[AArch64][SME2] Add multi-vector floating point min/max number intrinsics

[kmclaughlin]

Summary

Adds IR intrinsics for the following SME2 instructions:

• fmaxnm/fminnm (single, 2 & 4 vector)
• fmaxnm/fminnm (multi, 2 & 4 vector)

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

[david-arm]

LGTM!

llvm/include/llvm/IR/IntrinsicsAArch64.td

foreach instr = ["max", "min"] in {
       def int_aarch64_sve_ # ty # instr # _single_x4 : SME2_VG4_Multi_Single_Intrinsic;
 
       def int_aarch64_sve_ # ty # instr # _x2 : SME2_VG2_Multi_Multi_Intrinsic;
       def int_aarch64_sve_ # ty # instr # _x4 : SME2_VG4_Multi_Multi_Intrinsic;
     }
   }
 
   //
+  // Multi-vector floating point min/max number
+  //
+
+  foreach instr = ["fmaxnm", "fminnm"] in {
+    def int_aarch64_sve_ # instr # _single_x2 : SME2_VG2_Multi_Single_Intrinsic;
+    def int_aarch64_sve_ # instr # _single_x4 : SME2_VG4_Multi_Single_Intrinsic;
+
+    def int_aarch64_sve_ # instr # _x2 : SME2_VG2_Multi_Multi_Intrinsic;
+    def int_aarch64_sve_ # instr # _x4 : SME2_VG4_Multi_Multi_Intrinsic;
+  }
+
+  //
   // Multi-vector vertical dot-products
   //
 
   def int_aarch64_sme_fvdot_lane_za32_vg1x2 : SME2_Matrix_ArrayVector_VG2_Multi_Index_Intrinsic;
 
   foreach ty = ["s", "u"] in {
     def int_aarch64_sme_ #ty # vdot_lane_za32_vg1x2 : SME2_Matrix_ArrayVector_VG2_Multi_Index_Intrinsic;
     def int_aarch64_sme_ #ty # vdot_lane_za32_vg1x4 : SME2_Matrix_ArrayVector_VG4_Multi_Index_Intrinsic;

llvm/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp

case Intrinsic::aarch64_sve_umin_x4:
       return;
     case Intrinsic::aarch64_sve_fmin_x4:
       if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>(
               Node->getValueType(0),
               {0, AArch64::FMIN_VG4_4Z4Z_H, AArch64::FMIN_VG4_4Z4Z_S,
                AArch64::FMIN_VG4_4Z4Z_D}))
         SelectDestructiveMultiIntrinsic(Node, 4, true, Op);
       return;
+    case Intrinsic::aarch64_sve_fmaxnm_single_x2 :
+      if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>(
+              Node->getValueType(0),
+              {0, AArch64::FMAXNM_VG2_2ZZ_H, AArch64::FMAXNM_VG2_2ZZ_S,
+               AArch64::FMAXNM_VG2_2ZZ_D}))
+        SelectDestructiveMultiIntrinsic(Node, 2, false, Op);
+      return;
+    case Intrinsic::aarch64_sve_fmaxnm_single_x4 :
+      if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>(
+              Node->getValueType(0),
+              {0, AArch64::FMAXNM_VG4_4ZZ_H, AArch64::FMAXNM_VG4_4ZZ_S,
+               AArch64::FMAXNM_VG4_4ZZ_D}))
+        SelectDestructiveMultiIntrinsic(Node, 4, false, Op);
+      return;
+    case Intrinsic::aarch64_sve_fminnm_single_x2:
+      if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>(
+              Node->getValueType(0),
+              {0, AArch64::FMINNM_VG2_2ZZ_H, AArch64::FMINNM_VG2_2ZZ_S,
+               AArch64::FMINNM_VG2_2ZZ_D}))
+        SelectDestructiveMultiIntrinsic(Node, 2, false, Op);
+      return;
+    case Intrinsic::aarch64_sve_fminnm_single_x4:
+      if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>(
+              Node->getValueType(0),
+              {0, AArch64::FMINNM_VG4_4ZZ_H, AArch64::FMINNM_VG4_4ZZ_S,
+               AArch64::FMINNM_VG4_4ZZ_D}))
+        SelectDestructiveMultiIntrinsic(Node, 4, false, Op);
+      return;
+    case Intrinsic::aarch64_sve_fmaxnm_x2:
+      if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>(
+              Node->getValueType(0),
+              {0, AArch64::FMAXNM_VG2_2Z2Z_H, AArch64::FMAXNM_VG2_2Z2Z_S,
+               AArch64::FMAXNM_VG2_2Z2Z_D}))
+        SelectDestructiveMultiIntrinsic(Node, 2, true, Op);
+      return;
+    case Intrinsic::aarch64_sve_fmaxnm_x4:
+      if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>(
+              Node->getValueType(0),
+              {0, AArch64::FMAXNM_VG4_4Z4Z_H, AArch64::FMAXNM_VG4_4Z4Z_S,
+               AArch64::FMAXNM_VG4_4Z4Z_D}))
+        SelectDestructiveMultiIntrinsic(Node, 4, true, Op);
+      return;
+    case Intrinsic::aarch64_sve_fminnm_x2:
+      if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>(
+              Node->getValueType(0),
+              {0, AArch64::FMINNM_VG2_2Z2Z_H, AArch64::FMINNM_VG2_2Z2Z_S,
+               AArch64::FMINNM_VG2_2Z2Z_D}))
+        SelectDestructiveMultiIntrinsic(Node, 2, true, Op);
+      return;
+    case Intrinsic::aarch64_sve_fminnm_x4:
+      if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>(
+              Node->getValueType(0),
+              {0, AArch64::FMINNM_VG4_4Z4Z_H, AArch64::FMINNM_VG4_4Z4Z_S,
+               AArch64::FMINNM_VG4_4Z4Z_D}))
+        SelectDestructiveMultiIntrinsic(Node, 4, true, Op);
+      return;
     case Intrinsic::aarch64_sve_fcvts_x2:
       SelectCVTIntrinsic(Node, 2, AArch64::FCVTZS_2Z2Z_StoS);
       return;
     case Intrinsic::aarch64_sve_scvtf_x2:
       SelectCVTIntrinsic(Node, 2, AArch64::SCVTF_2Z2Z_StoS);
       return;
     case Intrinsic::aarch64_sve_fcvtu_x2:
       SelectCVTIntrinsic(Node, 2, AArch64::FCVTZU_2Z2Z_StoS);

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

 ; CHECK-NEXT:    fmax { z0.d - z3.d }, { z0.d - z3.d }, { z4.d - z7.d }
 ; CHECK-NEXT:    ret
   %res = call { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> }
               @llvm.aarch64.sve.fmax.x4.nxv2f64(<vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x double> %zdn3, <vscale x 2 x double> %zdn4,
                                                       <vscale x 2 x double> %zm1,  <vscale x 2 x double> %zm2,  <vscale x 2 x double> %zm3,  <vscale x 2 x double> %zm4)
   ret { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } %res
 }
 
+; FMAXNM (Single, x2)
+
+define { <vscale x 8 x half>, <vscale x 8 x half> }  @multi_vec_maxnm_single_x2_f16(<vscale x 8 x half> %dummy, <vscale x 8 x half> %zdn1, <vscale x 8 x half> %zdn2, <vscale x 8 x half> %zm) {
+; CHECK-LABEL: multi_vec_maxnm_single_x2_f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z5.d, z2.d
+; CHECK-NEXT:    mov z4.d, z1.d
+; CHECK-NEXT:    fmaxnm { z4.h, z5.h }, { z4.h, z5.h }, z3.h
+; CHECK-NEXT:    mov z0.d, z4.d
+; CHECK-NEXT:    mov z1.d, z5.d
+; CHECK-NEXT:    ret
+  %res = call { <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sve.fmaxnm.single.x2.nxv8f16(<vscale x 8 x half> %zdn1, <vscale x 8 x half> %zdn2, <vscale x 8 x half> %zm)
+  ret { <vscale x 8 x half>, <vscale x 8 x half> } %res
+}
+
+define { <vscale x 4 x float>, <vscale x 4 x float> }  @multi_vec_maxnm_single_x2_f32(<vscale x 8 x half> %dummy, <vscale x 4 x float> %zdn1, <vscale x 4 x float> %zdn2, <vscale x 4 x float> %zm) {
+; CHECK-LABEL: multi_vec_maxnm_single_x2_f32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z5.d, z2.d
+; CHECK-NEXT:    mov z4.d, z1.d
+; CHECK-NEXT:    fmaxnm { z4.s, z5.s }, { z4.s, z5.s }, z3.s
+; CHECK-NEXT:    mov z0.d, z4.d
+; CHECK-NEXT:    mov z1.d, z5.d
+; CHECK-NEXT:    ret
+  %res = call { <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sve.fmaxnm.single.x2.nxv4f32(<vscale x 4 x float> %zdn1, <vscale x 4 x float> %zdn2, <vscale x 4 x float> %zm)
+  ret { <vscale x 4 x float>, <vscale x 4 x float> } %res
+}
+
+define { <vscale x 2 x double>, <vscale x 2 x double> }  @multi_vec_maxnm_single_x2_f64(<vscale x 8 x half> %dummy, <vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x double> %zm) {
+; CHECK-LABEL: multi_vec_maxnm_single_x2_f64:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z5.d, z2.d
+; CHECK-NEXT:    mov z4.d, z1.d
+; CHECK-NEXT:    fmaxnm { z4.d, z5.d }, { z4.d, z5.d }, z3.d
+; CHECK-NEXT:    mov z0.d, z4.d
+; CHECK-NEXT:    mov z1.d, z5.d
+; CHECK-NEXT:    ret
+  %res = call { <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sve.fmaxnm.single.x2.nxv2f64(<vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x double> %zm)
+  ret { <vscale x 2 x double>, <vscale x 2 x double> } %res
+}
+
+; FMAXNM (Single, x4)
+
+define { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> }
+@multi_vec_maxnm_single_x4_f16(<vscale x 8 x half> %dummy, <vscale x 8 x half> %zdn1, <vscale x 8 x half> %zdn2, <vscale x 8 x half> %zdn3, <vscale x 8 x half> %zdn4, <vscale x 8 x half> %zm) {
+; CHECK-LABEL: multi_vec_maxnm_single_x4_f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z27.d, z4.d
+; CHECK-NEXT:    mov z26.d, z3.d
+; CHECK-NEXT:    mov z25.d, z2.d
+; CHECK-NEXT:    mov z24.d, z1.d
+; CHECK-NEXT:    fmaxnm { z24.h - z27.h }, { z24.h - z27.h }, z5.h
+; CHECK-NEXT:    mov z0.d, z24.d
+; CHECK-NEXT:    mov z1.d, z25.d
+; CHECK-NEXT:    mov z2.d, z26.d
+; CHECK-NEXT:    mov z3.d, z27.d
+; CHECK-NEXT:    ret
+  %res = call { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> }
+              @llvm.aarch64.sve.fmaxnm.single.x4.nxv8f16(<vscale x 8 x half> %zdn1, <vscale x 8 x half> %zdn2, <vscale x 8 x half> %zdn3, <vscale x 8 x half> %zdn4, <vscale x 8 x half> %zm)
+  ret { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } %res
+}
+
+define { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> }
+@multi_vec_maxnm_single_x4_f32(<vscale x 8 x half> %dummy, <vscale x 4 x float> %zdn1, <vscale x 4 x float> %zdn2, <vscale x 4 x float> %zdn3, <vscale x 4 x float> %zdn4, <vscale x 4 x float> %zm) {
+; CHECK-LABEL: multi_vec_maxnm_single_x4_f32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z27.d, z4.d
+; CHECK-NEXT:    mov z26.d, z3.d
+; CHECK-NEXT:    mov z25.d, z2.d
+; CHECK-NEXT:    mov z24.d, z1.d
+; CHECK-NEXT:    fmaxnm { z24.s - z27.s }, { z24.s - z27.s }, z5.s
+; CHECK-NEXT:    mov z0.d, z24.d
+; CHECK-NEXT:    mov z1.d, z25.d
+; CHECK-NEXT:    mov z2.d, z26.d
+; CHECK-NEXT:    mov z3.d, z27.d
+; CHECK-NEXT:    ret
+  %res = call { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> }
+              @llvm.aarch64.sve.fmaxnm.single.x4.nxv4f32(<vscale x 4 x float> %zdn1, <vscale x 4 x float> %zdn2, <vscale x 4 x float> %zdn3, <vscale x 4 x float> %zdn4, <vscale x 4 x float> %zm)
+  ret { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } %res
+}
+
+define { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> }
+@multi_vec_maxnm_single_x4_f64(<vscale x 8 x half> %dummy, <vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x double> %zdn3, <vscale x 2 x double> %zdn4, <vscale x 2 x double> %zm) {
+; CHECK-LABEL: multi_vec_maxnm_single_x4_f64:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z27.d, z4.d
+; CHECK-NEXT:    mov z26.d, z3.d
+; CHECK-NEXT:    mov z25.d, z2.d
+; CHECK-NEXT:    mov z24.d, z1.d
+; CHECK-NEXT:    fmaxnm { z24.d - z27.d }, { z24.d - z27.d }, z5.d
+; CHECK-NEXT:    mov z0.d, z24.d
+; CHECK-NEXT:    mov z1.d, z25.d
+; CHECK-NEXT:    mov z2.d, z26.d
+; CHECK-NEXT:    mov z3.d, z27.d
+; CHECK-NEXT:    ret
+  %res = call { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> }
+              @llvm.aarch64.sve.fmaxnm.single.x4.nxv2f64(<vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x double> %zdn3, <vscale x 2 x double> %zdn4, <vscale x 2 x double> %zm)
+  ret { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } %res
+}
+
+; FMAXNM (Multi, x2)
+
+define { <vscale x 8 x half>, <vscale x 8 x half> } @multi_vec_maxnm_x2_f16(<vscale x 8 x half> %dummy, <vscale x 8 x half> %zdn1, <vscale x 8 x half> %zdn2, <vscale x 8 x half> %zm1, <vscale x 8 x half> %zm2) {
+; CHECK-LABEL: multi_vec_maxnm_x2_f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z7.d, z4.d
+; CHECK-NEXT:    mov z5.d, z2.d
+; CHECK-NEXT:    mov z6.d, z3.d
+; CHECK-NEXT:    mov z4.d, z1.d
+; CHECK-NEXT:    fmaxnm { z4.h, z5.h }, { z4.h, z5.h }, { z6.h, z7.h }
+; CHECK-NEXT:    mov z0.d, z4.d
+; CHECK-NEXT:    mov z1.d, z5.d
+; CHECK-NEXT:    ret
+  %res = call { <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sve.fmaxnm.x2.nxv8f16(<vscale x 8 x half> %zdn1, <vscale x 8 x half> %zdn2, <vscale x 8 x half> %zm1, <vscale x 8 x half> %zm2)
+  ret { <vscale x 8 x half>, <vscale x 8 x half> } %res
+}
+
+define { <vscale x 4 x float>, <vscale x 4 x float> } @multi_vec_maxnm_x2_f32(<vscale x 8 x half> %dummy, <vscale x 4 x float> %zdn1, <vscale x 4 x float> %zdn2, <vscale x 4 x float> %zm1, <vscale x 4 x float> %zm2) {
+; CHECK-LABEL: multi_vec_maxnm_x2_f32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z7.d, z4.d
+; CHECK-NEXT:    mov z5.d, z2.d
+; CHECK-NEXT:    mov z6.d, z3.d
+; CHECK-NEXT:    mov z4.d, z1.d
+; CHECK-NEXT:    fmaxnm { z4.s, z5.s }, { z4.s, z5.s }, { z6.s, z7.s }
+; CHECK-NEXT:    mov z0.d, z4.d
+; CHECK-NEXT:    mov z1.d, z5.d
+; CHECK-NEXT:    ret
+  %res = call { <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sve.fmaxnm.x2.nxv4f32(<vscale x 4 x float> %zdn1, <vscale x 4 x float> %zdn2, <vscale x 4 x float> %zm1, <vscale x 4 x float> %zm2)
+  ret { <vscale x 4 x float>, <vscale x 4 x float> } %res
+}
+
+define { <vscale x 2 x double>, <vscale x 2 x double> } @multi_vec_maxnm_x2_f64(<vscale x 8 x half> %dummy, <vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x double> %zm1, <vscale x 2 x double> %zm2) {
+; CHECK-LABEL: multi_vec_maxnm_x2_f64:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z7.d, z4.d
+; CHECK-NEXT:    mov z5.d, z2.d
+; CHECK-NEXT:    mov z6.d, z3.d
+; CHECK-NEXT:    mov z4.d, z1.d
+; CHECK-NEXT:    fmaxnm { z4.d, z5.d }, { z4.d, z5.d }, { z6.d, z7.d }
+; CHECK-NEXT:    mov z0.d, z4.d
+; CHECK-NEXT:    mov z1.d, z5.d
+; CHECK-NEXT:    ret
+  %res = call { <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sve.fmaxnm.x2.nxv2f64(<vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x double> %zm1, <vscale x 2 x double> %zm2)
+  ret { <vscale x 2 x double>, <vscale x 2 x double> } %res
+}
+
+; FMAXNM (Multi, x4)
+
+define { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> }
+@multi_vec_maxnm_x4_f16(<vscale x 8 x half> %dummy, <vscale x 8 x half> %zdn1, <vscale x 8 x half> %zdn2, <vscale x 8 x half> %zdn3, <vscale x 8 x half> %zdn4, <vscale x 8 x half> %zm1, <vscale x 8 x half> %zm2, <vscale x 8 x half> %zm3, <vscale x 8 x half> %zm4) {
+; CHECK-LABEL: multi_vec_maxnm_x4_f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z30.d, z7.d
+; CHECK-NEXT:    ptrue p0.h
+; CHECK-NEXT:    mov z29.d, z6.d
+; CHECK-NEXT:    mov z27.d, z4.d
+; CHECK-NEXT:    mov z28.d, z5.d
+; CHECK-NEXT:    mov z26.d, z3.d
+; CHECK-NEXT:    ld1h { z31.h }, p0/z, [x0]
+; CHECK-NEXT:    mov z25.d, z2.d
+; CHECK-NEXT:    mov z24.d, z1.d
+; CHECK-NEXT:    fmaxnm { z24.h - z27.h }, { z24.h - z27.h }, { z28.h - z31.h }
+; CHECK-NEXT:    mov z0.d, z24.d
+; CHECK-NEXT:    mov z1.d, z25.d
+; CHECK-NEXT:    mov z2.d, z26.d
+; CHECK-NEXT:    mov z3.d, z27.d
+; CHECK-NEXT:    ret
+  %res = call { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> }
+              @llvm.aarch64.sve.fmaxnm.x4.nxv8f16(<vscale x 8 x half> %zdn1, <vscale x 8 x half> %zdn2, <vscale x 8 x half> %zdn3, <vscale x 8 x half> %zdn4, 
+                                                  <vscale x 8 x half> %zm1, <vscale x 8 x half> %zm2, <vscale x 8 x half> %zm3, <vscale x 8 x half> %zm4)
+  ret { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } %res
+}
+
+define { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> }
+@multi_vec_maxnm_x4_f32(<vscale x 8 x half> %dummy, <vscale x 4 x float> %zdn1, <vscale x 4 x float> %zdn2, <vscale x 4 x float> %zdn3, <vscale x 4 x float> %zdn4, <vscale x 4 x float> %zm1, <vscale x 4 x float> %zm2, <vscale x 4 x float> %zm3, <vscale x 4 x float> %zm4) {
+; CHECK-LABEL: multi_vec_maxnm_x4_f32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z30.d, z7.d
+; CHECK-NEXT:    ptrue p0.s
+; CHECK-NEXT:    mov z29.d, z6.d
+; CHECK-NEXT:    mov z27.d, z4.d
+; CHECK-NEXT:    mov z28.d, z5.d
+; CHECK-NEXT:    mov z26.d, z3.d
+; CHECK-NEXT:    ld1w { z31.s }, p0/z, [x0]
+; CHECK-NEXT:    mov z25.d, z2.d
+; CHECK-NEXT:    mov z24.d, z1.d
+; CHECK-NEXT:    fmaxnm { z24.s - z27.s }, { z24.s - z27.s }, { z28.s - z31.s }
+; CHECK-NEXT:    mov z0.d, z24.d
+; CHECK-NEXT:    mov z1.d, z25.d
+; CHECK-NEXT:    mov z2.d, z26.d
+; CHECK-NEXT:    mov z3.d, z27.d
+; CHECK-NEXT:    ret
+  %res = call { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> }
+              @llvm.aarch64.sve.fmaxnm.x4.nxv4f32(<vscale x 4 x float> %zdn1, <vscale x 4 x float> %zdn2, <vscale x 4 x float> %zdn3, <vscale x 4 x float> %zdn4,
+                                                  <vscale x 4 x float> %zm1, <vscale x 4 x float> %zm2, <vscale x 4 x float> %zm3, <vscale x 4 x float> %zm4)
+  ret { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } %res
+}
+
+define { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> }
+@multi_vec_maxnm_x4_f64(<vscale x 8 x half> %dummy, <vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x double> %zdn3, <vscale x 2 x double> %zdn4, <vscale x 2 x double> %zm1, <vscale x 2 x double> %zm2, <vscale x 2 x double> %zm3, <vscale x 2 x double> %zm4) {
+; CHECK-LABEL: multi_vec_maxnm_x4_f64:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z30.d, z7.d
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    mov z29.d, z6.d
+; CHECK-NEXT:    mov z27.d, z4.d
+; CHECK-NEXT:    mov z28.d, z5.d
+; CHECK-NEXT:    mov z26.d, z3.d
+; CHECK-NEXT:    ld1d { z31.d }, p0/z, [x0]
+; CHECK-NEXT:    mov z25.d, z2.d
+; CHECK-NEXT:    mov z24.d, z1.d
+; CHECK-NEXT:    fmaxnm { z24.d - z27.d }, { z24.d - z27.d }, { z28.d - z31.d }
+; CHECK-NEXT:    mov z0.d, z24.d
+; CHECK-NEXT:    mov z1.d, z25.d
+; CHECK-NEXT:    mov z2.d, z26.d
+; CHECK-NEXT:    mov z3.d, z27.d
+; CHECK-NEXT:    ret
+  %res = call { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> }
+              @llvm.aarch64.sve.fmaxnm.x4.nxv2f64(<vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x double> %zdn3, <vscale x 2 x double> %zdn4,
+                                                  <vscale x 2 x double> %zm1, <vscale x 2 x double> %zm2, <vscale x 2 x double> %zm3, <vscale x 2 x double> %zm4)
+  ret { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } %res
+}
+
 declare { <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sve.smax.single.x2.nxv16i8(<vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>)
 declare { <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sve.smax.single.x2.nxv8i16(<vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>)
 declare { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sve.smax.single.x2.nxv4i32(<vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>)
 declare { <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sve.smax.single.x2.nxv2i64(<vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>)
 
 declare { <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sve.umax.single.x2.nxv16i8(<vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>)
 declare { <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sve.umax.single.x2.nxv8i16(<vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>)
 declare { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sve.umax.single.x2.nxv4i32(<vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>)

  @llvm.aarch64.sve.umax.x4.nxv2i64(<vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>)
 
 declare { <vscale x 8 x half>, <vscale x 8 x half> , <vscale x 8 x half>, <vscale x 8 x half> }
  @llvm.aarch64.sve.fmax.x4.nxv8f16(<vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>)
 declare { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> }
  @llvm.aarch64.sve.fmax.x4.nxv4f32(<vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>)
 declare { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> }
  @llvm.aarch64.sve.fmax.x4.nxv2f64(<vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>)
+
+declare { <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sve.fmaxnm.single.x2.nxv8f16(<vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>)
+declare { <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sve.fmaxnm.single.x2.nxv4f32(<vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>)
+declare { <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sve.fmaxnm.single.x2.nxv2f64(<vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>)
+
+declare { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> }
+ @llvm.aarch64.sve.fmaxnm.single.x4.nxv8f16(<vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>)
+declare { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> }
+ @llvm.aarch64.sve.fmaxnm.single.x4.nxv4f32(<vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>)
+declare { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> }
+ @llvm.aarch64.sve.fmaxnm.single.x4.nxv2f64(<vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>)
+
+declare { <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sve.fmaxnm.x2.nxv8f16(<vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>)
+declare { <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sve.fmaxnm.x2.nxv4f32(<vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>)
+declare { <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sve.fmaxnm.x2.nxv2f64(<vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>)
+
+declare { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> }
+ @llvm.aarch64.sve.fmaxnm.x4.nxv8f16(<vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>)
+declare { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> }
+ @llvm.aarch64.sve.fmaxnm.x4.nxv4f32(<vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>)
+declare { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> }
+ @llvm.aarch64.sve.fmaxnm.x4.nxv2f64(<vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>)

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

-; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sme2 -mattr=+b16 -verify-machineinstrs < %s | FileCheck %s
+; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sme2 -verify-machineinstrs < %s | FileCheck %s
 
 ; SMIN (Single, x2)
 
 define { <vscale x 16 x i8>, <vscale x 16 x i8> } @multi_vec_min_single_x2_s8(<vscale x 16 x i8> %zdn1, <vscale x 16 x i8> %zdn2, <vscale x 16 x i8> %zm) {
 ; CHECK-LABEL: multi_vec_min_single_x2_s8:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    // kill: def $z1 killed $z1 killed $z0_z1 def $z0_z1
 ; CHECK-NEXT:    // kill: def $z0 killed $z0 killed $z0_z1 def $z0_z1

 ; CHECK-NEXT:    fmin { z0.d - z3.d }, { z0.d - z3.d }, { z4.d - z7.d }
 ; CHECK-NEXT:    ret
   %res = call { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> }
               @llvm.aarch64.sve.fmin.x4.nxv2f64(<vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x double> %zdn3, <vscale x 2 x double> %zdn4,
                                                       <vscale x 2 x double> %zm1,  <vscale x 2 x double> %zm2,  <vscale x 2 x double> %zm3,  <vscale x 2 x double> %zm4)
   ret { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } %res
 }
 
+; FMINNM (Single, x2)
+
+define { <vscale x 8 x half>, <vscale x 8 x half> }  @multi_vec_minnm_single_x2_f16(<vscale x 8 x half> %dummy, <vscale x 8 x half> %zdn1, <vscale x 8 x half> %zdn2, <vscale x 8 x half> %zm) {
+; CHECK-LABEL: multi_vec_minnm_single_x2_f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z5.d, z2.d
+; CHECK-NEXT:    mov z4.d, z1.d
+; CHECK-NEXT:    fminnm { z4.h, z5.h }, { z4.h, z5.h }, z3.h
+; CHECK-NEXT:    mov z0.d, z4.d
+; CHECK-NEXT:    mov z1.d, z5.d
+; CHECK-NEXT:    ret
+  %res = call { <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sve.fminnm.single.x2.nxv8f16(<vscale x 8 x half> %zdn1, <vscale x 8 x half> %zdn2, <vscale x 8 x half> %zm)
+  ret { <vscale x 8 x half>, <vscale x 8 x half> } %res
+}
+
+define { <vscale x 4 x float>, <vscale x 4 x float> }  @multi_vec_minnm_single_x2_f32(<vscale x 8 x half> %dummy, <vscale x 4 x float> %zdn1, <vscale x 4 x float> %zdn2, <vscale x 4 x float> %zm) {
+; CHECK-LABEL: multi_vec_minnm_single_x2_f32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z5.d, z2.d
+; CHECK-NEXT:    mov z4.d, z1.d
+; CHECK-NEXT:    fminnm { z4.s, z5.s }, { z4.s, z5.s }, z3.s
+; CHECK-NEXT:    mov z0.d, z4.d
+; CHECK-NEXT:    mov z1.d, z5.d
+; CHECK-NEXT:    ret
+  %res = call { <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sve.fminnm.single.x2.nxv4f32(<vscale x 4 x float> %zdn1, <vscale x 4 x float> %zdn2, <vscale x 4 x float> %zm)
+  ret { <vscale x 4 x float>, <vscale x 4 x float> } %res
+}
+
+define { <vscale x 2 x double>, <vscale x 2 x double> }  @multi_vec_minnm_single_x2_f64(<vscale x 8 x half> %dummy, <vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x double> %zm) {
+; CHECK-LABEL: multi_vec_minnm_single_x2_f64:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z5.d, z2.d
+; CHECK-NEXT:    mov z4.d, z1.d
+; CHECK-NEXT:    fminnm { z4.d, z5.d }, { z4.d, z5.d }, z3.d
+; CHECK-NEXT:    mov z0.d, z4.d
+; CHECK-NEXT:    mov z1.d, z5.d
+; CHECK-NEXT:    ret
+  %res = call { <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sve.fminnm.single.x2.nxv2f64(<vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x double> %zm)
+  ret { <vscale x 2 x double>, <vscale x 2 x double> } %res
+}
+
+; FMINNM (Single, x4)
+
+define { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> }
+@multi_vec_minnm_single_x4_f16(<vscale x 8 x half> %dummy, <vscale x 8 x half> %zdn1, <vscale x 8 x half> %zdn2, <vscale x 8 x half> %zdn3, <vscale x 8 x half> %zdn4, <vscale x 8 x half> %zm) {
+; CHECK-LABEL: multi_vec_minnm_single_x4_f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z27.d, z4.d
+; CHECK-NEXT:    mov z26.d, z3.d
+; CHECK-NEXT:    mov z25.d, z2.d
+; CHECK-NEXT:    mov z24.d, z1.d
+; CHECK-NEXT:    fminnm { z24.h - z27.h }, { z24.h - z27.h }, z5.h
+; CHECK-NEXT:    mov z0.d, z24.d
+; CHECK-NEXT:    mov z1.d, z25.d
+; CHECK-NEXT:    mov z2.d, z26.d
+; CHECK-NEXT:    mov z3.d, z27.d
+; CHECK-NEXT:    ret
+  %res = call { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> }
+              @llvm.aarch64.sve.fminnm.single.x4.nxv8f16(<vscale x 8 x half> %zdn1, <vscale x 8 x half> %zdn2, <vscale x 8 x half> %zdn3, <vscale x 8 x half> %zdn4, <vscale x 8 x half> %zm)
+  ret { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } %res
+}
+
+define { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> }
+@multi_vec_minnm_single_x4_f32(<vscale x 8 x half> %dummy, <vscale x 4 x float> %zdn1, <vscale x 4 x float> %zdn2, <vscale x 4 x float> %zdn3, <vscale x 4 x float> %zdn4, <vscale x 4 x float> %zm) {
+; CHECK-LABEL: multi_vec_minnm_single_x4_f32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z27.d, z4.d
+; CHECK-NEXT:    mov z26.d, z3.d
+; CHECK-NEXT:    mov z25.d, z2.d
+; CHECK-NEXT:    mov z24.d, z1.d
+; CHECK-NEXT:    fminnm { z24.s - z27.s }, { z24.s - z27.s }, z5.s
+; CHECK-NEXT:    mov z0.d, z24.d
+; CHECK-NEXT:    mov z1.d, z25.d
+; CHECK-NEXT:    mov z2.d, z26.d
+; CHECK-NEXT:    mov z3.d, z27.d
+; CHECK-NEXT:    ret
+  %res = call { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> }
+              @llvm.aarch64.sve.fminnm.single.x4.nxv4f32(<vscale x 4 x float> %zdn1, <vscale x 4 x float> %zdn2, <vscale x 4 x float> %zdn3, <vscale x 4 x float> %zdn4, <vscale x 4 x float> %zm)
+  ret { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } %res
+}
+
+define { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> }
+@multi_vec_minnm_single_x4_f64(<vscale x 8 x half> %dummy, <vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x double> %zdn3, <vscale x 2 x double> %zdn4, <vscale x 2 x double> %zm) {
+; CHECK-LABEL: multi_vec_minnm_single_x4_f64:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z27.d, z4.d
+; CHECK-NEXT:    mov z26.d, z3.d
+; CHECK-NEXT:    mov z25.d, z2.d
+; CHECK-NEXT:    mov z24.d, z1.d
+; CHECK-NEXT:    fminnm { z24.d - z27.d }, { z24.d - z27.d }, z5.d
+; CHECK-NEXT:    mov z0.d, z24.d
+; CHECK-NEXT:    mov z1.d, z25.d
+; CHECK-NEXT:    mov z2.d, z26.d
+; CHECK-NEXT:    mov z3.d, z27.d
+; CHECK-NEXT:    ret
+  %res = call { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> }
+              @llvm.aarch64.sve.fminnm.single.x4.nxv2f64(<vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x double> %zdn3, <vscale x 2 x double> %zdn4, <vscale x 2 x double> %zm)
+  ret { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } %res
+}
+
+; FMINNM (Multi, x2)
+
+define { <vscale x 8 x half>, <vscale x 8 x half> } @multi_vec_minnm_x2_f16(<vscale x 8 x half> %dummy, <vscale x 8 x half> %zdn1, <vscale x 8 x half> %zdn2, <vscale x 8 x half> %zm1, <vscale x 8 x half> %zm2) {
+; CHECK-LABEL: multi_vec_minnm_x2_f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z7.d, z4.d
+; CHECK-NEXT:    mov z5.d, z2.d
+; CHECK-NEXT:    mov z6.d, z3.d
+; CHECK-NEXT:    mov z4.d, z1.d
+; CHECK-NEXT:    fminnm { z4.h, z5.h }, { z4.h, z5.h }, { z6.h, z7.h }
+; CHECK-NEXT:    mov z0.d, z4.d
+; CHECK-NEXT:    mov z1.d, z5.d
+; CHECK-NEXT:    ret
+  %res = call { <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sve.fminnm.x2.nxv8f16(<vscale x 8 x half> %zdn1, <vscale x 8 x half> %zdn2, <vscale x 8 x half> %zm1, <vscale x 8 x half> %zm2)
+  ret { <vscale x 8 x half>, <vscale x 8 x half> } %res
+}
+
+define { <vscale x 4 x float>, <vscale x 4 x float> } @multi_vec_minnm_x2_f32(<vscale x 8 x half> %dummy, <vscale x 4 x float> %zdn1, <vscale x 4 x float> %zdn2, <vscale x 4 x float> %zm1, <vscale x 4 x float> %zm2) {
+; CHECK-LABEL: multi_vec_minnm_x2_f32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z7.d, z4.d
+; CHECK-NEXT:    mov z5.d, z2.d
+; CHECK-NEXT:    mov z6.d, z3.d
+; CHECK-NEXT:    mov z4.d, z1.d
+; CHECK-NEXT:    fminnm { z4.s, z5.s }, { z4.s, z5.s }, { z6.s, z7.s }
+; CHECK-NEXT:    mov z0.d, z4.d
+; CHECK-NEXT:    mov z1.d, z5.d
+; CHECK-NEXT:    ret
+  %res = call { <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sve.fminnm.x2.nxv4f32(<vscale x 4 x float> %zdn1, <vscale x 4 x float> %zdn2, <vscale x 4 x float> %zm1, <vscale x 4 x float> %zm2)
+  ret { <vscale x 4 x float>, <vscale x 4 x float> } %res
+}
+
+define { <vscale x 2 x double>, <vscale x 2 x double> } @multi_vec_minnm_x2_f64(<vscale x 8 x half> %dummy, <vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x double> %zm1, <vscale x 2 x double> %zm2) {
+; CHECK-LABEL: multi_vec_minnm_x2_f64:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z7.d, z4.d
+; CHECK-NEXT:    mov z5.d, z2.d
+; CHECK-NEXT:    mov z6.d, z3.d
+; CHECK-NEXT:    mov z4.d, z1.d
+; CHECK-NEXT:    fminnm { z4.d, z5.d }, { z4.d, z5.d }, { z6.d, z7.d }
+; CHECK-NEXT:    mov z0.d, z4.d
+; CHECK-NEXT:    mov z1.d, z5.d
+; CHECK-NEXT:    ret
+  %res = call { <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sve.fminnm.x2.nxv2f64(<vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x double> %zm1, <vscale x 2 x double> %zm2)
+  ret { <vscale x 2 x double>, <vscale x 2 x double> } %res
+}
+
+; FMINNM (Multi, x4)
+
+define { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> }
+@multi_vec_minnm_x4_f16(<vscale x 8 x half> %dummy, <vscale x 8 x half> %zdn1, <vscale x 8 x half> %zdn2, <vscale x 8 x half> %zdn3, <vscale x 8 x half> %zdn4, <vscale x 8 x half> %zm1, <vscale x 8 x half> %zm2, <vscale x 8 x half> %zm3, <vscale x 8 x half> %zm4) {
+; CHECK-LABEL: multi_vec_minnm_x4_f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z30.d, z7.d
+; CHECK-NEXT:    ptrue p0.h
+; CHECK-NEXT:    mov z29.d, z6.d
+; CHECK-NEXT:    mov z27.d, z4.d
+; CHECK-NEXT:    mov z28.d, z5.d
+; CHECK-NEXT:    mov z26.d, z3.d
+; CHECK-NEXT:    ld1h { z31.h }, p0/z, [x0]
+; CHECK-NEXT:    mov z25.d, z2.d
+; CHECK-NEXT:    mov z24.d, z1.d
+; CHECK-NEXT:    fminnm { z24.h - z27.h }, { z24.h - z27.h }, { z28.h - z31.h }
+; CHECK-NEXT:    mov z0.d, z24.d
+; CHECK-NEXT:    mov z1.d, z25.d
+; CHECK-NEXT:    mov z2.d, z26.d
+; CHECK-NEXT:    mov z3.d, z27.d
+; CHECK-NEXT:    ret
+  %res = call { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> }
+              @llvm.aarch64.sve.fminnm.x4.nxv8f16(<vscale x 8 x half> %zdn1, <vscale x 8 x half> %zdn2, <vscale x 8 x half> %zdn3, <vscale x 8 x half> %zdn4,
+                                                  <vscale x 8 x half> %zm1, <vscale x 8 x half> %zm2, <vscale x 8 x half> %zm3, <vscale x 8 x half> %zm4)
+  ret { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } %res
+}
+
+define { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> }
+@multi_vec_minnm_x4_f32(<vscale x 8 x half> %dummy, <vscale x 4 x float> %zdn1, <vscale x 4 x float> %zdn2, <vscale x 4 x float> %zdn3, <vscale x 4 x float> %zdn4, <vscale x 4 x float> %zm1, <vscale x 4 x float> %zm2, <vscale x 4 x float> %zm3, <vscale x 4 x float> %zm4) {
+; CHECK-LABEL: multi_vec_minnm_x4_f32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z30.d, z7.d
+; CHECK-NEXT:    ptrue p0.s
+; CHECK-NEXT:    mov z29.d, z6.d
+; CHECK-NEXT:    mov z27.d, z4.d
+; CHECK-NEXT:    mov z28.d, z5.d
+; CHECK-NEXT:    mov z26.d, z3.d
+; CHECK-NEXT:    ld1w { z31.s }, p0/z, [x0]
+; CHECK-NEXT:    mov z25.d, z2.d
+; CHECK-NEXT:    mov z24.d, z1.d
+; CHECK-NEXT:    fminnm { z24.s - z27.s }, { z24.s - z27.s }, { z28.s - z31.s }
+; CHECK-NEXT:    mov z0.d, z24.d
+; CHECK-NEXT:    mov z1.d, z25.d
+; CHECK-NEXT:    mov z2.d, z26.d
+; CHECK-NEXT:    mov z3.d, z27.d
+; CHECK-NEXT:    ret
+  %res = call { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> }
+              @llvm.aarch64.sve.fminnm.x4.nxv4f32(<vscale x 4 x float> %zdn1, <vscale x 4 x float> %zdn2, <vscale x 4 x float> %zdn3, <vscale x 4 x float> %zdn4,
+                                                  <vscale x 4 x float> %zm1, <vscale x 4 x float> %zm2, <vscale x 4 x float> %zm3, <vscale x 4 x float> %zm4)
+  ret { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } %res
+}
+
+define { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> }
+@multi_vec_minnm_x4_f64(<vscale x 8 x half> %dummy, <vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x double> %zdn3, <vscale x 2 x double> %zdn4, <vscale x 2 x double> %zm1, <vscale x 2 x double> %zm2, <vscale x 2 x double> %zm3, <vscale x 2 x double> %zm4) {
+; CHECK-LABEL: multi_vec_minnm_x4_f64:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z30.d, z7.d
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    mov z29.d, z6.d
+; CHECK-NEXT:    mov z27.d, z4.d
+; CHECK-NEXT:    mov z28.d, z5.d
+; CHECK-NEXT:    mov z26.d, z3.d
+; CHECK-NEXT:    ld1d { z31.d }, p0/z, [x0]
+; CHECK-NEXT:    mov z25.d, z2.d
+; CHECK-NEXT:    mov z24.d, z1.d
+; CHECK-NEXT:    fminnm { z24.d - z27.d }, { z24.d - z27.d }, { z28.d - z31.d }
+; CHECK-NEXT:    mov z0.d, z24.d
+; CHECK-NEXT:    mov z1.d, z25.d
+; CHECK-NEXT:    mov z2.d, z26.d
+; CHECK-NEXT:    mov z3.d, z27.d
+; CHECK-NEXT:    ret
+  %res = call { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> }
+              @llvm.aarch64.sve.fminnm.x4.nxv2f64(<vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x double> %zdn3, <vscale x 2 x double> %zdn4,
+                                                  <vscale x 2 x double> %zm1, <vscale x 2 x double> %zm2, <vscale x 2 x double> %zm3, <vscale x 2 x double> %zm4)
+  ret { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } %res
+}
+
 declare { <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sve.smin.single.x2.nxv16i8(<vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>)
 declare { <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sve.smin.single.x2.nxv8i16(<vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>)
 declare { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sve.smin.single.x2.nxv4i32(<vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>)
 declare { <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sve.smin.single.x2.nxv2i64(<vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>)
 
 declare { <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sve.umin.single.x2.nxv16i8(<vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>)
 declare { <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sve.umin.single.x2.nxv8i16(<vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>)
 declare { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sve.umin.single.x2.nxv4i32(<vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>)

  @llvm.aarch64.sve.umin.x4.nxv2i64(<vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>)
 
 declare { <vscale x 8 x half>, <vscale x 8 x half> , <vscale x 8 x half>, <vscale x 8 x half> }
  @llvm.aarch64.sve.fmin.x4.nxv8f16(<vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>)
 declare { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> }
  @llvm.aarch64.sve.fmin.x4.nxv4f32(<vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>)
 declare { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> }
  @llvm.aarch64.sve.fmin.x4.nxv2f64(<vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>)
+
+declare { <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sve.fminnm.single.x2.nxv8f16(<vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>)
+declare { <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sve.fminnm.single.x2.nxv4f32(<vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>)
+declare { <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sve.fminnm.single.x2.nxv2f64(<vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>)
+
+declare { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> }
+ @llvm.aarch64.sve.fminnm.single.x4.nxv8f16(<vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>)
+declare { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> }
+ @llvm.aarch64.sve.fminnm.single.x4.nxv4f32(<vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>)
+declare { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> }
+ @llvm.aarch64.sve.fminnm.single.x4.nxv2f64(<vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>)
+
+declare { <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sve.fminnm.x2.nxv8f16(<vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>)
+declare { <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sve.fminnm.x2.nxv4f32(<vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>)
+declare { <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sve.fminnm.x2.nxv2f64(<vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>)
+
+declare { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> }
+ @llvm.aarch64.sve.fminnm.x4.nxv8f16(<vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>)
+declare { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> }
+ @llvm.aarch64.sve.fminnm.x4.nxv4f32(<vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>)
+declare { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> }
+ @llvm.aarch64.sve.fminnm.x4.nxv2f64(<vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>)