Make the fast-isel code for literal 0.0 a bit shorter/faster, since 0.0 is common. rdar://problem/9303592 .

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@130338 91177308-0d34-0410-b5e6-96231b3b80d8

include/llvm/CodeGen/FastISel.h

@@ -317,6 +317,10 @@ protected:
     return 0;
   }
 
+  virtual unsigned TargetMaterializeFloatZero(const ConstantFP* CF) {
+    return 0;
+  }
+
 private:
   bool SelectBinaryOp(const User *I, unsigned ISDOpcode);
 

lib/CodeGen/SelectionDAG/FastISel.cpp

@@ -164,8 +164,12 @@ unsigned FastISel::materializeRegForValue(const Value *V, MVT VT) {
     Reg =
       getRegForValue(Constant::getNullValue(TD.getIntPtrType(V->getContext())));
   } else if (const ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
-    // Try to emit the constant directly.
-    Reg = FastEmit_f(VT, VT, ISD::ConstantFP, CF);
+    if (CF->isZero()) {
+      Reg = TargetMaterializeFloatZero(CF);
+    } else {
+      // Try to emit the constant directly.
+      Reg = FastEmit_f(VT, VT, ISD::ConstantFP, CF);
+    }
 
     if (!Reg) {
       // Try to emit the constant by using an integer constant with a cast.

lib/Target/X86/X86FastISel.cpp

@@ -124,6 +124,8 @@ private:
 
   unsigned TargetMaterializeAlloca(const AllocaInst *C);
 
+  unsigned TargetMaterializeFloatZero(const ConstantFP *CF);
+
   /// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is
   /// computed in an SSE register, not on the X87 floating point stack.
   bool isScalarFPTypeInSSEReg(EVT VT) const {
@@ -2049,6 +2051,45 @@ unsigned X86FastISel::TargetMaterializeAlloca(const AllocaInst *C) {
   return ResultReg;
 }
 
+unsigned X86FastISel::TargetMaterializeFloatZero(const ConstantFP *CF) {
+  MVT VT;
+  if (!isTypeLegal(CF->getType(), VT))
+    return false;
+
+  // Get opcode and regclass for the given zero.
+  unsigned Opc = 0;
+  const TargetRegisterClass *RC = NULL;
+  switch (VT.SimpleTy) {
+    default: return false;
+    case MVT::f32:
+      if (Subtarget->hasSSE1()) {
+        Opc = X86::FsFLD0SS;
+        RC  = X86::FR32RegisterClass;
+      } else {
+        Opc = X86::LD_Fp032;
+        RC  = X86::RFP32RegisterClass;
+      }
+      break;
+    case MVT::f64:
+      if (Subtarget->hasSSE2()) {
+        Opc = X86::FsFLD0SD;
+        RC  = X86::FR64RegisterClass;
+      } else {
+        Opc = X86::LD_Fp064;
+        RC  = X86::RFP64RegisterClass;
+      }
+      break;
+    case MVT::f80:
+      // No f80 support yet.
+      return false;
+  }
+
+  unsigned ResultReg = createResultReg(RC);
+  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), ResultReg);
+  return ResultReg;
+}
+
+
 /// TryToFoldLoad - The specified machine instr operand is a vreg, and that
 /// vreg is being provided by the specified load instruction.  If possible,
 /// try to fold the load as an operand to the instruction, returning true if

test/CodeGen/X86/fast-isel-x86-64.ll

@@ -225,3 +225,16 @@ if.else:                                          ; preds = %entry
 ; CHECK-NEXT: je 
 }
 
+; Check that 0.0 is materialized using pxor
+define void @test18(float* %p1) {
+  store float 0.0, float* %p1
+  ret void
+; CHECK: test18:
+; CHECK: pxor
+}
+define void @test19(double* %p1) {
+  store double 0.0, double* %p1
+  ret void
+; CHECK: test19:
+; CHECK: pxor
+}