-; SI: v_mov_b32_e32 [[IMMREG:v[0-9]+]], 0x7fffffff
-; SI: v_and_b32_e32 v[[FABS:[0-9]+]], {{s[0-9]+}}, [[IMMREG]]
-; SI: v_add_f64 {{v\[[0-9]+:[0-9]+\]}}, {{v\[[0-9]+:[0-9]+\]}}, -v{{\[[0-9]+}}:[[FABS]]{{\]}}
+; SI: v_add_f64 {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, -|v{{\[[0-9]+:[0-9]+\]}}|
 define void @fneg_fabs_fadd_f64(double addrspace(1)* %out, double %x, double %y) {
   %fabs = call double @llvm.fabs.f64(double %x)
   %fsub = fsub double -0.000000e+00, %fabs
Index: test/CodeGen/R600/fneg.f64.ll
===================================================================
--- test/CodeGen/R600/fneg.f64.ll
+++ test/CodeGen/R600/fneg.f64.ll
@@ -39,8 +39,7 @@
 ; unless the target returns true for isNegFree()
 
 ; FUNC-LABEL: {{^}}fneg_free_f64:
-; FIXME: Unnecessary copy to VGPRs
-; GCN: v_add_f64 {{v\[[0-9]+:[0-9]+\]}}, {{v\[[0-9]+:[0-9]+\]}}, -{{v\[[0-9]+:[0-9]+\]$}}
+; GCN: v_add_f64 {{v\[[0-9]+:[0-9]+\]}}, 0, -{{s\[[0-9]+:[0-9]+\]$}}
 define void @fneg_free_f64(double addrspace(1)* %out, i64 %in) {
   %bc = bitcast i64 %in to double
   %fsub = fsub double 0.0, %bc
Index: test/CodeGen/R600/fsub64.ll
===================================================================
--- test/CodeGen/R600/fsub64.ll
+++ test/CodeGen/R600/fsub64.ll
@@ -1,13 +1,107 @@
 ; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -check-prefix=SI %s
 ; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=SI %s
 
+declare double @llvm.fabs.f64(double) #0
+
 ; SI-LABEL: {{^}}fsub_f64:
 ; SI: v_add_f64 {{v\[[0-9]+:[0-9]+\], v\[[0-9]+:[0-9]+\], -v\[[0-9]+:[0-9]+\]}}
 define void @fsub_f64(double addrspace(1)* %out, double addrspace(1)* %in1,
                       double addrspace(1)* %in2) {
-   %r0 = load double addrspace(1)* %in1
-   %r1 = load double addrspace(1)* %in2
-   %r2 = fsub double %r0, %r1
-   store double %r2, double addrspace(1)* %out
-   ret void
+  %r0 = load double addrspace(1)* %in1
+  %r1 = load double addrspace(1)* %in2
+  %r2 = fsub double %r0, %r1
+  store double %r2, double addrspace(1)* %out
+  ret void
+}
+
+; SI-LABEL: {{^}}fsub_fabs_f64:
+; SI: v_add_f64 {{v\[[0-9]+:[0-9]+\], v\[[0-9]+:[0-9]+\], -\|v\[[0-9]+:[0-9]+\]\|}}
+define void @fsub_fabs_f64(double addrspace(1)* %out, double addrspace(1)* %in1,
+                           double addrspace(1)* %in2) {
+  %r0 = load double addrspace(1)* %in1
+  %r1 = load double addrspace(1)* %in2
+  %r1.fabs = call double @llvm.fabs.f64(double %r1) #0
+  %r2 = fsub double %r0, %r1.fabs
+  store double %r2, double addrspace(1)* %out
+  ret void
+}
+
+; SI-LABEL: {{^}}fsub_fabs_inv_f64:
+; SI: v_add_f64 {{v\[[0-9]+:[0-9]+\], |v\[[0-9]+:[0-9]+\]|, -v\[[0-9]+:[0-9]+\]}}
+define void @fsub_fabs_inv_f64(double addrspace(1)* %out, double addrspace(1)* %in1,
+                               double addrspace(1)* %in2) {
+  %r0 = load double addrspace(1)* %in1
+  %r1 = load double addrspace(1)* %in2
+  %r0.fabs = call double @llvm.fabs.f64(double %r0) #0
+  %r2 = fsub double %r0.fabs, %r1
+  store double %r2, double addrspace(1)* %out
+  ret void
+}
+
+; SI-LABEL: {{^}}s_fsub_f64:
+; SI: v_add_f64 {{v\[[0-9]+:[0-9]+\], s\[[0-9]+:[0-9]+\], -v\[[0-9]+:[0-9]+\]}}
+define void @s_fsub_f64(double addrspace(1)* %out, double %a, double %b) {
+  %sub = fsub double %a, %b
+  store double %sub, double addrspace(1)* %out
+  ret void
+}
+
+; SI-LABEL: {{^}}s_fsub_imm_f64:
+; SI: v_add_f64 {{v\[[0-9]+:[0-9]+\], 4.0, -s\[[0-9]+:[0-9]+\]}}
+define void @s_fsub_imm_f64(double addrspace(1)* %out, double %a, double %b) {
+  %sub = fsub double 4.0, %a
+  store double %sub, double addrspace(1)* %out
+  ret void
+}
+
+; SI-LABEL: {{^}}s_fsub_imm_inv_f64:
+; SI: v_add_f64 {{v\[[0-9]+:[0-9]+\], -4.0, s\[[0-9]+:[0-9]+\]}}
+define void @s_fsub_imm_inv_f64(double addrspace(1)* %out, double %a, double %b) {
+  %sub = fsub double %a, 4.0
+  store double %sub, double addrspace(1)* %out
+  ret void
+}
+
+; SI-LABEL: {{^}}s_fsub_self_f64:
+; SI: v_add_f64 {{v\[[0-9]+:[0-9]+\], s\[[0-9]+:[0-9]+\], -s\[[0-9]+:[0-9]+\]}}
+define void @s_fsub_self_f64(double addrspace(1)* %out, double %a) {
+  %sub = fsub double %a, %a
+  store double %sub, double addrspace(1)* %out
+  ret void
+}
+
+; SI-LABEL: {{^}}fsub_v2f64:
+; SI: v_add_f64 {{v\[[0-9]+:[0-9]+\], s\[[0-9]+:[0-9]+\], -v\[[0-9]+:[0-9]+\]}}
+; SI: v_add_f64 {{v\[[0-9]+:[0-9]+\], s\[[0-9]+:[0-9]+\], -v\[[0-9]+:[0-9]+\]}}
+define void @fsub_v2f64(<2 x double> addrspace(1)* %out, <2 x double> %a, <2 x double> %b) {
+  %sub = fsub <2 x double> %a, %b
+  store <2 x double> %sub, <2 x double> addrspace(1)* %out
+  ret void
+}
+
+; SI-LABEL: {{^}}fsub_v4f64:
+; SI: v_add_f64 {{v\[[0-9]+:[0-9]+\], v\[[0-9]+:[0-9]+\], -v\[[0-9]+:[0-9]+\]}}
+; SI: v_add_f64 {{v\[[0-9]+:[0-9]+\], v\[[0-9]+:[0-9]+\], -v\[[0-9]+:[0-9]+\]}}
+; SI: v_add_f64 {{v\[[0-9]+:[0-9]+\], v\[[0-9]+:[0-9]+\], -v\[[0-9]+:[0-9]+\]}}
+; SI: v_add_f64 {{v\[[0-9]+:[0-9]+\], v\[[0-9]+:[0-9]+\], -v\[[0-9]+:[0-9]+\]}}
+define void @fsub_v4f64(<4 x double> addrspace(1)* %out, <4 x double> addrspace(1)* %in) {
+  %b_ptr = getelementptr <4 x double> addrspace(1)* %in, i32 1
+  %a = load <4 x double> addrspace(1)* %in
+  %b = load <4 x double> addrspace(1)* %b_ptr
+  %result = fsub <4 x double> %a, %b
+  store <4 x double> %result, <4 x double> addrspace(1)* %out
+  ret void
 }
+
+; SI-LABEL: {{^}}s_fsub_v4f64:
+; SI: v_add_f64 {{v\[[0-9]+:[0-9]+\], s\[[0-9]+:[0-9]+\], -v\[[0-9]+:[0-9]+\]}}
+; SI: v_add_f64 {{v\[[0-9]+:[0-9]+\], s\[[0-9]+:[0-9]+\], -v\[[0-9]+:[0-9]+\]}}
+; SI: v_add_f64 {{v\[[0-9]+:[0-9]+\], s\[[0-9]+:[0-9]+\], -v\[[0-9]+:[0-9]+\]}}
+; SI: v_add_f64 {{v\[[0-9]+:[0-9]+\], s\[[0-9]+:[0-9]+\], -v\[[0-9]+:[0-9]+\]}}
+define void @s_fsub_v4f64(<4 x double> addrspace(1)* %out, <4 x double> %a, <4 x double> %b) {
+  %result = fsub <4 x double> %a, %b
+  store <4 x double> %result, <4 x double> addrspace(1)* %out, align 16
+  ret void
+}
+
+attributes #0 = { nounwind readnone }