[RISCV] Custom legalize vp.merge for mask vectors. (#120479)

The default legalization uses vmslt with a vector of XLen to compute a mask. This doesn't work if the type isn't legal. For fixed vectors it will scalarize. For scalable vectors it crashes the compiler. This patch uses an alternate strategy that promotes the i1 vector to an i8 vector and does the merge. I don't claim this to be the best lowering. I wrote it quickly almost 3 years ago when a crash was reported in our downstream. Fixes #120405.

Author: topperc

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

@@ -758,9 +758,9 @@ RISCVTargetLowering::RISCVTargetLowering(const TargetMachine &TM,
  Custom);
 
  setOperationAction(ISD::SELECT, VT, Custom);
- setOperationAction(
- {ISD::SELECT_CC, ISD::VSELECT, ISD::VP_MERGE, ISD::VP_SELECT}, VT,
-  Expand);
+ setOperationAction({ISD::SELECT_CC, ISD::VSELECT, ISD::VP_SELECT}, VT,
+  Expand);
+ setOperationAction(ISD::VP_MERGE, VT, Custom);
 
  setOperationAction({ISD::VP_CTTZ_ELTS, ISD::VP_CTTZ_ELTS_ZERO_UNDEF}, VT,
  Custom);
@@ -1237,6 +1237,8 @@ RISCVTargetLowering::RISCVTargetLowering(const TargetMachine &TM,
  ISD::VP_SETCC, ISD::VP_TRUNCATE},
  VT, Custom);
 
+ setOperationAction(ISD::VP_MERGE, VT, Custom);
+
  setOperationAction(ISD::EXPERIMENTAL_VP_SPLICE, VT, Custom);
  setOperationAction(ISD::EXPERIMENTAL_VP_REVERSE, VT, Custom);
  continue;
@@ -7492,8 +7494,11 @@ SDValue RISCVTargetLowering::LowerOperation(SDValue Op,
  return lowerSET_ROUNDING(Op, DAG);
  case ISD::EH_DWARF_CFA:
  return lowerEH_DWARF_CFA(Op, DAG);
- case ISD::VP_SELECT:
  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:
@@ -12078,6 +12083,65 @@ SDValue RISCVTargetLowering::lowerVPFPIntConvOp(SDValue Op,
  return convertFromScalableVector(VT, Result, DAG, Subtarget);
 }
 
+SDValue RISCVTargetLowering::lowerVPMergeMask(SDValue Op,
+ SelectionDAG &DAG) const {
+ SDLoc DL(Op);
+ MVT VT = Op.getSimpleValueType();
+ MVT XLenVT = Subtarget.getXLenVT();
+
+ SDValue Mask = Op.getOperand(0);
+ SDValue TrueVal = Op.getOperand(1);
+ SDValue FalseVal = Op.getOperand(2);
+ SDValue VL = Op.getOperand(3);
+
+ // Use default legalization if a vector of EVL type would be legal.
+ EVT EVLVecVT = EVT::getVectorVT(*DAG.getContext(), VL.getValueType(),
+ VT.getVectorElementCount());
+ if (isTypeLegal(EVLVecVT))
+ return SDValue();
+
+ MVT ContainerVT = VT;
+ if (VT.isFixedLengthVector()) {
+ ContainerVT = getContainerForFixedLengthVector(VT);
+ Mask = convertToScalableVector(ContainerVT, Mask, DAG, Subtarget);
+ TrueVal = convertToScalableVector(ContainerVT, TrueVal, DAG, Subtarget);
+ FalseVal = convertToScalableVector(ContainerVT, FalseVal, DAG, Subtarget);
+ }
+
+ // Promote to a vector of i8.
+ MVT PromotedVT = ContainerVT.changeVectorElementType(MVT::i8);
+
+ // Promote TrueVal and FalseVal using VLMax.
+ // FIXME: Is there a better way to do this?
+ SDValue VLMax = DAG.getRegister(RISCV::X0, XLenVT);
+ SDValue SplatOne = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, PromotedVT,
+ DAG.getUNDEF(PromotedVT),
+ DAG.getConstant(1, DL, XLenVT), VLMax);
+ SDValue SplatZero = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, PromotedVT,
+ DAG.getUNDEF(PromotedVT),
+ DAG.getConstant(0, DL, XLenVT), VLMax);
+ TrueVal = DAG.getNode(RISCVISD::VMERGE_VL, DL, PromotedVT, TrueVal, SplatOne,
+ SplatZero, DAG.getUNDEF(PromotedVT), VL);
+ // Any element past VL uses FalseVal, so use VLMax
+ FalseVal = DAG.getNode(RISCVISD::VMERGE_VL, DL, PromotedVT, FalseVal,
+ SplatOne, SplatZero, DAG.getUNDEF(PromotedVT), VLMax);
+
+ // VP_MERGE the two promoted values.
+ SDValue VPMerge = DAG.getNode(RISCVISD::VMERGE_VL, DL, PromotedVT, Mask,
+ TrueVal, FalseVal, FalseVal, VL);
+
+ // Convert back to mask.
+ SDValue TrueMask = DAG.getNode(RISCVISD::VMSET_VL, DL, ContainerVT, VL);
+ SDValue Result = DAG.getNode(
+ RISCVISD::SETCC_VL, DL, ContainerVT,
+ {VPMerge, DAG.getConstant(0, DL, PromotedVT), DAG.getCondCode(ISD::SETNE),
+ DAG.getUNDEF(getMaskTypeFor(ContainerVT)), TrueMask, VLMax});
+
+ if (VT.isFixedLengthVector())
+ Result = convertFromScalableVector(VT, Result, DAG, Subtarget);
+ return Result;
+}
+
 SDValue
 RISCVTargetLowering::lowerVPSpliceExperimental(SDValue Op,
  SelectionDAG &DAG) const {

llvm/lib/Target/RISCV/RISCVISelLowering.h

@@ -996,6 +996,7 @@ class RISCVTargetLowering : public TargetLowering {
  SDValue lowerLogicVPOp(SDValue Op, SelectionDAG &DAG) const;
  SDValue lowerVPExtMaskOp(SDValue Op, SelectionDAG &DAG) const;
  SDValue lowerVPSetCCMaskOp(SDValue Op, SelectionDAG &DAG) const;
+ SDValue lowerVPMergeMask(SDValue Op, SelectionDAG &DAG) const;
  SDValue lowerVPSplatExperimental(SDValue Op, SelectionDAG &DAG) const;
  SDValue lowerVPSpliceExperimental(SDValue Op, SelectionDAG &DAG) const;
  SDValue lowerVPReverseExperimental(SDValue Op, SelectionDAG &DAG) const;

llvm/test/CodeGen/RISCV/rvv/fixed-vectors-vpmerge.ll

@@ -58,6 +58,182 @@ define <4 x i1> @vpmerge_vv_v4i1(<4 x i1> %va, <4 x i1> %vb, <4 x i1> %m, i32 ze
  ret <4 x i1> %v
 }
 
+define <8 x i1> @vpmerge_vv_v8i1(<8 x i1> %va, <8 x i1> %vb, <8 x i1> %m, i32 zeroext %evl) {
+; RV32-LABEL: vpmerge_vv_v8i1:
+; RV32: # %bb.0:
+; RV32-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; RV32-NEXT: vid.v v10
+; RV32-NEXT: vmsltu.vx v12, v10, a0
+; RV32-NEXT: vmand.mm v9, v9, v12
+; RV32-NEXT: vmandn.mm v8, v8, v9
+; RV32-NEXT: vmand.mm v9, v0, v9
+; RV32-NEXT: vmor.mm v0, v9, v8
+; RV32-NEXT: ret
+;
+; RV64-LABEL: vpmerge_vv_v8i1:
+; RV64: # %bb.0:
+; RV64-NEXT: vsetivli zero, 8, e64, m4, ta, ma
+; RV64-NEXT: vid.v v12
+; RV64-NEXT: vmsltu.vx v10, v12, a0
+; RV64-NEXT: vmand.mm v9, v9, v10
+; RV64-NEXT: vmandn.mm v8, v8, v9
+; RV64-NEXT: vmand.mm v9, v0, v9
+; RV64-NEXT: vmor.mm v0, v9, v8
+; RV64-NEXT: ret
+;
+; RV32ZVFHMIN-LABEL: vpmerge_vv_v8i1:
+; RV32ZVFHMIN: # %bb.0:
+; RV32ZVFHMIN-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; RV32ZVFHMIN-NEXT: vid.v v10
+; RV32ZVFHMIN-NEXT: vmsltu.vx v12, v10, a0
+; RV32ZVFHMIN-NEXT: vmand.mm v9, v9, v12
+; RV32ZVFHMIN-NEXT: vmandn.mm v8, v8, v9
+; RV32ZVFHMIN-NEXT: vmand.mm v9, v0, v9
+; RV32ZVFHMIN-NEXT: vmor.mm v0, v9, v8
+; RV32ZVFHMIN-NEXT: ret
+;
+; RV64ZVFHMIN-LABEL: vpmerge_vv_v8i1:
+; RV64ZVFHMIN: # %bb.0:
+; RV64ZVFHMIN-NEXT: vsetivli zero, 8, e64, m4, ta, ma
+; RV64ZVFHMIN-NEXT: vid.v v12
+; RV64ZVFHMIN-NEXT: vmsltu.vx v10, v12, a0
+; RV64ZVFHMIN-NEXT: vmand.mm v9, v9, v10
+; RV64ZVFHMIN-NEXT: vmandn.mm v8, v8, v9
+; RV64ZVFHMIN-NEXT: vmand.mm v9, v0, v9
+; RV64ZVFHMIN-NEXT: vmor.mm v0, v9, v8
+; RV64ZVFHMIN-NEXT: ret
+ %v = call <8 x i1> @llvm.vp.merge.v8i1(<8 x i1> %m, <8 x i1> %va, <8 x i1> %vb, i32 %evl)
+ ret <8 x i1> %v
+}
+
+define <16 x i1> @vpmerge_vv_v16i1(<16 x i1> %va, <16 x i1> %vb, <16 x i1> %m, i32 zeroext %evl) {
+; RV32-LABEL: vpmerge_vv_v16i1:
+; RV32: # %bb.0:
+; RV32-NEXT: vsetivli zero, 16, e32, m4, ta, ma
+; RV32-NEXT: vid.v v12
+; RV32-NEXT: vmsltu.vx v10, v12, a0
+; RV32-NEXT: vmand.mm v9, v9, v10
+; RV32-NEXT: vmandn.mm v8, v8, v9
+; RV32-NEXT: vmand.mm v9, v0, v9
+; RV32-NEXT: vmor.mm v0, v9, v8
+; RV32-NEXT: ret
+;
+; RV64-LABEL: vpmerge_vv_v16i1:
+; RV64: # %bb.0:
+; RV64-NEXT: vsetivli zero, 16, e64, m8, ta, ma
+; RV64-NEXT: vid.v v16
+; RV64-NEXT: vmsltu.vx v10, v16, a0
+; RV64-NEXT: vmand.mm v9, v9, v10
+; RV64-NEXT: vmandn.mm v8, v8, v9
+; RV64-NEXT: vmand.mm v9, v0, v9
+; RV64-NEXT: vmor.mm v0, v9, v8
+; RV64-NEXT: ret
+;
+; RV32ZVFHMIN-LABEL: vpmerge_vv_v16i1:
+; RV32ZVFHMIN: # %bb.0:
+; RV32ZVFHMIN-NEXT: vsetivli zero, 16, e32, m4, ta, ma
+; RV32ZVFHMIN-NEXT: vid.v v12
+; RV32ZVFHMIN-NEXT: vmsltu.vx v10, v12, a0
+; RV32ZVFHMIN-NEXT: vmand.mm v9, v9, v10
+; RV32ZVFHMIN-NEXT: vmandn.mm v8, v8, v9
+; RV32ZVFHMIN-NEXT: vmand.mm v9, v0, v9
+; RV32ZVFHMIN-NEXT: vmor.mm v0, v9, v8
+; RV32ZVFHMIN-NEXT: ret
+;
+; RV64ZVFHMIN-LABEL: vpmerge_vv_v16i1:
+; RV64ZVFHMIN: # %bb.0:
+; RV64ZVFHMIN-NEXT: vsetivli zero, 16, e64, m8, ta, ma
+; RV64ZVFHMIN-NEXT: vid.v v16
+; RV64ZVFHMIN-NEXT: vmsltu.vx v10, v16, a0
+; RV64ZVFHMIN-NEXT: vmand.mm v9, v9, v10
+; RV64ZVFHMIN-NEXT: vmandn.mm v8, v8, v9
+; RV64ZVFHMIN-NEXT: vmand.mm v9, v0, v9
+; RV64ZVFHMIN-NEXT: vmor.mm v0, v9, v8
+; RV64ZVFHMIN-NEXT: ret
+ %v = call <16 x i1> @llvm.vp.merge.v16i1(<16 x i1> %m, <16 x i1> %va, <16 x i1> %vb, i32 %evl)
+ ret <16 x i1> %v
+}
+
+define <32 x i1> @vpmerge_vv_v32i1(<32 x i1> %va, <32 x i1> %vb, <32 x i1> %m, i32 zeroext %evl) {
+; RV32-LABEL: vpmerge_vv_v32i1:
+; RV32: # %bb.0:
+; RV32-NEXT: li a1, 32
+; RV32-NEXT: vsetvli zero, a1, e32, m8, ta, ma
+; RV32-NEXT: vid.v v16
+; RV32-NEXT: vmsltu.vx v10, v16, a0
+; RV32-NEXT: vmand.mm v9, v9, v10
+; RV32-NEXT: vmandn.mm v8, v8, v9
+; RV32-NEXT: vmand.mm v9, v0, v9
+; RV32-NEXT: vmor.mm v0, v9, v8
+; RV32-NEXT: ret
+;
+; RV64-LABEL: vpmerge_vv_v32i1:
+; RV64: # %bb.0:
+; RV64-NEXT: vsetvli a1, zero, e8, m2, ta, ma
+; RV64-NEXT: vmv.v.i v10, 0
+; RV64-NEXT: vsetvli zero, a0, e8, m2, ta, ma
+; RV64-NEXT: vmerge.vim v12, v10, 1, v0
+; RV64-NEXT: vmv1r.v v0, v8
+; RV64-NEXT: vsetvli a1, zero, e8, m2, ta, ma
+; RV64-NEXT: vmerge.vim v10, v10, 1, v0
+; RV64-NEXT: vmv1r.v v0, v9
+; RV64-NEXT: vsetvli zero, a0, e8, m2, tu, ma
+; RV64-NEXT: vmerge.vvm v10, v10, v12, v0
+; RV64-NEXT: vsetvli a0, zero, e8, m2, ta, ma
+; RV64-NEXT: vmsne.vi v0, v10, 0
+; RV64-NEXT: ret
+;
+; RV32ZVFHMIN-LABEL: vpmerge_vv_v32i1:
+; RV32ZVFHMIN: # %bb.0:
+; RV32ZVFHMIN-NEXT: li a1, 32
+; RV32ZVFHMIN-NEXT: vsetvli zero, a1, e32, m8, ta, ma
+; RV32ZVFHMIN-NEXT: vid.v v16
+; RV32ZVFHMIN-NEXT: vmsltu.vx v10, v16, a0
+; RV32ZVFHMIN-NEXT: vmand.mm v9, v9, v10
+; RV32ZVFHMIN-NEXT: vmandn.mm v8, v8, v9
+; RV32ZVFHMIN-NEXT: vmand.mm v9, v0, v9
+; RV32ZVFHMIN-NEXT: vmor.mm v0, v9, v8
+; RV32ZVFHMIN-NEXT: ret
+;
+; RV64ZVFHMIN-LABEL: vpmerge_vv_v32i1:
+; RV64ZVFHMIN: # %bb.0:
+; RV64ZVFHMIN-NEXT: vsetvli a1, zero, e8, m2, ta, ma
+; RV64ZVFHMIN-NEXT: vmv.v.i v10, 0
+; RV64ZVFHMIN-NEXT: vsetvli zero, a0, e8, m2, ta, ma
+; RV64ZVFHMIN-NEXT: vmerge.vim v12, v10, 1, v0
+; RV64ZVFHMIN-NEXT: vmv1r.v v0, v8
+; RV64ZVFHMIN-NEXT: vsetvli a1, zero, e8, m2, ta, ma
+; RV64ZVFHMIN-NEXT: vmerge.vim v10, v10, 1, v0
+; RV64ZVFHMIN-NEXT: vmv1r.v v0, v9
+; RV64ZVFHMIN-NEXT: vsetvli zero, a0, e8, m2, tu, ma
+; RV64ZVFHMIN-NEXT: vmerge.vvm v10, v10, v12, v0
+; RV64ZVFHMIN-NEXT: vsetvli a0, zero, e8, m2, ta, ma
+; RV64ZVFHMIN-NEXT: vmsne.vi v0, v10, 0
+; RV64ZVFHMIN-NEXT: ret
+ %v = call <32 x i1> @llvm.vp.merge.v32i1(<32 x i1> %m, <32 x i1> %va, <32 x i1> %vb, i32 %evl)
+ ret <32 x i1> %v
+}
+
+define <64 x i1> @vpmerge_vv_v64i1(<64 x i1> %va, <64 x i1> %vb, <64 x i1> %m, i32 zeroext %evl) {
+; CHECK-LABEL: vpmerge_vv_v64i1:
+; CHECK: # %bb.0:
+; CHECK-NEXT: vsetvli a1, zero, e8, m4, ta, ma
+; CHECK-NEXT: vmv.v.i v12, 0
+; CHECK-NEXT: vsetvli zero, a0, e8, m4, ta, ma
+; CHECK-NEXT: vmerge.vim v16, v12, 1, v0
+; CHECK-NEXT: vmv1r.v v0, v8
+; CHECK-NEXT: vsetvli a1, zero, e8, m4, ta, ma
+; CHECK-NEXT: vmerge.vim v12, v12, 1, v0
+; CHECK-NEXT: vmv1r.v v0, v9
+; CHECK-NEXT: vsetvli zero, a0, e8, m4, tu, ma
+; CHECK-NEXT: vmerge.vvm v12, v12, v16, v0
+; CHECK-NEXT: vsetvli a0, zero, e8, m4, ta, ma
+; CHECK-NEXT: vmsne.vi v0, v12, 0
+; CHECK-NEXT: ret
+ %v = call <64 x i1> @llvm.vp.merge.v64i1(<64 x i1> %m, <64 x i1> %va, <64 x i1> %vb, i32 %evl)
+ ret <64 x i1> %v
+}
+
 declare <2 x i8> @llvm.vp.merge.v2i8(<2 x i1>, <2 x i8>, <2 x i8>, i32)
 
 define <2 x i8> @vpmerge_vv_v2i8(<2 x i8> %va, <2 x i8> %vb, <2 x i1> %m, i32 zeroext %evl) {
@@ -1188,10 +1364,10 @@ define <32 x double> @vpmerge_vv_v32f64(<32 x double> %va, <32 x double> %vb, <3
 ; CHECK-NEXT: vle64.v v8, (a0)
 ; CHECK-NEXT: li a1, 16
 ; CHECK-NEXT: mv a0, a2
-; CHECK-NEXT: bltu a2, a1, .LBB79_2
+; CHECK-NEXT: bltu a2, a1, .LBB83_2
 ; CHECK-NEXT: # %bb.1:
 ; CHECK-NEXT: li a0, 16
-; CHECK-NEXT: .LBB79_2:
+; CHECK-NEXT: .LBB83_2:
 ; CHECK-NEXT: vsetvli zero, a0, e64, m8, tu, ma
 ; CHECK-NEXT: vmerge.vvm v8, v8, v16, v0
 ; CHECK-NEXT: addi a0, a2, -16
@@ -1221,10 +1397,10 @@ define <32 x double> @vpmerge_vf_v32f64(double %a, <32 x double> %vb, <32 x i1>
 ; CHECK: # %bb.0:
 ; CHECK-NEXT: li a2, 16
 ; CHECK-NEXT: mv a1, a0
-; CHECK-NEXT: bltu a0, a2, .LBB80_2
+; CHECK-NEXT: bltu a0, a2, .LBB84_2
 ; CHECK-NEXT: # %bb.1:
 ; CHECK-NEXT: li a1, 16
-; CHECK-NEXT: .LBB80_2:
+; CHECK-NEXT: .LBB84_2:
 ; CHECK-NEXT: vsetvli zero, a1, e64, m8, tu, ma
 ; CHECK-NEXT: vfmerge.vfm v8, v8, fa0, v0
 ; CHECK-NEXT: addi a1, a0, -16