[mips][msa] Splat.d endianness check

[smaksimovic]

Summary

Before this change, it was always the first element of a vector that got splatted since the lower 6 bits of vshf.d $wd were always zero for little endian.
Additionally, masking has been performed for vshf via which splat.d is created.

Vshf has a property where if its first operand's elements have either bit 6 or 7 set, destination element is set to zero.
Initially masked with 63 to avoid this property, which would result in generation of and.v + vshf.d in all cases.
Masking with one results in generating a single splati.d instruction when possible.

[smaksimovic]

Ping.

[smaksimovic]

Ping.

[smaksimovic]

Ping.

[sdardis]

LGTM.

lib/Target/Mips/MipsSEISelLowering.cpp

1442 ↗

(On Diff #95703)

Replace the 2 there with OpNr.

llvm/trunk/lib/Target/Mips/MipsSEISelLowering.cpp

SDValue Result = DAG.getNode(Opc, DL, ResTy, Vec, Idx,
                                DAG.getValueType(EltTy));
 
   return Result;
 }
 
 static SDValue lowerMSASplatZExt(SDValue Op, unsigned OpNr, SelectionDAG &DAG) {
   EVT ResVecTy = Op->getValueType(0);
   EVT ViaVecTy = ResVecTy;
+  bool BigEndian = !DAG.getSubtarget().getTargetTriple().isLittleEndian();
   SDLoc DL(Op);
 
   // When ResVecTy == MVT::v2i64, LaneA is the upper 32 bits of the lane and
   // LaneB is the lower 32-bits. Otherwise LaneA and LaneB are alternating
   // lanes.
-  SDValue LaneA;
-  SDValue LaneB = Op->getOperand(2);
+  SDValue LaneA = Op->getOperand(OpNr);
+  SDValue LaneB;
 
   if (ResVecTy == MVT::v2i64) {
-    LaneA = DAG.getConstant(0, DL, MVT::i32);
+    LaneB = DAG.getConstant(0, DL, MVT::i32);
     ViaVecTy = MVT::v4i32;
+    if(BigEndian)
+      std::swap(LaneA, LaneB);
   } else
-    LaneA = LaneB;
+    LaneB = LaneA;
 
   SDValue Ops[16] = { LaneA, LaneB, LaneA, LaneB, LaneA, LaneB, LaneA, LaneB,
                       LaneA, LaneB, LaneA, LaneB, LaneA, LaneB, LaneA, LaneB };
 
   SDValue Result = DAG.getBuildVector(
       ViaVecTy, DL, makeArrayRef(Ops, ViaVecTy.getVectorNumElements()));
 
-  if (ViaVecTy != ResVecTy)
-    Result = DAG.getNode(ISD::BITCAST, DL, ResVecTy, Result);
+  if (ViaVecTy != ResVecTy) {
+    SDValue One = DAG.getConstant(1, DL, ViaVecTy);
+    Result = DAG.getNode(ISD::BITCAST, DL, ResVecTy,
+                         DAG.getNode(ISD::AND, DL, ViaVecTy, Result, One));
+  }
 
   return Result;
 }
 
 static SDValue lowerMSASplatImm(SDValue Op, unsigned ImmOp, SelectionDAG &DAG,
                                 bool IsSigned = false) {
   return DAG.getConstant(
       APInt(Op->getValueType(0).getScalarType().getSizeInBits(),

llvm/trunk/test/CodeGen/Mips/msa/3r_splat.ll

 ;        difficult to match. This is because 64-bit values cannot be stored in
 ;        GPR32.
 ; MIPS64-DAG: lw   [[R1:\$[0-9]+]], %got(llvm_mips_splat_d_ARG1)(
 ; MIPS64-DAG: lw   [[R2:\$[0-9]+]], %got(llvm_mips_splat_d_RES)(
 ; MIPS64-DAG: ld.d [[R3:\$w[0-9]+]], 0([[R1]])
 ; MIPS64-DAG: splat.d [[R4:\$w[0-9]+]], [[R3]][$4]
 ; MIPS64-DAG: st.d [[R4]], 0([[R2]])
 ; MIPS32: .size llvm_mips_splat_d_test
+
+define void @llvm_mips_splat_d_arg_test(i32 %arg) {
+entry:
+  %0 = tail call <2 x i64> @llvm.mips.splat.d(<2 x i64> <i64 12720328, i64 10580959>, i32 %arg)
+  store volatile <2 x i64> %0, <2 x i64>* @llvm_mips_splat_d_RES
+  ret void
+}
+; CHECK-LABEL: llvm_mips_splat_d_arg_test
+; CHECK: ldi.w [[R1:\$w[0-9]+]], 1 
+; CHECK: and.v [[R2:\$w[0-9]+]], {{\$w[0-9]+}}, [[R1]]
+; CHECK: vshf.d [[R2]], {{.*}}
+
+define void @llvm_mips_splat_d_imm_test() {
+entry:
+  %0 = tail call <2 x i64> @llvm.mips.splat.d(<2 x i64> <i64 12720328, i64 10580959>, i32 76)
+  store volatile<2 x i64> %0, <2 x i64>* @llvm_mips_splat_d_RES
+  ret void
+}
+; CHECK-LABEL: llvm_mips_splat_d_imm_test
+; CHECK: splati. d {{.*}}, {{.*}}[0]
+; CHECK-NOT: vshf.d