Add vmov.f32 to materialize f32 immediate splats which cannot be handled by

integer variants. rdar://10437054


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@144608 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/ARM/ARMISelLowering.cpp

@@ -912,6 +912,7 @@ const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
   case ARMISD::VGETLANEs:     return "ARMISD::VGETLANEs";
   case ARMISD::VMOVIMM:       return "ARMISD::VMOVIMM";
   case ARMISD::VMVNIMM:       return "ARMISD::VMVNIMM";
+  case ARMISD::VMOVFPIMM:     return "ARMISD::VMOVFPIMM";
   case ARMISD::VDUP:          return "ARMISD::VDUP";
   case ARMISD::VDUPLANE:      return "ARMISD::VDUPLANE";
   case ARMISD::VEXT:          return "ARMISD::VEXT";
@@ -3986,6 +3987,16 @@ SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
         SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
         return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
       }
+
+      // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
+      if (VT == MVT::v2f32 || VT == MVT::v4f32) {
+        ConstantFPSDNode *C = cast<ConstantFPSDNode>(Op.getOperand(0));
+        int ImmVal = ARM_AM::getFP32Imm(C->getValueAPF());
+        if (ImmVal != -1) {
+          SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
+          return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
+        }
+      }
     }
   }
 

lib/Target/ARM/ARMISelLowering.h

@@ -146,6 +146,9 @@ namespace llvm {
       VMOVIMM,
       VMVNIMM,
 
+      // Vector move f32 immediate:
+      VMOVFPIMM,
+
       // Vector duplicate:
       VDUP,
       VDUPLANE,

lib/Target/ARM/ARMInstrNEON.td

@@ -39,6 +39,10 @@ def nImmVMOVI32 : Operand<i32> {
   let PrintMethod = "printNEONModImmOperand";
   let ParserMatchClass = nImmVMOVI32AsmOperand;
 }
+def nImmVMOVF32 : Operand<i32> {
+  let PrintMethod = "printFPImmOperand";
+  let ParserMatchClass = FPImmOperand;
+}
 def nImmSplatI64AsmOperand : AsmOperandClass { let Name = "NEONi64splat"; }
 def nImmSplatI64 : Operand<i32> {
   let PrintMethod = "printNEONModImmOperand";
@@ -173,6 +177,7 @@ def NEONvgetlanes : SDNode<"ARMISD::VGETLANEs", SDTARMVGETLN>;
 def SDTARMVMOVIMM : SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVT<1, i32>]>;
 def NEONvmovImm   : SDNode<"ARMISD::VMOVIMM", SDTARMVMOVIMM>;
 def NEONvmvnImm   : SDNode<"ARMISD::VMVNIMM", SDTARMVMOVIMM>;
+def NEONvmovFPImm : SDNode<"ARMISD::VMOVFPIMM", SDTARMVMOVIMM>;
 
 def SDTARMVORRIMM : SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0, 1>,
                                            SDTCisVT<2, i32>]>;
@@ -4513,6 +4518,15 @@ def VMOVv2i64 : N1ModImm<1, 0b000, 0b1110, 0, 1, 1, 1, (outs QPR:$Vd),
                          (ins nImmSplatI64:$SIMM), IIC_VMOVImm,
                          "vmov", "i64", "$Vd, $SIMM", "",
                          [(set QPR:$Vd, (v2i64 (NEONvmovImm timm:$SIMM)))]>;
+
+def VMOVv2f32 : N1ModImm<1, 0b000, 0b1111, 0, 0, 0, 1, (outs DPR:$Vd),
+                         (ins nImmVMOVF32:$SIMM), IIC_VMOVImm,
+                         "vmov", "f32", "$Vd, $SIMM", "",
+                         [(set DPR:$Vd, (v2f32 (NEONvmovFPImm timm:$SIMM)))]>;
+def VMOVv4f32 : N1ModImm<1, 0b000, 0b1111, 0, 1, 0, 1, (outs QPR:$Vd),
+                         (ins nImmVMOVF32:$SIMM), IIC_VMOVImm,
+                         "vmov", "f32", "$Vd, $SIMM", "",
+                         [(set QPR:$Vd, (v4f32 (NEONvmovFPImm timm:$SIMM)))]>;
 } // isReMaterializable
 
 //   VMOV     : Vector Get Lane (move scalar to ARM core register)

test/CodeGen/ARM/vmov.ll

@@ -353,3 +353,21 @@ define void @noTruncStore(<4 x i32>* %a, <4 x i16>* %b) nounwind {
   store <4 x i16> %tmp2, <4 x i16>* %b, align 8
   ret void
 }
+
+; Use vmov.f32 to materialize f32 immediate splats
+; rdar://10437054
+define void @v_mov_v2f32(<2 x float>* nocapture %p) nounwind {
+entry:
+;CHECK: v_mov_v2f32:
+;CHECK: vmov.f32 d{{.*}}, #-1.600000e+01
+  store <2 x float> <float -1.600000e+01, float -1.600000e+01>, <2 x float>* %p, align 4
+  ret void
+}
+
+define void @v_mov_v4f32(<4 x float>* nocapture %p) nounwind {
+entry:
+;CHECK: v_mov_v4f32:
+;CHECK: vmov.f32 q{{.*}}, #3.100000e+01
+  store <4 x float> <float 3.100000e+01, float 3.100000e+01, float 3.100000e+01, float 3.100000e+01>, <4 x float>* %p, align 4
+  ret void
+}

test/CodeGen/Thumb2/machine-licm.ll

@@ -51,12 +51,11 @@ return:                                           ; preds = %bb, %entry
 define void @t2(i8* %ptr1, i8* %ptr2) nounwind {
 entry:
 ; CHECK: t2:
-; CHECK: mov.w [[R3:r[0-9]+]], #1065353216
-; CHECK: vdup.32 q{{.*}}, [[R3]]
+; CHECK: vmov.f32 q{{.*}}, #1.000000e+00
   br i1 undef, label %bb1, label %bb2
 
 bb1:
-; CHECK-NEXT: %bb1
+; CHECK: %bb1
   %indvar = phi i32 [ %indvar.next, %bb1 ], [ 0, %entry ]
   %tmp1 = shl i32 %indvar, 2
   %gep1 = getelementptr i8* %ptr1, i32 %tmp1

utils/TableGen/EDEmitter.cpp

@@ -607,6 +607,7 @@ static int ARMFlagFromOpName(LiteralConstantEmitter *type,
   IMM("nImmSplatI32");
   IMM("nImmSplatI64");
   IMM("nImmVMOVI32");
+  IMM("nImmVMOVF32");
   IMM("imm0_7");
   IMM("imm0_15");
   IMM("imm0_255");