[RISCV] Move the rest of Zfa FLI instruction handling to lowerConstantFP. (llvm#109217)

We already moved the fneg case. This moves the rest so we can drop the
custom isel.

Author: topperc

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

@@ -889,29 +889,6 @@ void RISCVDAGToDAGISel::Select(SDNode *Node) {
   }
   case ISD::ConstantFP: {
     const APFloat &APF = cast<ConstantFPSDNode>(Node)->getValueAPF();
-    int FPImm = static_cast<const RISCVTargetLowering *>(TLI)->getLegalZfaFPImm(
-        APF, VT);
-    if (FPImm >= 0) {
-      unsigned Opc;
-      switch (VT.SimpleTy) {
-      default:
-        llvm_unreachable("Unexpected size");
-      case MVT::f16:
-        Opc = RISCV::FLI_H;
-        break;
-      case MVT::f32:
-        Opc = RISCV::FLI_S;
-        break;
-      case MVT::f64:
-        Opc = RISCV::FLI_D;
-        break;
-      }
-      SDNode *Res = CurDAG->getMachineNode(
-          Opc, DL, VT, CurDAG->getTargetConstant(FPImm, DL, XLenVT));
-
-      ReplaceNode(Node, Res);
-      return;
-    }
 
     bool NegZeroF64 = APF.isNegZero() && VT == MVT::f64;
     SDValue Imm;
@@ -3552,13 +3529,6 @@ bool RISCVDAGToDAGISel::selectScalarFPAsInt(SDValue N, SDValue &Imm) {
 
   MVT VT = CFP->getSimpleValueType(0);
 
-  // Even if this FPImm requires an additional FNEG (i.e. the second element of
-  // the returned pair is true) we still prefer FLI + FNEG over immediate
-  // materialization as the latter might generate a longer instruction sequence.
-  if (static_cast<const RISCVTargetLowering *>(TLI)->getLegalZfaFPImm(APF,
-                                                                      VT) >= 0)
-    return false;
-
   MVT XLenVT = Subtarget->getXLenVT();
   if (VT == MVT::f64 && !Subtarget->is64Bit()) {
     assert(APF.isNegZero() && "Unexpected constant.");

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

@@ -5802,22 +5802,29 @@ SDValue RISCVTargetLowering::lowerConstantFP(SDValue Op,
   MVT VT = Op.getSimpleValueType();
   const APFloat &Imm = cast<ConstantFPSDNode>(Op)->getValueAPF();
 
-  if (getLegalZfaFPImm(Imm, VT) >= 0)
-    return Op;
+  // Can this constant be selected by a Zfa FLI instruction?
+  bool Negate = false;
+  int Index = getLegalZfaFPImm(Imm, VT);
 
-  if (!Imm.isNegative())
-    return SDValue();
+  // If the constant is negative, try negating.
+  if (Index < 0 && Imm.isNegative()) {
+    Index = getLegalZfaFPImm(-Imm, VT);
+    Negate = true;
+  }
 
-  int Index = getLegalZfaFPImm(-Imm, VT);
+  // If we couldn't find a FLI lowering, fall back to generic code.
   if (Index < 0)
     return SDValue();
 
   // Emit an FLI+FNEG. We use a custom node to hide from constant folding.
   SDLoc DL(Op);
   SDValue Const =
-      DAG.getNode(RISCVISD::FLI, Op, VT,
+      DAG.getNode(RISCVISD::FLI, DL, VT,
                   DAG.getTargetConstant(Index, DL, Subtarget.getXLenVT()));
-  return DAG.getNode(ISD::FNEG, Op, VT, Const);
+  if (!Negate)
+    return Const;
+
+  return DAG.getNode(ISD::FNEG, DL, VT, Const);
 }
 
 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,

llvm/test/CodeGen/RISCV/double-zfa.ll

@@ -371,3 +371,31 @@ define double @fma_neg_addend_multiplicand(double %x) nounwind {
   %a = call double @llvm.fma.f32(double %x, double -0.5, double -0.25)
   ret double %a
 }
+
+define double @select_loadfpimm(double %x) nounwind {
+; RV32IDZFA-LABEL: select_loadfpimm:
+; RV32IDZFA:       # %bb.0: # %entry
+; RV32IDZFA-NEXT:    fcvt.d.w fa5, zero
+; RV32IDZFA-NEXT:    fle.d a0, fa5, fa0
+; RV32IDZFA-NEXT:    fli.d fa0, 0.5
+; RV32IDZFA-NEXT:    bnez a0, .LBB35_2
+; RV32IDZFA-NEXT:  # %bb.1:
+; RV32IDZFA-NEXT:    fneg.d fa0, fa0
+; RV32IDZFA-NEXT:  .LBB35_2: # %entry
+; RV32IDZFA-NEXT:    ret
+;
+; RV64DZFA-LABEL: select_loadfpimm:
+; RV64DZFA:       # %bb.0: # %entry
+; RV64DZFA-NEXT:    fmv.d.x fa5, zero
+; RV64DZFA-NEXT:    fle.d a0, fa5, fa0
+; RV64DZFA-NEXT:    fli.d fa0, 0.5
+; RV64DZFA-NEXT:    bnez a0, .LBB35_2
+; RV64DZFA-NEXT:  # %bb.1:
+; RV64DZFA-NEXT:    fneg.d fa0, fa0
+; RV64DZFA-NEXT:  .LBB35_2: # %entry
+; RV64DZFA-NEXT:    ret
+entry:
+  %cmp = fcmp ult double %x, 0.000000e+00
+  %sel = select i1 %cmp, double -5.000000e-01, double 5.000000e-01
+  ret double %sel
+}

llvm/test/CodeGen/RISCV/float-zfa.ll

@@ -310,3 +310,20 @@ define float @fma_neg_addend_multiplicand(float %x) nounwind {
   %a = call float @llvm.fma.f32(float %x, float -0.5, float -0.25)
   ret float %a
 }
+
+define float @select_loadfpimm(float %x) nounwind {
+; CHECK-LABEL: select_loadfpimm:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    fmv.w.x fa5, zero
+; CHECK-NEXT:    fle.s a0, fa5, fa0
+; CHECK-NEXT:    fli.s fa0, 0.5
+; CHECK-NEXT:    bnez a0, .LBB30_2
+; CHECK-NEXT:  # %bb.1:
+; CHECK-NEXT:    fneg.s fa0, fa0
+; CHECK-NEXT:  .LBB30_2: # %entry
+; CHECK-NEXT:    ret
+entry:
+  %cmp = fcmp ult float %x, 0.000000e+00
+  %sel = select i1 %cmp, float -5.000000e-01, float 5.000000e-01
+  ret float %sel
+}

llvm/test/CodeGen/RISCV/half-zfa.ll

@@ -314,3 +314,33 @@ define half @fma_neg_addend_multiplicand(half %x) nounwind {
   %a = call half @llvm.fma.f32(half %x, half -0.5, half -0.25)
   ret half %a
 }
+
+define half @select_loadfpimm(half %x) nounwind {
+; CHECK-LABEL: select_loadfpimm:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    fmv.h.x fa5, zero
+; CHECK-NEXT:    fle.h a0, fa5, fa0
+; CHECK-NEXT:    fli.h fa0, 0.5
+; CHECK-NEXT:    bnez a0, .LBB16_2
+; CHECK-NEXT:  # %bb.1:
+; CHECK-NEXT:    fneg.h fa0, fa0
+; CHECK-NEXT:  .LBB16_2: # %entry
+; CHECK-NEXT:    ret
+;
+; ZFHMIN-LABEL: select_loadfpimm:
+; ZFHMIN:       # %bb.0: # %entry
+; ZFHMIN-NEXT:    fcvt.s.h fa5, fa0
+; ZFHMIN-NEXT:    fmv.w.x fa4, zero
+; ZFHMIN-NEXT:    fle.s a0, fa4, fa5
+; ZFHMIN-NEXT:    xori a0, a0, 1
+; ZFHMIN-NEXT:    slli a0, a0, 1
+; ZFHMIN-NEXT:    lui a1, %hi(.LCPI16_0)
+; ZFHMIN-NEXT:    addi a1, a1, %lo(.LCPI16_0)
+; ZFHMIN-NEXT:    add a0, a1, a0
+; ZFHMIN-NEXT:    flh fa0, 0(a0)
+; ZFHMIN-NEXT:    ret
+entry:
+  %cmp = fcmp ult half %x, 0.000000e+00
+  %sel = select i1 %cmp, half -5.000000e-01, half 5.000000e-01
+  ret half %sel
+}