[mips][msa] Pattern match the splat.d instruction

[smaksimovic]

Summary

Introduced a new pattern for matching splat.d explicitly.

Both splat.d and splati.d can now be generated from the @llvm.mips.splat.d
intrinsic depending on whether an immediate value has been passed.

[sdardis]

LGTM. Some minor nits inlined.

lib/Target/Mips/MipsMSAInstrInfo.td

195–202

Nit on the formatting here:

def vsplati64_splat_d : PatFrag<(ops node:$e0),
                                (v2i64 (bitconvert
                                         (v4i32 (and
                                           (v4i32 (build_vector node:$e0,
                                                                node:$e0,
                                                                node:$e0,
                                                                node:$e0)),
                                           vsplati64_imm_eq_1))))>;

lib/Target/Mips/MipsSEISelLowering.cpp

1346

In case of the index being passed as an immediate value....

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

107

This line can be dropped.

123

This line can be dropped.

lib/Target/Mips/MipsMSAInstrInfo.td

 def vsplati16 : PatFrag<(ops node:$e0),
                         (v8i16 (build_vector node:$e0, node:$e0,
                                              node:$e0, node:$e0,
                                              node:$e0, node:$e0,
                                              node:$e0, node:$e0))>;
 def vsplati32 : PatFrag<(ops node:$e0),
                         (v4i32 (build_vector node:$e0, node:$e0,
                                              node:$e0, node:$e0))>;
+
+def vsplati64_imm_eq_1 : PatLeaf<(bitconvert (v4i32 (build_vector))), [{
+  APInt Imm;
+  SDNode *BV = N->getOperand(0).getNode();
+  EVT EltTy = N->getValueType(0).getVectorElementType();
+
+  return selectVSplat(BV, Imm, EltTy.getSizeInBits()) &&
+         Imm.getBitWidth() == EltTy.getSizeInBits() && Imm == 1;
+}]>;
+
 def vsplati64 : PatFrag<(ops node:$e0),
                         (v2i64 (build_vector node:$e0, node:$e0))>;
+
+def vsplati64_splat_d : PatFrag<(ops node:$e0),
+                        (v2i64 (bitconvert
+                                 (v4i32 (and
+                                   (v4i32 (build_vector node:$e0, node:$e0,
+                                                       node:$e0, node:$e0)),
+                                    vsplati64_imm_eq_1))))>;
+
 def vsplatf32 : PatFrag<(ops node:$e0),
                         (v4f32 (build_vector node:$e0, node:$e0,
                                              node:$e0, node:$e0))>;
 def vsplatf64 : PatFrag<(ops node:$e0),
                         (v2f64 (build_vector node:$e0, node:$e0))>;
 
 def vsplati8_elt  : PatFrag<(ops node:$v, node:$i),
                             (MipsVSHF (vsplati8 node:$i), node:$v, node:$v)>;
 def vsplati16_elt : PatFrag<(ops node:$v, node:$i),
                             (MipsVSHF (vsplati16 node:$i), node:$v, node:$v)>;
 def vsplati32_elt : PatFrag<(ops node:$v, node:$i),
                             (MipsVSHF (vsplati32 node:$i), node:$v, node:$v)>;
 def vsplati64_elt : PatFrag<(ops node:$v, node:$i),
-                            (MipsVSHF (vsplati64 node:$i), node:$v, node:$v)>;
+                            (MipsVSHF (vsplati64_splat_d node:$i),
+                                      node:$v, node:$v)>;
 
 class SplatPatLeaf<Operand opclass, dag frag, code pred = [{}],
                    SDNodeXForm xform = NOOP_SDNodeXForm>
   : PatLeaf<frag, pred, xform> {
   Operand OpClass = opclass;
 }
 
 class SplatComplexPattern<Operand opclass, ValueType ty, int numops, string fn,

 def vsplat_imm_eq_1 : PatLeaf<(build_vector), [{
   APInt Imm;
   EVT EltTy = N->getValueType(0).getVectorElementType();
 
   return selectVSplat(N, Imm, EltTy.getSizeInBits()) &&
          Imm.getBitWidth() == EltTy.getSizeInBits() && Imm == 1;
 }]>;
 
-def vsplati64_imm_eq_1 : PatLeaf<(bitconvert (v4i32 (build_vector))), [{
-  APInt Imm;
-  SDNode *BV = N->getOperand(0).getNode();
-  EVT EltTy = N->getValueType(0).getVectorElementType();
-
-  return selectVSplat(BV, Imm, EltTy.getSizeInBits()) &&
-         Imm.getBitWidth() == EltTy.getSizeInBits() && Imm == 1;
-}]>;
-
 def vbclr_b : PatFrag<(ops node:$ws, node:$wt),
                       (and node:$ws, (xor (shl vsplat_imm_eq_1, node:$wt),
                                           immAllOnesV))>;
 def vbclr_h : PatFrag<(ops node:$ws, node:$wt),
                       (and node:$ws, (xor (shl vsplat_imm_eq_1, node:$wt),
                                           immAllOnesV))>;
 def vbclr_w : PatFrag<(ops node:$ws, node:$wt),
                       (and node:$ws, (xor (shl vsplat_imm_eq_1, node:$wt),

lib/Target/Mips/MipsSEISelLowering.cpp

static SDValue lowerMSASplatZExt(SDValue Op, unsigned OpNr, SelectionDAG &DAG) {
 
   // 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 = Op->getOperand(OpNr);
   SDValue LaneB;
 
   if (ResVecTy == MVT::v2i64) {
+    // In case of index passed as an immediate value, set the upper lane to 0
+    // so that the splati.d instruction can be matched.
+    if (isa<ConstantSDNode>(LaneA))
       LaneB = DAG.getConstant(0, DL, MVT::i32);
+    // Having the index passed in a register, set the upper lane to the same
+    // value as the lower - this results in the BUILD_VECTOR node not being
+    // expanded through stack. This way we are able to pattern match the set of
+    // nodes created here to splat.d.
+    else
+      LaneB = LaneA;
     ViaVecTy = MVT::v4i32;
     if(BigEndian)
       std::swap(LaneA, LaneB);
   } else
     LaneB = LaneA;
 
   SDValue Ops[16] = { LaneA, LaneB, LaneA, LaneB, LaneA, LaneB, LaneA, LaneB,
                       LaneA, LaneB, LaneA, LaneB, LaneA, LaneB, LaneA, LaneB };

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

 ; Test the MSA splat intrinsics that are encoded with the 3R instruction
 ; format.
 
-; RUN: llc -march=mips -mattr=+msa,+fp64 -relocation-model=pic < %s | \
+; RUN: llc -march=mips -mcpu=mips32r5 -mattr=+msa,+fp64 -relocation-model=pic < %s | \
 ; RUN:     FileCheck -check-prefix=MIPS32 %s
-; RUN: llc -march=mipsel -mattr=+msa,+fp64 -relocation-model=pic < %s | \
+; RUN: llc -march=mipsel -mcpu=mips32r5 -mattr=+msa,+fp64 -relocation-model=pic < %s | \
 ; RUN:     FileCheck -check-prefix=MIPS32 %s
 
 @llvm_mips_splat_b_ARG1 = global <16 x i8> <i8 0, i8 1, i8 2, i8 3, i8 4, i8 5, i8 6, i8 7, i8 8, i8 9, i8 10, i8 11, i8 12, i8 13, i8 14, i8 15>, align 16
 @llvm_mips_splat_b_RES  = global <16 x i8> <i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0>, align 16
 
 define void @llvm_mips_splat_b_test(i32 %a) nounwind {
 entry:
   %0 = load <16 x i8>, <16 x i8>* @llvm_mips_splat_b_ARG1
entry:
   %1 = tail call <2 x i64> @llvm.mips.splat.d(<2 x i64> %0, i32 %a)
   store <2 x i64> %1, <2 x i64>* @llvm_mips_splat_d_RES
   ret void
 }
 
 declare <2 x i64> @llvm.mips.splat.d(<2 x i64>, i32) nounwind
 
 ; MIPS32: llvm_mips_splat_d_test:
-; FIXME: This test is currently disabled for MIPS32 because the indices are
-;        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-DAG: lw   [[R1:\$[0-9]+]], %got(llvm_mips_splat_d_ARG1)(
+; MIPS32-DAG: lw   [[R2:\$[0-9]+]], %got(llvm_mips_splat_d_RES)(
+; MIPS32-DAG: ld.d [[R3:\$w[0-9]+]], 0([[R1]])
+; MIPS32-DAG: splat.d [[R4:\$w[0-9]+]], [[R3]][$4]
+; MIPS32-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]], {{.*}}
+; MIPS32-LABEL: llvm_mips_splat_d_arg_test
+; MIPS32-DAG: lw      [[R0:\$[0-9]+]], %got(
+; MIPS32-DAG: addiu   [[R1:\$[0-9]+]], [[R0]], %lo(
+; MIPS32-DAG: lw      [[R2:\$[0-9]+]], %got(llvm_mips_splat_d_RES)(
+; MIPS32-DAG: ld.d    [[R3:\$w[0-9]+]], 0([[R1]])
+; MIPS32-DAG: splat.d [[R4:\$w[0-9]+]], [[R3]][$4]
+; MIPS32-DAG: st.d    [[R4]], 0([[R2]])
+; MIPS32-NOT: vshf.d
+; MIPS32: .size llvm_mips_splat_d_arg_test
 
 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
+; MIPS32-LABEL: llvm_mips_splat_d_imm_test
+; MIPS32-DAG: lw       [[R0:\$[0-9]+]], %got(
+; MIPS32-DAG: addiu    [[R1:\$[0-9]+]], [[R0]], %lo(
+; MIPS32-DAG: lw       [[R2:\$[0-9]+]], %got(llvm_mips_splat_d_RES)(
+; MIPS32-DAG: ld.d     [[R3:\$w[0-9]+]], 0([[R1]])
+; MIPS32-DAG: splati.d [[R4:\$w[0-9]+]], [[R3]][0]
+; MIPS32-DAG: st.d     [[R4]], 0([[R2]])
+; MIPS32-NOT: vshf.d
+; MIPS32: .size llvm_mips_splat_d_imm_test