[LoongArch][CodeGen] Implement 128-bit and 256-bit vector shuffle.

llvm/lib/Target/LoongArch/LoongArchISelLowering.h

@@ -120,6 +120,16 @@ enum NodeType : unsigned {
 
   // Vector Shuffle
   VREPLVE,
+  VSHUF,
+  VPICKEV,
+  VPICKOD,
+  VPACKEV,
+  VPACKOD,
+  VILVL,
+  VILVH,
+  VSHUF4I,
+  VREPLVEI,
+  XVPERMI,
 
   // Extended vector element extraction
   VPICK_SEXT_ELT,

llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td

@@ -10,6 +10,8 @@
 //
 //===----------------------------------------------------------------------===//
 
+def loongarch_xvpermi: SDNode<"LoongArchISD::XVPERMI", SDT_loongArchV1RUimm>;
+
 def lasxsplati8
   : PatFrag<(ops node:$e0),
             (v32i8 (build_vector node:$e0, node:$e0, node:$e0, node:$e0,
@@ -1575,6 +1577,134 @@ def : Pat<(loongarch_vreplve v8i32:$xj, GRLenVT:$rk),
 def : Pat<(loongarch_vreplve v4i64:$xj, GRLenVT:$rk),
           (XVREPLVE_D v4i64:$xj, GRLenVT:$rk)>;
 
+// XVSHUF_{B/H/W/D}
+def : Pat<(loongarch_vshuf v32i8:$xa, v32i8:$xj, v32i8:$xk),
+          (XVSHUF_B v32i8:$xj, v32i8:$xk, v32i8:$xa)>;
+def : Pat<(loongarch_vshuf v16i16:$xd, v16i16:$xj, v16i16:$xk),
+          (XVSHUF_H v16i16:$xd, v16i16:$xj, v16i16:$xk)>;
+def : Pat<(loongarch_vshuf v8i32:$xd, v8i32:$xj, v8i32:$xk),
+          (XVSHUF_W v8i32:$xd, v8i32:$xj, v8i32:$xk)>;
+def : Pat<(loongarch_vshuf v4i64:$xd, v4i64:$xj, v4i64:$xk),
+          (XVSHUF_D v4i64:$xd, v4i64:$xj, v4i64:$xk)>;
+def : Pat<(loongarch_vshuf v8i32:$xd, v8f32:$xj, v8f32:$xk),
+          (XVSHUF_W v8i32:$xd, v8f32:$xj, v8f32:$xk)>;
+def : Pat<(loongarch_vshuf v4i64:$xd, v4f64:$xj, v4f64:$xk),
+          (XVSHUF_D v4i64:$xd, v4f64:$xj, v4f64:$xk)>;
+
+// XVPICKEV_{B/H/W/D}
+def : Pat<(loongarch_vpickev v32i8:$xj, v32i8:$xk),
+          (XVPICKEV_B v32i8:$xj, v32i8:$xk)>;
+def : Pat<(loongarch_vpickev v16i16:$xj, v16i16:$xk),
+          (XVPICKEV_H v16i16:$xj, v16i16:$xk)>;
+def : Pat<(loongarch_vpickev v8i32:$xj, v8i32:$xk),
+          (XVPICKEV_W v8i32:$xj, v8i32:$xk)>;
+def : Pat<(loongarch_vpickev v4i64:$xj, v4i64:$xk),
+          (XVPICKEV_D v4i64:$xj, v4i64:$xk)>;
+def : Pat<(loongarch_vpickev v8f32:$xj, v8f32:$xk),
+          (XVPICKEV_W v8f32:$xj, v8f32:$xk)>;
+def : Pat<(loongarch_vpickev v4f64:$xj, v4f64:$xk),
+          (XVPICKEV_D v4f64:$xj, v4f64:$xk)>;
+
+// XVPICKOD_{B/H/W/D}
+def : Pat<(loongarch_vpickod v32i8:$xj, v32i8:$xk),
+          (XVPICKOD_B v32i8:$xj, v32i8:$xk)>;
+def : Pat<(loongarch_vpickod v16i16:$xj, v16i16:$xk),
+          (XVPICKOD_H v16i16:$xj, v16i16:$xk)>;
+def : Pat<(loongarch_vpickod v8i32:$xj, v8i32:$xk),
+          (XVPICKOD_W v8i32:$xj, v8i32:$xk)>;
+def : Pat<(loongarch_vpickod v4i64:$xj, v4i64:$xk),
+          (XVPICKOD_D v4i64:$xj, v4i64:$xk)>;
+def : Pat<(loongarch_vpickod v8f32:$xj, v8f32:$xk),
+          (XVPICKOD_W v8f32:$xj, v8f32:$xk)>;
+def : Pat<(loongarch_vpickod v4f64:$xj, v4f64:$xk),
+          (XVPICKOD_D v4f64:$xj, v4f64:$xk)>;
+
+// XVPACKEV_{B/H/W/D}
+def : Pat<(loongarch_vpackev v32i8:$xj, v32i8:$xk),
+          (XVPACKEV_B v32i8:$xj, v32i8:$xk)>;
+def : Pat<(loongarch_vpackev v16i16:$xj, v16i16:$xk),
+          (XVPACKEV_H v16i16:$xj, v16i16:$xk)>;
+def : Pat<(loongarch_vpackev v8i32:$xj, v8i32:$xk),
+          (XVPACKEV_W v8i32:$xj, v8i32:$xk)>;
+def : Pat<(loongarch_vpackev v4i64:$xj, v4i64:$xk),
+          (XVPACKEV_D v4i64:$xj, v4i64:$xk)>;
+def : Pat<(loongarch_vpackev v8f32:$xj, v8f32:$xk),
+          (XVPACKEV_W v8f32:$xj, v8f32:$xk)>;
+def : Pat<(loongarch_vpackev v4f64:$xj, v4f64:$xk),
+          (XVPACKEV_D v4f64:$xj, v4f64:$xk)>;
+
+// XVPACKOD_{B/H/W/D}
+def : Pat<(loongarch_vpackod v32i8:$xj, v32i8:$xk),
+          (XVPACKOD_B v32i8:$xj, v32i8:$xk)>;
+def : Pat<(loongarch_vpackod v16i16:$xj, v16i16:$xk),
+          (XVPACKOD_H v16i16:$xj, v16i16:$xk)>;
+def : Pat<(loongarch_vpackod v8i32:$xj, v8i32:$xk),
+          (XVPACKOD_W v8i32:$xj, v8i32:$xk)>;
+def : Pat<(loongarch_vpackod v4i64:$xj, v4i64:$xk),
+          (XVPACKOD_D v4i64:$xj, v4i64:$xk)>;
+def : Pat<(loongarch_vpackod v8f32:$xj, v8f32:$xk),
+          (XVPACKOD_W v8f32:$xj, v8f32:$xk)>;
+def : Pat<(loongarch_vpackod v4f64:$xj, v4f64:$xk),
+          (XVPACKOD_D v4f64:$xj, v4f64:$xk)>;
+
+// XVILVL_{B/H/W/D}
+def : Pat<(loongarch_vilvl v32i8:$xj, v32i8:$xk),
+          (XVILVL_B v32i8:$xj, v32i8:$xk)>;
+def : Pat<(loongarch_vilvl v16i16:$xj, v16i16:$xk),
+          (XVILVL_H v16i16:$xj, v16i16:$xk)>;
+def : Pat<(loongarch_vilvl v8i32:$xj, v8i32:$xk),
+          (XVILVL_W v8i32:$xj, v8i32:$xk)>;
+def : Pat<(loongarch_vilvl v4i64:$xj, v4i64:$xk),
+          (XVILVL_D v4i64:$xj, v4i64:$xk)>;
+def : Pat<(loongarch_vilvl v8f32:$xj, v8f32:$xk),
+          (XVILVL_W v8f32:$xj, v8f32:$xk)>;
+def : Pat<(loongarch_vilvl v4f64:$xj, v4f64:$xk),
+          (XVILVL_D v4f64:$xj, v4f64:$xk)>;
+
+// XVILVH_{B/H/W/D}
+def : Pat<(loongarch_vilvh v32i8:$xj, v32i8:$xk),
+          (XVILVH_B v32i8:$xj, v32i8:$xk)>;
+def : Pat<(loongarch_vilvh v16i16:$xj, v16i16:$xk),
+          (XVILVH_H v16i16:$xj, v16i16:$xk)>;
+def : Pat<(loongarch_vilvh v8i32:$xj, v8i32:$xk),
+          (XVILVH_W v8i32:$xj, v8i32:$xk)>;
+def : Pat<(loongarch_vilvh v4i64:$xj, v4i64:$xk),
+          (XVILVH_D v4i64:$xj, v4i64:$xk)>;
+def : Pat<(loongarch_vilvh v8f32:$xj, v8f32:$xk),
+          (XVILVH_W v8f32:$xj, v8f32:$xk)>;
+def : Pat<(loongarch_vilvh v4f64:$xj, v4f64:$xk),
+          (XVILVH_D v4f64:$xj, v4f64:$xk)>;
+
+// XVSHUF4I_{B/H/W}
+def : Pat<(loongarch_vshuf4i v32i8:$xj, immZExt8:$ui8),
+          (XVSHUF4I_B v32i8:$xj, immZExt8:$ui8)>;
+def : Pat<(loongarch_vshuf4i v16i16:$xj, immZExt8:$ui8),
+        (XVSHUF4I_H v16i16:$xj, immZExt8:$ui8)>;
+def : Pat<(loongarch_vshuf4i v8i32:$xj, immZExt8:$ui8),
+        (XVSHUF4I_W v8i32:$xj, immZExt8:$ui8)>;
+def : Pat<(loongarch_vshuf4i v8f32:$xj, immZExt8:$ui8),
+        (XVSHUF4I_W v8f32:$xj, immZExt8:$ui8)>;
+
+// XVREPL128VEI_{B/H/W/D}
+def : Pat<(loongarch_vreplvei v32i8:$xj, immZExt4:$ui4),
+          (XVREPL128VEI_B v32i8:$xj, immZExt4:$ui4)>;
+def : Pat<(loongarch_vreplvei v16i16:$xj, immZExt3:$ui3),
+        (XVREPL128VEI_H v16i16:$xj, immZExt3:$ui3)>;
+def : Pat<(loongarch_vreplvei v8i32:$xj, immZExt2:$ui2),
+        (XVREPL128VEI_W v8i32:$xj, immZExt2:$ui2)>;
+def : Pat<(loongarch_vreplvei v4i64:$xj, immZExt1:$ui1),
+        (XVREPL128VEI_D v4i64:$xj, immZExt1:$ui1)>;
+def : Pat<(loongarch_vreplvei v8f32:$xj, immZExt2:$ui2),
+        (XVREPL128VEI_W v8f32:$xj, immZExt2:$ui2)>;
+def : Pat<(loongarch_vreplvei v4f64:$xj, immZExt1:$ui1),
+        (XVREPL128VEI_D v4f64:$xj, immZExt1:$ui1)>;
+
+// XVPERMI_D
+def : Pat<(loongarch_xvpermi v4i64:$xj, immZExt8: $ui8),
+          (XVPERMI_D v4i64:$xj, immZExt8: $ui8)>;
+def : Pat<(loongarch_xvpermi v4f64:$xj, immZExt8: $ui8),
+          (XVPERMI_D v4f64:$xj, immZExt8: $ui8)>;
+
 // XVREPLVE0_{W/D}
 def : Pat<(lasxsplatf32 FPR32:$fj),
           (XVREPLVE0_W (SUBREG_TO_REG (i64 0), FPR32:$fj, sub_32))>;

llvm/lib/Target/LoongArch/LoongArchLSXInstrInfo.td

@@ -15,6 +15,15 @@ def SDT_LoongArchVreplve : SDTypeProfile<1, 2, [SDTCisInt<0>, SDTCisVec<0>,
                                          SDTCisSameAs<0, 1>, SDTCisInt<2>]>;
 def SDT_LoongArchVecCond : SDTypeProfile<1, 1, [SDTCisInt<0>, SDTCisVec<1>]>;
 
+def SDT_LoongArchVShuf : SDTypeProfile<1, 3, [SDTCisVec<0>,
+                                         SDTCisInt<1>, SDTCisVec<1>,
+                                         SDTCisSameAs<0, 2>,
+                                         SDTCisSameAs<2, 3>]>;
+def SDT_LoongArchV2R : SDTypeProfile<1, 2, [SDTCisVec<0>,
+                                         SDTCisSameAs<0, 1>, SDTCisSameAs<1, 2>]>;
+def SDT_loongArchV1RUimm: SDTypeProfile<1, 2, [SDTCisVec<0>,
+                                         SDTCisSameAs<0,1>, SDTCisVT<2, i64>]>;
+
 // Target nodes.
 def loongarch_vreplve : SDNode<"LoongArchISD::VREPLVE", SDT_LoongArchVreplve>;
 def loongarch_vall_nonzero : SDNode<"LoongArchISD::VALL_NONZERO",
@@ -31,6 +40,23 @@ def loongarch_vpick_sext_elt : SDNode<"LoongArchISD::VPICK_SEXT_ELT",
 def loongarch_vpick_zext_elt : SDNode<"LoongArchISD::VPICK_ZEXT_ELT",
                                       SDTypeProfile<1, 3, [SDTCisPtrTy<2>]>>;
 
+def loongarch_vshuf: SDNode<"LoongArchISD::VSHUF", SDT_LoongArchVShuf>;
+def loongarch_vpickev: SDNode<"LoongArchISD::VPICKEV", SDT_LoongArchV2R>;
+def loongarch_vpickod: SDNode<"LoongArchISD::VPICKOD", SDT_LoongArchV2R>;
+def loongarch_vpackev: SDNode<"LoongArchISD::VPACKEV", SDT_LoongArchV2R>;
+def loongarch_vpackod: SDNode<"LoongArchISD::VPACKOD", SDT_LoongArchV2R>;
+def loongarch_vilvl: SDNode<"LoongArchISD::VILVL", SDT_LoongArchV2R>;
+def loongarch_vilvh: SDNode<"LoongArchISD::VILVH", SDT_LoongArchV2R>;
+
+def loongarch_vshuf4i: SDNode<"LoongArchISD::VSHUF4I", SDT_loongArchV1RUimm>;
+def loongarch_vreplvei: SDNode<"LoongArchISD::VREPLVEI", SDT_loongArchV1RUimm>;
+
+def immZExt1 : ImmLeaf<i64, [{return isUInt<1>(Imm);}]>;
+def immZExt2 : ImmLeaf<i64, [{return isUInt<2>(Imm);}]>;
+def immZExt3 : ImmLeaf<i64, [{return isUInt<3>(Imm);}]>;
+def immZExt4 : ImmLeaf<i64, [{return isUInt<4>(Imm);}]>;
+def immZExt8 : ImmLeaf<i64, [{return isUInt<8>(Imm);}]>;
+
 class VecCond<SDPatternOperator OpNode, ValueType TyNode,
               RegisterClass RC = LSX128>
     : Pseudo<(outs GPR:$rd), (ins RC:$vj),
@@ -1682,6 +1708,128 @@ def : Pat<(loongarch_vreplve v4i32:$vj, GRLenVT:$rk),
 def : Pat<(loongarch_vreplve v2i64:$vj, GRLenVT:$rk),
           (VREPLVE_D v2i64:$vj, GRLenVT:$rk)>;
 
+// VSHUF_{B/H/W/D}
+def : Pat<(loongarch_vshuf v16i8:$va, v16i8:$vj, v16i8:$vk),
+          (VSHUF_B v16i8:$vj, v16i8:$vk, v16i8:$va)>;
+def : Pat<(loongarch_vshuf v8i16:$vd, v8i16:$vj, v8i16:$vk),
+          (VSHUF_H v8i16:$vd, v8i16:$vj, v8i16:$vk)>;
+def : Pat<(loongarch_vshuf v4i32:$vd, v4i32:$vj, v4i32:$vk),
+          (VSHUF_W v4i32:$vd, v4i32:$vj, v4i32:$vk)>;
+def : Pat<(loongarch_vshuf v2i64:$vd, v2i64:$vj, v2i64:$vk),
+          (VSHUF_D v2i64:$vd, v2i64:$vj, v2i64:$vk)>;
+def : Pat<(loongarch_vshuf v4i32:$vd, v4f32:$vj, v4f32:$vk),
+          (VSHUF_W v4i32:$vd, v4f32:$vj, v4f32:$vk)>;
+def : Pat<(loongarch_vshuf v2i64:$vd, v2f64:$vj, v2f64:$vk),
+          (VSHUF_D v2i64:$vd, v2f64:$vj, v2f64:$vk)>;
+
+// VPICKEV_{B/H/W/D}
+def : Pat<(loongarch_vpickev v16i8:$vj, v16i8:$vk),
+          (VPICKEV_B v16i8:$vj, v16i8:$vk)>;
+def : Pat<(loongarch_vpickev v8i16:$vj, v8i16:$vk),
+          (VPICKEV_H v8i16:$vj, v8i16:$vk)>;
+def : Pat<(loongarch_vpickev v4i32:$vj, v4i32:$vk),
+          (VPICKEV_W v4i32:$vj, v4i32:$vk)>;
+def : Pat<(loongarch_vpickev v2i64:$vj, v2i64:$vk),
+          (VPICKEV_D v2i64:$vj, v2i64:$vk)>;
+def : Pat<(loongarch_vpickev v4f32:$vj, v4f32:$vk),
+          (VPICKEV_W v4f32:$vj, v4f32:$vk)>;
+def : Pat<(loongarch_vpickev v2f64:$vj, v2f64:$vk),
+          (VPICKEV_D v2f64:$vj, v2f64:$vk)>;
+
+// VPICKOD_{B/H/W/D}
+def : Pat<(loongarch_vpickod v16i8:$vj, v16i8:$vk),
+          (VPICKOD_B v16i8:$vj, v16i8:$vk)>;
+def : Pat<(loongarch_vpickod v8i16:$vj, v8i16:$vk),
+          (VPICKOD_H v8i16:$vj, v8i16:$vk)>;
+def : Pat<(loongarch_vpickod v4i32:$vj, v4i32:$vk),
+          (VPICKOD_W v4i32:$vj, v4i32:$vk)>;
+def : Pat<(loongarch_vpickod v2i64:$vj, v2i64:$vk),
+          (VPICKOD_D v2i64:$vj, v2i64:$vk)>;
+def : Pat<(loongarch_vpickod v4f32:$vj, v4f32:$vk),
+          (VPICKOD_W v4f32:$vj, v4f32:$vk)>;
+def : Pat<(loongarch_vpickod v2f64:$vj, v2f64:$vk),
+          (VPICKOD_D v2f64:$vj, v2f64:$vk)>;
+
+// VPACKEV_{B/H/W/D}
+def : Pat<(loongarch_vpackev v16i8:$vj, v16i8:$vk),
+          (VPACKEV_B v16i8:$vj, v16i8:$vk)>;
+def : Pat<(loongarch_vpackev v8i16:$vj, v8i16:$vk),
+          (VPACKEV_H v8i16:$vj, v8i16:$vk)>;
+def : Pat<(loongarch_vpackev v4i32:$vj, v4i32:$vk),
+          (VPACKEV_W v4i32:$vj, v4i32:$vk)>;
+def : Pat<(loongarch_vpackev v2i64:$vj, v2i64:$vk),
+          (VPACKEV_D v2i64:$vj, v2i64:$vk)>;
+def : Pat<(loongarch_vpackev v4f32:$vj, v4f32:$vk),
+          (VPACKEV_W v4f32:$vj, v4f32:$vk)>;
+def : Pat<(loongarch_vpackev v2f64:$vj, v2f64:$vk),
+          (VPACKEV_D v2f64:$vj, v2f64:$vk)>;
+
+// VPACKOD_{B/H/W/D}
+def : Pat<(loongarch_vpackod v16i8:$vj, v16i8:$vk),
+          (VPACKOD_B v16i8:$vj, v16i8:$vk)>;
+def : Pat<(loongarch_vpackod v8i16:$vj, v8i16:$vk),
+          (VPACKOD_H v8i16:$vj, v8i16:$vk)>;
+def : Pat<(loongarch_vpackod v4i32:$vj, v4i32:$vk),
+          (VPACKOD_W v4i32:$vj, v4i32:$vk)>;
+def : Pat<(loongarch_vpackod v2i64:$vj, v2i64:$vk),
+          (VPACKOD_D v2i64:$vj, v2i64:$vk)>;
+def : Pat<(loongarch_vpackod v4f32:$vj, v4f32:$vk),
+          (VPACKOD_W v4f32:$vj, v4f32:$vk)>;
+def : Pat<(loongarch_vpackod v2f64:$vj, v2f64:$vk),
+          (VPACKOD_D v2f64:$vj, v2f64:$vk)>;
+
+// VILVL_{B/H/W/D}
+def : Pat<(loongarch_vilvl v16i8:$vj, v16i8:$vk),
+          (VILVL_B v16i8:$vj, v16i8:$vk)>;
+def : Pat<(loongarch_vilvl v8i16:$vj, v8i16:$vk),
+          (VILVL_H v8i16:$vj, v8i16:$vk)>;
+def : Pat<(loongarch_vilvl v4i32:$vj, v4i32:$vk),
+          (VILVL_W v4i32:$vj, v4i32:$vk)>;
+def : Pat<(loongarch_vilvl v2i64:$vj, v2i64:$vk),
+          (VILVL_D v2i64:$vj, v2i64:$vk)>;
+def : Pat<(loongarch_vilvl v4f32:$vj, v4f32:$vk),
+          (VILVL_W v4f32:$vj, v4f32:$vk)>;
+def : Pat<(loongarch_vilvl v2f64:$vj, v2f64:$vk),
+          (VILVL_D v2f64:$vj, v2f64:$vk)>;
+
+// VILVH_{B/H/W/D}
+def : Pat<(loongarch_vilvh v16i8:$vj, v16i8:$vk),
+          (VILVH_B v16i8:$vj, v16i8:$vk)>;
+def : Pat<(loongarch_vilvh v8i16:$vj, v8i16:$vk),
+          (VILVH_H v8i16:$vj, v8i16:$vk)>;
+def : Pat<(loongarch_vilvh v4i32:$vj, v4i32:$vk),
+          (VILVH_W v4i32:$vj, v4i32:$vk)>;
+def : Pat<(loongarch_vilvh v2i64:$vj, v2i64:$vk),
+          (VILVH_D v2i64:$vj, v2i64:$vk)>;
+def : Pat<(loongarch_vilvh v4f32:$vj, v4f32:$vk),
+          (VILVH_W v4f32:$vj, v4f32:$vk)>;
+def : Pat<(loongarch_vilvh v2f64:$vj, v2f64:$vk),
+          (VILVH_D v2f64:$vj, v2f64:$vk)>;
+
+// VSHUF4I_{B/H/W}
+def : Pat<(loongarch_vshuf4i v16i8:$vj, immZExt8:$ui8),
+          (VSHUF4I_B v16i8:$vj, immZExt8:$ui8)>;
+def : Pat<(loongarch_vshuf4i v8i16:$vj, immZExt8:$ui8),
+        (VSHUF4I_H v8i16:$vj, immZExt8:$ui8)>;
+def : Pat<(loongarch_vshuf4i v4i32:$vj, immZExt8:$ui8),
+        (VSHUF4I_W v4i32:$vj, immZExt8:$ui8)>;
+def : Pat<(loongarch_vshuf4i v4f32:$vj, immZExt8:$ui8),
+        (VSHUF4I_W v4f32:$vj, immZExt8:$ui8)>;
+
+// VREPLVEI_{B/H/W/D}
+def : Pat<(loongarch_vreplvei v16i8:$vj, immZExt4:$ui4),
+          (VREPLVEI_B v16i8:$vj, immZExt4:$ui4)>;
+def : Pat<(loongarch_vreplvei v8i16:$vj, immZExt3:$ui3),
+        (VREPLVEI_H v8i16:$vj, immZExt3:$ui3)>;
+def : Pat<(loongarch_vreplvei v4i32:$vj, immZExt2:$ui2),
+        (VREPLVEI_W v4i32:$vj, immZExt2:$ui2)>;
+def : Pat<(loongarch_vreplvei v2i64:$vj, immZExt1:$ui1),
+        (VREPLVEI_D v2i64:$vj, immZExt1:$ui1)>;
+def : Pat<(loongarch_vreplvei v4f32:$vj, immZExt2:$ui2),
+        (VREPLVEI_W v4f32:$vj, immZExt2:$ui2)>;
+def : Pat<(loongarch_vreplvei v2f64:$vj, immZExt1:$ui1),
+        (VREPLVEI_D v2f64:$vj, immZExt1:$ui1)>;
+
 // VREPLVEI_{W/D}
 def : Pat<(lsxsplatf32 FPR32:$fj),
           (VREPLVEI_W (SUBREG_TO_REG (i64 0), FPR32:$fj, sub_32), 0)>;