R600/SI: FMA is faster than fmul and fadd for f64

llvm-svn: 188136

3 files changed, 50 insertions, 0 deletions

diff --git a/llvm/lib/Target/R600/SIISelLowering.cpp b/llvm/lib/Target/R600/SIISelLowering.cpp
index b714fc19267..a76e6ee3145 100644
--- a/llvm/lib/Target/R600/SIISelLowering.cpp
+++ b/llvm/lib/Target/R600/SIISelLowering.cpp
@@ -338,6 +338,24 @@ MVT SITargetLowering::getScalarShiftAmountTy(EVT VT) const {
   return MVT::i32;
 }
 
+bool SITargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
+  VT = VT.getScalarType();
+
+  if (!VT.isSimple())
+    return false;
+
+  switch (VT.getSimpleVT().SimpleTy) {
+  case MVT::f32:
+    return false; /* There is V_MAD_F32 for f32 */
+  case MVT::f64:
+    return true;
+  default:
+    break;
+  }
+
+  return false;
+}
+
 //===----------------------------------------------------------------------===//
 // Custom DAG Lowering Operations
 //===----------------------------------------------------------------------===//
diff --git a/llvm/lib/Target/R600/SIISelLowering.h b/llvm/lib/Target/R600/SIISelLowering.h
index b4202c475d4..effbf1f85de 100644
--- a/llvm/lib/Target/R600/SIISelLowering.h
+++ b/llvm/lib/Target/R600/SIISelLowering.h
@@ -55,6 +55,7 @@ public:
                                               MachineBasicBlock * BB) const;
   virtual EVT getSetCCResultType(LLVMContext &Context, EVT VT) const;
   virtual MVT getScalarShiftAmountTy(EVT VT) const;
+  virtual bool isFMAFasterThanFMulAndFAdd(EVT VT) const;
   virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
   virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
   virtual SDNode *PostISelFolding(MachineSDNode *N, SelectionDAG &DAG) const;
diff --git a/llvm/test/CodeGen/R600/fmuladd.ll b/llvm/test/CodeGen/R600/fmuladd.ll
new file mode 100644
index 00000000000..ac379f44aaa
--- /dev/null
+++ b/llvm/test/CodeGen/R600/fmuladd.ll
@@ -0,0 +1,31 @@
+; RUN: llc < %s -march=r600 -mcpu=SI | FileCheck %s
+
+; CHECK: @fmuladd_f32
+; CHECK: V_MAD_F32 {{VGPR[0-9]+, VGPR[0-9]+, VGPR[0-9]+, VGPR[0-9]+}}
+
+define void @fmuladd_f32(float addrspace(1)* %out, float addrspace(1)* %in1,
+                         float addrspace(1)* %in2, float addrspace(1)* %in3) {
+   %r0 = load float addrspace(1)* %in1
+   %r1 = load float addrspace(1)* %in2
+   %r2 = load float addrspace(1)* %in3
+   %r3 = tail call float @llvm.fmuladd.f32(float %r0, float %r1, float %r2)
+   store float %r3, float addrspace(1)* %out
+   ret void
+}
+
+declare float @llvm.fmuladd.f32(float, float, float)
+
+; CHECK: @fmuladd_f64
+; CHECK: V_FMA_F64 {{VGPR[0-9]+_VGPR[0-9]+, VGPR[0-9]+_VGPR[0-9]+, VGPR[0-9]+_VGPR[0-9]+, VGPR[0-9]+_VGPR[0-9]+}}
+
+define void @fmuladd_f64(double addrspace(1)* %out, double addrspace(1)* %in1,
+                         double addrspace(1)* %in2, double addrspace(1)* %in3) {
+   %r0 = load double addrspace(1)* %in1
+   %r1 = load double addrspace(1)* %in2
+   %r2 = load double addrspace(1)* %in3
+   %r3 = tail call double @llvm.fmuladd.f64(double %r0, double %r1, double %r2)
+   store double %r3, double addrspace(1)* %out
+   ret void
+}
+
+declare double @llvm.fmuladd.f64(double, double, double)