[RISCV] Lower VP_SELECT constant false to use vmerge.vxm/vmerge.vim

Currently, when the false path of a vp_select is a splat vector,
it is lowered to a vmv_v_x/vmv_v_i. The vmv is hoisted out of
the loop and the whole copy in loop body by MachineLICM.

By inverting the mask register and swapping the true and false values
in the vp_select, we can eliminate some instructions inside the loop.

corrent: https://godbolt.org/z/EnGMn3xeM
expected similar form: https://godbolt.org/z/nWhGM6Ej5

Author: ChunyuLiao

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

@@ -8168,11 +8168,17 @@ SDValue RISCVTargetLowering::LowerOperation(SDValue Op,
     return lowerRESET_FPENV(Op, DAG);
   case ISD::EH_DWARF_CFA:
     return lowerEH_DWARF_CFA(Op, DAG);
+  case ISD::VP_SELECT:
+    if (SDValue Op2 = Op.getOperand(2);
+        Op2.hasOneUse() && (Op2.getOpcode() == ISD::SPLAT_VECTOR ||
+                            Op2.getOpcode() == ISD::SPLAT_VECTOR_PARTS))
+      return lowerVPSelectConstantFalse(Op, DAG);
+    else
+      return lowerVPOp(Op, DAG);
   case ISD::VP_MERGE:
     if (Op.getSimpleValueType().getVectorElementType() == MVT::i1)
       return lowerVPMergeMask(Op, DAG);
     [[fallthrough]];
-  case ISD::VP_SELECT:
   case ISD::VP_ADD:
   case ISD::VP_SUB:
   case ISD::VP_MUL:
@@ -13175,6 +13181,20 @@ SDValue RISCVTargetLowering::lowerVPFPIntConvOp(SDValue Op,
   return convertFromScalableVector(VT, Result, DAG, Subtarget);
 }
 
+SDValue
+RISCVTargetLowering::lowerVPSelectConstantFalse(SDValue Op,
+                                                SelectionDAG &DAG) const {
+  SDLoc DL(Op);
+  MVT VT = Op.getSimpleValueType();
+  SDValue TrueVal = Op.getOperand(1);
+  SDValue FalseVal = Op.getOperand(2);
+  SDValue VL = Op.getOperand(3);
+  SDValue Op0 = Op.getOperand(0);
+  SDValue Val = DAG.getLogicalNOT(DL, Op0, Op0.getValueType());
+  return DAG.getNode(RISCVISD::VMERGE_VL, DL, VT, Val, FalseVal, TrueVal,
+                     DAG.getUNDEF(VT), VL);
+}
+
 SDValue RISCVTargetLowering::lowerVPMergeMask(SDValue Op,
                                               SelectionDAG &DAG) const {
   SDLoc DL(Op);

llvm/lib/Target/RISCV/RISCVISelLowering.h

@@ -486,6 +486,7 @@ class RISCVTargetLowering : public TargetLowering {
   SDValue getTLSDescAddr(GlobalAddressSDNode *N, SelectionDAG &DAG) const;
 
   SDValue lowerConstantFP(SDValue Op, SelectionDAG &DAG) const;
+  SDValue lowerVPSelectConstantFalse(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerConstantPool(SDValue Op, SelectionDAG &DAG) const;

llvm/test/CodeGen/RISCV/rvv/masked-load-int.ll

@@ -34,10 +34,10 @@ define <vscale x 1 x i8> @masked_load_passthru_nxv1i8(ptr %a, <vscale x 1 x i1>
 ; ZVE32:       # %bb.0:
 ; ZVE32-NEXT:    csrr a1, vlenb
 ; ZVE32-NEXT:    srli a1, a1, 3
-; ZVE32-NEXT:    vsetvli a2, zero, e8, mf4, ta, ma
-; ZVE32-NEXT:    vmv.v.i v8, 0
-; ZVE32-NEXT:    vsetvli zero, a1, e8, mf4, ta, mu
+; ZVE32-NEXT:    vsetvli zero, a1, e8, mf4, ta, ma
 ; ZVE32-NEXT:    vle8.v v8, (a0), v0.t
+; ZVE32-NEXT:    vmnot.m v0, v0
+; ZVE32-NEXT:    vmerge.vim v8, v8, 0, v0
 ; ZVE32-NEXT:    ret
   %load = call <vscale x 1 x i8> @llvm.masked.load.nxv1i8(ptr %a, i32 1, <vscale x 1 x i1> %mask, <vscale x 1 x i8> zeroinitializer)
   ret <vscale x 1 x i8> %load

llvm/test/CodeGen/RISCV/rvv/vselect-vp.ll

@@ -483,10 +483,10 @@ define <vscale x 2 x i64> @select_nxv2i64_constant_true(<vscale x 2 x i1> %a, <v
 define <vscale x 2 x i64> @select_nxv2i64_constant_false(<vscale x 2 x i1> %a, <vscale x 2 x i64> %b, i32 zeroext %evl) {
 ; CHECK-LABEL: select_nxv2i64_constant_false:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    li a1, 100
 ; CHECK-NEXT:    vsetvli zero, a0, e64, m2, ta, ma
-; CHECK-NEXT:    vmv.v.x v10, a1
-; CHECK-NEXT:    vmerge.vvm v8, v10, v8, v0
+; CHECK-NEXT:    vmnot.m v0, v0
+; CHECK-NEXT:    li a0, 100
+; CHECK-NEXT:    vmerge.vxm v8, v8, a0, v0
 ; CHECK-NEXT:    ret
   %v = call <vscale x 2 x i64> @llvm.vp.select.nxv2i64(<vscale x 2 x i1> %a, <vscale x 2 x i64> %b, <vscale x 2 x i64> splat (i64 100), i32 %evl)
   ret <vscale x 2 x i64> %v