[AArch64][SVE] Add SVE2 intrinsics for polynomial arithmetic

[kmclaughlin]

Summary

Implements the following intrinsics:

• @llvm.aarch64.sve.eorbt
• @llvm.aarch64.sve.eortb
• @llvm.aarch64.sve.pmullb.pair
• @llvm.aarch64.sve.pmullt.pair

[efriedma]

LGTM

llvm/include/llvm/IR/IntrinsicsAArch64.td

 def int_aarch64_sve_sqdmlslbt   : SVE2_3VectorArg_Long_Intrinsic;
 
 // SVE2 ADDSUB Long Unpredicated.
 def int_aarch64_sve_adclb       : AdvSIMD_3VectorArg_Intrinsic;
 def int_aarch64_sve_adclt       : AdvSIMD_3VectorArg_Intrinsic;
 def int_aarch64_sve_sbclb       : AdvSIMD_3VectorArg_Intrinsic;
 def int_aarch64_sve_sbclt       : AdvSIMD_3VectorArg_Intrinsic;
 
+//
+// SVE2 - Polynomial arithmetic
+//
+
+def int_aarch64_sve_eorbt       : AdvSIMD_3VectorArg_Intrinsic;
+def int_aarch64_sve_eortb       : AdvSIMD_3VectorArg_Intrinsic;
+def int_aarch64_sve_pmullb_pair : AdvSIMD_2VectorArg_Intrinsic;
+def int_aarch64_sve_pmullt_pair : AdvSIMD_2VectorArg_Intrinsic;
+
 }

llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td

let Predicates = [HasSVE2] in {
 
   // SVE2 integer multiply long
   defm SQDMULLB_ZZZ : sve2_wide_int_arith_long<0b11000, "sqdmullb", int_aarch64_sve_sqdmullb>;
   defm SQDMULLT_ZZZ : sve2_wide_int_arith_long<0b11001, "sqdmullt", int_aarch64_sve_sqdmullt>;
   defm SMULLB_ZZZ   : sve2_wide_int_arith_long<0b11100, "smullb",   int_aarch64_sve_smullb>;
   defm SMULLT_ZZZ   : sve2_wide_int_arith_long<0b11101, "smullt",   int_aarch64_sve_smullt>;
   defm UMULLB_ZZZ   : sve2_wide_int_arith_long<0b11110, "umullb",   int_aarch64_sve_umullb>;
   defm UMULLT_ZZZ   : sve2_wide_int_arith_long<0b11111, "umullt",   int_aarch64_sve_umullt>;
-  defm PMULLB_ZZZ   : sve2_pmul_long<0b0, "pmullb">;
-  defm PMULLT_ZZZ   : sve2_pmul_long<0b1, "pmullt">;
+  defm PMULLB_ZZZ   : sve2_pmul_long<0b0, "pmullb", int_aarch64_sve_pmullb_pair>;
+  defm PMULLT_ZZZ   : sve2_pmul_long<0b1, "pmullt", int_aarch64_sve_pmullt_pair>;
 
   // SVE2 bitwise shift and insert
   defm SRI_ZZI : sve2_int_bin_shift_imm_right<0b0, "sri", int_aarch64_sve_sri>;
   defm SLI_ZZI : sve2_int_bin_shift_imm_left< 0b1, "sli", int_aarch64_sve_sli>;
 
   // SVE2 bitwise shift right and accumulate
   defm SSRA_ZZI  : sve2_int_bin_accum_shift_imm_right<0b00, "ssra",  int_aarch64_sve_ssra>;
   defm USRA_ZZI  : sve2_int_bin_accum_shift_imm_right<0b01, "usra",  int_aarch64_sve_usra>;
let Predicates = [HasSVE2] in {
   defm UQXTNT_ZZ  : sve2_int_sat_extract_narrow_top<0b01, "uqxtnt",  int_aarch64_sve_uqxtnt>;
   defm SQXTUNT_ZZ : sve2_int_sat_extract_narrow_top<0b10, "sqxtunt", int_aarch64_sve_sqxtunt>;
 
   // SVE2 character match
   defm MATCH_PPzZZ  : sve2_char_match<0b0, "match",  int_aarch64_sve_match>;
   defm NMATCH_PPzZZ : sve2_char_match<0b1, "nmatch", int_aarch64_sve_nmatch>;
 
   // SVE2 bitwise exclusive-or interleaved
-  defm EORBT_ZZZ : sve2_bitwise_xor_interleaved<0b0, "eorbt">;
-  defm EORTB_ZZZ : sve2_bitwise_xor_interleaved<0b1, "eortb">;
+  defm EORBT_ZZZ : sve2_bitwise_xor_interleaved<0b0, "eorbt", int_aarch64_sve_eorbt>;
+  defm EORTB_ZZZ : sve2_bitwise_xor_interleaved<0b1, "eortb", int_aarch64_sve_eortb>;
 
   // SVE2 bitwise shift left long
   defm SSHLLB_ZZI : sve2_bitwise_shift_left_long<0b00, "sshllb", int_aarch64_sve_sshllb>;
   defm SSHLLT_ZZI : sve2_bitwise_shift_left_long<0b01, "sshllt", int_aarch64_sve_sshllt>;
   defm USHLLB_ZZI : sve2_bitwise_shift_left_long<0b10, "ushllb", int_aarch64_sve_ushllb>;
   defm USHLLT_ZZI : sve2_bitwise_shift_left_long<0b11, "ushllt", int_aarch64_sve_ushllt>;
 
   // SVE2 integer add/subtract interleaved long
let Predicates = [HasSVE2AES] in {
 
   // SVE2 crypto unary operations
   def AESMC_ZZ_B  : sve2_crypto_unary_op<0b0, "aesmc">;
   def AESIMC_ZZ_B : sve2_crypto_unary_op<0b1, "aesimc">;
 
   // PMULLB and PMULLT instructions which operate with 64-bit source and
   // 128-bit destination elements are enabled with crypto extensions, similar
   // to NEON PMULL2 instruction.
-  def PMULLB_ZZZ_Q : sve2_wide_int_arith<0b00, 0b11010, "pmullb",
-                                         ZPR128, ZPR64, ZPR64>;
-  def PMULLT_ZZZ_Q : sve2_wide_int_arith<0b00, 0b11011, "pmullt",
-                                         ZPR128, ZPR64, ZPR64>;
+  defm PMULLB_ZZZ_Q : sve2_wide_int_arith_pmul<0b00, 0b11010, "pmullb", int_aarch64_sve_pmullb_pair>;
+  defm PMULLT_ZZZ_Q : sve2_wide_int_arith_pmul<0b00, 0b11011, "pmullt", int_aarch64_sve_pmullt_pair>;
 }
 
 let Predicates = [HasSVE2SM4] in {
   // SVE2 crypto constructive binary operations
   def SM4EKEY_ZZZ_S : sve2_crypto_cons_bin_op<0b0, "sm4ekey", ZPR32>;
   // SVE2 crypto destructive binary operations
   def SM4E_ZZZ_S : sve2_crypto_des_bin_op<0b10, "sm4e", ZPR32>;
 }

llvm/lib/Target/AArch64/SVEInstrFormats.td

multiclass sve2_wide_int_arith_wide<bits<3> opc, string asm,
   def _S : sve2_wide_int_arith<0b10, { 0b10, opc }, asm, ZPR32, ZPR32, ZPR16>;
   def _D : sve2_wide_int_arith<0b11, { 0b10, opc }, asm, ZPR64, ZPR64, ZPR32>;
 
   def : SVE_2_Op_Pat<nxv8i16, op, nxv8i16, nxv16i8, !cast<Instruction>(NAME # _H)>;
   def : SVE_2_Op_Pat<nxv4i32, op, nxv4i32, nxv8i16, !cast<Instruction>(NAME # _S)>;
   def : SVE_2_Op_Pat<nxv2i64, op, nxv2i64, nxv4i32, !cast<Instruction>(NAME # _D)>;
 }
 
-multiclass sve2_pmul_long<bits<1> opc, string asm> {
+multiclass sve2_wide_int_arith_pmul<bits<2> sz, bits<5> opc, string asm,
+                                     SDPatternOperator op> {
+  def NAME : sve2_wide_int_arith<sz, opc, asm, ZPR128, ZPR64, ZPR64>;
+
+  // To avoid using 128 bit elements in the IR, the pattern below works with
+  // llvm intrinsics with the _pair suffix, to reflect that
+  // _Q is implemented as a pair of _D.
+  def : SVE_2_Op_Pat<nxv2i64, op, nxv2i64, nxv2i64, !cast<Instruction>(NAME)>;
+}
+
+multiclass sve2_pmul_long<bits<1> opc, string asm, SDPatternOperator op> {
   def _H : sve2_wide_int_arith<0b01, {0b1101, opc}, asm, ZPR16, ZPR8, ZPR8>;
   def _D : sve2_wide_int_arith<0b11, {0b1101, opc}, asm, ZPR64, ZPR32, ZPR32>;
+
+  // To avoid using 128 bit elements in the IR, the patterns below work with
+  // llvm intrinsics with the _pair suffix, to reflect that
+  // _H is implemented as a pair of _B and _D is implemented as a pair of _S.
+  def : SVE_2_Op_Pat<nxv16i8, op, nxv16i8, nxv16i8, !cast<Instruction>(NAME # _H)>;
+  def : SVE_2_Op_Pat<nxv4i32, op, nxv4i32, nxv4i32, !cast<Instruction>(NAME # _D)>;
 }
 
 //===----------------------------------------------------------------------===//
 // SVE2 Misc Group
 //===----------------------------------------------------------------------===//
 
 class sve2_misc<bits<2> sz, bits<4> opc, string asm,
                 ZPRRegOp zprty1, ZPRRegOp zprty2>
: I<(outs zprty1:$Zd), (ins zprty1:$_Zd, zprty2:$Zn, zprty2:$Zm),
   let Inst{9-5}   = Zn;
   let Inst{4-0}   = Zd;
 
   let Constraints = "$Zd = $_Zd";
   let DestructiveInstType = DestructiveOther;
   let ElementSize = ElementSizeNone;
 }
 
-multiclass sve2_bitwise_xor_interleaved<bit opc, string asm> {
+multiclass sve2_bitwise_xor_interleaved<bit opc, string asm,
+                                        SDPatternOperator op> {
   def _B : sve2_bitwise_xor_interleaved<0b00, opc, asm, ZPR8,  ZPR8>;
   def _H : sve2_bitwise_xor_interleaved<0b01, opc, asm, ZPR16, ZPR16>;
   def _S : sve2_bitwise_xor_interleaved<0b10, opc, asm, ZPR32, ZPR32>;
   def _D : sve2_bitwise_xor_interleaved<0b11, opc, asm, ZPR64, ZPR64>;
+
+  def : SVE_3_Op_Pat<nxv16i8, op, nxv16i8, nxv16i8, nxv16i8, !cast<Instruction>(NAME # _B)>;
+  def : SVE_3_Op_Pat<nxv8i16, op, nxv8i16, nxv8i16, nxv8i16, !cast<Instruction>(NAME # _H)>;
+  def : SVE_3_Op_Pat<nxv4i32, op, nxv4i32, nxv4i32, nxv4i32, !cast<Instruction>(NAME # _S)>;
+  def : SVE_3_Op_Pat<nxv2i64, op, nxv2i64, nxv2i64, nxv2i64, !cast<Instruction>(NAME # _D)>;
 }
 
 class sve2_bitwise_shift_left_long<bits<3> tsz8_64, bits<2> opc, string asm,
                                    ZPRRegOp zprty1, ZPRRegOp zprty2,
                                    Operand immtype>
 : I<(outs zprty1:$Zd), (ins zprty2:$Zn, immtype:$imm),
   asm, "\t$Zd, $Zn, $imm",
   "", []>, Sched<[]> {

llvm/test/CodeGen/AArch64/sve2-intrinsics-polynomial-arithmetic-128.ll

+; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sve2-aes -asm-verbose=0 < %s | FileCheck %s
+
+;
+; PMULLB
+;
+
+define <vscale x 2 x i64> @pmullb_i64(<vscale x 2 x i64> %a, <vscale x 2 x i64> %b) {
+; CHECK-LABEL: pmullb_i64:
+; CHECK: pmullb z0.q, z0.d, z1.d
+; CHECK-NEXT: ret
+  %out = call <vscale x 2 x i64> @llvm.aarch64.sve.pmullb.pair.nxv2i64(<vscale x 2 x i64> %a,
+                                                                       <vscale x 2 x i64> %b)
+  ret <vscale x 2 x i64> %out
+}
+
+;
+; PMULLT
+;
+
+define <vscale x 2 x i64> @pmullt_i64(<vscale x 2 x i64> %a, <vscale x 2 x i64> %b) {
+; CHECK-LABEL: pmullt_i64:
+; CHECK: pmullt z0.q, z0.d, z1.d
+; CHECK-NEXT: ret
+  %out = call <vscale x 2 x i64> @llvm.aarch64.sve.pmullt.pair.nxv2i64(<vscale x 2 x i64> %a,
+                                                                       <vscale x 2 x i64> %b)
+  ret <vscale x 2 x i64> %out
+}
+
+declare <vscale x 2 x i64> @llvm.aarch64.sve.pmullb.pair.nxv2i64(<vscale x 2 x i64>, <vscale x 2 x i64>)
+
+declare <vscale x 2 x i64> @llvm.aarch64.sve.pmullt.pair.nxv2i64(<vscale x 2 x i64>, <vscale x 2 x i64>)

llvm/test/CodeGen/AArch64/sve2-intrinsics-polynomial-arithmetic.ll

+; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sve2 -asm-verbose=0 < %s | FileCheck %s
+
+;
+; EORBT
+;
+
+define <vscale x 16 x i8> @eorbt_i8(<vscale x 16 x i8> %a, <vscale x 16 x i8> %b, <vscale x 16 x i8> %c) {
+; CHECK-LABEL: eorbt_i8:
+; CHECK: eorbt z0.b, z1.b, z2.b
+; CHECK-NEXT: ret
+  %out = call <vscale x 16 x i8> @llvm.aarch64.sve.eorbt.nxv16i8(<vscale x 16 x i8> %a,
+                                                                 <vscale x 16 x i8> %b,
+                                                                 <vscale x 16 x i8> %c)
+  ret <vscale x 16 x i8> %out
+}
+
+define <vscale x 8 x i16> @eorbt_i16(<vscale x 8 x i16> %a, <vscale x 8 x i16> %b, <vscale x 8 x i16> %c) {
+; CHECK-LABEL: eorbt_i16:
+; CHECK: eorbt z0.h, z1.h, z2.h
+; CHECK-NEXT: ret
+  %out = call <vscale x 8 x i16> @llvm.aarch64.sve.eorbt.nxv8i16(<vscale x 8 x i16> %a,
+                                                                 <vscale x 8 x i16> %b,
+                                                                 <vscale x 8 x i16> %c)
+  ret <vscale x 8 x i16> %out
+}
+
+define <vscale x 4 x i32> @eorbt_i32(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b, <vscale x 4 x i32> %c) {
+; CHECK-LABEL: eorbt_i32:
+; CHECK: eorbt z0.s, z1.s, z2.s
+; CHECK-NEXT: ret
+  %out = call <vscale x 4 x i32> @llvm.aarch64.sve.eorbt.nxv4i32(<vscale x 4 x i32> %a,
+                                                                 <vscale x 4 x i32> %b,
+                                                                 <vscale x 4 x i32> %c)
+  ret <vscale x 4 x i32> %out
+}
+
+define <vscale x 2 x i64> @eorbt_i64(<vscale x 2 x i64> %a, <vscale x 2 x i64> %b, <vscale x 2 x i64> %c) {
+; CHECK-LABEL: eorbt_i64:
+; CHECK: eorbt z0.d, z1.d, z2.d
+; CHECK-NEXT: ret
+  %out = call <vscale x 2 x i64> @llvm.aarch64.sve.eorbt.nxv2i64(<vscale x 2 x i64> %a,
+                                                                 <vscale x 2 x i64> %b,
+                                                                 <vscale x 2 x i64> %c)
+  ret <vscale x 2 x i64> %out
+}
+
+;
+; EORTB
+;
+
+define <vscale x 16 x i8> @eortb_i8(<vscale x 16 x i8> %a, <vscale x 16 x i8> %b, <vscale x 16 x i8> %c) {
+; CHECK-LABEL: eortb_i8:
+; CHECK: eortb z0.b, z1.b, z2.b
+; CHECK-NEXT: ret
+  %out = call <vscale x 16 x i8> @llvm.aarch64.sve.eortb.nxv16i8(<vscale x 16 x i8> %a,
+                                                                 <vscale x 16 x i8> %b,
+                                                                 <vscale x 16 x i8> %c)
+  ret <vscale x 16 x i8> %out
+}
+
+define <vscale x 8 x i16> @eortb_i16(<vscale x 8 x i16> %a, <vscale x 8 x i16> %b, <vscale x 8 x i16> %c) {
+; CHECK-LABEL: eortb_i16:
+; CHECK: eortb z0.h, z1.h, z2.h
+; CHECK-NEXT: ret
+  %out = call <vscale x 8 x i16> @llvm.aarch64.sve.eortb.nxv8i16(<vscale x 8 x i16> %a,
+                                                                 <vscale x 8 x i16> %b,
+                                                                 <vscale x 8 x i16> %c)
+  ret <vscale x 8 x i16> %out
+}
+
+define <vscale x 4 x i32> @eortb_i32(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b, <vscale x 4 x i32> %c) {
+; CHECK-LABEL: eortb_i32:
+; CHECK: eortb z0.s, z1.s, z2.s
+; CHECK-NEXT: ret
+  %out = call <vscale x 4 x i32> @llvm.aarch64.sve.eortb.nxv4i32(<vscale x 4 x i32> %a,
+                                                                 <vscale x 4 x i32> %b,
+                                                                 <vscale x 4 x i32> %c)
+  ret <vscale x 4 x i32> %out
+}
+
+define <vscale x 2 x i64> @eortb_i64(<vscale x 2 x i64> %a, <vscale x 2 x i64> %b, <vscale x 2 x i64> %c) {
+; CHECK-LABEL: eortb_i64:
+; CHECK: eortb z0.d, z1.d, z2.d
+; CHECK-NEXT: ret
+  %out = call <vscale x 2 x i64> @llvm.aarch64.sve.eortb.nxv2i64(<vscale x 2 x i64> %a,
+                                                                 <vscale x 2 x i64> %b,
+                                                                 <vscale x 2 x i64> %c)
+  ret <vscale x 2 x i64> %out
+}
+
+;
+; PMULLB
+;
+
+define <vscale x 16 x i8> @pmullb_i8(<vscale x 16 x i8> %a, <vscale x 16 x i8> %b) {
+; CHECK-LABEL: pmullb_i8:
+; CHECK: pmullb z0.h, z0.b, z1.b
+; CHECK-NEXT: ret
+  %out = call <vscale x 16 x i8> @llvm.aarch64.sve.pmullb.pair.nxv16i8(<vscale x 16 x i8> %a,
+                                                                       <vscale x 16 x i8> %b)
+  ret <vscale x 16 x i8> %out
+}
+
+define <vscale x 4 x i32> @pmullb_i32(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b) {
+; CHECK-LABEL: pmullb_i32:
+; CHECK: pmullb z0.d, z0.s, z1.s
+; CHECK-NEXT: ret
+  %out = call <vscale x 4 x i32> @llvm.aarch64.sve.pmullb.pair.nxv4i32(<vscale x 4 x i32> %a,
+                                                                       <vscale x 4 x i32> %b)
+  ret <vscale x 4 x i32> %out
+}
+
+;
+; PMULLT
+;
+
+define <vscale x 16 x i8> @pmullt_i8(<vscale x 16 x i8> %a, <vscale x 16 x i8> %b) {
+; CHECK-LABEL: pmullt_i8:
+; CHECK: pmullt z0.h, z0.b, z1.b
+; CHECK-NEXT: ret
+  %out = call <vscale x 16 x i8> @llvm.aarch64.sve.pmullt.pair.nxv16i8(<vscale x 16 x i8> %a,
+                                                                       <vscale x 16 x i8> %b)
+  ret <vscale x 16 x i8> %out
+}
+
+define <vscale x 4 x i32> @pmullt_i32(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b) {
+; CHECK-LABEL: pmullt_i32:
+; CHECK: pmullt z0.d, z0.s, z1.s
+; CHECK-NEXT: ret
+  %out = call <vscale x 4 x i32> @llvm.aarch64.sve.pmullt.pair.nxv4i32(<vscale x 4 x i32> %a,
+                                                                       <vscale x 4 x i32> %b)
+  ret <vscale x 4 x i32> %out
+}
+
+declare <vscale x 16 x i8> @llvm.aarch64.sve.eorbt.nxv16i8(<vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>)
+declare <vscale x 8 x i16> @llvm.aarch64.sve.eorbt.nxv8i16(<vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>)
+declare <vscale x 4 x i32> @llvm.aarch64.sve.eorbt.nxv4i32(<vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>)
+declare <vscale x 2 x i64> @llvm.aarch64.sve.eorbt.nxv2i64(<vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>)
+
+declare <vscale x 16 x i8> @llvm.aarch64.sve.eortb.nxv16i8(<vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>)
+declare <vscale x 8 x i16> @llvm.aarch64.sve.eortb.nxv8i16(<vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>)
+declare <vscale x 4 x i32> @llvm.aarch64.sve.eortb.nxv4i32(<vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>)
+declare <vscale x 2 x i64> @llvm.aarch64.sve.eortb.nxv2i64(<vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>)
+
+declare <vscale x 16 x i8> @llvm.aarch64.sve.pmullb.pair.nxv16i8(<vscale x 16 x i8>, <vscale x 16 x i8>)
+declare <vscale x 4 x i32> @llvm.aarch64.sve.pmullb.pair.nxv4i32(<vscale x 4 x i32>, <vscale x 4 x i32>)
+
+declare <vscale x 16 x i8> @llvm.aarch64.sve.pmullt.pair.nxv16i8(<vscale x 16 x i8>, <vscale x 16 x i8>)
+declare <vscale x 4 x i32> @llvm.aarch64.sve.pmullt.pair.nxv4i32(<vscale x 4 x i32>, <vscale x 4 x i32>)