; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
; RUN: llc < %s -mtriple=s390x-linux-gnu | FileCheck %s
; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z16 | FileCheck %s --check-prefix=VECTOR
;
; Test some fp16 vector operations, which must be scalarized. With less than
; 8 elements there should only be operations emitted for the used elements.

%Ty0 = type <8 x half>
define void @fun0(ptr %Src, ptr %Dst) {
; CHECK-LABEL: fun0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    stmg %r13, %r15, 104(%r15)
; CHECK-NEXT:    .cfi_offset %r13, -56
; CHECK-NEXT:    .cfi_offset %r14, -48
; CHECK-NEXT:    .cfi_offset %r15, -40
; CHECK-NEXT:    aghi %r15, -288
; CHECK-NEXT:    .cfi_def_cfa_offset 448
; CHECK-NEXT:    std %f8, 280(%r15) # 8-byte Spill
; CHECK-NEXT:    std %f9, 272(%r15) # 8-byte Spill
; CHECK-NEXT:    std %f10, 264(%r15) # 8-byte Spill
; CHECK-NEXT:    std %f11, 256(%r15) # 8-byte Spill
; CHECK-NEXT:    std %f12, 248(%r15) # 8-byte Spill
; CHECK-NEXT:    std %f13, 240(%r15) # 8-byte Spill
; CHECK-NEXT:    std %f14, 232(%r15) # 8-byte Spill
; CHECK-NEXT:    std %f15, 224(%r15) # 8-byte Spill
; CHECK-NEXT:    .cfi_offset %f8, -168
; CHECK-NEXT:    .cfi_offset %f9, -176
; CHECK-NEXT:    .cfi_offset %f10, -184
; CHECK-NEXT:    .cfi_offset %f11, -192
; CHECK-NEXT:    .cfi_offset %f12, -200
; CHECK-NEXT:    .cfi_offset %f13, -208
; CHECK-NEXT:    .cfi_offset %f14, -216
; CHECK-NEXT:    .cfi_offset %f15, -224
; CHECK-NEXT:    lgh %r0, 14(%r2)
; CHECK-NEXT:    lgr %r13, %r3
; CHECK-NEXT:    lgh %r1, 12(%r2)
; CHECK-NEXT:    sllg %r0, %r0, 48
; CHECK-NEXT:    stg %r0, 216(%r15) # 8-byte Spill
; CHECK-NEXT:    lgh %r0, 10(%r2)
; CHECK-NEXT:    sllg %r1, %r1, 48
; CHECK-NEXT:    stg %r1, 208(%r15) # 8-byte Spill
; CHECK-NEXT:    lgh %r1, 8(%r2)
; CHECK-NEXT:    sllg %r0, %r0, 48
; CHECK-NEXT:    stg %r0, 200(%r15) # 8-byte Spill
; CHECK-NEXT:    lgh %r0, 6(%r2)
; CHECK-NEXT:    sllg %r1, %r1, 48
; CHECK-NEXT:    stg %r1, 192(%r15) # 8-byte Spill
; CHECK-NEXT:    lgh %r1, 4(%r2)
; CHECK-NEXT:    sllg %r0, %r0, 48
; CHECK-NEXT:    stg %r0, 176(%r15) # 8-byte Spill
; CHECK-NEXT:    lgh %r0, 2(%r2)
; CHECK-NEXT:    sllg %r1, %r1, 48
; CHECK-NEXT:    stg %r1, 160(%r15) # 8-byte Spill
; CHECK-NEXT:    lgh %r1, 0(%r2)
; CHECK-NEXT:    sllg %r0, %r0, 48
; CHECK-NEXT:    ldgr %f8, %r0
; CHECK-NEXT:    lgh %r0, 30(%r2)
; CHECK-NEXT:    sllg %r1, %r1, 48
; CHECK-NEXT:    ldgr %f13, %r1
; CHECK-NEXT:    lgh %r1, 28(%r2)
; CHECK-NEXT:    sllg %r0, %r0, 48
; CHECK-NEXT:    stg %r0, 184(%r15) # 8-byte Spill
; CHECK-NEXT:    lgh %r0, 26(%r2)
; CHECK-NEXT:    sllg %r1, %r1, 48
; CHECK-NEXT:    stg %r1, 168(%r15) # 8-byte Spill
; CHECK-NEXT:    lgh %r1, 24(%r2)
; CHECK-NEXT:    sllg %r0, %r0, 48
; CHECK-NEXT:    lgh %r3, 22(%r2)
; CHECK-NEXT:    ldgr %f10, %r0
; CHECK-NEXT:    sllg %r0, %r1, 48
; CHECK-NEXT:    ldgr %f11, %r0
; CHECK-NEXT:    sllg %r0, %r3, 48
; CHECK-NEXT:    lgh %r1, 20(%r2)
; CHECK-NEXT:    ldgr %f12, %r0
; CHECK-NEXT:    lgh %r0, 18(%r2)
; CHECK-NEXT:    lgh %r2, 16(%r2)
; CHECK-NEXT:    sllg %r1, %r1, 48
; CHECK-NEXT:    ldgr %f14, %r1
; CHECK-NEXT:    sllg %r0, %r0, 48
; CHECK-NEXT:    sllg %r1, %r2, 48
; CHECK-NEXT:    ldgr %f0, %r1
; CHECK-NEXT:    ldgr %f15, %r0
; CHECK-NEXT:    # kill: def $f0h killed $f0h killed $f0d
; CHECK-NEXT:    brasl %r14, __extendhfsf2@PLT
; CHECK-NEXT:    ler %f9, %f0
; CHECK-NEXT:    ler %f0, %f13
; CHECK-NEXT:    brasl %r14, __extendhfsf2@PLT
; CHECK-NEXT:    aebr %f0, %f9
; CHECK-NEXT:    brasl %r14, __truncsfhf2@PLT
; CHECK-NEXT:    ler %f13, %f0
; CHECK-NEXT:    ler %f0, %f15
; CHECK-NEXT:    brasl %r14, __extendhfsf2@PLT
; CHECK-NEXT:    ler %f9, %f0
; CHECK-NEXT:    ler %f0, %f8
; CHECK-NEXT:    brasl %r14, __extendhfsf2@PLT
; CHECK-NEXT:    aebr %f0, %f9
; CHECK-NEXT:    brasl %r14, __truncsfhf2@PLT
; CHECK-NEXT:    ler %f8, %f0
; CHECK-NEXT:    ler %f0, %f14
; CHECK-NEXT:    brasl %r14, __extendhfsf2@PLT
; CHECK-NEXT:    ler %f9, %f0
; CHECK-NEXT:    ld %f0, 160(%r15) # 8-byte Reload
; CHECK-NEXT:    # kill: def $f0h killed $f0h killed $f0d
; CHECK-NEXT:    brasl %r14, __extendhfsf2@PLT
; CHECK-NEXT:    aebr %f0, %f9
; CHECK-NEXT:    brasl %r14, __truncsfhf2@PLT
; CHECK-NEXT:    ler %f9, %f0
; CHECK-NEXT:    ler %f0, %f12
; CHECK-NEXT:    brasl %r14, __extendhfsf2@PLT
; CHECK-NEXT:    ler %f12, %f0
; CHECK-NEXT:    ld %f0, 176(%r15) # 8-byte Reload
; CHECK-NEXT:    # kill: def $f0h killed $f0h killed $f0d
; CHECK-NEXT:    brasl %r14, __extendhfsf2@PLT
; CHECK-NEXT:    aebr %f0, %f12
; CHECK-NEXT:    brasl %r14, __truncsfhf2@PLT
; CHECK-NEXT:    ler %f14, %f0
; CHECK-NEXT:    ler %f0, %f11
; CHECK-NEXT:    brasl %r14, __extendhfsf2@PLT
; CHECK-NEXT:    ler %f11, %f0
; CHECK-NEXT:    ld %f0, 192(%r15) # 8-byte Reload
; CHECK-NEXT:    # kill: def $f0h killed $f0h killed $f0d
; CHECK-NEXT:    brasl %r14, __extendhfsf2@PLT
; CHECK-NEXT:    aebr %f0, %f11
; CHECK-NEXT:    brasl %r14, __truncsfhf2@PLT
; CHECK-NEXT:    ler %f11, %f0
; CHECK-NEXT:    ler %f0, %f10
; CHECK-NEXT:    brasl %r14, __extendhfsf2@PLT
; CHECK-NEXT:    ler %f10, %f0
; CHECK-NEXT:    ld %f0, 200(%r15) # 8-byte Reload
; CHECK-NEXT:    # kill: def $f0h killed $f0h killed $f0d
; CHECK-NEXT:    brasl %r14, __extendhfsf2@PLT
; CHECK-NEXT:    aebr %f0, %f10
; CHECK-NEXT:    brasl %r14, __truncsfhf2@PLT
; CHECK-NEXT:    ler %f10, %f0
; CHECK-NEXT:    ld %f0, 168(%r15) # 8-byte Reload
; CHECK-NEXT:    # kill: def $f0h killed $f0h killed $f0d
; CHECK-NEXT:    brasl %r14, __extendhfsf2@PLT
; CHECK-NEXT:    ler %f12, %f0
; CHECK-NEXT:    ld %f0, 208(%r15) # 8-byte Reload
; CHECK-NEXT:    # kill: def $f0h killed $f0h killed $f0d
; CHECK-NEXT:    brasl %r14, __extendhfsf2@PLT
; CHECK-NEXT:    aebr %f0, %f12
; CHECK-NEXT:    brasl %r14, __truncsfhf2@PLT
; CHECK-NEXT:    ler %f12, %f0
; CHECK-NEXT:    ld %f0, 184(%r15) # 8-byte Reload
; CHECK-NEXT:    # kill: def $f0h killed $f0h killed $f0d
; CHECK-NEXT:    brasl %r14, __extendhfsf2@PLT
; CHECK-NEXT:    ler %f15, %f0
; CHECK-NEXT:    ld %f0, 216(%r15) # 8-byte Reload
; CHECK-NEXT:    # kill: def $f0h killed $f0h killed $f0d
; CHECK-NEXT:    brasl %r14, __extendhfsf2@PLT
; CHECK-NEXT:    aebr %f0, %f15
; CHECK-NEXT:    brasl %r14, __truncsfhf2@PLT
; CHECK-NEXT:    # kill: def $f0h killed $f0h def $f0d
; CHECK-NEXT:    lgdr %r0, %f0
; CHECK-NEXT:    srlg %r0, %r0, 48
; CHECK-NEXT:    sth %r0, 14(%r13)
; CHECK-NEXT:    lgdr %r0, %f12
; CHECK-NEXT:    srlg %r0, %r0, 48
; CHECK-NEXT:    sth %r0, 12(%r13)
; CHECK-NEXT:    lgdr %r0, %f10
; CHECK-NEXT:    srlg %r0, %r0, 48
; CHECK-NEXT:    sth %r0, 10(%r13)
; CHECK-NEXT:    lgdr %r0, %f11
; CHECK-NEXT:    srlg %r0, %r0, 48
; CHECK-NEXT:    sth %r0, 8(%r13)
; CHECK-NEXT:    lgdr %r0, %f14
; CHECK-NEXT:    srlg %r0, %r0, 48
; CHECK-NEXT:    sth %r0, 6(%r13)
; CHECK-NEXT:    lgdr %r0, %f9
; CHECK-NEXT:    srlg %r0, %r0, 48
; CHECK-NEXT:    sth %r0, 4(%r13)
; CHECK-NEXT:    lgdr %r0, %f8
; CHECK-NEXT:    srlg %r0, %r0, 48
; CHECK-NEXT:    sth %r0, 2(%r13)
; CHECK-NEXT:    lgdr %r0, %f13
; CHECK-NEXT:    srlg %r0, %r0, 48
; CHECK-NEXT:    sth %r0, 0(%r13)
; CHECK-NEXT:    ld %f8, 280(%r15) # 8-byte Reload
; CHECK-NEXT:    ld %f9, 272(%r15) # 8-byte Reload
; CHECK-NEXT:    ld %f10, 264(%r15) # 8-byte Reload
; CHECK-NEXT:    ld %f11, 256(%r15) # 8-byte Reload
; CHECK-NEXT:    ld %f12, 248(%r15) # 8-byte Reload
; CHECK-NEXT:    ld %f13, 240(%r15) # 8-byte Reload
; CHECK-NEXT:    ld %f14, 232(%r15) # 8-byte Reload
; CHECK-NEXT:    ld %f15, 224(%r15) # 8-byte Reload
; CHECK-NEXT:    lmg %r13, %r15, 392(%r15)
; CHECK-NEXT:    br %r14
;
; VECTOR-LABEL: fun0:
; VECTOR:       # %bb.0:
; VECTOR-NEXT:    stmg %r13, %r15, 104(%r15)
; VECTOR-NEXT:    .cfi_offset %r13, -56
; VECTOR-NEXT:    .cfi_offset %r14, -48
; VECTOR-NEXT:    .cfi_offset %r15, -40
; VECTOR-NEXT:    aghi %r15, -248
; VECTOR-NEXT:    .cfi_def_cfa_offset 408
; VECTOR-NEXT:    std %f8, 240(%r15) # 8-byte Spill
; VECTOR-NEXT:    .cfi_offset %f8, -168
; VECTOR-NEXT:    vl %v0, 16(%r2), 3
; VECTOR-NEXT:    mvc 160(16,%r15), 0(%r2) # 16-byte Folded Spill
; VECTOR-NEXT:    lgr %r13, %r3
; VECTOR-NEXT:    vst %v0, 176(%r15), 3 # 16-byte Spill
; VECTOR-NEXT:    vreph %v0, %v0, 7
; VECTOR-NEXT:    # kill: def $f0h killed $f0h killed $v0
; VECTOR-NEXT:    brasl %r14, __extendhfsf2@PLT
; VECTOR-NEXT:    ldr %f8, %f0
; VECTOR-NEXT:    vl %v0, 160(%r15), 3 # 16-byte Reload
; VECTOR-NEXT:    vreph %v0, %v0, 7
; VECTOR-NEXT:    # kill: def $f0h killed $f0h killed $v0
; VECTOR-NEXT:    brasl %r14, __extendhfsf2@PLT
; VECTOR-NEXT:    aebr %f0, %f8
; VECTOR-NEXT:    brasl %r14, __truncsfhf2@PLT
; VECTOR-NEXT:    # kill: def $f0h killed $f0h def $v0
; VECTOR-NEXT:    vst %v0, 208(%r15), 3 # 16-byte Spill
; VECTOR-NEXT:    vl %v0, 176(%r15), 3 # 16-byte Reload
; VECTOR-NEXT:    vreph %v0, %v0, 6
; VECTOR-NEXT:    # kill: def $f0h killed $f0h killed $v0
; VECTOR-NEXT:    brasl %r14, __extendhfsf2@PLT
; VECTOR-NEXT:    ldr %f8, %f0
; VECTOR-NEXT:    vl %v0, 160(%r15), 3 # 16-byte Reload
; VECTOR-NEXT:    vreph %v0, %v0, 6
; VECTOR-NEXT:    # kill: def $f0h killed $f0h killed $v0
; VECTOR-NEXT:    brasl %r14, __extendhfsf2@PLT
; VECTOR-NEXT:    aebr %f0, %f8
; VECTOR-NEXT:    brasl %r14, __truncsfhf2@PLT
; VECTOR-NEXT:    vl %v1, 208(%r15), 3 # 16-byte Reload
; VECTOR-NEXT:    # kill: def $f0h killed $f0h def $v0
; VECTOR-NEXT:    vmrhh %v0, %v0, %v1
; VECTOR-NEXT:    vst %v0, 208(%r15), 3 # 16-byte Spill
; VECTOR-NEXT:    vl %v0, 176(%r15), 3 # 16-byte Reload
; VECTOR-NEXT:    vreph %v0, %v0, 5
; VECTOR-NEXT:    # kill: def $f0h killed $f0h killed $v0
; VECTOR-NEXT:    brasl %r14, __extendhfsf2@PLT
; VECTOR-NEXT:    ldr %f8, %f0
; VECTOR-NEXT:    vl %v0, 160(%r15), 3 # 16-byte Reload
; VECTOR-NEXT:    vreph %v0, %v0, 5
; VECTOR-NEXT:    # kill: def $f0h killed $f0h killed $v0
; VECTOR-NEXT:    brasl %r14, __extendhfsf2@PLT
; VECTOR-NEXT:    aebr %f0, %f8
; VECTOR-NEXT:    brasl %r14, __truncsfhf2@PLT
; VECTOR-NEXT:    # kill: def $f0h killed $f0h def $v0
; VECTOR-NEXT:    vst %v0, 192(%r15), 3 # 16-byte Spill
; VECTOR-NEXT:    vl %v0, 176(%r15), 3 # 16-byte Reload
; VECTOR-NEXT:    vreph %v0, %v0, 4
; VECTOR-NEXT:    # kill: def $f0h killed $f0h killed $v0
; VECTOR-NEXT:    brasl %r14, __extendhfsf2@PLT
; VECTOR-NEXT:    ldr %f8, %f0
; VECTOR-NEXT:    vl %v0, 160(%r15), 3 # 16-byte Reload
; VECTOR-NEXT:    vreph %v0, %v0, 4
; VECTOR-NEXT:    # kill: def $f0h killed $f0h killed $v0
; VECTOR-NEXT:    brasl %r14, __extendhfsf2@PLT
; VECTOR-NEXT:    aebr %f0, %f8
; VECTOR-NEXT:    brasl %r14, __truncsfhf2@PLT
; VECTOR-NEXT:    vl %v1, 192(%r15), 3 # 16-byte Reload
; VECTOR-NEXT:    # kill: def $f0h killed $f0h def $v0
; VECTOR-NEXT:    vmrhh %v0, %v0, %v1
; VECTOR-NEXT:    vl %v1, 208(%r15), 3 # 16-byte Reload
; VECTOR-NEXT:    vmrhf %v0, %v0, %v1
; VECTOR-NEXT:    vst %v0, 208(%r15), 3 # 16-byte Spill
; VECTOR-NEXT:    vl %v0, 176(%r15), 3 # 16-byte Reload
; VECTOR-NEXT:    vreph %v0, %v0, 3
; VECTOR-NEXT:    # kill: def $f0h killed $f0h killed $v0
; VECTOR-NEXT:    brasl %r14, __extendhfsf2@PLT
; VECTOR-NEXT:    ldr %f8, %f0
; VECTOR-NEXT:    vl %v0, 160(%r15), 3 # 16-byte Reload
; VECTOR-NEXT:    vreph %v0, %v0, 3
; VECTOR-NEXT:    # kill: def $f0h killed $f0h killed $v0
; VECTOR-NEXT:    brasl %r14, __extendhfsf2@PLT
; VECTOR-NEXT:    aebr %f0, %f8
; VECTOR-NEXT:    brasl %r14, __truncsfhf2@PLT
; VECTOR-NEXT:    # kill: def $f0h killed $f0h def $v0
; VECTOR-NEXT:    vst %v0, 192(%r15), 3 # 16-byte Spill
; VECTOR-NEXT:    vl %v0, 176(%r15), 3 # 16-byte Reload
; VECTOR-NEXT:    vreph %v0, %v0, 2
; VECTOR-NEXT:    # kill: def $f0h killed $f0h killed $v0
; VECTOR-NEXT:    brasl %r14, __extendhfsf2@PLT
; VECTOR-NEXT:    ldr %f8, %f0
; VECTOR-NEXT:    vl %v0, 160(%r15), 3 # 16-byte Reload
; VECTOR-NEXT:    vreph %v0, %v0, 2
; VECTOR-NEXT:    # kill: def $f0h killed $f0h killed $v0
; VECTOR-NEXT:    brasl %r14, __extendhfsf2@PLT
; VECTOR-NEXT:    aebr %f0, %f8
; VECTOR-NEXT:    brasl %r14, __truncsfhf2@PLT
; VECTOR-NEXT:    vl %v1, 192(%r15), 3 # 16-byte Reload
; VECTOR-NEXT:    # kill: def $f0h killed $f0h def $v0
; VECTOR-NEXT:    vmrhh %v0, %v0, %v1
; VECTOR-NEXT:    vst %v0, 192(%r15), 3 # 16-byte Spill
; VECTOR-NEXT:    vl %v0, 176(%r15), 3 # 16-byte Reload
; VECTOR-NEXT:    # kill: def $f0h killed $f0h killed $v0
; VECTOR-NEXT:    brasl %r14, __extendhfsf2@PLT
; VECTOR-NEXT:    ldr %f8, %f0
; VECTOR-NEXT:    vl %v0, 160(%r15), 3 # 16-byte Reload
; VECTOR-NEXT:    # kill: def $f0h killed $f0h killed $v0
; VECTOR-NEXT:    brasl %r14, __extendhfsf2@PLT
; VECTOR-NEXT:    aebr %f0, %f8
; VECTOR-NEXT:    brasl %r14, __truncsfhf2@PLT
; VECTOR-NEXT:    # kill: def $f0h killed $f0h def $v0
; VECTOR-NEXT:    vst %v0, 224(%r15), 3 # 16-byte Spill
; VECTOR-NEXT:    vl %v0, 176(%r15), 3 # 16-byte Reload
; VECTOR-NEXT:    vreph %v0, %v0, 1
; VECTOR-NEXT:    # kill: def $f0h killed $f0h killed $v0
; VECTOR-NEXT:    brasl %r14, __extendhfsf2@PLT
; VECTOR-NEXT:    ldr %f8, %f0
; VECTOR-NEXT:    vl %v0, 160(%r15), 3 # 16-byte Reload
; VECTOR-NEXT:    vreph %v0, %v0, 1
; VECTOR-NEXT:    # kill: def $f0h killed $f0h killed $v0
; VECTOR-NEXT:    brasl %r14, __extendhfsf2@PLT
; VECTOR-NEXT:    aebr %f0, %f8
; VECTOR-NEXT:    brasl %r14, __truncsfhf2@PLT
; VECTOR-NEXT:    vl %v1, 224(%r15), 3 # 16-byte Reload
; VECTOR-NEXT:    # kill: def $f0h killed $f0h def $v0
; VECTOR-NEXT:    vmrhh %v0, %v1, %v0
; VECTOR-NEXT:    vl %v1, 192(%r15), 3 # 16-byte Reload
; VECTOR-NEXT:    vmrhf %v0, %v0, %v1
; VECTOR-NEXT:    vl %v1, 208(%r15), 3 # 16-byte Reload
; VECTOR-NEXT:    ld %f8, 240(%r15) # 8-byte Reload
; VECTOR-NEXT:    vmrhg %v0, %v0, %v1
; VECTOR-NEXT:    vst %v0, 0(%r13), 3
; VECTOR-NEXT:    lmg %r13, %r15, 352(%r15)
; VECTOR-NEXT:    br %r14
  %LHS = load %Ty0, ptr %Src
  %S2 = getelementptr %Ty0, ptr %Src, i32 1
  %RHS = load %Ty0, ptr %S2
  %Res = fadd %Ty0 %LHS, %RHS
  store %Ty0 %Res, ptr %Dst
  ret void
}

%Ty1 = type <4 x half>
define void @fun1(ptr %Src, ptr %Dst) {
; CHECK-LABEL: fun1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    stmg %r13, %r15, 104(%r15)
; CHECK-NEXT:    .cfi_offset %r13, -56
; CHECK-NEXT:    .cfi_offset %r14, -48
; CHECK-NEXT:    .cfi_offset %r15, -40
; CHECK-NEXT:    aghi %r15, -224
; CHECK-NEXT:    .cfi_def_cfa_offset 384
; CHECK-NEXT:    std %f8, 216(%r15) # 8-byte Spill
; CHECK-NEXT:    std %f9, 208(%r15) # 8-byte Spill
; CHECK-NEXT:    std %f10, 200(%r15) # 8-byte Spill
; CHECK-NEXT:    std %f11, 192(%r15) # 8-byte Spill
; CHECK-NEXT:    std %f12, 184(%r15) # 8-byte Spill
; CHECK-NEXT:    std %f13, 176(%r15) # 8-byte Spill
; CHECK-NEXT:    std %f14, 168(%r15) # 8-byte Spill
; CHECK-NEXT:    std %f15, 160(%r15) # 8-byte Spill
; CHECK-NEXT:    .cfi_offset %f8, -168
; CHECK-NEXT:    .cfi_offset %f9, -176
; CHECK-NEXT:    .cfi_offset %f10, -184
; CHECK-NEXT:    .cfi_offset %f11, -192
; CHECK-NEXT:    .cfi_offset %f12, -200
; CHECK-NEXT:    .cfi_offset %f13, -208
; CHECK-NEXT:    .cfi_offset %f14, -216
; CHECK-NEXT:    .cfi_offset %f15, -224
; CHECK-NEXT:    lgh %r0, 6(%r2)
; CHECK-NEXT:    lgr %r13, %r3
; CHECK-NEXT:    lgh %r1, 4(%r2)
; CHECK-NEXT:    sllg %r0, %r0, 48
; CHECK-NEXT:    ldgr %f8, %r0
; CHECK-NEXT:    lgh %r0, 2(%r2)
; CHECK-NEXT:    sllg %r1, %r1, 48
; CHECK-NEXT:    ldgr %f9, %r1
; CHECK-NEXT:    lgh %r1, 0(%r2)
; CHECK-NEXT:    sllg %r0, %r0, 48
; CHECK-NEXT:    lgh %r3, 14(%r2)
; CHECK-NEXT:    ldgr %f12, %r0
; CHECK-NEXT:    sllg %r0, %r1, 48
; CHECK-NEXT:    ldgr %f10, %r0
; CHECK-NEXT:    sllg %r0, %r3, 48
; CHECK-NEXT:    lgh %r1, 12(%r2)
; CHECK-NEXT:    ldgr %f11, %r0
; CHECK-NEXT:    lgh %r0, 10(%r2)
; CHECK-NEXT:    lgh %r2, 8(%r2)
; CHECK-NEXT:    sllg %r1, %r1, 48
; CHECK-NEXT:    ldgr %f13, %r1
; CHECK-NEXT:    sllg %r0, %r0, 48
; CHECK-NEXT:    sllg %r1, %r2, 48
; CHECK-NEXT:    ldgr %f0, %r1
; CHECK-NEXT:    ldgr %f14, %r0
; CHECK-NEXT:    # kill: def $f0h killed $f0h killed $f0d
; CHECK-NEXT:    brasl %r14, __extendhfsf2@PLT
; CHECK-NEXT:    ler %f15, %f0
; CHECK-NEXT:    ler %f0, %f10
; CHECK-NEXT:    brasl %r14, __extendhfsf2@PLT
; CHECK-NEXT:    sebr %f0, %f15
; CHECK-NEXT:    brasl %r14, __truncsfhf2@PLT
; CHECK-NEXT:    ler %f10, %f0
; CHECK-NEXT:    ler %f0, %f14
; CHECK-NEXT:    brasl %r14, __extendhfsf2@PLT
; CHECK-NEXT:    ler %f14, %f0
; CHECK-NEXT:    ler %f0, %f12
; CHECK-NEXT:    brasl %r14, __extendhfsf2@PLT
; CHECK-NEXT:    sebr %f0, %f14
; CHECK-NEXT:    brasl %r14, __truncsfhf2@PLT
; CHECK-NEXT:    ler %f12, %f0
; CHECK-NEXT:    ler %f0, %f13
; CHECK-NEXT:    brasl %r14, __extendhfsf2@PLT
; CHECK-NEXT:    ler %f13, %f0
; CHECK-NEXT:    ler %f0, %f9
; CHECK-NEXT:    brasl %r14, __extendhfsf2@PLT
; CHECK-NEXT:    sebr %f0, %f13
; CHECK-NEXT:    brasl %r14, __truncsfhf2@PLT
; CHECK-NEXT:    ler %f9, %f0
; CHECK-NEXT:    ler %f0, %f11
; CHECK-NEXT:    brasl %r14, __extendhfsf2@PLT
; CHECK-NEXT:    ler %f11, %f0
; CHECK-NEXT:    ler %f0, %f8
; CHECK-NEXT:    brasl %r14, __extendhfsf2@PLT
; CHECK-NEXT:    sebr %f0, %f11
; CHECK-NEXT:    brasl %r14, __truncsfhf2@PLT
; CHECK-NEXT:    # kill: def $f0h killed $f0h def $f0d
; CHECK-NEXT:    lgdr %r0, %f0
; CHECK-NEXT:    srlg %r0, %r0, 48
; CHECK-NEXT:    sth %r0, 6(%r13)
; CHECK-NEXT:    lgdr %r0, %f9
; CHECK-NEXT:    srlg %r0, %r0, 48
; CHECK-NEXT:    sth %r0, 4(%r13)
; CHECK-NEXT:    lgdr %r0, %f12
; CHECK-NEXT:    srlg %r0, %r0, 48
; CHECK-NEXT:    sth %r0, 2(%r13)
; CHECK-NEXT:    lgdr %r0, %f10
; CHECK-NEXT:    srlg %r0, %r0, 48
; CHECK-NEXT:    sth %r0, 0(%r13)
; CHECK-NEXT:    ld %f8, 216(%r15) # 8-byte Reload
; CHECK-NEXT:    ld %f9, 208(%r15) # 8-byte Reload
; CHECK-NEXT:    ld %f10, 200(%r15) # 8-byte Reload
; CHECK-NEXT:    ld %f11, 192(%r15) # 8-byte Reload
; CHECK-NEXT:    ld %f12, 184(%r15) # 8-byte Reload
; CHECK-NEXT:    ld %f13, 176(%r15) # 8-byte Reload
; CHECK-NEXT:    ld %f14, 168(%r15) # 8-byte Reload
; CHECK-NEXT:    ld %f15, 160(%r15) # 8-byte Reload
; CHECK-NEXT:    lmg %r13, %r15, 328(%r15)
; CHECK-NEXT:    br %r14
;
; VECTOR-LABEL: fun1:
; VECTOR:       # %bb.0:
; VECTOR-NEXT:    stmg %r13, %r15, 104(%r15)
; VECTOR-NEXT:    .cfi_offset %r13, -56
; VECTOR-NEXT:    .cfi_offset %r14, -48
; VECTOR-NEXT:    .cfi_offset %r15, -40
; VECTOR-NEXT:    aghi %r15, -224
; VECTOR-NEXT:    .cfi_def_cfa_offset 384
; VECTOR-NEXT:    std %f8, 216(%r15) # 8-byte Spill
; VECTOR-NEXT:    std %f9, 208(%r15) # 8-byte Spill
; VECTOR-NEXT:    std %f10, 200(%r15) # 8-byte Spill
; VECTOR-NEXT:    std %f11, 192(%r15) # 8-byte Spill
; VECTOR-NEXT:    std %f12, 184(%r15) # 8-byte Spill
; VECTOR-NEXT:    std %f13, 176(%r15) # 8-byte Spill
; VECTOR-NEXT:    std %f14, 168(%r15) # 8-byte Spill
; VECTOR-NEXT:    std %f15, 160(%r15) # 8-byte Spill
; VECTOR-NEXT:    .cfi_offset %f8, -168
; VECTOR-NEXT:    .cfi_offset %f9, -176
; VECTOR-NEXT:    .cfi_offset %f10, -184
; VECTOR-NEXT:    .cfi_offset %f11, -192
; VECTOR-NEXT:    .cfi_offset %f12, -200
; VECTOR-NEXT:    .cfi_offset %f13, -208
; VECTOR-NEXT:    .cfi_offset %f14, -216
; VECTOR-NEXT:    .cfi_offset %f15, -224
; VECTOR-NEXT:    vlreph %v0, 8(%r2)
; VECTOR-NEXT:    vlreph %v8, 6(%r2)
; VECTOR-NEXT:    vlreph %v9, 4(%r2)
; VECTOR-NEXT:    vlreph %v10, 2(%r2)
; VECTOR-NEXT:    lgr %r13, %r3
; VECTOR-NEXT:    vlreph %v11, 0(%r2)
; VECTOR-NEXT:    vlreph %v12, 14(%r2)
; VECTOR-NEXT:    vlreph %v13, 12(%r2)
; VECTOR-NEXT:    vlreph %v14, 10(%r2)
; VECTOR-NEXT:    brasl %r14, __extendhfsf2@PLT
; VECTOR-NEXT:    ldr %f15, %f0
; VECTOR-NEXT:    ldr %f0, %f11
; VECTOR-NEXT:    brasl %r14, __extendhfsf2@PLT
; VECTOR-NEXT:    sebr %f0, %f15
; VECTOR-NEXT:    brasl %r14, __truncsfhf2@PLT
; VECTOR-NEXT:    ldr %f11, %f0
; VECTOR-NEXT:    ldr %f0, %f14
; VECTOR-NEXT:    brasl %r14, __extendhfsf2@PLT
; VECTOR-NEXT:    ldr %f14, %f0
; VECTOR-NEXT:    ldr %f0, %f10
; VECTOR-NEXT:    brasl %r14, __extendhfsf2@PLT
; VECTOR-NEXT:    sebr %f0, %f14
; VECTOR-NEXT:    brasl %r14, __truncsfhf2@PLT
; VECTOR-NEXT:    ldr %f10, %f0
; VECTOR-NEXT:    ldr %f0, %f13
; VECTOR-NEXT:    brasl %r14, __extendhfsf2@PLT
; VECTOR-NEXT:    ldr %f13, %f0
; VECTOR-NEXT:    ldr %f0, %f9
; VECTOR-NEXT:    brasl %r14, __extendhfsf2@PLT
; VECTOR-NEXT:    sebr %f0, %f13
; VECTOR-NEXT:    brasl %r14, __truncsfhf2@PLT
; VECTOR-NEXT:    ldr %f9, %f0
; VECTOR-NEXT:    ldr %f0, %f12
; VECTOR-NEXT:    brasl %r14, __extendhfsf2@PLT
; VECTOR-NEXT:    ldr %f12, %f0
; VECTOR-NEXT:    ldr %f0, %f8
; VECTOR-NEXT:    brasl %r14, __extendhfsf2@PLT
; VECTOR-NEXT:    sebr %f0, %f12
; VECTOR-NEXT:    brasl %r14, __truncsfhf2@PLT
; VECTOR-NEXT:    vsteh %v9, 4(%r13), 0
; VECTOR-NEXT:    vsteh %v10, 2(%r13), 0
; VECTOR-NEXT:    vsteh %v11, 0(%r13), 0
; VECTOR-NEXT:    ld %f8, 216(%r15) # 8-byte Reload
; VECTOR-NEXT:    ld %f9, 208(%r15) # 8-byte Reload
; VECTOR-NEXT:    ld %f10, 200(%r15) # 8-byte Reload
; VECTOR-NEXT:    ld %f11, 192(%r15) # 8-byte Reload
; VECTOR-NEXT:    ld %f12, 184(%r15) # 8-byte Reload
; VECTOR-NEXT:    ld %f13, 176(%r15) # 8-byte Reload
; VECTOR-NEXT:    ld %f14, 168(%r15) # 8-byte Reload
; VECTOR-NEXT:    ld %f15, 160(%r15) # 8-byte Reload
; VECTOR-NEXT:    vsteh %v0, 6(%r13), 0
; VECTOR-NEXT:    lmg %r13, %r15, 328(%r15)
; VECTOR-NEXT:    br %r14
  %LHS = load %Ty1, ptr %Src
  %S2 = getelementptr %Ty1, ptr %Src, i32 1
  %RHS = load %Ty1, ptr %S2
  %Res = fsub %Ty1 %LHS, %RHS
  store %Ty1 %Res, ptr %Dst
  ret void
}