[AArch64][SVE] Add missing patterns for unpredicated subr intrinsics

[paulwalker-arm]

My first thought is if you lowered "unpredicated" int_aarch64_sve_subr much like we do for int_aarch64_sve_sub, would we then get the isel improvements for free?

[bsmith]

In D109369#2987058, @paulwalker-arm wrote:

My first thought is if you lowered "unpredicated" int_aarch64_sve_subr much like we do for int_aarch64_sve_sub, would we then get the isel improvements for free?

I did consider this, however we'd have to add a new SUBR node for it, and given it won't be used elsewhere, unlike lowering the add, it didn't seem worth it. Additionally we wouldn't quite get the isel improvements for free as unpredicted subr is expressed in terms of sub, not subr, hence we'd still have to add extra patterns, they just wouldn't need to care about predicates.

That said, now that I think about it again, perhaps we could lower int_aarch64_sve_subr to a normal sub with reversed operands, which then I think would give us the isel for free, unless this is what you meant?

[paulwalker-arm]

In D109369#2987120, @bsmith wrote:

In D109369#2987058, @paulwalker-arm wrote:

That said, now that I think about it again, perhaps we could lower int_aarch64_sve_subr to a normal sub with reversed operands, which then I think would give us the isel for free, unless this is what you meant?

Yes, this is exactly what I meant. I can see that we already lower int_aarch64_sve_sub to ISD::SUB and we have SUBR patterns for sub so the missing link looks like just lowering int_aarch64_sve_subr to ISD::SUB as well.

[bsmith]

• Lower subr intrinsics to a sub node with reversed operands rather than adding patterns for the intrinsic.

llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td

 let Predicates = [HasSVEorStreamingSVE] in {
   defm ORR_ZPmZ : sve_int_bin_pred_log<0b000, "orr", int_aarch64_sve_orr>;
   defm EOR_ZPmZ : sve_int_bin_pred_log<0b001, "eor", int_aarch64_sve_eor>;
   defm AND_ZPmZ : sve_int_bin_pred_log<0b010, "and", int_aarch64_sve_and>;
   defm BIC_ZPmZ : sve_int_bin_pred_log<0b011, "bic", int_aarch64_sve_bic>;
 
   defm ADD_ZI   : sve_int_arith_imm0<0b000, "add", add>;
   defm SUB_ZI   : sve_int_arith_imm0<0b001, "sub", sub>;
-  defm SUBR_ZI  : sve_int_arith_imm0_subr<0b011, "subr", sub>;
+  defm SUBR_ZI  : sve_int_arith_imm0_subr<0b011, "subr", sub, int_aarch64_sve_subr>;
   defm SQADD_ZI : sve_int_arith_imm0<0b100, "sqadd", saddsat>;
   defm UQADD_ZI : sve_int_arith_imm0<0b101, "uqadd", uaddsat>;
   defm SQSUB_ZI : sve_int_arith_imm0<0b110, "sqsub", ssubsat>;
   defm UQSUB_ZI : sve_int_arith_imm0<0b111, "uqsub", usubsat>;
 
   defm MAD_ZPmZZ : sve_int_mladdsub_vvv_pred<0b0, "mad", int_aarch64_sve_mad>;
   defm MSB_ZPmZZ : sve_int_mladdsub_vvv_pred<0b1, "msb", int_aarch64_sve_msb>;
   defm MLA_ZPmZZ : sve_int_mlas_vvv_pred<0b0, "mla", int_aarch64_sve_mla, add, AArch64mul_p_oneuse>;

llvm/lib/Target/AArch64/SVEInstrFormats.td

class SVE_1_Op_Imm_OptLsl_Reverse_Pat<ValueType vt, SDPatternOperator op, ZPRRegOp zprty,
   : Pat<(vt (op (vt (AArch64dup (it (cpx i32:$imm, i32:$shift)))), (vt zprty:$Op1))),
         (inst $Op1, i32:$imm, i32:$shift)>;
 
 class SVE_1_Op_Imm_OptLsl_Pat<ValueType vt, SDPatternOperator op, ZPRRegOp zprty,
                               ValueType it, ComplexPattern cpx, Instruction inst>
   : Pat<(vt (op (vt zprty:$Op1), (vt (AArch64dup (it (cpx i32:$imm, i32:$shift)))))),
         (inst $Op1, i32:$imm, i32:$shift)>;
 
+class SVE_1_Op_Imm_OptLsl_Pat_All_Active<ValueType vt, ValueType pt, SDPatternOperator op, ZPRRegOp zprty,
+                                         ValueType it, ComplexPattern cpx, Instruction inst>
+  : Pat<(vt (op (pt (SVEAllActive)), (vt zprty:$Op1), (vt (AArch64dup (it (cpx i32:$imm, i32:$shift)))))),
+        (inst $Op1, i32:$imm, i32:$shift)>;
+
 class SVE_1_Op_Imm_Arith_All_Active<ValueType vt, ValueType pt, SDPatternOperator op,
                                   ZPRRegOp zprty, ValueType it, ComplexPattern cpx, Instruction inst>
   : Pat<(vt (op (pt (SVEAllActive)), (vt zprty:$Op1), (vt (AArch64dup (it (cpx i32:$imm)))))),
         (inst $Op1, i32:$imm)>;
 
 class SVE_1_Op_Imm_Log_Pat<ValueType vt, SDPatternOperator op, ZPRRegOp zprty,
                            ValueType it, ComplexPattern cpx, Instruction inst>
   : Pat<(vt (op (vt zprty:$Op1), (vt (AArch64dup (it (cpx i64:$imm)))))),
multiclass sve_int_arith_imm0<bits<3> opc, string asm, SDPatternOperator op> {
   def _D : sve_int_arith_imm0<0b11, opc, asm, ZPR64, addsub_imm8_opt_lsl_i64>;
 
   def : SVE_1_Op_Imm_OptLsl_Pat<nxv16i8, op, ZPR8,  i32, SVEAddSubImm8Pat,  !cast<Instruction>(NAME # _B)>;
   def : SVE_1_Op_Imm_OptLsl_Pat<nxv8i16, op, ZPR16, i32, SVEAddSubImm16Pat, !cast<Instruction>(NAME # _H)>;
   def : SVE_1_Op_Imm_OptLsl_Pat<nxv4i32, op, ZPR32, i32, SVEAddSubImm32Pat, !cast<Instruction>(NAME # _S)>;
   def : SVE_1_Op_Imm_OptLsl_Pat<nxv2i64, op, ZPR64, i64, SVEAddSubImm64Pat, !cast<Instruction>(NAME # _D)>;
 }
 
-multiclass sve_int_arith_imm0_subr<bits<3> opc, string asm, SDPatternOperator op> {
+multiclass sve_int_arith_imm0_subr<bits<3> opc, string asm, SDPatternOperator op, SDPatternOperator int_op> {
   def _B : sve_int_arith_imm0<0b00, opc, asm, ZPR8,  addsub_imm8_opt_lsl_i8>;
   def _H : sve_int_arith_imm0<0b01, opc, asm, ZPR16, addsub_imm8_opt_lsl_i16>;
   def _S : sve_int_arith_imm0<0b10, opc, asm, ZPR32, addsub_imm8_opt_lsl_i32>;
   def _D : sve_int_arith_imm0<0b11, opc, asm, ZPR64, addsub_imm8_opt_lsl_i64>;
 
   def : SVE_1_Op_Imm_OptLsl_Reverse_Pat<nxv16i8, op, ZPR8,  i32, SVEAddSubImm8Pat,  !cast<Instruction>(NAME # _B)>;
   def : SVE_1_Op_Imm_OptLsl_Reverse_Pat<nxv8i16, op, ZPR16, i32, SVEAddSubImm16Pat, !cast<Instruction>(NAME # _H)>;
   def : SVE_1_Op_Imm_OptLsl_Reverse_Pat<nxv4i32, op, ZPR32, i32, SVEAddSubImm32Pat, !cast<Instruction>(NAME # _S)>;
   def : SVE_1_Op_Imm_OptLsl_Reverse_Pat<nxv2i64, op, ZPR64, i64, SVEAddSubImm64Pat, !cast<Instruction>(NAME # _D)>;
+
+  def : SVE_1_Op_Imm_OptLsl_Pat_All_Active<nxv16i8, nxv16i1, int_op, ZPR8,  i32, SVEAddSubImm8Pat,  !cast<Instruction>(NAME # _B)>;
+  def : SVE_1_Op_Imm_OptLsl_Pat_All_Active<nxv8i16, nxv8i1,  int_op, ZPR16, i32, SVEAddSubImm16Pat, !cast<Instruction>(NAME # _H)>;
+  def : SVE_1_Op_Imm_OptLsl_Pat_All_Active<nxv4i32, nxv4i1,  int_op, ZPR32, i32, SVEAddSubImm32Pat, !cast<Instruction>(NAME # _S)>;
+  def : SVE_1_Op_Imm_OptLsl_Pat_All_Active<nxv2i64, nxv2i1,  int_op, ZPR64, i64, SVEAddSubImm64Pat, !cast<Instruction>(NAME # _D)>;
 }
 
 class sve_int_arith_imm<bits<2> sz8_64, bits<6> opc, string asm,
                         ZPRRegOp zprty, Operand immtype>
 : I<(outs zprty:$Zdn), (ins zprty:$_Zdn, immtype:$imm),
   asm, "\t$Zdn, $_Zdn, $imm",
   "",
   []>, Sched<[]> {

llvm/test/CodeGen/AArch64/sve-intrinsics-int-arith-imm.ll

; CHECK-NEXT: ret
   %pg.s = tail call <vscale x 4 x i1> @llvm.aarch64.sve.convert.from.svbool.nxv4i1(<vscale x 16 x i1> %pg.b)
   %b = tail call <vscale x 4 x i32> @llvm.aarch64.sve.dup.x.nxv4i32(i32 1)
   %out = tail call <vscale x 4 x i32> @llvm.aarch64.sve.sub.nxv4i32(<vscale x 4 x i1> %pg.s,
                                                                     <vscale x 4 x i32> %a,
                                                                     <vscale x 4 x i32> %b)
   ret <vscale x 4 x i32> %out
 }
 
+; SUBR
+
+define <vscale x 16 x i8> @subr_i8(<vscale x 16 x i8> %a) {
+; CHECK-LABEL: subr_i8:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    subr z0.b, z0.b, #127
+; CHECK-NEXT:    ret
+  %pg = call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 31)
+  %elt = insertelement <vscale x 16 x i8> undef, i8 127, i32 0
+  %splat = shufflevector <vscale x 16 x i8> %elt, <vscale x 16 x i8> undef, <vscale x 16 x i32> zeroinitializer
+  %out = call <vscale x 16 x i8> @llvm.aarch64.sve.subr.nxv16i8(<vscale x 16 x i1> %pg,
+                                                                <vscale x 16 x i8> %a,
+                                                                <vscale x 16 x i8> %splat)
+  ret <vscale x 16 x i8> %out
+}
+
+define <vscale x 8 x i16> @subr_i16(<vscale x 8 x i16> %a) {
+; CHECK-LABEL: subr_i16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    subr z0.h, z0.h, #127
+; CHECK-NEXT:    ret
+  %pg = call <vscale x 8 x i1> @llvm.aarch64.sve.ptrue.nxv8i1(i32 31)
+  %elt = insertelement <vscale x 8 x i16> undef, i16 127, i32 0
+  %splat = shufflevector <vscale x 8 x i16> %elt, <vscale x 8 x i16> undef, <vscale x 8 x i32> zeroinitializer
+  %out = call <vscale x 8 x i16> @llvm.aarch64.sve.subr.nxv8i16(<vscale x 8 x i1> %pg,
+                                                                <vscale x 8 x i16> %a,
+                                                                <vscale x 8 x i16> %splat)
+  ret <vscale x 8 x i16> %out
+}
+
+define <vscale x 8 x i16> @subr_i16_out_of_range(<vscale x 8 x i16> %a) {
+; CHECK-LABEL: subr_i16_out_of_range:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov w8, #257
+; CHECK-NEXT:    ptrue p0.h
+; CHECK-NEXT:    mov z1.h, w8
+; CHECK-NEXT:    subr z0.h, p0/m, z0.h, z1.h
+; CHECK-NEXT:    ret
+  %pg = call <vscale x 8 x i1> @llvm.aarch64.sve.ptrue.nxv8i1(i32 31)
+  %elt = insertelement <vscale x 8 x i16> undef, i16 257, i32 0
+  %splat = shufflevector <vscale x 8 x i16> %elt, <vscale x 8 x i16> undef, <vscale x 8 x i32> zeroinitializer
+  %out = call <vscale x 8 x i16> @llvm.aarch64.sve.subr.nxv8i16(<vscale x 8 x i1> %pg,
+                                                                <vscale x 8 x i16> %a,
+                                                                <vscale x 8 x i16> %splat)
+  ret <vscale x 8 x i16> %out
+}
+
+define <vscale x 4 x i32> @subr_i32(<vscale x 4 x i32> %a) {
+; CHECK-LABEL: subr_i32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    subr z0.s, z0.s, #127
+; CHECK-NEXT:    ret
+  %pg = call <vscale x 4 x i1> @llvm.aarch64.sve.ptrue.nxv4i1(i32 31)
+  %elt = insertelement <vscale x 4 x i32> undef, i32 127, i32 0
+  %splat = shufflevector <vscale x 4 x i32> %elt, <vscale x 4 x i32> undef, <vscale x 4 x i32> zeroinitializer
+  %out = call <vscale x 4 x i32> @llvm.aarch64.sve.subr.nxv4i32(<vscale x 4 x i1> %pg,
+                                                                <vscale x 4 x i32> %a,
+                                                                <vscale x 4 x i32> %splat)
+  ret <vscale x 4 x i32> %out
+}
+
+define <vscale x 4 x i32> @subr_i32_out_of_range(<vscale x 4 x i32> %a) {
+; CHECK-LABEL: subr_i32_out_of_range:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov w8, #257
+; CHECK-NEXT:    ptrue p0.s
+; CHECK-NEXT:    mov z1.s, w8
+; CHECK-NEXT:    subr z0.s, p0/m, z0.s, z1.s
+; CHECK-NEXT:    ret
+  %pg = call <vscale x 4 x i1> @llvm.aarch64.sve.ptrue.nxv4i1(i32 31)
+  %elt = insertelement <vscale x 4 x i32> undef, i32 257, i32 0
+  %splat = shufflevector <vscale x 4 x i32> %elt, <vscale x 4 x i32> undef, <vscale x 4 x i32> zeroinitializer
+  %out = call <vscale x 4 x i32> @llvm.aarch64.sve.subr.nxv4i32(<vscale x 4 x i1> %pg,
+                                                                <vscale x 4 x i32> %a,
+                                                                <vscale x 4 x i32> %splat)
+  ret <vscale x 4 x i32> %out
+}
+
+define <vscale x 2 x i64> @subr_i64(<vscale x 2 x i64> %a) {
+; CHECK-LABEL: subr_i64:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    subr z0.d, z0.d, #127
+; CHECK-NEXT:    ret
+  %pg = call <vscale x 2 x i1> @llvm.aarch64.sve.ptrue.nxv2i1(i32 31)
+  %elt = insertelement <vscale x 2 x i64> undef, i64 127, i64 0
+  %splat = shufflevector <vscale x 2 x i64> %elt, <vscale x 2 x i64> undef, <vscale x 2 x i32> zeroinitializer
+  %out = call <vscale x 2 x i64> @llvm.aarch64.sve.subr.nxv2i64(<vscale x 2 x i1> %pg,
+                                                                <vscale x 2 x i64> %a,
+                                                                <vscale x 2 x i64> %splat)
+  ret <vscale x 2 x i64> %out
+}
+
+define <vscale x 2 x i64> @subr_i64_out_of_range(<vscale x 2 x i64> %a) {
+; CHECK-LABEL: subr_i64_out_of_range:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov w8, #257
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    mov z1.d, x8
+; CHECK-NEXT:    subr z0.d, p0/m, z0.d, z1.d
+; CHECK-NEXT:    ret
+  %pg = call <vscale x 2 x i1> @llvm.aarch64.sve.ptrue.nxv2i1(i32 31)
+  %elt = insertelement <vscale x 2 x i64> undef, i64 257, i64 0
+  %splat = shufflevector <vscale x 2 x i64> %elt, <vscale x 2 x i64> undef, <vscale x 2 x i32> zeroinitializer
+  %out = call <vscale x 2 x i64> @llvm.aarch64.sve.subr.nxv2i64(<vscale x 2 x i1> %pg,
+                                                                <vscale x 2 x i64> %a,
+                                                                <vscale x 2 x i64> %splat)
+  ret <vscale x 2 x i64> %out
+}
+
+; As subr_i32 but where pg is i8 based and thus compatible for i32.
+define <vscale x 4 x i32> @subr_i32_ptrue_all_b(<vscale x 4 x i32> %a) #0 {
+; CHECK-LABEL: subr_i32_ptrue_all_b:
+; CHECK: subr z0.s, z0.s, #1
+; CHECK-NEXT: ret
+  %pg.b = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 31)
+  %pg.s = tail call <vscale x 4 x i1> @llvm.aarch64.sve.convert.from.svbool.nxv4i1(<vscale x 16 x i1> %pg.b)
+  %b = tail call <vscale x 4 x i32> @llvm.aarch64.sve.dup.x.nxv4i32(i32 1)
+  %out = tail call <vscale x 4 x i32> @llvm.aarch64.sve.subr.nxv4i32(<vscale x 4 x i1> %pg.s,
+                                                                     <vscale x 4 x i32> %a,
+                                                                     <vscale x 4 x i32> %b)
+  ret <vscale x 4 x i32> %out
+}
+
+; As subr_i32 but where pg is i16 based and thus compatible for i32.
+define <vscale x 4 x i32> @subr_i32_ptrue_all_h(<vscale x 4 x i32> %a) #0 {
+; CHECK-LABEL: subr_i32_ptrue_all_h:
+; CHECK: subr z0.s, z0.s, #1
+; CHECK-NEXT: ret
+  %pg.h = tail call <vscale x 8 x i1> @llvm.aarch64.sve.ptrue.nxv8i1(i32 31)
+  %pg.b = tail call <vscale x 16 x i1> @llvm.aarch64.sve.convert.to.svbool.nxv8i1(<vscale x 8 x i1> %pg.h)
+  %pg.s = tail call <vscale x 4 x i1> @llvm.aarch64.sve.convert.from.svbool.nxv4i1(<vscale x 16 x i1> %pg.b)
+  %b = tail call <vscale x 4 x i32> @llvm.aarch64.sve.dup.x.nxv4i32(i32 1)
+  %out = tail call <vscale x 4 x i32> @llvm.aarch64.sve.subr.nxv4i32(<vscale x 4 x i1> %pg.s,
+                                                                     <vscale x 4 x i32> %a,
+                                                                     <vscale x 4 x i32> %b)
+  ret <vscale x 4 x i32> %out
+}
+
+; As subr_i32 but where pg is i64 based, which is not compatibile for i32 and
+; thus inactive lanes are important and the immediate form cannot be used.
+define <vscale x 4 x i32> @subr_i32_ptrue_all_d(<vscale x 4 x i32> %a) #0 {
+; CHECK-LABEL: subr_i32_ptrue_all_d:
+; CHECK-DAG: ptrue [[PG:p[0-9]+]].d
+; CHECK-DAG: mov [[DUP:z[0-9]+]].s, #1
+; CHECK-DAG: subr z0.s, [[PG]]/m, z0.s, [[DUP]].s
+; CHECK-NEXT: ret
+  %pg.d = tail call <vscale x 2 x i1> @llvm.aarch64.sve.ptrue.nxv2i1(i32 31)
+  %pg.b = tail call <vscale x 16 x i1> @llvm.aarch64.sve.convert.to.svbool.nxv2i1(<vscale x 2 x i1> %pg.d)
+  %pg.s = tail call <vscale x 4 x i1> @llvm.aarch64.sve.convert.from.svbool.nxv4i1(<vscale x 16 x i1> %pg.b)
+  %b = tail call <vscale x 4 x i32> @llvm.aarch64.sve.dup.x.nxv4i32(i32 1)
+  %out = tail call <vscale x 4 x i32> @llvm.aarch64.sve.subr.nxv4i32(<vscale x 4 x i1> %pg.s,
+                                                                     <vscale x 4 x i32> %a,
+                                                                     <vscale x 4 x i32> %b)
+  ret <vscale x 4 x i32> %out
+}
+
 ; SMAX
 
 define <vscale x 16 x i8> @smax_i8(<vscale x 16 x i8> %a) {
 ; CHECK-LABEL: smax_i8:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    smax z0.b, z0.b, #-128
 ; CHECK-NEXT:    ret
   %pg = call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 31)

 declare <vscale x 4 x i32> @llvm.aarch64.sve.add.nxv4i32(<vscale x 4 x i1>, <vscale x 4 x i32>, <vscale x 4 x i32>)
 declare <vscale x 2 x i64> @llvm.aarch64.sve.add.nxv2i64(<vscale x 2 x i1>, <vscale x 2 x i64>, <vscale x 2 x i64>)
 
 declare <vscale x 16 x i8> @llvm.aarch64.sve.sub.nxv16i8(<vscale x 16 x i1>, <vscale x 16 x i8>, <vscale x 16 x i8>)
 declare <vscale x 8 x i16> @llvm.aarch64.sve.sub.nxv8i16(<vscale x 8 x i1>, <vscale x 8 x i16>, <vscale x 8 x i16>)
 declare <vscale x 4 x i32> @llvm.aarch64.sve.sub.nxv4i32(<vscale x 4 x i1>, <vscale x 4 x i32>, <vscale x 4 x i32>)
 declare <vscale x 2 x i64> @llvm.aarch64.sve.sub.nxv2i64(<vscale x 2 x i1>, <vscale x 2 x i64>, <vscale x 2 x i64>)
 
+declare <vscale x 16 x i8> @llvm.aarch64.sve.subr.nxv16i8(<vscale x 16 x i1>, <vscale x 16 x i8>, <vscale x 16 x i8>)
+declare <vscale x 8 x i16> @llvm.aarch64.sve.subr.nxv8i16(<vscale x 8 x i1>, <vscale x 8 x i16>, <vscale x 8 x i16>)
+declare <vscale x 4 x i32> @llvm.aarch64.sve.subr.nxv4i32(<vscale x 4 x i1>, <vscale x 4 x i32>, <vscale x 4 x i32>)
+declare <vscale x 2 x i64> @llvm.aarch64.sve.subr.nxv2i64(<vscale x 2 x i1>, <vscale x 2 x i64>, <vscale x 2 x i64>)
+
 declare <vscale x 16 x i8> @llvm.aarch64.sve.sqadd.x.nxv16i8(<vscale x 16 x i8>, <vscale x 16 x i8>)
 declare <vscale x 8 x i16> @llvm.aarch64.sve.sqadd.x.nxv8i16(<vscale x 8 x i16>, <vscale x 8 x i16>)
 declare <vscale x 4 x i32> @llvm.aarch64.sve.sqadd.x.nxv4i32(<vscale x 4 x i32>, <vscale x 4 x i32>)
 declare <vscale x 2 x i64> @llvm.aarch64.sve.sqadd.x.nxv2i64(<vscale x 2 x i64>, <vscale x 2 x i64>)
 
 declare <vscale x 16 x i8> @llvm.aarch64.sve.sqsub.x.nxv16i8(<vscale x 16 x i8>, <vscale x 16 x i8>)
 declare <vscale x 8 x i16> @llvm.aarch64.sve.sqsub.x.nxv8i16(<vscale x 8 x i16>, <vscale x 8 x i16>)
 declare <vscale x 4 x i32> @llvm.aarch64.sve.sqsub.x.nxv4i32(<vscale x 4 x i32>, <vscale x 4 x i32>)