[AArch64][SVE] Add SVE2 intrinsics for complex integer dot product

[kmclaughlin]

Summary

Implements the following intrinsics:

• @llvm.aarch64.sve.cdot
• @llvm.aarch64.sve.cdot.lane

[efriedma]

llvm/include/llvm/IR/IntrinsicsAArch64.td

1111

Missing ImmArg?

1121

Missing ImmArg<4>?

[kmclaughlin]

• Removed the AdvSIMD_SVE_CDOT_Intrinsic class
• Added ImmArg<4> to AdvSIMD_SVE_CDOT_LANE_Intrinsic

[kmclaughlin]

Thanks for reviewing this, @efriedma!

llvm/include/llvm/IR/IntrinsicsAArch64.td

1111

Replaced this with AdvSIMD_SVE_DOT_Indexed_Intrinsic, which has the ImmArg property but is otherwise identical

[efriedma]

LGTM

llvm/include/llvm/IR/IntrinsicsAArch64.td

let TargetPrefix = "aarch64" in {  // All intrinsics start with "llvm.aarch64.".
   class SVE2_3VectorArg_Indexed_Intrinsic
     : Intrinsic<[llvm_anyvector_ty],
                 [LLVMMatchType<0>,
                  LLVMSubdivide2VectorType<0>,
                  LLVMSubdivide2VectorType<0>,
                  llvm_i64_ty],
                 [IntrNoMem, ImmArg<3>]>;
 
+  class AdvSIMD_SVE_CDOT_Intrinsic
+    : Intrinsic<[llvm_anyvector_ty],
+                [LLVMMatchType<0>,
+                 LLVMSubdivide4VectorType<0>,
+                 LLVMSubdivide4VectorType<0>,
+                 llvm_i32_ty],
+                [IntrNoMem]>;
+
+  class AdvSIMD_SVE_CDOT_LANE_Intrinsic
+    : Intrinsic<[llvm_anyvector_ty],
+                [LLVMMatchType<0>,
+                 LLVMSubdivide4VectorType<0>,
+                 LLVMSubdivide4VectorType<0>,
+                 llvm_i32_ty,
+                 llvm_i32_ty],
+                [IntrNoMem, ImmArg<3>]>;
+
   // NOTE: There is no relationship between these intrinsics beyond an attempt
   // to reuse currently identical class definitions.
   class AdvSIMD_SVE_LOGB_Intrinsic  : AdvSIMD_SVE_CNT_Intrinsic;
 
   // This class of intrinsics are not intended to be useful within LLVM IR but
   // are instead here to support some of the more regid parts of the ACLE.
   class Builtin_SVCVT<string name, LLVMType OUT, LLVMType IN>
   : GCCBuiltin<"__builtin_sve_" # name>,

 def int_aarch64_sve_uhadd         : AdvSIMD_Pred2VectorArg_Intrinsic;
 def int_aarch64_sve_uhsub         : AdvSIMD_Pred2VectorArg_Intrinsic;
 def int_aarch64_sve_uhsubr        : AdvSIMD_Pred2VectorArg_Intrinsic;
 def int_aarch64_sve_urecpe        : AdvSIMD_Merged1VectorArg_Intrinsic;
 def int_aarch64_sve_urhadd        : AdvSIMD_Pred2VectorArg_Intrinsic;
 def int_aarch64_sve_ursqrte       : AdvSIMD_Merged1VectorArg_Intrinsic;
 
 //
+// SVE2 - Widening complex integer dot product
+//
+
+def int_aarch64_sve_cdot      : AdvSIMD_SVE_CDOT_Intrinsic;
+def int_aarch64_sve_cdot_lane : AdvSIMD_SVE_CDOT_LANE_Intrinsic;
+
+//
 // SVE2 - Non-widening pairwise arithmetic
 //
 
 def int_aarch64_sve_addp    : AdvSIMD_Pred2VectorArg_Intrinsic;
 def int_aarch64_sve_faddp   : AdvSIMD_Pred2VectorArg_Intrinsic;
 def int_aarch64_sve_fmaxp   : AdvSIMD_Pred2VectorArg_Intrinsic;
 def int_aarch64_sve_fmaxnmp : AdvSIMD_Pred2VectorArg_Intrinsic;
 def int_aarch64_sve_fminp   : AdvSIMD_Pred2VectorArg_Intrinsic;

llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td

def : Pat<(nxv16i8 (int_aarch64_sve_umulh (nxv16i1 (AArch64ptrue 31)), nxv16i8:$Op1, nxv16i8:$Op2)),
             (UMULH_ZZZ_B $Op1, $Op2)>;
   def : Pat<(nxv8i16 (int_aarch64_sve_umulh (nxv8i1 (AArch64ptrue 31)), nxv8i16:$Op1, nxv8i16:$Op2)),
             (UMULH_ZZZ_H $Op1, $Op2)>;
   def : Pat<(nxv4i32 (int_aarch64_sve_umulh (nxv4i1 (AArch64ptrue 31)), nxv4i32:$Op1, nxv4i32:$Op2)),
             (UMULH_ZZZ_S $Op1, $Op2)>;
   def : Pat<(nxv2i64 (int_aarch64_sve_umulh (nxv2i1 (AArch64ptrue 31)), nxv2i64:$Op1, nxv2i64:$Op2)),
             (UMULH_ZZZ_D $Op1, $Op2)>;
   // SVE2 complex integer dot product (indexed)
-  defm CDOT_ZZZI : sve2_cintx_dot_by_indexed_elem<"cdot">;
+  defm CDOT_ZZZI : sve2_cintx_dot_by_indexed_elem<"cdot", int_aarch64_sve_cdot_lane>;
 
   // SVE2 complex integer dot product
-  defm CDOT_ZZZ : sve2_cintx_dot<"cdot">;
+  defm CDOT_ZZZ : sve2_cintx_dot<"cdot", int_aarch64_sve_cdot>;
 
   // SVE2 complex integer multiply-add (indexed)
   defm CMLA_ZZZI      : sve2_cmla_by_indexed_elem<0b0, "cmla">;
   // SVE2 complex saturating multiply-add (indexed)
   defm SQRDCMLAH_ZZZI : sve2_cmla_by_indexed_elem<0b1, "sqrdcmlah">;
 
   // SVE2 complex integer multiply-add
   defm CMLA_ZZZ      : sve2_int_cmla<0b0, "cmla">;

llvm/lib/Target/AArch64/SVEInstrFormats.td

: I<(outs zprty1:$Zda), (ins zprty1:$_Zda, zprty2:$Zn, zprty2:$Zm,
   let Inst{9-5}   = Zn;
   let Inst{4-0}   = Zda;
 
   let Constraints = "$Zda = $_Zda";
   let DestructiveInstType = Destructive;
   let ElementSize = ElementSizeNone;
 }
 
-multiclass sve2_cintx_dot<string asm> {
+multiclass sve2_cintx_dot<string asm, SDPatternOperator op> {
   def _S : sve2_complex_int_arith<0b10, 0b0001, asm, ZPR32, ZPR8>;
   def _D : sve2_complex_int_arith<0b11, 0b0001, asm, ZPR64, ZPR16>;
+
+  def : Pat<(nxv4i32 (op (nxv4i32 ZPR32:$Op1), (nxv16i8 ZPR8:$Op2), (nxv16i8 ZPR8:$Op3),
+                         (i32 complexrotateop:$imm))),
+            (!cast<Instruction>(NAME # "_S") ZPR32:$Op1, ZPR8:$Op2, ZPR8:$Op3, complexrotateop:$imm)>;
+  def : Pat<(nxv2i64 (op (nxv2i64 ZPR64:$Op1), (nxv8i16 ZPR16:$Op2), (nxv8i16 ZPR16:$Op3),
+                         (i32 complexrotateop:$imm))),
+            (!cast<Instruction>(NAME # "_D") ZPR64:$Op1, ZPR16:$Op2, ZPR16:$Op3, complexrotateop:$imm)>;
 }
 
 //===----------------------------------------------------------------------===//
 // SVE2 Complex Multiply-Add Group
 //===----------------------------------------------------------------------===//
 
 multiclass sve2_int_cmla<bit opc, string asm> {
   def _B : sve2_complex_int_arith<0b00, { 0b001, opc }, asm, ZPR8, ZPR8>;
: I<(outs zprty1:$Zda), (ins zprty1:$_Zda, zprty2:$Zn, zprty3:$Zm, itype:$iop,
   let Inst{9-5}   = Zn;
   let Inst{4-0}   = Zda;
 
   let Constraints = "$Zda = $_Zda";
   let DestructiveInstType = Destructive;
   let ElementSize = ElementSizeNone;
 }
 
-multiclass sve2_cintx_dot_by_indexed_elem<string asm> {
-  def _S : sve2_complex_int_arith_indexed<0b10, 0b0100, asm, ZPR32, ZPR8, ZPR3b8, VectorIndexS> {
+multiclass sve2_cintx_dot_by_indexed_elem<string asm, SDPatternOperator op> {
+  def _S : sve2_complex_int_arith_indexed<0b10, 0b0100, asm, ZPR32, ZPR8, ZPR3b8, VectorIndexS32b> {
     bits<2> iop;
     bits<3> Zm;
     let Inst{20-19} = iop;
     let Inst{18-16} = Zm;
   }
-  def _D : sve2_complex_int_arith_indexed<0b11, 0b0100, asm, ZPR64, ZPR16, ZPR4b16, VectorIndexD> {
+  def _D : sve2_complex_int_arith_indexed<0b11, 0b0100, asm, ZPR64, ZPR16, ZPR4b16, VectorIndexD32b> {
     bit iop;
     bits<4> Zm;
     let Inst{20} = iop;
     let Inst{19-16} = Zm;
   }
+
+  def : Pat<(nxv4i32 (op (nxv4i32 ZPR32:$Op1), (nxv16i8 ZPR8:$Op2), (nxv16i8 ZPR8:$Op3),
+                         (i32 VectorIndexS32b_timm:$idx), (i32 complexrotateop:$imm))),
+            (!cast<Instruction>(NAME # "_S") ZPR32:$Op1, ZPR8:$Op2, ZPR8:$Op3, VectorIndexS32b_timm:$idx, complexrotateop:$imm)>;
+  def : Pat<(nxv2i64 (op (nxv2i64 ZPR64:$Op1), (nxv8i16 ZPR16:$Op2), (nxv8i16 ZPR16:$Op3),
+                         (i32 VectorIndexD32b_timm:$idx), (i32 complexrotateop:$imm))),
+            (!cast<Instruction>(NAME # "_D") ZPR64:$Op1, ZPR16:$Op2, ZPR16:$Op3, VectorIndexD32b_timm:$idx, complexrotateop:$imm)>;
 }
 
 //===----------------------------------------------------------------------===//
 // SVE2 Complex Multiply-Add - Indexed Group
 //===----------------------------------------------------------------------===//
 
 multiclass sve2_cmla_by_indexed_elem<bit opc, string asm> {
   def _H : sve2_complex_int_arith_indexed<0b10, { 0b011, opc }, asm, ZPR16, ZPR16, ZPR3b16, VectorIndexS> {

llvm/test/CodeGen/AArch64/sve2-intrinsics-complex-dot.ll

+; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sve2 < %s | FileCheck %s
+
+
+;
+; CDOT
+;
+
+define <vscale x 4 x i32> @cdot_s(<vscale x 4 x i32> %a, <vscale x 16 x i8> %b, <vscale x 16 x i8> %c) {
+; CHECK-LABEL: cdot_s:
+; CHECK: cdot z0.s, z1.b, z2.b, #0
+; CHECK-NEXT: ret
+  %out = call <vscale x 4 x i32> @llvm.aarch64.sve.cdot.nxv4i32(<vscale x 4 x i32> %a,
+                                                                <vscale x 16 x i8> %b,
+                                                                <vscale x 16 x i8> %c,
+                                                                i32 0)
+  ret <vscale x 4 x i32> %out
+}
+
+define <vscale x 2 x i64> @cdot_d(<vscale x 2 x i64> %a, <vscale x 8 x i16> %b, <vscale x 8 x i16> %c) {
+; CHECK-LABEL: cdot_d:
+; CHECK: cdot z0.d, z1.h, z2.h, #90
+; CHECK-NEXT: ret
+  %out = call <vscale x 2 x i64> @llvm.aarch64.sve.cdot.nxv2i64(<vscale x 2 x i64> %a,
+                                                                <vscale x 8 x i16> %b,
+                                                                <vscale x 8 x i16> %c,
+                                                                i32 90)
+  ret <vscale x 2 x i64> %out
+}
+
+;
+; CDOT(indexed)
+;
+
+define <vscale x 4 x i32> @cdot_s_idx(<vscale x 4 x i32> %a, <vscale x 16 x i8> %b, <vscale x 16 x i8> %c) {
+; CHECK-LABEL: cdot_s_idx:
+; CHECK: cdot z0.s, z1.b, z2.b[0], #180
+; CHECK-NEXT: ret
+  %out = call <vscale x 4 x i32> @llvm.aarch64.sve.cdot.lane.nxv4i32(<vscale x 4 x i32> %a,
+                                                                     <vscale x 16 x i8> %b,
+                                                                     <vscale x 16 x i8> %c,
+                                                                     i32 0, i32 180)
+  ret <vscale x 4 x i32> %out
+}
+
+
+define <vscale x 2 x i64> @cdot_d_idx(<vscale x 2 x i64> %a, <vscale x 8 x i16> %b, <vscale x 8 x i16> %c) {
+; CHECK-LABEL: cdot_d_idx:
+; CHECK: cdot z0.d, z1.h, z2.h[1], #270
+; CHECK-NEXT: ret
+  %out = call <vscale x 2 x i64> @llvm.aarch64.sve.cdot.lane.nxv2i64(<vscale x 2 x i64> %a,
+                                                                     <vscale x 8 x i16> %b,
+                                                                     <vscale x 8 x i16> %c,
+                                                                     i32 1, i32 270)
+  ret <vscale x 2 x i64> %out
+}
+
+
+declare <vscale x 4 x i32> @llvm.aarch64.sve.cdot.nxv4i32(<vscale x 4 x i32>, <vscale x 16 x i8>, <vscale x 16 x i8>, i32)
+declare <vscale x 2 x i64> @llvm.aarch64.sve.cdot.nxv2i64(<vscale x 2 x i64>, <vscale x 8 x i16>, <vscale x 8 x i16>, i32)
+declare <vscale x 4 x i32> @llvm.aarch64.sve.cdot.lane.nxv4i32(<vscale x 4 x i32>, <vscale x 16 x i8>, <vscale x 16 x i8>, i32, i32)
+declare <vscale x 2 x i64> @llvm.aarch64.sve.cdot.lane.nxv2i64(<vscale x 2 x i64>, <vscale x 8 x i16>, <vscale x 8 x i16>, i32, i32)