diff --git a/lib/Target/R600/SIISelLowering.cpp b/lib/Target/R600/SIISelLowering.cpp
index 12677f0e062..71aec8400d7 100644
--- a/lib/Target/R600/SIISelLowering.cpp
+++ b/lib/Target/R600/SIISelLowering.cpp
@@ -636,6 +636,17 @@ MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
   return BB;
 }
 
+bool SITargetLowering::enableAggressiveFMAFusion(EVT VT) const {
+  // This currently forces unfolding various combinations of fsub into fma with
+  // free fneg'd operands. As long as we have fast FMA (controlled by
+  // isFMAFasterThanFMulAndFAdd), we should perform these.
+
+  // When fma is quarter rate, for f64 where add / sub are at best half rate,
+  // most of these combines appear to be cycle neutral but save on instruction
+  // count / code size.
+  return true;
+}
+
 EVT SITargetLowering::getSetCCResultType(LLVMContext &Ctx, EVT VT) const {
   if (!VT.isVector()) {
     return MVT::i1;
@@ -647,6 +658,21 @@ MVT SITargetLowering::getScalarShiftAmountTy(EVT VT) const {
   return MVT::i32;
 }
 
+// Answering this is somewhat tricky and depends on the specific device which
+// have different rates for fma or all f64 operations.
+//
+// v_fma_f64 and v_mul_f64 always take the same number of cycles as each other
+// regardless of which device (although the number of cycles differs between
+// devices), so it is always profitable for f64.
+//
+// v_fma_f32 takes 4 or 16 cycles depending on the device, so it is profitable
+// only on full rate devices. Normally, we should prefer selecting v_mad_f32
+// which we can always do even without fused FP ops since it returns the same
+// result as the separate operations and since it is always full
+// rate. Therefore, we lie and report that it is not faster for f32. v_mad_f32
+// however does not support denormals, so we do report fma as faster if we have
+// a fast fma device and require denormals.
+//
 bool SITargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
   VT = VT.getScalarType();
 
@@ -655,7 +681,10 @@ bool SITargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
 
   switch (VT.getSimpleVT().SimpleTy) {
   case MVT::f32:
-    return Subtarget->hasFastFMAF32();
+    // This is as fast on some subtargets. However, we always have full rate f32
+    // mad available which returns the same result as the separate operations
+    // which we should prefer over fma.
+    return false;
   case MVT::f64:
     return true;
   default:
diff --git a/lib/Target/R600/SIISelLowering.h b/lib/Target/R600/SIISelLowering.h
index 876fd8c9f36..bb003f46f4b 100644
--- a/lib/Target/R600/SIISelLowering.h
+++ b/lib/Target/R600/SIISelLowering.h
@@ -95,6 +95,7 @@ public:
 
   MachineBasicBlock * EmitInstrWithCustomInserter(MachineInstr * MI,
                                       MachineBasicBlock * BB) const override;
+  bool enableAggressiveFMAFusion(EVT VT) const override;
   EVT getSetCCResultType(LLVMContext &Context, EVT VT) const override;
   MVT getScalarShiftAmountTy(EVT VT) const override;
   bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
diff --git a/test/CodeGen/R600/fma-combine.ll b/test/CodeGen/R600/fma-combine.ll
new file mode 100644
index 00000000000..9aac90cb953
--- /dev/null
+++ b/test/CodeGen/R600/fma-combine.ll
@@ -0,0 +1,368 @@
+; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs -fp-contract=fast < %s | FileCheck -check-prefix=SI-FASTFMAF -check-prefix=SI -check-prefix=FUNC %s
+; RUN: llc -march=amdgcn -mcpu=verde -verify-machineinstrs -fp-contract=fast < %s | FileCheck -check-prefix=SI-SLOWFMAF -check-prefix=SI -check-prefix=FUNC %s
+
+declare i32 @llvm.r600.read.tidig.x() #0
+declare double @llvm.fabs.f64(double) #0
+declare double @llvm.fma.f64(double, double, double) #0
+declare float @llvm.fma.f32(float, float, float) #0
+
+; (fadd (fmul x, y), z) -> (fma x, y, z)
+; FUNC-LABEL: {{^}}combine_to_fma_f64_0:
+; SI-DAG: buffer_load_dwordx2 [[A:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: buffer_load_dwordx2 [[B:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
+; SI-DAG: buffer_load_dwordx2 [[C:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}}
+; SI: v_fma_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[A]], [[B]], [[C]]
+; SI: buffer_store_dwordx2 [[RESULT]]
+define void @combine_to_fma_f64_0(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) #1 {
+  %tid = tail call i32 @llvm.r600.read.tidig.x() #0
+  %gep.0 = getelementptr double addrspace(1)* %in, i32 %tid
+  %gep.1 = getelementptr double addrspace(1)* %gep.0, i32 1
+  %gep.2 = getelementptr double addrspace(1)* %gep.0, i32 2
+  %gep.out = getelementptr double addrspace(1)* %out, i32 %tid
+
+  %a = load double addrspace(1)* %gep.0
+  %b = load double addrspace(1)* %gep.1
+  %c = load double addrspace(1)* %gep.2
+
+  %mul = fmul double %a, %b
+  %fma = fadd double %mul, %c
+  store double %fma, double addrspace(1)* %gep.out
+  ret void
+}
+
+; (fadd (fmul x, y), z) -> (fma x, y, z)
+; FUNC-LABEL: {{^}}combine_to_fma_f64_0_2use:
+; SI-DAG: buffer_load_dwordx2 [[A:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: buffer_load_dwordx2 [[B:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
+; SI-DAG: buffer_load_dwordx2 [[C:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}}
+; SI-DAG: buffer_load_dwordx2 [[D:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:24{{$}}
+; SI-DAG: v_fma_f64 [[RESULT0:v\[[0-9]+:[0-9]+\]]], [[A]], [[B]], [[C]]
+; SI-DAG: v_fma_f64 [[RESULT1:v\[[0-9]+:[0-9]+\]]], [[A]], [[B]], [[D]]
+; SI-DAG: buffer_store_dwordx2 [[RESULT0]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: buffer_store_dwordx2 [[RESULT1]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
+; SI: s_endpgm
+define void @combine_to_fma_f64_0_2use(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) #1 {
+  %tid = tail call i32 @llvm.r600.read.tidig.x() #0
+  %gep.0 = getelementptr double addrspace(1)* %in, i32 %tid
+  %gep.1 = getelementptr double addrspace(1)* %gep.0, i32 1
+  %gep.2 = getelementptr double addrspace(1)* %gep.0, i32 2
+  %gep.3 = getelementptr double addrspace(1)* %gep.0, i32 3
+  %gep.out.0 = getelementptr double addrspace(1)* %out, i32 %tid
+  %gep.out.1 = getelementptr double addrspace(1)* %gep.out.0, i32 1
+
+  %a = load double addrspace(1)* %gep.0
+  %b = load double addrspace(1)* %gep.1
+  %c = load double addrspace(1)* %gep.2
+  %d = load double addrspace(1)* %gep.3
+
+  %mul = fmul double %a, %b
+  %fma0 = fadd double %mul, %c
+  %fma1 = fadd double %mul, %d
+  store double %fma0, double addrspace(1)* %gep.out.0
+  store double %fma1, double addrspace(1)* %gep.out.1
+  ret void
+}
+
+; (fadd x, (fmul y, z)) -> (fma y, z, x)
+; FUNC-LABEL: {{^}}combine_to_fma_f64_1:
+; SI-DAG: buffer_load_dwordx2 [[A:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: buffer_load_dwordx2 [[B:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
+; SI-DAG: buffer_load_dwordx2 [[C:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}}
+; SI: v_fma_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[A]], [[B]], [[C]]
+; SI: buffer_store_dwordx2 [[RESULT]]
+define void @combine_to_fma_f64_1(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) #1 {
+  %tid = tail call i32 @llvm.r600.read.tidig.x() #0
+  %gep.0 = getelementptr double addrspace(1)* %in, i32 %tid
+  %gep.1 = getelementptr double addrspace(1)* %gep.0, i32 1
+  %gep.2 = getelementptr double addrspace(1)* %gep.0, i32 2
+  %gep.out = getelementptr double addrspace(1)* %out, i32 %tid
+
+  %a = load double addrspace(1)* %gep.0
+  %b = load double addrspace(1)* %gep.1
+  %c = load double addrspace(1)* %gep.2
+
+  %mul = fmul double %a, %b
+  %fma = fadd double %c, %mul
+  store double %fma, double addrspace(1)* %gep.out
+  ret void
+}
+
+; (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
+; FUNC-LABEL: {{^}}combine_to_fma_fsub_0_f64:
+; SI-DAG: buffer_load_dwordx2 [[A:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: buffer_load_dwordx2 [[B:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
+; SI-DAG: buffer_load_dwordx2 [[C:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}}
+; SI: v_fma_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[A]], [[B]], -[[C]]
+; SI: buffer_store_dwordx2 [[RESULT]]
+define void @combine_to_fma_fsub_0_f64(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) #1 {
+  %tid = tail call i32 @llvm.r600.read.tidig.x() #0
+  %gep.0 = getelementptr double addrspace(1)* %in, i32 %tid
+  %gep.1 = getelementptr double addrspace(1)* %gep.0, i32 1
+  %gep.2 = getelementptr double addrspace(1)* %gep.0, i32 2
+  %gep.out = getelementptr double addrspace(1)* %out, i32 %tid
+
+  %a = load double addrspace(1)* %gep.0
+  %b = load double addrspace(1)* %gep.1
+  %c = load double addrspace(1)* %gep.2
+
+  %mul = fmul double %a, %b
+  %fma = fsub double %mul, %c
+  store double %fma, double addrspace(1)* %gep.out
+  ret void
+}
+
+; (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
+; FUNC-LABEL: {{^}}combine_to_fma_fsub_f64_0_2use:
+; SI-DAG: buffer_load_dwordx2 [[A:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: buffer_load_dwordx2 [[B:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
+; SI-DAG: buffer_load_dwordx2 [[C:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}}
+; SI-DAG: buffer_load_dwordx2 [[D:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:24{{$}}
+; SI-DAG: v_fma_f64 [[RESULT0:v\[[0-9]+:[0-9]+\]]], [[A]], [[B]], -[[C]]
+; SI-DAG: v_fma_f64 [[RESULT1:v\[[0-9]+:[0-9]+\]]], [[A]], [[B]], -[[D]]
+; SI-DAG: buffer_store_dwordx2 [[RESULT0]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: buffer_store_dwordx2 [[RESULT1]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
+; SI: s_endpgm
+define void @combine_to_fma_fsub_f64_0_2use(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) #1 {
+  %tid = tail call i32 @llvm.r600.read.tidig.x() #0
+  %gep.0 = getelementptr double addrspace(1)* %in, i32 %tid
+  %gep.1 = getelementptr double addrspace(1)* %gep.0, i32 1
+  %gep.2 = getelementptr double addrspace(1)* %gep.0, i32 2
+  %gep.3 = getelementptr double addrspace(1)* %gep.0, i32 3
+  %gep.out.0 = getelementptr double addrspace(1)* %out, i32 %tid
+  %gep.out.1 = getelementptr double addrspace(1)* %gep.out.0, i32 1
+
+  %a = load double addrspace(1)* %gep.0
+  %b = load double addrspace(1)* %gep.1
+  %c = load double addrspace(1)* %gep.2
+  %d = load double addrspace(1)* %gep.3
+
+  %mul = fmul double %a, %b
+  %fma0 = fsub double %mul, %c
+  %fma1 = fsub double %mul, %d
+  store double %fma0, double addrspace(1)* %gep.out.0
+  store double %fma1, double addrspace(1)* %gep.out.1
+  ret void
+}
+
+; (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
+; FUNC-LABEL: {{^}}combine_to_fma_fsub_1_f64:
+; SI-DAG: buffer_load_dwordx2 [[A:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: buffer_load_dwordx2 [[B:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
+; SI-DAG: buffer_load_dwordx2 [[C:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}}
+; SI: v_fma_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], -[[A]], [[B]], [[C]]
+; SI: buffer_store_dwordx2 [[RESULT]]
+define void @combine_to_fma_fsub_1_f64(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) #1 {
+  %tid = tail call i32 @llvm.r600.read.tidig.x() #0
+  %gep.0 = getelementptr double addrspace(1)* %in, i32 %tid
+  %gep.1 = getelementptr double addrspace(1)* %gep.0, i32 1
+  %gep.2 = getelementptr double addrspace(1)* %gep.0, i32 2
+  %gep.out = getelementptr double addrspace(1)* %out, i32 %tid
+
+  %a = load double addrspace(1)* %gep.0
+  %b = load double addrspace(1)* %gep.1
+  %c = load double addrspace(1)* %gep.2
+
+  %mul = fmul double %a, %b
+  %fma = fsub double %c, %mul
+  store double %fma, double addrspace(1)* %gep.out
+  ret void
+}
+
+; (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
+; FUNC-LABEL: {{^}}combine_to_fma_fsub_1_f64_2use:
+; SI-DAG: buffer_load_dwordx2 [[A:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: buffer_load_dwordx2 [[B:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
+; SI-DAG: buffer_load_dwordx2 [[C:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}}
+; SI-DAG: buffer_load_dwordx2 [[D:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:24{{$}}
+; SI-DAG: v_fma_f64 [[RESULT0:v\[[0-9]+:[0-9]+\]]], -[[A]], [[B]], [[C]]
+; SI-DAG: v_fma_f64 [[RESULT1:v\[[0-9]+:[0-9]+\]]], -[[A]], [[B]], [[D]]
+; SI-DAG: buffer_store_dwordx2 [[RESULT0]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: buffer_store_dwordx2 [[RESULT1]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
+; SI: s_endpgm
+define void @combine_to_fma_fsub_1_f64_2use(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) #1 {
+  %tid = tail call i32 @llvm.r600.read.tidig.x() #0
+  %gep.0 = getelementptr double addrspace(1)* %in, i32 %tid
+  %gep.1 = getelementptr double addrspace(1)* %gep.0, i32 1
+  %gep.2 = getelementptr double addrspace(1)* %gep.0, i32 2
+  %gep.3 = getelementptr double addrspace(1)* %gep.0, i32 3
+  %gep.out.0 = getelementptr double addrspace(1)* %out, i32 %tid
+  %gep.out.1 = getelementptr double addrspace(1)* %gep.out.0, i32 1
+
+  %a = load double addrspace(1)* %gep.0
+  %b = load double addrspace(1)* %gep.1
+  %c = load double addrspace(1)* %gep.2
+  %d = load double addrspace(1)* %gep.3
+
+  %mul = fmul double %a, %b
+  %fma0 = fsub double %c, %mul
+  %fma1 = fsub double %d, %mul
+  store double %fma0, double addrspace(1)* %gep.out.0
+  store double %fma1, double addrspace(1)* %gep.out.1
+  ret void
+}
+
+; (fsub (fneg (fmul x, y)), z) -> (fma (fneg x), y, (fneg z))
+; FUNC-LABEL: {{^}}combine_to_fma_fsub_2_f64:
+; SI-DAG: buffer_load_dwordx2 [[A:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: buffer_load_dwordx2 [[B:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
+; SI-DAG: buffer_load_dwordx2 [[C:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}}
+; SI: v_fma_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], -[[A]], [[B]], -[[C]]
+; SI: buffer_store_dwordx2 [[RESULT]]
+define void @combine_to_fma_fsub_2_f64(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) #1 {
+  %tid = tail call i32 @llvm.r600.read.tidig.x() #0
+  %gep.0 = getelementptr double addrspace(1)* %in, i32 %tid
+  %gep.1 = getelementptr double addrspace(1)* %gep.0, i32 1
+  %gep.2 = getelementptr double addrspace(1)* %gep.0, i32 2
+  %gep.out = getelementptr double addrspace(1)* %out, i32 %tid
+
+  %a = load double addrspace(1)* %gep.0
+  %b = load double addrspace(1)* %gep.1
+  %c = load double addrspace(1)* %gep.2
+
+  %mul = fmul double %a, %b
+  %mul.neg = fsub double -0.0, %mul
+  %fma = fsub double %mul.neg, %c
+
+  store double %fma, double addrspace(1)* %gep.out
+  ret void
+}
+
+; (fsub (fneg (fmul x, y)), z) -> (fma (fneg x), y, (fneg z))
+; FUNC-LABEL: {{^}}combine_to_fma_fsub_2_f64_2uses_neg:
+; SI-DAG: buffer_load_dwordx2 [[A:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: buffer_load_dwordx2 [[B:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
+; SI-DAG: buffer_load_dwordx2 [[C:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}}
+; SI-DAG: v_fma_f64 [[RESULT0:v\[[0-9]+:[0-9]+\]]], -[[A]], [[B]], -[[C]]
+; SI-DAG: v_fma_f64 [[RESULT1:v\[[0-9]+:[0-9]+\]]], -[[A]], [[B]], -[[D]]
+; SI-DAG: buffer_store_dwordx2 [[RESULT0]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: buffer_store_dwordx2 [[RESULT1]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
+; SI: s_endpgm
+define void @combine_to_fma_fsub_2_f64_2uses_neg(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) #1 {
+  %tid = tail call i32 @llvm.r600.read.tidig.x() #0
+  %gep.0 = getelementptr double addrspace(1)* %in, i32 %tid
+  %gep.1 = getelementptr double addrspace(1)* %gep.0, i32 1
+  %gep.2 = getelementptr double addrspace(1)* %gep.0, i32 2
+  %gep.3 = getelementptr double addrspace(1)* %gep.0, i32 3
+  %gep.out.0 = getelementptr double addrspace(1)* %out, i32 %tid
+  %gep.out.1 = getelementptr double addrspace(1)* %gep.out.0, i32 1
+
+  %a = load double addrspace(1)* %gep.0
+  %b = load double addrspace(1)* %gep.1
+  %c = load double addrspace(1)* %gep.2
+  %d = load double addrspace(1)* %gep.3
+
+  %mul = fmul double %a, %b
+  %mul.neg = fsub double -0.0, %mul
+  %fma0 = fsub double %mul.neg, %c
+  %fma1 = fsub double %mul.neg, %d
+
+  store double %fma0, double addrspace(1)* %gep.out.0
+  store double %fma1, double addrspace(1)* %gep.out.1
+  ret void
+}
+
+; (fsub (fneg (fmul x, y)), z) -> (fma (fneg x), y, (fneg z))
+; FUNC-LABEL: {{^}}combine_to_fma_fsub_2_f64_2uses_mul:
+; SI-DAG: buffer_load_dwordx2 [[A:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: buffer_load_dwordx2 [[B:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
+; SI-DAG: buffer_load_dwordx2 [[C:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}}
+; SI-DAG: v_fma_f64 [[RESULT0:v\[[0-9]+:[0-9]+\]]], -[[A]], [[B]], -[[C]]
+; SI-DAG: v_fma_f64 [[RESULT1:v\[[0-9]+:[0-9]+\]]], [[A]], [[B]], -[[D]]
+; SI-DAG: buffer_store_dwordx2 [[RESULT0]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: buffer_store_dwordx2 [[RESULT1]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
+; SI: s_endpgm
+define void @combine_to_fma_fsub_2_f64_2uses_mul(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) #1 {
+  %tid = tail call i32 @llvm.r600.read.tidig.x() #0
+  %gep.0 = getelementptr double addrspace(1)* %in, i32 %tid
+  %gep.1 = getelementptr double addrspace(1)* %gep.0, i32 1
+  %gep.2 = getelementptr double addrspace(1)* %gep.0, i32 2
+  %gep.3 = getelementptr double addrspace(1)* %gep.0, i32 3
+  %gep.out.0 = getelementptr double addrspace(1)* %out, i32 %tid
+  %gep.out.1 = getelementptr double addrspace(1)* %gep.out.0, i32 1
+
+  %a = load double addrspace(1)* %gep.0
+  %b = load double addrspace(1)* %gep.1
+  %c = load double addrspace(1)* %gep.2
+  %d = load double addrspace(1)* %gep.3
+
+  %mul = fmul double %a, %b
+  %mul.neg = fsub double -0.0, %mul
+  %fma0 = fsub double %mul.neg, %c
+  %fma1 = fsub double %mul, %d
+
+  store double %fma0, double addrspace(1)* %gep.out.0
+  store double %fma1, double addrspace(1)* %gep.out.1
+  ret void
+}
+
+; fold (fsub (fma x, y, (fmul u, v)), z) -> (fma x, y (fma u, v, (fneg z)))
+
+; FUNC-LABEL: {{^}}aggressive_combine_to_fma_fsub_0_f64:
+; SI-DAG: buffer_load_dwordx2 [[X:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: buffer_load_dwordx2 [[Y:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
+; SI-DAG: buffer_load_dwordx2 [[Z:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}}
+; SI-DAG: buffer_load_dwordx2 [[U:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:24{{$}}
+; SI-DAG: buffer_load_dwordx2 [[V:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:32{{$}}
+; SI: v_fma_f64 [[FMA0:v\[[0-9]+:[0-9]+\]]], [[U]], [[V]], -[[Z]]
+; SI: v_fma_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[X]], [[Y]], [[FMA0]]
+; SI: buffer_store_dwordx2 [[RESULT]]
+define void @aggressive_combine_to_fma_fsub_0_f64(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) #1 {
+  %tid = tail call i32 @llvm.r600.read.tidig.x() #0
+  %gep.0 = getelementptr double addrspace(1)* %in, i32 %tid
+  %gep.1 = getelementptr double addrspace(1)* %gep.0, i32 1
+  %gep.2 = getelementptr double addrspace(1)* %gep.0, i32 2
+  %gep.3 = getelementptr double addrspace(1)* %gep.0, i32 3
+  %gep.4 = getelementptr double addrspace(1)* %gep.0, i32 4
+  %gep.out = getelementptr double addrspace(1)* %out, i32 %tid
+
+  %x = load double addrspace(1)* %gep.0
+  %y = load double addrspace(1)* %gep.1
+  %z = load double addrspace(1)* %gep.2
+  %u = load double addrspace(1)* %gep.3
+  %v = load double addrspace(1)* %gep.4
+
+  %tmp0 = fmul double %u, %v
+  %tmp1 = call double @llvm.fma.f64(double %x, double %y, double %tmp0) #0
+  %tmp2 = fsub double %tmp1, %z
+
+  store double %tmp2, double addrspace(1)* %gep.out
+  ret void
+}
+
+; fold (fsub x, (fma y, z, (fmul u, v)))
+;   -> (fma (fneg y), z, (fma (fneg u), v, x))
+
+; FUNC-LABEL: {{^}}aggressive_combine_to_fma_fsub_1_f64:
+; SI-DAG: buffer_load_dwordx2 [[X:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: buffer_load_dwordx2 [[Y:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
+; SI-DAG: buffer_load_dwordx2 [[Z:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}}
+; SI-DAG: buffer_load_dwordx2 [[U:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:24{{$}}
+; SI-DAG: buffer_load_dwordx2 [[V:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:32{{$}}
+; SI: v_fma_f64 [[FMA0:v\[[0-9]+:[0-9]+\]]], -[[U]], [[V]], [[X]]
+; SI: v_fma_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], -[[Y]], [[Z]], [[FMA0]]
+; SI: buffer_store_dwordx2 [[RESULT]]
+define void @aggressive_combine_to_fma_fsub_1_f64(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) #1 {
+  %tid = tail call i32 @llvm.r600.read.tidig.x() #0
+  %gep.0 = getelementptr double addrspace(1)* %in, i32 %tid
+  %gep.1 = getelementptr double addrspace(1)* %gep.0, i32 1
+  %gep.2 = getelementptr double addrspace(1)* %gep.0, i32 2
+  %gep.3 = getelementptr double addrspace(1)* %gep.0, i32 3
+  %gep.4 = getelementptr double addrspace(1)* %gep.0, i32 4
+  %gep.out = getelementptr double addrspace(1)* %out, i32 %tid
+
+  %x = load double addrspace(1)* %gep.0
+  %y = load double addrspace(1)* %gep.1
+  %z = load double addrspace(1)* %gep.2
+  %u = load double addrspace(1)* %gep.3
+  %v = load double addrspace(1)* %gep.4
+
+  %tmp0 = fmul double %u, %v
+  %tmp1 = call double @llvm.fma.f64(double %y, double %z, double %tmp0) #0
+  %tmp2 = fsub double %x, %tmp1
+
+  store double %tmp2, double addrspace(1)* %gep.out
+  ret void
+}
+
+attributes #0 = { nounwind readnone }
+attributes #1 = { nounwind }