[RISCV] use vwadd.vx for extended splat

Author: sun-jacobi

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

@@ -13597,7 +13597,8 @@ struct NodeExtensionHelper {
 
   /// Check if this instance represents a splat.
   bool isSplat() const {
-    return OrigOperand.getOpcode() == RISCVISD::VMV_V_X_VL;
+    return OrigOperand.getOpcode() == RISCVISD::VMV_V_X_VL ||
+           OrigOperand.getOpcode() == ISD::SPLAT_VECTOR;
   }
 
   /// Get the extended opcode.
@@ -13641,6 +13642,8 @@ struct NodeExtensionHelper {
     case RISCVISD::VZEXT_VL:
     case RISCVISD::FP_EXTEND_VL:
       return DAG.getNode(ExtOpc, DL, NarrowVT, Source, Mask, VL);
+    case ISD::SPLAT_VECTOR:
+      return DAG.getSplat(NarrowVT, DL, Source.getOperand(0));
     case RISCVISD::VMV_V_X_VL:
       return DAG.getNode(RISCVISD::VMV_V_X_VL, DL, NarrowVT,
                          DAG.getUNDEF(NarrowVT), Source.getOperand(1), VL);
@@ -13776,6 +13779,47 @@ struct NodeExtensionHelper {
   /// Check if this node needs to be fully folded or extended for all users.
   bool needToPromoteOtherUsers() const { return EnforceOneUse; }
 
+  void fillUpExtensionSupportForSplat(SDNode *Root, SelectionDAG &DAG,
+                                      const RISCVSubtarget &Subtarget) {
+    unsigned Opc = OrigOperand.getOpcode();
+    MVT VT = OrigOperand.getSimpleValueType();
+
+    assert((Opc == ISD::SPLAT_VECTOR || Opc == RISCVISD::VMV_V_X_VL) &&
+           "Unexpected Opcode");
+
+    // The pasthru must be undef for tail agnostic.
+    if (Opc == RISCVISD::VMV_V_X_VL && !OrigOperand.getOperand(0).isUndef())
+      return;
+
+    // Get the scalar value.
+    SDValue Op = Opc == ISD::SPLAT_VECTOR ? OrigOperand.getOperand(0)
+                                          : OrigOperand.getOperand(1);
+
+    // See if we have enough sign bits or zero bits in the scalar to use a
+    // widening opcode by splatting to smaller element size.
+    unsigned EltBits = VT.getScalarSizeInBits();
+    unsigned ScalarBits = Op.getValueSizeInBits();
+    // Make sure we're getting all element bits from the scalar register.
+    // FIXME: Support implicit sign extension of vmv.v.x?
+    if (ScalarBits < EltBits)
+      return;
+
+    unsigned NarrowSize = VT.getScalarSizeInBits() / 2;
+    // If the narrow type cannot be expressed with a legal VMV,
+    // this is not a valid candidate.
+    if (NarrowSize < 8)
+      return;
+
+    if (DAG.ComputeMaxSignificantBits(Op) <= NarrowSize)
+      SupportsSExt = true;
+
+    if (DAG.MaskedValueIsZero(Op,
+                              APInt::getBitsSetFrom(ScalarBits, NarrowSize)))
+      SupportsZExt = true;
+
+    EnforceOneUse = false;
+  }
+
   /// Helper method to set the various fields of this struct based on the
   /// type of \p Root.
   void fillUpExtensionSupport(SDNode *Root, SelectionDAG &DAG,
@@ -13814,43 +13858,10 @@ struct NodeExtensionHelper {
     case RISCVISD::FP_EXTEND_VL:
       SupportsFPExt = true;
       break;
-    case RISCVISD::VMV_V_X_VL: {
-      // Historically, we didn't care about splat values not disappearing during
-      // combines.
-      EnforceOneUse = false;
-
-      // The operand is a splat of a scalar.
-
-      // The pasthru must be undef for tail agnostic.
-      if (!OrigOperand.getOperand(0).isUndef())
-        break;
-
-      // Get the scalar value.
-      SDValue Op = OrigOperand.getOperand(1);
-
-      // See if we have enough sign bits or zero bits in the scalar to use a
-      // widening opcode by splatting to smaller element size.
-      MVT VT = Root->getSimpleValueType(0);
-      unsigned EltBits = VT.getScalarSizeInBits();
-      unsigned ScalarBits = Op.getValueSizeInBits();
-      // Make sure we're getting all element bits from the scalar register.
-      // FIXME: Support implicit sign extension of vmv.v.x?
-      if (ScalarBits < EltBits)
-        break;
-
-      unsigned NarrowSize = VT.getScalarSizeInBits() / 2;
-      // If the narrow type cannot be expressed with a legal VMV,
-      // this is not a valid candidate.
-      if (NarrowSize < 8)
-        break;
-
-      if (DAG.ComputeMaxSignificantBits(Op) <= NarrowSize)
-        SupportsSExt = true;
-      if (DAG.MaskedValueIsZero(Op,
-                                APInt::getBitsSetFrom(ScalarBits, NarrowSize)))
-        SupportsZExt = true;
+    case ISD::SPLAT_VECTOR:
+    case RISCVISD::VMV_V_X_VL:
+      fillUpExtensionSupportForSplat(Root, DAG, Subtarget);
       break;
-    }
     default:
       break;
     }

llvm/test/CodeGen/RISCV/rvv/ctlz-sdnode.ll

@@ -1229,22 +1229,6 @@ define <vscale x 1 x i64> @ctlz_nxv1i64(<vscale x 1 x i64> %va) {
 ; RV64I-NEXT:    vsrl.vx v8, v8, a0
 ; RV64I-NEXT:    ret
 ;
-; CHECK-F-LABEL: ctlz_nxv1i64:
-; CHECK-F:       # %bb.0:
-; CHECK-F-NEXT:    li a0, 190
-; CHECK-F-NEXT:    vsetvli a1, zero, e64, m1, ta, ma
-; CHECK-F-NEXT:    vmv.v.x v9, a0
-; CHECK-F-NEXT:    vsetvli zero, zero, e32, mf2, ta, ma
-; CHECK-F-NEXT:    fsrmi a0, 1
-; CHECK-F-NEXT:    vfncvt.f.xu.w v10, v8
-; CHECK-F-NEXT:    vsrl.vi v8, v10, 23
-; CHECK-F-NEXT:    vwsubu.wv v9, v9, v8
-; CHECK-F-NEXT:    li a1, 64
-; CHECK-F-NEXT:    vsetvli zero, zero, e64, m1, ta, ma
-; CHECK-F-NEXT:    vminu.vx v8, v9, a1
-; CHECK-F-NEXT:    fsrm a0
-; CHECK-F-NEXT:    ret
-;
 ; CHECK-D-LABEL: ctlz_nxv1i64:
 ; CHECK-D:       # %bb.0:
 ; CHECK-D-NEXT:    vsetvli a0, zero, e64, m1, ta, ma
@@ -1370,22 +1354,6 @@ define <vscale x 2 x i64> @ctlz_nxv2i64(<vscale x 2 x i64> %va) {
 ; RV64I-NEXT:    vsrl.vx v8, v8, a0
 ; RV64I-NEXT:    ret
 ;
-; CHECK-F-LABEL: ctlz_nxv2i64:
-; CHECK-F:       # %bb.0:
-; CHECK-F-NEXT:    li a0, 190
-; CHECK-F-NEXT:    vsetvli a1, zero, e64, m2, ta, ma
-; CHECK-F-NEXT:    vmv.v.x v10, a0
-; CHECK-F-NEXT:    vsetvli zero, zero, e32, m1, ta, ma
-; CHECK-F-NEXT:    fsrmi a0, 1
-; CHECK-F-NEXT:    vfncvt.f.xu.w v12, v8
-; CHECK-F-NEXT:    vsrl.vi v8, v12, 23
-; CHECK-F-NEXT:    vwsubu.wv v10, v10, v8
-; CHECK-F-NEXT:    li a1, 64
-; CHECK-F-NEXT:    vsetvli zero, zero, e64, m2, ta, ma
-; CHECK-F-NEXT:    vminu.vx v8, v10, a1
-; CHECK-F-NEXT:    fsrm a0
-; CHECK-F-NEXT:    ret
-;
 ; CHECK-D-LABEL: ctlz_nxv2i64:
 ; CHECK-D:       # %bb.0:
 ; CHECK-D-NEXT:    vsetvli a0, zero, e64, m2, ta, ma
@@ -1511,22 +1479,6 @@ define <vscale x 4 x i64> @ctlz_nxv4i64(<vscale x 4 x i64> %va) {
 ; RV64I-NEXT:    vsrl.vx v8, v8, a0
 ; RV64I-NEXT:    ret
 ;
-; CHECK-F-LABEL: ctlz_nxv4i64:
-; CHECK-F:       # %bb.0:
-; CHECK-F-NEXT:    li a0, 190
-; CHECK-F-NEXT:    vsetvli a1, zero, e64, m4, ta, ma
-; CHECK-F-NEXT:    vmv.v.x v12, a0
-; CHECK-F-NEXT:    vsetvli zero, zero, e32, m2, ta, ma
-; CHECK-F-NEXT:    fsrmi a0, 1
-; CHECK-F-NEXT:    vfncvt.f.xu.w v16, v8
-; CHECK-F-NEXT:    vsrl.vi v8, v16, 23
-; CHECK-F-NEXT:    vwsubu.wv v12, v12, v8
-; CHECK-F-NEXT:    li a1, 64
-; CHECK-F-NEXT:    vsetvli zero, zero, e64, m4, ta, ma
-; CHECK-F-NEXT:    vminu.vx v8, v12, a1
-; CHECK-F-NEXT:    fsrm a0
-; CHECK-F-NEXT:    ret
-;
 ; CHECK-D-LABEL: ctlz_nxv4i64:
 ; CHECK-D:       # %bb.0:
 ; CHECK-D-NEXT:    vsetvli a0, zero, e64, m4, ta, ma
@@ -1652,22 +1604,6 @@ define <vscale x 8 x i64> @ctlz_nxv8i64(<vscale x 8 x i64> %va) {
 ; RV64I-NEXT:    vsrl.vx v8, v8, a0
 ; RV64I-NEXT:    ret
 ;
-; CHECK-F-LABEL: ctlz_nxv8i64:
-; CHECK-F:       # %bb.0:
-; CHECK-F-NEXT:    li a0, 190
-; CHECK-F-NEXT:    vsetvli a1, zero, e64, m8, ta, ma
-; CHECK-F-NEXT:    vmv.v.x v16, a0
-; CHECK-F-NEXT:    vsetvli zero, zero, e32, m4, ta, ma
-; CHECK-F-NEXT:    fsrmi a0, 1
-; CHECK-F-NEXT:    vfncvt.f.xu.w v24, v8
-; CHECK-F-NEXT:    vsrl.vi v8, v24, 23
-; CHECK-F-NEXT:    vwsubu.wv v16, v16, v8
-; CHECK-F-NEXT:    li a1, 64
-; CHECK-F-NEXT:    vsetvli zero, zero, e64, m8, ta, ma
-; CHECK-F-NEXT:    vminu.vx v8, v16, a1
-; CHECK-F-NEXT:    fsrm a0
-; CHECK-F-NEXT:    ret
-;
 ; CHECK-D-LABEL: ctlz_nxv8i64:
 ; CHECK-D:       # %bb.0:
 ; CHECK-D-NEXT:    vsetvli a0, zero, e64, m8, ta, ma
@@ -2835,20 +2771,6 @@ define <vscale x 1 x i64> @ctlz_zero_undef_nxv1i64(<vscale x 1 x i64> %va) {
 ; RV64I-NEXT:    vsrl.vx v8, v8, a0
 ; RV64I-NEXT:    ret
 ;
-; CHECK-F-LABEL: ctlz_zero_undef_nxv1i64:
-; CHECK-F:       # %bb.0:
-; CHECK-F-NEXT:    li a0, 190
-; CHECK-F-NEXT:    vsetvli a1, zero, e64, m1, ta, ma
-; CHECK-F-NEXT:    vmv.v.x v9, a0
-; CHECK-F-NEXT:    vsetvli zero, zero, e32, mf2, ta, ma
-; CHECK-F-NEXT:    fsrmi a0, 1
-; CHECK-F-NEXT:    vfncvt.f.xu.w v10, v8
-; CHECK-F-NEXT:    vsrl.vi v8, v10, 23
-; CHECK-F-NEXT:    vwsubu.wv v9, v9, v8
-; CHECK-F-NEXT:    fsrm a0
-; CHECK-F-NEXT:    vmv1r.v v8, v9
-; CHECK-F-NEXT:    ret
-;
 ; CHECK-D-LABEL: ctlz_zero_undef_nxv1i64:
 ; CHECK-D:       # %bb.0:
 ; CHECK-D-NEXT:    vsetvli a0, zero, e64, m1, ta, ma
@@ -2971,20 +2893,6 @@ define <vscale x 2 x i64> @ctlz_zero_undef_nxv2i64(<vscale x 2 x i64> %va) {
 ; RV64I-NEXT:    vsrl.vx v8, v8, a0
 ; RV64I-NEXT:    ret
 ;
-; CHECK-F-LABEL: ctlz_zero_undef_nxv2i64:
-; CHECK-F:       # %bb.0:
-; CHECK-F-NEXT:    li a0, 190
-; CHECK-F-NEXT:    vsetvli a1, zero, e64, m2, ta, ma
-; CHECK-F-NEXT:    vmv.v.x v10, a0
-; CHECK-F-NEXT:    vsetvli zero, zero, e32, m1, ta, ma
-; CHECK-F-NEXT:    fsrmi a0, 1
-; CHECK-F-NEXT:    vfncvt.f.xu.w v12, v8
-; CHECK-F-NEXT:    vsrl.vi v8, v12, 23
-; CHECK-F-NEXT:    vwsubu.wv v10, v10, v8
-; CHECK-F-NEXT:    fsrm a0
-; CHECK-F-NEXT:    vmv2r.v v8, v10
-; CHECK-F-NEXT:    ret
-;
 ; CHECK-D-LABEL: ctlz_zero_undef_nxv2i64:
 ; CHECK-D:       # %bb.0:
 ; CHECK-D-NEXT:    vsetvli a0, zero, e64, m2, ta, ma
@@ -3107,20 +3015,6 @@ define <vscale x 4 x i64> @ctlz_zero_undef_nxv4i64(<vscale x 4 x i64> %va) {
 ; RV64I-NEXT:    vsrl.vx v8, v8, a0
 ; RV64I-NEXT:    ret
 ;
-; CHECK-F-LABEL: ctlz_zero_undef_nxv4i64:
-; CHECK-F:       # %bb.0:
-; CHECK-F-NEXT:    li a0, 190
-; CHECK-F-NEXT:    vsetvli a1, zero, e64, m4, ta, ma
-; CHECK-F-NEXT:    vmv.v.x v12, a0
-; CHECK-F-NEXT:    vsetvli zero, zero, e32, m2, ta, ma
-; CHECK-F-NEXT:    fsrmi a0, 1
-; CHECK-F-NEXT:    vfncvt.f.xu.w v16, v8
-; CHECK-F-NEXT:    vsrl.vi v8, v16, 23
-; CHECK-F-NEXT:    vwsubu.wv v12, v12, v8
-; CHECK-F-NEXT:    fsrm a0
-; CHECK-F-NEXT:    vmv4r.v v8, v12
-; CHECK-F-NEXT:    ret
-;
 ; CHECK-D-LABEL: ctlz_zero_undef_nxv4i64:
 ; CHECK-D:       # %bb.0:
 ; CHECK-D-NEXT:    vsetvli a0, zero, e64, m4, ta, ma
@@ -3243,20 +3137,6 @@ define <vscale x 8 x i64> @ctlz_zero_undef_nxv8i64(<vscale x 8 x i64> %va) {
 ; RV64I-NEXT:    vsrl.vx v8, v8, a0
 ; RV64I-NEXT:    ret
 ;
-; CHECK-F-LABEL: ctlz_zero_undef_nxv8i64:
-; CHECK-F:       # %bb.0:
-; CHECK-F-NEXT:    vmv8r.v v16, v8
-; CHECK-F-NEXT:    li a0, 190
-; CHECK-F-NEXT:    vsetvli a1, zero, e64, m8, ta, ma
-; CHECK-F-NEXT:    vmv.v.x v8, a0
-; CHECK-F-NEXT:    vsetvli zero, zero, e32, m4, ta, ma
-; CHECK-F-NEXT:    fsrmi a0, 1
-; CHECK-F-NEXT:    vfncvt.f.xu.w v24, v16
-; CHECK-F-NEXT:    vsrl.vi v16, v24, 23
-; CHECK-F-NEXT:    vwsubu.wv v8, v8, v16
-; CHECK-F-NEXT:    fsrm a0
-; CHECK-F-NEXT:    ret
-;
 ; CHECK-D-LABEL: ctlz_zero_undef_nxv8i64:
 ; CHECK-D:       # %bb.0:
 ; CHECK-D-NEXT:    vsetvli a0, zero, e64, m8, ta, ma