[SME2][AArch64] Add multi-single multiply-add long long intrinsics

[kmclaughlin]

Summary

Adds intrinsics for the following SME2 instructions:

• smlall (1, 2 & 4 vectors)
• umlall (1, 2 & 4 vectors)
• smlsll (1, 2 & 4 vectors)
• umlsll (1, 2 & 4 vectors)
• sumlall (2 & 4 vectors)
• usmlall (1, 2 & 4 vectors)

NOTE: These intrinsics are still in development and are subject to future changes.

[david-arm]

llvm/lib/Target/AArch64/SMEInstrFormats.td

2757

This is just a suggestion, but in theory you could reduce the amount of duplication by creating a multiclass sme2_mla_ll_array_vg24_single like this:

multiclass sme2_mla_ll_array_vg24_single<string mnemonic, bits<5> op,
                                        MatrixOperand matrix_ty,
                                        RegisterOperand multi_vector_ty,
                                        ZPRRegOp zpr_ty, ValueType vt, SDPatternOperator intrinsic> {
  def NAME: sme2_mla_ll_array_vg24_single<op, matrix_ty, multi_vector_ty,
                                        zpr_ty, mnemonic>, SMEPseudo2Instr<NAME, 1>;

  def NAME # _PSEUDO : sme2_za_array_2op_multi_single_pseudo<NAME, uimm1s4range, multi_vector_ty, zpr_ty, SMEMatrixArray>;

  def : InstAlias<mnemonic # "\t$ZAd[$Rv, $imm], $Zn, $Zm",
                 (!cast<Instruction>(NAME) matrix_ty:$ZAd,  MatrixIndexGPR32Op8_11:$Rv, uimm1s4range:$imm, multi_vector_ty:$Zn, zpr_ty:$Zm), 0>;
}

Then for each of vg2 and vg4 you just need:

multiclass sme2_mla_ll_array_vg2_single<string mnemonic, bits<5> op,
                                        MatrixOperand matrix_ty,
                                        RegisterOperand multi_vector_ty,
                                        ZPRRegOp zpr_ty, ValueType vt, SDPatternOperator intrinsic> {
  defm : sme2_mla_ll_array_vg24_single;

  def : SME2_ZA_TwoOp_VG2_Multi_Single_Pat<NAME, intrinsic, uimm1s4range, zpr_ty, vt, tileslicerange1s4>;
}

Please feel free to ignore this suggestion if you think it doesn't improve things. I wouldn't hold up the patch for it!

llvm/test/CodeGen/AArch64/sme2-intrinsics-mlall.ll

3

nit: I don't think we need the -mattr=+sve flag here, since sme2 implies it.

[kmclaughlin]

• Added a sme2_mla_ll_array_vg24_single multiclass to reduce duplication in sme2_mla_ll_array_vg2_single & sme2_mla_ll_array_vg4_single
• Removed -mattr=+sve from sme2-intrinsics-mlall.ll

[david-arm]

LGTM! Muy bueno!