; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
; RUN: llc -mattr=+sve,+bf16             < %s | FileCheck %s --check-prefixes=CHECK,NOB16B16
; RUN: llc -mattr=+sve,+bf16,+sve-b16b16 < %s | FileCheck %s --check-prefixes=CHECK,B16B16

target triple = "aarch64-unknown-linux-gnu"

;
; FABS
;

define <4 x bfloat> @fabs_v4bf16(<4 x bfloat> %a) {
; CHECK-LABEL: fabs_v4bf16:
; CHECK:       // %bb.0:
; CHECK-NEXT:    fabs v0.4h, v0.4h
; CHECK-NEXT:    ret
  %res = call <4 x bfloat> @llvm.fabs.v4bf16(<4 x bfloat> %a)
  ret <4 x bfloat> %res
}

define <8 x bfloat> @fabs_v8bf16(<8 x bfloat> %a) {
; CHECK-LABEL: fabs_v8bf16:
; CHECK:       // %bb.0:
; CHECK-NEXT:    fabs v0.8h, v0.8h
; CHECK-NEXT:    ret
  %res = call <8 x bfloat> @llvm.fabs.v8bf16(<8 x bfloat> %a)
  ret <8 x bfloat> %res
}

;
; FADD
;

define <4 x bfloat> @fadd_v4bf16(<4 x bfloat> %a, <4 x bfloat> %b) {
; NOB16B16-LABEL: fadd_v4bf16:
; NOB16B16:       // %bb.0:
; NOB16B16-NEXT:    shll v1.4s, v1.4h, #16
; NOB16B16-NEXT:    shll v0.4s, v0.4h, #16
; NOB16B16-NEXT:    fadd v0.4s, v0.4s, v1.4s
; NOB16B16-NEXT:    bfcvtn v0.4h, v0.4s
; NOB16B16-NEXT:    ret
;
; B16B16-LABEL: fadd_v4bf16:
; B16B16:       // %bb.0:
; B16B16-NEXT:    ptrue p0.h, vl4
; B16B16-NEXT:    // kill: def $d0 killed $d0 def $z0
; B16B16-NEXT:    // kill: def $d1 killed $d1 def $z1
; B16B16-NEXT:    bfadd z0.h, p0/m, z0.h, z1.h
; B16B16-NEXT:    // kill: def $d0 killed $d0 killed $z0
; B16B16-NEXT:    ret
  %res = fadd <4 x bfloat> %a, %b
  ret <4 x bfloat> %res
}

define <8 x bfloat> @fadd_v8bf16(<8 x bfloat> %a, <8 x bfloat> %b) {
; NOB16B16-LABEL: fadd_v8bf16:
; NOB16B16:       // %bb.0:
; NOB16B16-NEXT:    shll v2.4s, v1.4h, #16
; NOB16B16-NEXT:    shll v3.4s, v0.4h, #16
; NOB16B16-NEXT:    shll2 v1.4s, v1.8h, #16
; NOB16B16-NEXT:    shll2 v0.4s, v0.8h, #16
; NOB16B16-NEXT:    fadd v2.4s, v3.4s, v2.4s
; NOB16B16-NEXT:    fadd v1.4s, v0.4s, v1.4s
; NOB16B16-NEXT:    bfcvtn v0.4h, v2.4s
; NOB16B16-NEXT:    bfcvtn2 v0.8h, v1.4s
; NOB16B16-NEXT:    ret
;
; B16B16-LABEL: fadd_v8bf16:
; B16B16:       // %bb.0:
; B16B16-NEXT:    ptrue p0.h, vl8
; B16B16-NEXT:    // kill: def $q0 killed $q0 def $z0
; B16B16-NEXT:    // kill: def $q1 killed $q1 def $z1
; B16B16-NEXT:    bfadd z0.h, p0/m, z0.h, z1.h
; B16B16-NEXT:    // kill: def $q0 killed $q0 killed $z0
; B16B16-NEXT:    ret
  %res = fadd <8 x bfloat> %a, %b
  ret <8 x bfloat> %res
}

;
; FDIV
;

define <4 x bfloat> @fdiv_v4bf16(<4 x bfloat> %a, <4 x bfloat> %b) {
; CHECK-LABEL: fdiv_v4bf16:
; CHECK:       // %bb.0:
; CHECK-NEXT:    shll v1.4s, v1.4h, #16
; CHECK-NEXT:    shll v0.4s, v0.4h, #16
; CHECK-NEXT:    fdiv v0.4s, v0.4s, v1.4s
; CHECK-NEXT:    bfcvtn v0.4h, v0.4s
; CHECK-NEXT:    ret
  %res = fdiv <4 x bfloat> %a, %b
  ret <4 x bfloat> %res
}

define <8 x bfloat> @fdiv_v8bf16(<8 x bfloat> %a, <8 x bfloat> %b) {
; CHECK-LABEL: fdiv_v8bf16:
; CHECK:       // %bb.0:
; CHECK-NEXT:    shll v2.4s, v1.4h, #16
; CHECK-NEXT:    shll v3.4s, v0.4h, #16
; CHECK-NEXT:    shll2 v1.4s, v1.8h, #16
; CHECK-NEXT:    shll2 v0.4s, v0.8h, #16
; CHECK-NEXT:    fdiv v2.4s, v3.4s, v2.4s
; CHECK-NEXT:    fdiv v1.4s, v0.4s, v1.4s
; CHECK-NEXT:    bfcvtn v0.4h, v2.4s
; CHECK-NEXT:    bfcvtn2 v0.8h, v1.4s
; CHECK-NEXT:    ret
  %res = fdiv <8 x bfloat> %a, %b
  ret <8 x bfloat> %res
}

;
; FMAX
;

define <4 x bfloat> @fmax_v4bf16(<4 x bfloat> %a, <4 x bfloat> %b) {
; NOB16B16-LABEL: fmax_v4bf16:
; NOB16B16:       // %bb.0:
; NOB16B16-NEXT:    // kill: def $d1 killed $d1 def $q1
; NOB16B16-NEXT:    // kill: def $d0 killed $d0 def $q0
; NOB16B16-NEXT:    mov h2, v1.h[1]
; NOB16B16-NEXT:    mov h3, v0.h[1]
; NOB16B16-NEXT:    mov h4, v1.h[2]
; NOB16B16-NEXT:    shll v5.4s, v1.4h, #16
; NOB16B16-NEXT:    shll v6.4s, v0.4h, #16
; NOB16B16-NEXT:    mov h7, v0.h[2]
; NOB16B16-NEXT:    mov h1, v1.h[3]
; NOB16B16-NEXT:    shll v2.4s, v2.4h, #16
; NOB16B16-NEXT:    shll v3.4s, v3.4h, #16
; NOB16B16-NEXT:    shll v4.4s, v4.4h, #16
; NOB16B16-NEXT:    shll v1.4s, v1.4h, #16
; NOB16B16-NEXT:    fmax s2, s3, s2
; NOB16B16-NEXT:    fmax s3, s6, s5
; NOB16B16-NEXT:    shll v5.4s, v7.4h, #16
; NOB16B16-NEXT:    mov h6, v0.h[3]
; NOB16B16-NEXT:    fmax s4, s5, s4
; NOB16B16-NEXT:    bfcvt h2, s2
; NOB16B16-NEXT:    bfcvt h0, s3
; NOB16B16-NEXT:    shll v3.4s, v6.4h, #16
; NOB16B16-NEXT:    mov v0.h[1], v2.h[0]
; NOB16B16-NEXT:    bfcvt h2, s4
; NOB16B16-NEXT:    fmax s1, s3, s1
; NOB16B16-NEXT:    mov v0.h[2], v2.h[0]
; NOB16B16-NEXT:    bfcvt h1, s1
; NOB16B16-NEXT:    mov v0.h[3], v1.h[0]
; NOB16B16-NEXT:    // kill: def $d0 killed $d0 killed $q0
; NOB16B16-NEXT:    ret
;
; B16B16-LABEL: fmax_v4bf16:
; B16B16:       // %bb.0:
; B16B16-NEXT:    ptrue p0.h, vl4
; B16B16-NEXT:    // kill: def $d0 killed $d0 def $z0
; B16B16-NEXT:    // kill: def $d1 killed $d1 def $z1
; B16B16-NEXT:    bfmax z0.h, p0/m, z0.h, z1.h
; B16B16-NEXT:    // kill: def $d0 killed $d0 killed $z0
; B16B16-NEXT:    ret
  %res = call <4 x bfloat> @llvm.maximum.v4bf16(<4 x bfloat> %a, <4 x bfloat> %b)
  ret <4 x bfloat> %res
}

define <8 x bfloat> @fmax_v8bf16(<8 x bfloat> %a, <8 x bfloat> %b) {
; NOB16B16-LABEL: fmax_v8bf16:
; NOB16B16:       // %bb.0:
; NOB16B16-NEXT:    mov h2, v1.h[1]
; NOB16B16-NEXT:    mov h3, v0.h[1]
; NOB16B16-NEXT:    shll v4.4s, v1.4h, #16
; NOB16B16-NEXT:    shll v5.4s, v0.4h, #16
; NOB16B16-NEXT:    mov h6, v1.h[2]
; NOB16B16-NEXT:    mov h7, v0.h[2]
; NOB16B16-NEXT:    mov h16, v1.h[3]
; NOB16B16-NEXT:    shll v2.4s, v2.4h, #16
; NOB16B16-NEXT:    shll v3.4s, v3.4h, #16
; NOB16B16-NEXT:    fmax s4, s5, s4
; NOB16B16-NEXT:    mov h5, v0.h[3]
; NOB16B16-NEXT:    shll v6.4s, v6.4h, #16
; NOB16B16-NEXT:    shll v7.4s, v7.4h, #16
; NOB16B16-NEXT:    fmax s3, s3, s2
; NOB16B16-NEXT:    bfcvt h2, s4
; NOB16B16-NEXT:    fmax s4, s7, s6
; NOB16B16-NEXT:    shll v6.4s, v16.4h, #16
; NOB16B16-NEXT:    shll v5.4s, v5.4h, #16
; NOB16B16-NEXT:    mov h7, v1.h[4]
; NOB16B16-NEXT:    mov h16, v0.h[4]
; NOB16B16-NEXT:    bfcvt h3, s3
; NOB16B16-NEXT:    fmax s5, s5, s6
; NOB16B16-NEXT:    bfcvt h4, s4
; NOB16B16-NEXT:    mov h6, v0.h[5]
; NOB16B16-NEXT:    shll v7.4s, v7.4h, #16
; NOB16B16-NEXT:    shll v16.4s, v16.4h, #16
; NOB16B16-NEXT:    mov v2.h[1], v3.h[0]
; NOB16B16-NEXT:    mov h3, v1.h[5]
; NOB16B16-NEXT:    bfcvt h5, s5
; NOB16B16-NEXT:    fmax s7, s16, s7
; NOB16B16-NEXT:    mov h16, v0.h[6]
; NOB16B16-NEXT:    shll v6.4s, v6.4h, #16
; NOB16B16-NEXT:    mov h0, v0.h[7]
; NOB16B16-NEXT:    mov v2.h[2], v4.h[0]
; NOB16B16-NEXT:    mov h4, v1.h[6]
; NOB16B16-NEXT:    shll v3.4s, v3.4h, #16
; NOB16B16-NEXT:    mov h1, v1.h[7]
; NOB16B16-NEXT:    shll v0.4s, v0.4h, #16
; NOB16B16-NEXT:    fmax s3, s6, s3
; NOB16B16-NEXT:    shll v6.4s, v16.4h, #16
; NOB16B16-NEXT:    mov v2.h[3], v5.h[0]
; NOB16B16-NEXT:    bfcvt h5, s7
; NOB16B16-NEXT:    shll v4.4s, v4.4h, #16
; NOB16B16-NEXT:    shll v1.4s, v1.4h, #16
; NOB16B16-NEXT:    fmax s4, s6, s4
; NOB16B16-NEXT:    bfcvt h3, s3
; NOB16B16-NEXT:    mov v2.h[4], v5.h[0]
; NOB16B16-NEXT:    fmax s0, s0, s1
; NOB16B16-NEXT:    mov v2.h[5], v3.h[0]
; NOB16B16-NEXT:    bfcvt h3, s4
; NOB16B16-NEXT:    bfcvt h0, s0
; NOB16B16-NEXT:    mov v2.h[6], v3.h[0]
; NOB16B16-NEXT:    mov v2.h[7], v0.h[0]
; NOB16B16-NEXT:    mov v0.16b, v2.16b
; NOB16B16-NEXT:    ret
;
; B16B16-LABEL: fmax_v8bf16:
; B16B16:       // %bb.0:
; B16B16-NEXT:    ptrue p0.h, vl8
; B16B16-NEXT:    // kill: def $q0 killed $q0 def $z0
; B16B16-NEXT:    // kill: def $q1 killed $q1 def $z1
; B16B16-NEXT:    bfmax z0.h, p0/m, z0.h, z1.h
; B16B16-NEXT:    // kill: def $q0 killed $q0 killed $z0
; B16B16-NEXT:    ret
  %res = call <8 x bfloat> @llvm.maximum.v8bf16(<8 x bfloat> %a, <8 x bfloat> %b)
  ret <8 x bfloat> %res
}

;
; FMAXNM
;

define <4 x bfloat> @fmaxnm_v4bf16(<4 x bfloat> %a, <4 x bfloat> %b) {
; NOB16B16-LABEL: fmaxnm_v4bf16:
; NOB16B16:       // %bb.0:
; NOB16B16-NEXT:    // kill: def $d1 killed $d1 def $q1
; NOB16B16-NEXT:    // kill: def $d0 killed $d0 def $q0
; NOB16B16-NEXT:    mov h2, v1.h[1]
; NOB16B16-NEXT:    mov h3, v0.h[1]
; NOB16B16-NEXT:    mov h4, v1.h[2]
; NOB16B16-NEXT:    shll v5.4s, v1.4h, #16
; NOB16B16-NEXT:    shll v6.4s, v0.4h, #16
; NOB16B16-NEXT:    mov h7, v0.h[2]
; NOB16B16-NEXT:    mov h1, v1.h[3]
; NOB16B16-NEXT:    shll v2.4s, v2.4h, #16
; NOB16B16-NEXT:    shll v3.4s, v3.4h, #16
; NOB16B16-NEXT:    shll v4.4s, v4.4h, #16
; NOB16B16-NEXT:    shll v1.4s, v1.4h, #16
; NOB16B16-NEXT:    fmaxnm s2, s3, s2
; NOB16B16-NEXT:    fmaxnm s3, s6, s5
; NOB16B16-NEXT:    shll v5.4s, v7.4h, #16
; NOB16B16-NEXT:    mov h6, v0.h[3]
; NOB16B16-NEXT:    fmaxnm s4, s5, s4
; NOB16B16-NEXT:    bfcvt h2, s2
; NOB16B16-NEXT:    bfcvt h0, s3
; NOB16B16-NEXT:    shll v3.4s, v6.4h, #16
; NOB16B16-NEXT:    mov v0.h[1], v2.h[0]
; NOB16B16-NEXT:    bfcvt h2, s4
; NOB16B16-NEXT:    fmaxnm s1, s3, s1
; NOB16B16-NEXT:    mov v0.h[2], v2.h[0]
; NOB16B16-NEXT:    bfcvt h1, s1
; NOB16B16-NEXT:    mov v0.h[3], v1.h[0]
; NOB16B16-NEXT:    // kill: def $d0 killed $d0 killed $q0
; NOB16B16-NEXT:    ret
;
; B16B16-LABEL: fmaxnm_v4bf16:
; B16B16:       // %bb.0:
; B16B16-NEXT:    ptrue p0.h, vl4
; B16B16-NEXT:    // kill: def $d0 killed $d0 def $z0
; B16B16-NEXT:    // kill: def $d1 killed $d1 def $z1
; B16B16-NEXT:    bfmaxnm z0.h, p0/m, z0.h, z1.h
; B16B16-NEXT:    // kill: def $d0 killed $d0 killed $z0
; B16B16-NEXT:    ret
  %res = call <4 x bfloat> @llvm.maxnum.v4bf16(<4 x bfloat> %a, <4 x bfloat> %b)
  ret <4 x bfloat> %res
}

define <8 x bfloat> @fmaxnm_v8bf16(<8 x bfloat> %a, <8 x bfloat> %b) {
; NOB16B16-LABEL: fmaxnm_v8bf16:
; NOB16B16:       // %bb.0:
; NOB16B16-NEXT:    mov h2, v1.h[1]
; NOB16B16-NEXT:    mov h3, v0.h[1]
; NOB16B16-NEXT:    shll v4.4s, v1.4h, #16
; NOB16B16-NEXT:    shll v5.4s, v0.4h, #16
; NOB16B16-NEXT:    mov h6, v1.h[2]
; NOB16B16-NEXT:    mov h7, v0.h[2]
; NOB16B16-NEXT:    mov h16, v1.h[3]
; NOB16B16-NEXT:    shll v2.4s, v2.4h, #16
; NOB16B16-NEXT:    shll v3.4s, v3.4h, #16
; NOB16B16-NEXT:    fmaxnm s4, s5, s4
; NOB16B16-NEXT:    mov h5, v0.h[3]
; NOB16B16-NEXT:    shll v6.4s, v6.4h, #16
; NOB16B16-NEXT:    shll v7.4s, v7.4h, #16
; NOB16B16-NEXT:    fmaxnm s3, s3, s2
; NOB16B16-NEXT:    bfcvt h2, s4
; NOB16B16-NEXT:    fmaxnm s4, s7, s6
; NOB16B16-NEXT:    shll v6.4s, v16.4h, #16
; NOB16B16-NEXT:    shll v5.4s, v5.4h, #16
; NOB16B16-NEXT:    mov h7, v1.h[4]
; NOB16B16-NEXT:    mov h16, v0.h[4]
; NOB16B16-NEXT:    bfcvt h3, s3
; NOB16B16-NEXT:    fmaxnm s5, s5, s6
; NOB16B16-NEXT:    bfcvt h4, s4
; NOB16B16-NEXT:    mov h6, v0.h[5]
; NOB16B16-NEXT:    shll v7.4s, v7.4h, #16
; NOB16B16-NEXT:    shll v16.4s, v16.4h, #16
; NOB16B16-NEXT:    mov v2.h[1], v3.h[0]
; NOB16B16-NEXT:    mov h3, v1.h[5]
; NOB16B16-NEXT:    bfcvt h5, s5
; NOB16B16-NEXT:    fmaxnm s7, s16, s7
; NOB16B16-NEXT:    mov h16, v0.h[6]
; NOB16B16-NEXT:    shll v6.4s, v6.4h, #16
; NOB16B16-NEXT:    mov h0, v0.h[7]
; NOB16B16-NEXT:    mov v2.h[2], v4.h[0]
; NOB16B16-NEXT:    mov h4, v1.h[6]
; NOB16B16-NEXT:    shll v3.4s, v3.4h, #16
; NOB16B16-NEXT:    mov h1, v1.h[7]
; NOB16B16-NEXT:    shll v0.4s, v0.4h, #16
; NOB16B16-NEXT:    fmaxnm s3, s6, s3
; NOB16B16-NEXT:    shll v6.4s, v16.4h, #16
; NOB16B16-NEXT:    mov v2.h[3], v5.h[0]
; NOB16B16-NEXT:    bfcvt h5, s7
; NOB16B16-NEXT:    shll v4.4s, v4.4h, #16
; NOB16B16-NEXT:    shll v1.4s, v1.4h, #16
; NOB16B16-NEXT:    fmaxnm s4, s6, s4
; NOB16B16-NEXT:    bfcvt h3, s3
; NOB16B16-NEXT:    mov v2.h[4], v5.h[0]
; NOB16B16-NEXT:    fmaxnm s0, s0, s1
; NOB16B16-NEXT:    mov v2.h[5], v3.h[0]
; NOB16B16-NEXT:    bfcvt h3, s4
; NOB16B16-NEXT:    bfcvt h0, s0
; NOB16B16-NEXT:    mov v2.h[6], v3.h[0]
; NOB16B16-NEXT:    mov v2.h[7], v0.h[0]
; NOB16B16-NEXT:    mov v0.16b, v2.16b
; NOB16B16-NEXT:    ret
;
; B16B16-LABEL: fmaxnm_v8bf16:
; B16B16:       // %bb.0:
; B16B16-NEXT:    ptrue p0.h, vl8
; B16B16-NEXT:    // kill: def $q0 killed $q0 def $z0
; B16B16-NEXT:    // kill: def $q1 killed $q1 def $z1
; B16B16-NEXT:    bfmaxnm z0.h, p0/m, z0.h, z1.h
; B16B16-NEXT:    // kill: def $q0 killed $q0 killed $z0
; B16B16-NEXT:    ret
  %res = call <8 x bfloat> @llvm.maxnum.v8bf16(<8 x bfloat> %a, <8 x bfloat> %b)
  ret <8 x bfloat> %res
}

;
; FMIN
;

define <4 x bfloat> @fmin_v4bf16(<4 x bfloat> %a, <4 x bfloat> %b) {
; NOB16B16-LABEL: fmin_v4bf16:
; NOB16B16:       // %bb.0:
; NOB16B16-NEXT:    // kill: def $d1 killed $d1 def $q1
; NOB16B16-NEXT:    // kill: def $d0 killed $d0 def $q0
; NOB16B16-NEXT:    mov h2, v1.h[1]
; NOB16B16-NEXT:    mov h3, v0.h[1]
; NOB16B16-NEXT:    mov h4, v1.h[2]
; NOB16B16-NEXT:    shll v5.4s, v1.4h, #16
; NOB16B16-NEXT:    shll v6.4s, v0.4h, #16
; NOB16B16-NEXT:    mov h7, v0.h[2]
; NOB16B16-NEXT:    mov h1, v1.h[3]
; NOB16B16-NEXT:    shll v2.4s, v2.4h, #16
; NOB16B16-NEXT:    shll v3.4s, v3.4h, #16
; NOB16B16-NEXT:    shll v4.4s, v4.4h, #16
; NOB16B16-NEXT:    shll v1.4s, v1.4h, #16
; NOB16B16-NEXT:    fmin s2, s3, s2
; NOB16B16-NEXT:    fmin s3, s6, s5
; NOB16B16-NEXT:    shll v5.4s, v7.4h, #16
; NOB16B16-NEXT:    mov h6, v0.h[3]
; NOB16B16-NEXT:    fmin s4, s5, s4
; NOB16B16-NEXT:    bfcvt h2, s2
; NOB16B16-NEXT:    bfcvt h0, s3
; NOB16B16-NEXT:    shll v3.4s, v6.4h, #16
; NOB16B16-NEXT:    mov v0.h[1], v2.h[0]
; NOB16B16-NEXT:    bfcvt h2, s4
; NOB16B16-NEXT:    fmin s1, s3, s1
; NOB16B16-NEXT:    mov v0.h[2], v2.h[0]
; NOB16B16-NEXT:    bfcvt h1, s1
; NOB16B16-NEXT:    mov v0.h[3], v1.h[0]
; NOB16B16-NEXT:    // kill: def $d0 killed $d0 killed $q0
; NOB16B16-NEXT:    ret
;
; B16B16-LABEL: fmin_v4bf16:
; B16B16:       // %bb.0:
; B16B16-NEXT:    ptrue p0.h, vl4
; B16B16-NEXT:    // kill: def $d0 killed $d0 def $z0
; B16B16-NEXT:    // kill: def $d1 killed $d1 def $z1
; B16B16-NEXT:    bfmin z0.h, p0/m, z0.h, z1.h
; B16B16-NEXT:    // kill: def $d0 killed $d0 killed $z0
; B16B16-NEXT:    ret
  %res = call <4 x bfloat> @llvm.minimum.v4bf16(<4 x bfloat> %a, <4 x bfloat> %b)
  ret <4 x bfloat> %res
}

define <8 x bfloat> @fmin_v8bf16(<8 x bfloat> %a, <8 x bfloat> %b) {
; NOB16B16-LABEL: fmin_v8bf16:
; NOB16B16:       // %bb.0:
; NOB16B16-NEXT:    mov h2, v1.h[1]
; NOB16B16-NEXT:    mov h3, v0.h[1]
; NOB16B16-NEXT:    shll v4.4s, v1.4h, #16
; NOB16B16-NEXT:    shll v5.4s, v0.4h, #16
; NOB16B16-NEXT:    mov h6, v1.h[2]
; NOB16B16-NEXT:    mov h7, v0.h[2]
; NOB16B16-NEXT:    mov h16, v1.h[3]
; NOB16B16-NEXT:    shll v2.4s, v2.4h, #16
; NOB16B16-NEXT:    shll v3.4s, v3.4h, #16
; NOB16B16-NEXT:    fmin s4, s5, s4
; NOB16B16-NEXT:    mov h5, v0.h[3]
; NOB16B16-NEXT:    shll v6.4s, v6.4h, #16
; NOB16B16-NEXT:    shll v7.4s, v7.4h, #16
; NOB16B16-NEXT:    fmin s3, s3, s2
; NOB16B16-NEXT:    bfcvt h2, s4
; NOB16B16-NEXT:    fmin s4, s7, s6
; NOB16B16-NEXT:    shll v6.4s, v16.4h, #16
; NOB16B16-NEXT:    shll v5.4s, v5.4h, #16
; NOB16B16-NEXT:    mov h7, v1.h[4]
; NOB16B16-NEXT:    mov h16, v0.h[4]
; NOB16B16-NEXT:    bfcvt h3, s3
; NOB16B16-NEXT:    fmin s5, s5, s6
; NOB16B16-NEXT:    bfcvt h4, s4
; NOB16B16-NEXT:    mov h6, v0.h[5]
; NOB16B16-NEXT:    shll v7.4s, v7.4h, #16
; NOB16B16-NEXT:    shll v16.4s, v16.4h, #16
; NOB16B16-NEXT:    mov v2.h[1], v3.h[0]
; NOB16B16-NEXT:    mov h3, v1.h[5]
; NOB16B16-NEXT:    bfcvt h5, s5
; NOB16B16-NEXT:    fmin s7, s16, s7
; NOB16B16-NEXT:    mov h16, v0.h[6]
; NOB16B16-NEXT:    shll v6.4s, v6.4h, #16
; NOB16B16-NEXT:    mov h0, v0.h[7]
; NOB16B16-NEXT:    mov v2.h[2], v4.h[0]
; NOB16B16-NEXT:    mov h4, v1.h[6]
; NOB16B16-NEXT:    shll v3.4s, v3.4h, #16
; NOB16B16-NEXT:    mov h1, v1.h[7]
; NOB16B16-NEXT:    shll v0.4s, v0.4h, #16
; NOB16B16-NEXT:    fmin s3, s6, s3
; NOB16B16-NEXT:    shll v6.4s, v16.4h, #16
; NOB16B16-NEXT:    mov v2.h[3], v5.h[0]
; NOB16B16-NEXT:    bfcvt h5, s7
; NOB16B16-NEXT:    shll v4.4s, v4.4h, #16
; NOB16B16-NEXT:    shll v1.4s, v1.4h, #16
; NOB16B16-NEXT:    fmin s4, s6, s4
; NOB16B16-NEXT:    bfcvt h3, s3
; NOB16B16-NEXT:    mov v2.h[4], v5.h[0]
; NOB16B16-NEXT:    fmin s0, s0, s1
; NOB16B16-NEXT:    mov v2.h[5], v3.h[0]
; NOB16B16-NEXT:    bfcvt h3, s4
; NOB16B16-NEXT:    bfcvt h0, s0
; NOB16B16-NEXT:    mov v2.h[6], v3.h[0]
; NOB16B16-NEXT:    mov v2.h[7], v0.h[0]
; NOB16B16-NEXT:    mov v0.16b, v2.16b
; NOB16B16-NEXT:    ret
;
; B16B16-LABEL: fmin_v8bf16:
; B16B16:       // %bb.0:
; B16B16-NEXT:    ptrue p0.h, vl8
; B16B16-NEXT:    // kill: def $q0 killed $q0 def $z0
; B16B16-NEXT:    // kill: def $q1 killed $q1 def $z1
; B16B16-NEXT:    bfmin z0.h, p0/m, z0.h, z1.h
; B16B16-NEXT:    // kill: def $q0 killed $q0 killed $z0
; B16B16-NEXT:    ret
  %res = call <8 x bfloat> @llvm.minimum.v8bf16(<8 x bfloat> %a, <8 x bfloat> %b)
  ret <8 x bfloat> %res
}

;
; FMINNM
;

define <4 x bfloat> @fminnm_v4bf16(<4 x bfloat> %a, <4 x bfloat> %b) {
; NOB16B16-LABEL: fminnm_v4bf16:
; NOB16B16:       // %bb.0:
; NOB16B16-NEXT:    // kill: def $d1 killed $d1 def $q1
; NOB16B16-NEXT:    // kill: def $d0 killed $d0 def $q0
; NOB16B16-NEXT:    mov h2, v1.h[1]
; NOB16B16-NEXT:    mov h3, v0.h[1]
; NOB16B16-NEXT:    mov h4, v1.h[2]
; NOB16B16-NEXT:    shll v5.4s, v1.4h, #16
; NOB16B16-NEXT:    shll v6.4s, v0.4h, #16
; NOB16B16-NEXT:    mov h7, v0.h[2]
; NOB16B16-NEXT:    mov h1, v1.h[3]
; NOB16B16-NEXT:    shll v2.4s, v2.4h, #16
; NOB16B16-NEXT:    shll v3.4s, v3.4h, #16
; NOB16B16-NEXT:    shll v4.4s, v4.4h, #16
; NOB16B16-NEXT:    shll v1.4s, v1.4h, #16
; NOB16B16-NEXT:    fminnm s2, s3, s2
; NOB16B16-NEXT:    fminnm s3, s6, s5
; NOB16B16-NEXT:    shll v5.4s, v7.4h, #16
; NOB16B16-NEXT:    mov h6, v0.h[3]
; NOB16B16-NEXT:    fminnm s4, s5, s4
; NOB16B16-NEXT:    bfcvt h2, s2
; NOB16B16-NEXT:    bfcvt h0, s3
; NOB16B16-NEXT:    shll v3.4s, v6.4h, #16
; NOB16B16-NEXT:    mov v0.h[1], v2.h[0]
; NOB16B16-NEXT:    bfcvt h2, s4
; NOB16B16-NEXT:    fminnm s1, s3, s1
; NOB16B16-NEXT:    mov v0.h[2], v2.h[0]
; NOB16B16-NEXT:    bfcvt h1, s1
; NOB16B16-NEXT:    mov v0.h[3], v1.h[0]
; NOB16B16-NEXT:    // kill: def $d0 killed $d0 killed $q0
; NOB16B16-NEXT:    ret
;
; B16B16-LABEL: fminnm_v4bf16:
; B16B16:       // %bb.0:
; B16B16-NEXT:    ptrue p0.h, vl4
; B16B16-NEXT:    // kill: def $d0 killed $d0 def $z0
; B16B16-NEXT:    // kill: def $d1 killed $d1 def $z1
; B16B16-NEXT:    bfminnm z0.h, p0/m, z0.h, z1.h
; B16B16-NEXT:    // kill: def $d0 killed $d0 killed $z0
; B16B16-NEXT:    ret
  %res = call <4 x bfloat> @llvm.minnum.v4bf16(<4 x bfloat> %a, <4 x bfloat> %b)
  ret <4 x bfloat> %res
}

define <8 x bfloat> @fminnm_v8bf16(<8 x bfloat> %a, <8 x bfloat> %b) {
; NOB16B16-LABEL: fminnm_v8bf16:
; NOB16B16:       // %bb.0:
; NOB16B16-NEXT:    mov h2, v1.h[1]
; NOB16B16-NEXT:    mov h3, v0.h[1]
; NOB16B16-NEXT:    shll v4.4s, v1.4h, #16
; NOB16B16-NEXT:    shll v5.4s, v0.4h, #16
; NOB16B16-NEXT:    mov h6, v1.h[2]
; NOB16B16-NEXT:    mov h7, v0.h[2]
; NOB16B16-NEXT:    mov h16, v1.h[3]
; NOB16B16-NEXT:    shll v2.4s, v2.4h, #16
; NOB16B16-NEXT:    shll v3.4s, v3.4h, #16
; NOB16B16-NEXT:    fminnm s4, s5, s4
; NOB16B16-NEXT:    mov h5, v0.h[3]
; NOB16B16-NEXT:    shll v6.4s, v6.4h, #16
; NOB16B16-NEXT:    shll v7.4s, v7.4h, #16
; NOB16B16-NEXT:    fminnm s3, s3, s2
; NOB16B16-NEXT:    bfcvt h2, s4
; NOB16B16-NEXT:    fminnm s4, s7, s6
; NOB16B16-NEXT:    shll v6.4s, v16.4h, #16
; NOB16B16-NEXT:    shll v5.4s, v5.4h, #16
; NOB16B16-NEXT:    mov h7, v1.h[4]
; NOB16B16-NEXT:    mov h16, v0.h[4]
; NOB16B16-NEXT:    bfcvt h3, s3
; NOB16B16-NEXT:    fminnm s5, s5, s6
; NOB16B16-NEXT:    bfcvt h4, s4
; NOB16B16-NEXT:    mov h6, v0.h[5]
; NOB16B16-NEXT:    shll v7.4s, v7.4h, #16
; NOB16B16-NEXT:    shll v16.4s, v16.4h, #16
; NOB16B16-NEXT:    mov v2.h[1], v3.h[0]
; NOB16B16-NEXT:    mov h3, v1.h[5]
; NOB16B16-NEXT:    bfcvt h5, s5
; NOB16B16-NEXT:    fminnm s7, s16, s7
; NOB16B16-NEXT:    mov h16, v0.h[6]
; NOB16B16-NEXT:    shll v6.4s, v6.4h, #16
; NOB16B16-NEXT:    mov h0, v0.h[7]
; NOB16B16-NEXT:    mov v2.h[2], v4.h[0]
; NOB16B16-NEXT:    mov h4, v1.h[6]
; NOB16B16-NEXT:    shll v3.4s, v3.4h, #16
; NOB16B16-NEXT:    mov h1, v1.h[7]
; NOB16B16-NEXT:    shll v0.4s, v0.4h, #16
; NOB16B16-NEXT:    fminnm s3, s6, s3
; NOB16B16-NEXT:    shll v6.4s, v16.4h, #16
; NOB16B16-NEXT:    mov v2.h[3], v5.h[0]
; NOB16B16-NEXT:    bfcvt h5, s7
; NOB16B16-NEXT:    shll v4.4s, v4.4h, #16
; NOB16B16-NEXT:    shll v1.4s, v1.4h, #16
; NOB16B16-NEXT:    fminnm s4, s6, s4
; NOB16B16-NEXT:    bfcvt h3, s3
; NOB16B16-NEXT:    mov v2.h[4], v5.h[0]
; NOB16B16-NEXT:    fminnm s0, s0, s1
; NOB16B16-NEXT:    mov v2.h[5], v3.h[0]
; NOB16B16-NEXT:    bfcvt h3, s4
; NOB16B16-NEXT:    bfcvt h0, s0
; NOB16B16-NEXT:    mov v2.h[6], v3.h[0]
; NOB16B16-NEXT:    mov v2.h[7], v0.h[0]
; NOB16B16-NEXT:    mov v0.16b, v2.16b
; NOB16B16-NEXT:    ret
;
; B16B16-LABEL: fminnm_v8bf16:
; B16B16:       // %bb.0:
; B16B16-NEXT:    ptrue p0.h, vl8
; B16B16-NEXT:    // kill: def $q0 killed $q0 def $z0
; B16B16-NEXT:    // kill: def $q1 killed $q1 def $z1
; B16B16-NEXT:    bfminnm z0.h, p0/m, z0.h, z1.h
; B16B16-NEXT:    // kill: def $q0 killed $q0 killed $z0
; B16B16-NEXT:    ret
  %res = call <8 x bfloat> @llvm.minnum.v8bf16(<8 x bfloat> %a, <8 x bfloat> %b)
  ret <8 x bfloat> %res
}

;
; FMLA
;

define <4 x bfloat> @fmla_v4bf16(<4 x bfloat> %a, <4 x bfloat> %b, <4 x bfloat> %c) {
; NOB16B16-LABEL: fmla_v4bf16:
; NOB16B16:       // %bb.0:
; NOB16B16-NEXT:    // kill: def $d2 killed $d2 def $q2
; NOB16B16-NEXT:    // kill: def $d1 killed $d1 def $q1
; NOB16B16-NEXT:    // kill: def $d0 killed $d0 def $q0
; NOB16B16-NEXT:    mov h3, v2.h[1]
; NOB16B16-NEXT:    mov h4, v1.h[1]
; NOB16B16-NEXT:    mov h5, v0.h[1]
; NOB16B16-NEXT:    shll v6.4s, v2.4h, #16
; NOB16B16-NEXT:    shll v7.4s, v1.4h, #16
; NOB16B16-NEXT:    shll v16.4s, v0.4h, #16
; NOB16B16-NEXT:    mov h17, v2.h[2]
; NOB16B16-NEXT:    mov h18, v1.h[2]
; NOB16B16-NEXT:    mov h19, v0.h[2]
; NOB16B16-NEXT:    mov h2, v2.h[3]
; NOB16B16-NEXT:    mov h1, v1.h[3]
; NOB16B16-NEXT:    fmadd s6, s16, s7, s6
; NOB16B16-NEXT:    shll v3.4s, v3.4h, #16
; NOB16B16-NEXT:    shll v4.4s, v4.4h, #16
; NOB16B16-NEXT:    shll v5.4s, v5.4h, #16
; NOB16B16-NEXT:    mov h16, v0.h[3]
; NOB16B16-NEXT:    shll v7.4s, v19.4h, #16
; NOB16B16-NEXT:    shll v2.4s, v2.4h, #16
; NOB16B16-NEXT:    shll v1.4s, v1.4h, #16
; NOB16B16-NEXT:    fmadd s3, s5, s4, s3
; NOB16B16-NEXT:    shll v4.4s, v17.4h, #16
; NOB16B16-NEXT:    shll v5.4s, v18.4h, #16
; NOB16B16-NEXT:    bfcvt h0, s6
; NOB16B16-NEXT:    fmadd s4, s7, s5, s4
; NOB16B16-NEXT:    shll v5.4s, v16.4h, #16
; NOB16B16-NEXT:    bfcvt h3, s3
; NOB16B16-NEXT:    fmadd s1, s5, s1, s2
; NOB16B16-NEXT:    mov v0.h[1], v3.h[0]
; NOB16B16-NEXT:    bfcvt h3, s4
; NOB16B16-NEXT:    bfcvt h1, s1
; NOB16B16-NEXT:    mov v0.h[2], v3.h[0]
; NOB16B16-NEXT:    mov v0.h[3], v1.h[0]
; NOB16B16-NEXT:    // kill: def $d0 killed $d0 killed $q0
; NOB16B16-NEXT:    ret
;
; B16B16-LABEL: fmla_v4bf16:
; B16B16:       // %bb.0:
; B16B16-NEXT:    ptrue p0.h, vl4
; B16B16-NEXT:    // kill: def $d0 killed $d0 def $z0
; B16B16-NEXT:    // kill: def $d2 killed $d2 def $z2
; B16B16-NEXT:    // kill: def $d1 killed $d1 def $z1
; B16B16-NEXT:    bfmla z0.h, p0/m, z1.h, z2.h
; B16B16-NEXT:    // kill: def $d0 killed $d0 killed $z0
; B16B16-NEXT:    ret
  %res = call <4 x bfloat> @llvm.fma.v4bf16(<4 x bfloat> %a, <4 x bfloat> %b, <4 x bfloat> %c)
  ret <4 x bfloat> %res
}

define <8 x bfloat> @fmla_v8bf16(<8 x bfloat> %a, <8 x bfloat> %b, <8 x bfloat> %c) {
; NOB16B16-LABEL: fmla_v8bf16:
; NOB16B16:       // %bb.0:
; NOB16B16-NEXT:    mov h3, v2.h[1]
; NOB16B16-NEXT:    mov h4, v1.h[1]
; NOB16B16-NEXT:    mov h5, v0.h[1]
; NOB16B16-NEXT:    shll v6.4s, v2.4h, #16
; NOB16B16-NEXT:    shll v7.4s, v1.4h, #16
; NOB16B16-NEXT:    shll v16.4s, v0.4h, #16
; NOB16B16-NEXT:    mov h17, v2.h[2]
; NOB16B16-NEXT:    mov h18, v1.h[2]
; NOB16B16-NEXT:    mov h19, v0.h[2]
; NOB16B16-NEXT:    mov h20, v2.h[3]
; NOB16B16-NEXT:    mov h21, v1.h[3]
; NOB16B16-NEXT:    fmadd s6, s16, s7, s6
; NOB16B16-NEXT:    shll v3.4s, v3.4h, #16
; NOB16B16-NEXT:    shll v4.4s, v4.4h, #16
; NOB16B16-NEXT:    shll v5.4s, v5.4h, #16
; NOB16B16-NEXT:    mov h7, v0.h[3]
; NOB16B16-NEXT:    shll v16.4s, v19.4h, #16
; NOB16B16-NEXT:    mov h19, v0.h[4]
; NOB16B16-NEXT:    fmadd s4, s5, s4, s3
; NOB16B16-NEXT:    shll v3.4s, v17.4h, #16
; NOB16B16-NEXT:    shll v5.4s, v18.4h, #16
; NOB16B16-NEXT:    mov h17, v2.h[4]
; NOB16B16-NEXT:    mov h18, v1.h[4]
; NOB16B16-NEXT:    shll v7.4s, v7.4h, #16
; NOB16B16-NEXT:    shll v19.4s, v19.4h, #16
; NOB16B16-NEXT:    fmadd s5, s16, s5, s3
; NOB16B16-NEXT:    bfcvt h3, s6
; NOB16B16-NEXT:    shll v6.4s, v20.4h, #16
; NOB16B16-NEXT:    bfcvt h4, s4
; NOB16B16-NEXT:    shll v16.4s, v21.4h, #16
; NOB16B16-NEXT:    shll v17.4s, v17.4h, #16
; NOB16B16-NEXT:    shll v18.4s, v18.4h, #16
; NOB16B16-NEXT:    fmadd s6, s7, s16, s6
; NOB16B16-NEXT:    bfcvt h5, s5
; NOB16B16-NEXT:    mov h7, v1.h[5]
; NOB16B16-NEXT:    mov v3.h[1], v4.h[0]
; NOB16B16-NEXT:    mov h4, v2.h[5]
; NOB16B16-NEXT:    mov h16, v0.h[5]
; NOB16B16-NEXT:    fmadd s17, s19, s18, s17
; NOB16B16-NEXT:    mov h18, v2.h[6]
; NOB16B16-NEXT:    mov h19, v1.h[6]
; NOB16B16-NEXT:    mov h2, v2.h[7]
; NOB16B16-NEXT:    mov h1, v1.h[7]
; NOB16B16-NEXT:    bfcvt h6, s6
; NOB16B16-NEXT:    shll v7.4s, v7.4h, #16
; NOB16B16-NEXT:    mov v3.h[2], v5.h[0]
; NOB16B16-NEXT:    mov h5, v0.h[6]
; NOB16B16-NEXT:    shll v4.4s, v4.4h, #16
; NOB16B16-NEXT:    shll v16.4s, v16.4h, #16
; NOB16B16-NEXT:    mov h0, v0.h[7]
; NOB16B16-NEXT:    shll v2.4s, v2.4h, #16
; NOB16B16-NEXT:    shll v1.4s, v1.4h, #16
; NOB16B16-NEXT:    fmadd s4, s16, s7, s4
; NOB16B16-NEXT:    mov v3.h[3], v6.h[0]
; NOB16B16-NEXT:    bfcvt h6, s17
; NOB16B16-NEXT:    shll v7.4s, v18.4h, #16
; NOB16B16-NEXT:    shll v16.4s, v19.4h, #16
; NOB16B16-NEXT:    shll v5.4s, v5.4h, #16
; NOB16B16-NEXT:    shll v0.4s, v0.4h, #16
; NOB16B16-NEXT:    fmadd s5, s5, s16, s7
; NOB16B16-NEXT:    mov v3.h[4], v6.h[0]
; NOB16B16-NEXT:    bfcvt h4, s4
; NOB16B16-NEXT:    fmadd s0, s0, s1, s2
; NOB16B16-NEXT:    mov v3.h[5], v4.h[0]
; NOB16B16-NEXT:    bfcvt h4, s5
; NOB16B16-NEXT:    bfcvt h0, s0
; NOB16B16-NEXT:    mov v3.h[6], v4.h[0]
; NOB16B16-NEXT:    mov v3.h[7], v0.h[0]
; NOB16B16-NEXT:    mov v0.16b, v3.16b
; NOB16B16-NEXT:    ret
;
; B16B16-LABEL: fmla_v8bf16:
; B16B16:       // %bb.0:
; B16B16-NEXT:    ptrue p0.h, vl8
; B16B16-NEXT:    // kill: def $q0 killed $q0 def $z0
; B16B16-NEXT:    // kill: def $q2 killed $q2 def $z2
; B16B16-NEXT:    // kill: def $q1 killed $q1 def $z1
; B16B16-NEXT:    bfmla z0.h, p0/m, z1.h, z2.h
; B16B16-NEXT:    // kill: def $q0 killed $q0 killed $z0
; B16B16-NEXT:    ret
  %res = call <8 x bfloat> @llvm.fma.v8bf16(<8 x bfloat> %a, <8 x bfloat> %b, <8 x bfloat> %c)
  ret <8 x bfloat> %res
}

;
; FMUL
;

define <4 x bfloat> @fmul_v4bf16(<4 x bfloat> %a, <4 x bfloat> %b) {
; NOB16B16-LABEL: fmul_v4bf16:
; NOB16B16:       // %bb.0:
; NOB16B16-NEXT:    shll v1.4s, v1.4h, #16
; NOB16B16-NEXT:    shll v0.4s, v0.4h, #16
; NOB16B16-NEXT:    fmul v0.4s, v0.4s, v1.4s
; NOB16B16-NEXT:    bfcvtn v0.4h, v0.4s
; NOB16B16-NEXT:    ret
;
; B16B16-LABEL: fmul_v4bf16:
; B16B16:       // %bb.0:
; B16B16-NEXT:    ptrue p0.h, vl4
; B16B16-NEXT:    // kill: def $d0 killed $d0 def $z0
; B16B16-NEXT:    // kill: def $d1 killed $d1 def $z1
; B16B16-NEXT:    bfmul z0.h, p0/m, z0.h, z1.h
; B16B16-NEXT:    // kill: def $d0 killed $d0 killed $z0
; B16B16-NEXT:    ret
  %res = fmul <4 x bfloat> %a, %b
  ret <4 x bfloat> %res
}

define <8 x bfloat> @fmul_v8bf16(<8 x bfloat> %a, <8 x bfloat> %b) {
; NOB16B16-LABEL: fmul_v8bf16:
; NOB16B16:       // %bb.0:
; NOB16B16-NEXT:    movi v2.4s, #128, lsl #24
; NOB16B16-NEXT:    movi v3.4s, #128, lsl #24
; NOB16B16-NEXT:    ptrue p0.s, vl4
; NOB16B16-NEXT:    bfmlalb v2.4s, v0.8h, v1.8h
; NOB16B16-NEXT:    bfmlalt v3.4s, v0.8h, v1.8h
; NOB16B16-NEXT:    bfcvt z2.h, p0/m, z2.s
; NOB16B16-NEXT:    bfcvtnt z2.h, p0/m, z3.s
; NOB16B16-NEXT:    mov v0.16b, v2.16b
; NOB16B16-NEXT:    ret
;
; B16B16-LABEL: fmul_v8bf16:
; B16B16:       // %bb.0:
; B16B16-NEXT:    ptrue p0.h, vl8
; B16B16-NEXT:    // kill: def $q0 killed $q0 def $z0
; B16B16-NEXT:    // kill: def $q1 killed $q1 def $z1
; B16B16-NEXT:    bfmul z0.h, p0/m, z0.h, z1.h
; B16B16-NEXT:    // kill: def $q0 killed $q0 killed $z0
; B16B16-NEXT:    ret
  %res = fmul <8 x bfloat> %a, %b
  ret <8 x bfloat> %res
}

; Note: Unlike fmul_v8bf16 this case can use +0.0 (not -0.0) for the
; accumulator (NOB16B16), since fabs does not preserve the zero sign.
define <8 x bfloat> @abs_fmul_v8bf16(<8 x bfloat> %a, <8 x bfloat> %b) {
; NOB16B16-LABEL: abs_fmul_v8bf16:
; NOB16B16:       // %bb.0:
; NOB16B16-NEXT:    movi v2.2d, #0000000000000000
; NOB16B16-NEXT:    movi v3.2d, #0000000000000000
; NOB16B16-NEXT:    ptrue p0.s, vl4
; NOB16B16-NEXT:    bfmlalb v2.4s, v0.8h, v1.8h
; NOB16B16-NEXT:    bfmlalt v3.4s, v0.8h, v1.8h
; NOB16B16-NEXT:    bfcvt z2.h, p0/m, z2.s
; NOB16B16-NEXT:    bfcvtnt z2.h, p0/m, z3.s
; NOB16B16-NEXT:    fabs v0.8h, v2.8h
; NOB16B16-NEXT:    ret
;
; B16B16-LABEL: abs_fmul_v8bf16:
; B16B16:       // %bb.0:
; B16B16-NEXT:    ptrue p0.h, vl8
; B16B16-NEXT:    // kill: def $q0 killed $q0 def $z0
; B16B16-NEXT:    // kill: def $q1 killed $q1 def $z1
; B16B16-NEXT:    bfmul z0.h, p0/m, z0.h, z1.h
; B16B16-NEXT:    fabs v0.8h, v0.8h
; B16B16-NEXT:    ret
  %mul = fmul <8 x bfloat> %a, %b
  %res = call <8 x bfloat> @llvm.fabs.v8bf16(<8 x bfloat> %mul)
  ret <8 x bfloat> %res
}

;
; FNEG
;

define <4 x bfloat> @fneg_v4bf16(<4 x bfloat> %a) {
; CHECK-LABEL: fneg_v4bf16:
; CHECK:       // %bb.0:
; CHECK-NEXT:    fneg v0.4h, v0.4h
; CHECK-NEXT:    ret
  %res = fneg <4 x bfloat> %a
  ret <4 x bfloat> %res
}

define <8 x bfloat> @fneg_v8bf16(<8 x bfloat> %a) {
; CHECK-LABEL: fneg_v8bf16:
; CHECK:       // %bb.0:
; CHECK-NEXT:    fneg v0.8h, v0.8h
; CHECK-NEXT:    ret
  %res = fneg <8 x bfloat> %a
  ret <8 x bfloat> %res
}

;
; FSQRT
;

define <4 x bfloat> @fsqrt_v4bf16(<4 x bfloat> %a) {
; CHECK-LABEL: fsqrt_v4bf16:
; CHECK:       // %bb.0:
; CHECK-NEXT:    // kill: def $d0 killed $d0 def $q0
; CHECK-NEXT:    mov h1, v0.h[1]
; CHECK-NEXT:    shll v2.4s, v0.4h, #16
; CHECK-NEXT:    mov h3, v0.h[2]
; CHECK-NEXT:    mov h0, v0.h[3]
; CHECK-NEXT:    fsqrt s2, s2
; CHECK-NEXT:    shll v1.4s, v1.4h, #16
; CHECK-NEXT:    shll v3.4s, v3.4h, #16
; CHECK-NEXT:    shll v0.4s, v0.4h, #16
; CHECK-NEXT:    fsqrt s1, s1
; CHECK-NEXT:    bfcvt h1, s1
; CHECK-NEXT:    fsqrt s3, s3
; CHECK-NEXT:    fsqrt s4, s0
; CHECK-NEXT:    bfcvt h0, s2
; CHECK-NEXT:    mov v0.h[1], v1.h[0]
; CHECK-NEXT:    bfcvt h1, s3
; CHECK-NEXT:    mov v0.h[2], v1.h[0]
; CHECK-NEXT:    bfcvt h1, s4
; CHECK-NEXT:    mov v0.h[3], v1.h[0]
; CHECK-NEXT:    // kill: def $d0 killed $d0 killed $q0
; CHECK-NEXT:    ret
  %res = call <4 x bfloat> @llvm.sqrt.v4bf16(<4 x bfloat> %a)
  ret <4 x bfloat> %res
}

define <8 x bfloat> @fsqrt_v8bf16(<8 x bfloat> %a) {
; CHECK-LABEL: fsqrt_v8bf16:
; CHECK:       // %bb.0:
; CHECK-NEXT:    mov h1, v0.h[1]
; CHECK-NEXT:    shll v2.4s, v0.4h, #16
; CHECK-NEXT:    mov h3, v0.h[2]
; CHECK-NEXT:    mov h4, v0.h[3]
; CHECK-NEXT:    mov h5, v0.h[4]
; CHECK-NEXT:    mov h6, v0.h[5]
; CHECK-NEXT:    mov h7, v0.h[6]
; CHECK-NEXT:    mov h0, v0.h[7]
; CHECK-NEXT:    fsqrt s2, s2
; CHECK-NEXT:    shll v1.4s, v1.4h, #16
; CHECK-NEXT:    shll v3.4s, v3.4h, #16
; CHECK-NEXT:    shll v4.4s, v4.4h, #16
; CHECK-NEXT:    shll v5.4s, v5.4h, #16
; CHECK-NEXT:    shll v6.4s, v6.4h, #16
; CHECK-NEXT:    shll v7.4s, v7.4h, #16
; CHECK-NEXT:    shll v16.4s, v0.4h, #16
; CHECK-NEXT:    bfcvt h0, s2
; CHECK-NEXT:    fsqrt s1, s1
; CHECK-NEXT:    bfcvt h1, s1
; CHECK-NEXT:    mov v0.h[1], v1.h[0]
; CHECK-NEXT:    fsqrt s3, s3
; CHECK-NEXT:    bfcvt h1, s3
; CHECK-NEXT:    mov v0.h[2], v1.h[0]
; CHECK-NEXT:    fsqrt s4, s4
; CHECK-NEXT:    bfcvt h1, s4
; CHECK-NEXT:    mov v0.h[3], v1.h[0]
; CHECK-NEXT:    fsqrt s5, s5
; CHECK-NEXT:    bfcvt h1, s5
; CHECK-NEXT:    mov v0.h[4], v1.h[0]
; CHECK-NEXT:    fsqrt s6, s6
; CHECK-NEXT:    bfcvt h1, s6
; CHECK-NEXT:    mov v0.h[5], v1.h[0]
; CHECK-NEXT:    fsqrt s7, s7
; CHECK-NEXT:    bfcvt h1, s7
; CHECK-NEXT:    mov v0.h[6], v1.h[0]
; CHECK-NEXT:    fsqrt s2, s16
; CHECK-NEXT:    bfcvt h1, s2
; CHECK-NEXT:    mov v0.h[7], v1.h[0]
; CHECK-NEXT:    ret
  %res = call <8 x bfloat> @llvm.sqrt.v8bf16(<8 x bfloat> %a)
  ret <8 x bfloat> %res
}

;
; FSUB
;

define <4 x bfloat> @fsub_v4bf16(<4 x bfloat> %a, <4 x bfloat> %b) {
; NOB16B16-LABEL: fsub_v4bf16:
; NOB16B16:       // %bb.0:
; NOB16B16-NEXT:    shll v1.4s, v1.4h, #16
; NOB16B16-NEXT:    shll v0.4s, v0.4h, #16
; NOB16B16-NEXT:    fsub v0.4s, v0.4s, v1.4s
; NOB16B16-NEXT:    bfcvtn v0.4h, v0.4s
; NOB16B16-NEXT:    ret
;
; B16B16-LABEL: fsub_v4bf16:
; B16B16:       // %bb.0:
; B16B16-NEXT:    ptrue p0.h, vl4
; B16B16-NEXT:    // kill: def $d0 killed $d0 def $z0
; B16B16-NEXT:    // kill: def $d1 killed $d1 def $z1
; B16B16-NEXT:    bfsub z0.h, p0/m, z0.h, z1.h
; B16B16-NEXT:    // kill: def $d0 killed $d0 killed $z0
; B16B16-NEXT:    ret
  %res = fsub <4 x bfloat> %a, %b
  ret <4 x bfloat> %res
}

define <8 x bfloat> @fsub_v8bf16(<8 x bfloat> %a, <8 x bfloat> %b) {
; NOB16B16-LABEL: fsub_v8bf16:
; NOB16B16:       // %bb.0:
; NOB16B16-NEXT:    shll v2.4s, v1.4h, #16
; NOB16B16-NEXT:    shll v3.4s, v0.4h, #16
; NOB16B16-NEXT:    shll2 v1.4s, v1.8h, #16
; NOB16B16-NEXT:    shll2 v0.4s, v0.8h, #16
; NOB16B16-NEXT:    fsub v2.4s, v3.4s, v2.4s
; NOB16B16-NEXT:    fsub v1.4s, v0.4s, v1.4s
; NOB16B16-NEXT:    bfcvtn v0.4h, v2.4s
; NOB16B16-NEXT:    bfcvtn2 v0.8h, v1.4s
; NOB16B16-NEXT:    ret
;
; B16B16-LABEL: fsub_v8bf16:
; B16B16:       // %bb.0:
; B16B16-NEXT:    ptrue p0.h, vl8
; B16B16-NEXT:    // kill: def $q0 killed $q0 def $z0
; B16B16-NEXT:    // kill: def $q1 killed $q1 def $z1
; B16B16-NEXT:    bfsub z0.h, p0/m, z0.h, z1.h
; B16B16-NEXT:    // kill: def $q0 killed $q0 killed $z0
; B16B16-NEXT:    ret
  %res = fsub <8 x bfloat> %a, %b
  ret <8 x bfloat> %res
}

define <4 x float> @partial_reduce_to_v4f32(<4 x float> %acc, <8 x bfloat> %a, <8 x bfloat> %b) {
; CHECK-LABEL: partial_reduce_to_v4f32:
; CHECK:       // %bb.0: // %entry
; CHECK-NEXT:    bfmlalb v0.4s, v1.8h, v2.8h
; CHECK-NEXT:    bfmlalt v0.4s, v1.8h, v2.8h
; CHECK-NEXT:    ret
entry:
  %a.wide = fpext <8 x bfloat> %a to <8 x float>
  %b.wide = fpext <8 x bfloat> %b to <8 x float>
  %mult = fmul <8 x float> %a.wide, %b.wide
  %partial.reduce = call <4 x float> @llvm.vector.partial.reduce.fadd(<4 x float> %acc, <8 x float> %mult)
  ret <4 x float> %partial.reduce
}