llvm/test/CodeGen/SPARC/2011-01-19-DelaySlot.ll

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
;RUN: llc -mtriple=sparc < %s -verify-machineinstrs | FileCheck %s
;RUN: llc -mtriple=sparc -O0 < %s -verify-machineinstrs | FileCheck %s -check-prefix=UNOPT

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

define i32 @test(i32 %a) #0 {
; CHECK-LABEL: test:
; CHECK:       ! %bb.0: ! %entry
; CHECK-NEXT:    save %sp, -96, %sp
; CHECK-NEXT:    call bar
; CHECK-NEXT:    mov %i0, %o0
; CHECK-NEXT:    ret
; CHECK-NEXT:    restore %g0, %o0, %o0
;
; UNOPT-LABEL: test:
; UNOPT:       ! %bb.0: ! %entry
; UNOPT-NEXT:    save %sp, -96, %sp
; UNOPT-NEXT:    call bar
; UNOPT-NEXT:    mov %i0, %o0
; UNOPT-NEXT:    ret
; UNOPT-NEXT:    restore %g0, %o0, %o0
entry:
  %0 = tail call i32 @bar(i32 %a) nounwind
  ret i32 %0
}

define i32 @test_jmpl(ptr nocapture %f, i32 %a, i32 %b) #0 {
; CHECK-LABEL: test_jmpl:
; CHECK:       ! %bb.0: ! %entry
; CHECK-NEXT:    save %sp, -96, %sp
; CHECK-NEXT:    mov %i2, %o1
; CHECK-NEXT:    call %i0
; CHECK-NEXT:    mov %i1, %o0
; CHECK-NEXT:    ret
; CHECK-NEXT:    restore %g0, %o0, %o0
;
; UNOPT-LABEL: test_jmpl:
; UNOPT:       ! %bb.0: ! %entry
; UNOPT-NEXT:    save %sp, -96, %sp
; UNOPT-NEXT:    mov %i2, %o1
; UNOPT-NEXT:    call %i0
; UNOPT-NEXT:    mov %i1, %o0
; UNOPT-NEXT:    ret
; UNOPT-NEXT:    restore %g0, %o0, %o0
entry:
  %0 = tail call i32 %f(i32 %a, i32 %b) nounwind
  ret i32 %0
}

define i32 @test_loop(i32 %a, i32 %b) nounwind readnone {
; CHECK-LABEL: test_loop:
; CHECK:       ! %bb.0: ! %entry
; CHECK-NEXT:    cmp %o1, 1
; CHECK-NEXT:    bl .LBB2_3
; CHECK-NEXT:    nop
; CHECK-NEXT:  ! %bb.1: ! %bb.preheader
; CHECK-NEXT:    mov %g0, %o2
; CHECK-NEXT:    mov %g0, %o3
; CHECK-NEXT:    mov 1, %o4
; CHECK-NEXT:  .LBB2_2: ! %bb
; CHECK-NEXT:    ! =>This Inner Loop Header: Depth=1
; CHECK-NEXT:    andn %o4, %o3, %o5
; CHECK-NEXT:    sll %o2, %o5, %o5
; CHECK-NEXT:    add %o5, %o0, %o0
; CHECK-NEXT:    add %o3, 1, %o3
; CHECK-NEXT:    cmp %o1, %o3
; CHECK-NEXT:    bne .LBB2_2
; CHECK-NEXT:    add %o2, %o1, %o2
; CHECK-NEXT:  .LBB2_3: ! %bb5
; CHECK-NEXT:    retl
; CHECK-NEXT:    nop
;
; UNOPT-LABEL: test_loop:
; UNOPT:       ! %bb.0: ! %entry
; UNOPT-NEXT:    add %sp, -104, %sp
; UNOPT-NEXT:    mov %o1, %o2
; UNOPT-NEXT:    st %o2, [%sp+92] ! 4-byte Folded Spill
; UNOPT-NEXT:    mov %o0, %o1
; UNOPT-NEXT:    mov %g0, %o0
; UNOPT-NEXT:    cmp %o2, 1
; UNOPT-NEXT:    st %o1, [%sp+96] ! 4-byte Folded Spill
; UNOPT-NEXT:    bl .LBB2_2
; UNOPT-NEXT:    st %o0, [%sp+100]
; UNOPT-NEXT:    ba .LBB2_1
; UNOPT-NEXT:    nop
; UNOPT-NEXT:  .LBB2_1: ! %bb
; UNOPT-NEXT:    ! =>This Inner Loop Header: Depth=1
; UNOPT-NEXT:    ld [%sp+100], %o0 ! 4-byte Folded Reload
; UNOPT-NEXT:    ld [%sp+96], %o3 ! 4-byte Folded Reload
; UNOPT-NEXT:    ld [%sp+92], %o2 ! 4-byte Folded Reload
; UNOPT-NEXT:    smul %o0, %o2, %o1
; UNOPT-NEXT:    mov 1, %o4
; UNOPT-NEXT:    andn %o4, %o0, %o4
; UNOPT-NEXT:    sll %o1, %o4, %o1
; UNOPT-NEXT:    add %o1, %o3, %o1
; UNOPT-NEXT:    add %o0, 1, %o0
; UNOPT-NEXT:    cmp %o0, %o2
; UNOPT-NEXT:    st %o1, [%sp+96] ! 4-byte Folded Spill
; UNOPT-NEXT:    bne .LBB2_1
; UNOPT-NEXT:    st %o0, [%sp+100]
; UNOPT-NEXT:    ba .LBB2_2
; UNOPT-NEXT:    nop
; UNOPT-NEXT:  .LBB2_2: ! %bb5
; UNOPT-NEXT:    ld [%sp+96], %o0 ! 4-byte Folded Reload
; UNOPT-NEXT:    retl
; UNOPT-NEXT:    add %sp, 104, %sp
entry:
  %0 = icmp sgt i32 %b, 0
  br i1 %0, label %bb, label %bb5

bb:                                               ; preds = %entry, %bb
  %a_addr.18 = phi i32 [ %a_addr.0, %bb ], [ %a, %entry ]
  %1 = phi i32 [ %3, %bb ], [ 0, %entry ]
  %tmp9 = mul i32 %1, %b
  %2 = and i32 %1, 1
  %tmp = xor i32 %2, 1
  %.pn = shl i32 %tmp9, %tmp
  %a_addr.0 = add i32 %.pn, %a_addr.18
  %3 = add nsw i32 %1, 1
  %exitcond = icmp eq i32 %3, %b
  br i1 %exitcond, label %bb5, label %bb

bb5:                                              ; preds = %bb, %entry
  %a_addr.1.lcssa = phi i32 [ %a, %entry ], [ %a_addr.0, %bb ]
  ret i32 %a_addr.1.lcssa
}

define i32 @test_inlineasm(i32 %a) #0 {
; CHECK-LABEL: test_inlineasm:
; CHECK:       ! %bb.0: ! %entry
; CHECK-NEXT:    save %sp, -96, %sp
; CHECK-NEXT:    mov %i0, %o0
; CHECK-NEXT:    cmp %i0, -1
; CHECK-NEXT:    !APP
; CHECK-NEXT:    sethi 0, %g0
; CHECK-NEXT:    !NO_APP
; CHECK-NEXT:    ble .LBB3_2
; CHECK-NEXT:    nop
; CHECK-NEXT:  ! %bb.1: ! %bb1
; CHECK-NEXT:    call bar
; CHECK-NEXT:    nop
; CHECK-NEXT:    ba .LBB3_3
; CHECK-NEXT:    nop
; CHECK-NEXT:  .LBB3_2: ! %bb
; CHECK-NEXT:    call foo
; CHECK-NEXT:    nop
; CHECK-NEXT:  .LBB3_3: ! %bb
; CHECK-NEXT:    ret
; CHECK-NEXT:    restore %g0, %o0, %o0
;
; UNOPT-LABEL: test_inlineasm:
; UNOPT:       ! %bb.0: ! %entry
; UNOPT-NEXT:    save %sp, -96, %sp
; UNOPT-NEXT:    st %i0, [%fp+-4] ! 4-byte Folded Spill
; UNOPT-NEXT:    !APP
; UNOPT-NEXT:    sethi 0, %g0
; UNOPT-NEXT:    !NO_APP
; UNOPT-NEXT:    cmp %i0, -1
; UNOPT-NEXT:    bg .LBB3_2
; UNOPT-NEXT:    nop
; UNOPT-NEXT:    ba .LBB3_1
; UNOPT-NEXT:    nop
; UNOPT-NEXT:  .LBB3_1: ! %bb
; UNOPT-NEXT:    call foo
; UNOPT-NEXT:    ld [%fp+-4], %o0
; UNOPT-NEXT:    ret
; UNOPT-NEXT:    restore %g0, %o0, %o0
; UNOPT-NEXT:  .LBB3_2: ! %bb1
; UNOPT-NEXT:    call bar
; UNOPT-NEXT:    ld [%fp+-4], %o0
; UNOPT-NEXT:    ret
; UNOPT-NEXT:    restore %g0, %o0, %o0
entry:
  tail call void asm sideeffect "sethi 0, %g0", ""() nounwind
  %0 = icmp slt i32 %a, 0
  br i1 %0, label %bb, label %bb1

bb:                                               ; preds = %entry
  %1 = tail call i32 (...) @foo(i32 %a) nounwind
  ret i32 %1

bb1:                                              ; preds = %entry
  %2 = tail call i32 @bar(i32 %a) nounwind
  ret i32 %2
}

declare i32 @foo(...)

declare i32 @bar(i32)


define i32 @test_implicit_def() #0 {
; CHECK-LABEL: test_implicit_def:
; CHECK:       ! %bb.0: ! %entry
; CHECK-NEXT:    save %sp, -96, %sp
; CHECK-NEXT:    call func
; CHECK-NEXT:    nop
; CHECK-NEXT:    ret
; CHECK-NEXT:    restore %g0, %g0, %o0
;
; UNOPT-LABEL: test_implicit_def:
; UNOPT:       ! %bb.0: ! %entry
; UNOPT-NEXT:    save %sp, -96, %sp
; UNOPT-NEXT:    ! implicit-def: $o0
; UNOPT-NEXT:    call func
; UNOPT-NEXT:    nop
; UNOPT-NEXT:    ret
; UNOPT-NEXT:    restore %g0, %g0, %o0
entry:
  %0 = tail call i32 @func(ptr undef) nounwind
  ret i32 0
}

define i32 @prevent_o7_in_call_delay_slot(i32 %i0) #0 {
; CHECK-LABEL: prevent_o7_in_call_delay_slot:
; CHECK:       ! %bb.0: ! %entry
; CHECK-NEXT:    save %sp, -96, %sp
; CHECK-NEXT:    add %i0, 2, %o5
; CHECK-NEXT:    add %i0, 3, %o7
; CHECK-NEXT:    !APP
; CHECK-NEXT:    !NO_APP
; CHECK-NEXT:    add %o5, %o7, %o0
; CHECK-NEXT:    call bar
; CHECK-NEXT:    nop
; CHECK-NEXT:    ret
; CHECK-NEXT:    restore %g0, %o0, %o0
;
; UNOPT-LABEL: prevent_o7_in_call_delay_slot:
; UNOPT:       ! %bb.0: ! %entry
; UNOPT-NEXT:    save %sp, -104, %sp
; UNOPT-NEXT:    add %i0, 2, %o5
; UNOPT-NEXT:    st %o5, [%fp+-4] ! 4-byte Folded Spill
; UNOPT-NEXT:    add %i0, 3, %o7
; UNOPT-NEXT:    st %o7, [%fp+-8] ! 4-byte Folded Spill
; UNOPT-NEXT:    !APP
; UNOPT-NEXT:    !NO_APP
; UNOPT-NEXT:    ld [%fp+-8], %i1 ! 4-byte Folded Reload
; UNOPT-NEXT:    ld [%fp+-4], %i0 ! 4-byte Folded Reload
; UNOPT-NEXT:    call bar
; UNOPT-NEXT:    add %i0, %i1, %o0
; UNOPT-NEXT:    ret
; UNOPT-NEXT:    restore %g0, %o0, %o0
entry:
  %0 = add nsw i32 %i0, 2
  %1 = add nsw i32 %i0, 3
  tail call void asm sideeffect "", "r,r,~{l0},~{l1},~{l2},~{l3},~{l4},~{l5},~{l6},~{l7},~{i0},~{i1},~{i2},~{i3},~{i4},~{i5},~{i6},~{i7},~{o0},~{o1},~{o2},~{o3},~{o4},~{o6},~{g1},~{g2},~{g3},~{g4},~{g5},~{g6},~{g7}"(i32 %0, i32 %1)
  %2 = add nsw i32 %0, %1
  %3 = tail call i32 @bar(i32 %2)
  ret i32 %3
}


declare i32 @func(ptr)


define i32 @restore_add(i32 %a, i32 %b) {
; CHECK-LABEL: restore_add:
; CHECK:         .cfi_startproc
; CHECK-NEXT:  ! %bb.0: ! %entry
; CHECK-NEXT:    save %sp, -96, %sp
; CHECK-NEXT:    .cfi_def_cfa_register %fp
; CHECK-NEXT:    .cfi_window_save
; CHECK-NEXT:    .cfi_register %o7, %i7
; CHECK-NEXT:    call bar
; CHECK-NEXT:    mov %i0, %o0
; CHECK-NEXT:    ret
; CHECK-NEXT:    restore %o0, %i1, %o0
;
; UNOPT-LABEL: restore_add:
; UNOPT:         .cfi_startproc
; UNOPT-NEXT:  ! %bb.0: ! %entry
; UNOPT-NEXT:    save %sp, -96, %sp
; UNOPT-NEXT:    .cfi_def_cfa_register %fp
; UNOPT-NEXT:    .cfi_window_save
; UNOPT-NEXT:    .cfi_register %o7, %i7
; UNOPT-NEXT:    call bar
; UNOPT-NEXT:    mov %i0, %o0
; UNOPT-NEXT:    ret
; UNOPT-NEXT:    restore %o0, %i1, %o0
entry:
  %0 = tail call i32 @bar(i32 %a) nounwind
  %1 = add nsw i32 %0, %b
  ret i32 %1
}

define i32 @restore_add_imm(i32 %a) {
; CHECK-LABEL: restore_add_imm:
; CHECK:         .cfi_startproc
; CHECK-NEXT:  ! %bb.0: ! %entry
; CHECK-NEXT:    save %sp, -96, %sp
; CHECK-NEXT:    .cfi_def_cfa_register %fp
; CHECK-NEXT:    .cfi_window_save
; CHECK-NEXT:    .cfi_register %o7, %i7
; CHECK-NEXT:    call bar
; CHECK-NEXT:    mov %i0, %o0
; CHECK-NEXT:    ret
; CHECK-NEXT:    restore %o0, 20, %o0
;
; UNOPT-LABEL: restore_add_imm:
; UNOPT:         .cfi_startproc
; UNOPT-NEXT:  ! %bb.0: ! %entry
; UNOPT-NEXT:    save %sp, -96, %sp
; UNOPT-NEXT:    .cfi_def_cfa_register %fp
; UNOPT-NEXT:    .cfi_window_save
; UNOPT-NEXT:    .cfi_register %o7, %i7
; UNOPT-NEXT:    call bar
; UNOPT-NEXT:    mov %i0, %o0
; UNOPT-NEXT:    ret
; UNOPT-NEXT:    restore %o0, 20, %o0
entry:
  %0 = tail call i32 @bar(i32 %a) nounwind
  %1 = add nsw i32 %0, 20
  ret i32 %1
}

define i32 @restore_or(i32 %a) #0 {
; CHECK-LABEL: restore_or:
; CHECK:       ! %bb.0: ! %entry
; CHECK-NEXT:    save %sp, -96, %sp
; CHECK-NEXT:    call bar
; CHECK-NEXT:    mov %i0, %o0
; CHECK-NEXT:    ret
; CHECK-NEXT:    restore %g0, %o0, %o0
;
; UNOPT-LABEL: restore_or:
; UNOPT:       ! %bb.0: ! %entry
; UNOPT-NEXT:    save %sp, -96, %sp
; UNOPT-NEXT:    call bar
; UNOPT-NEXT:    mov %i0, %o0
; UNOPT-NEXT:    ret
; UNOPT-NEXT:    restore %g0, %o0, %o0
entry:
  %0 = tail call i32 @bar(i32 %a) nounwind
  ret i32 %0
}

define i32 @restore_or_imm(i32 %a) {
; CHECK-LABEL: restore_or_imm:
; CHECK:         .cfi_startproc
; CHECK-NEXT:  ! %bb.0: ! %entry
; CHECK-NEXT:    save %sp, -96, %sp
; CHECK-NEXT:    .cfi_def_cfa_register %fp
; CHECK-NEXT:    .cfi_window_save
; CHECK-NEXT:    .cfi_register %o7, %i7
; CHECK-NEXT:    call bar
; CHECK-NEXT:    mov %i0, %o0
; CHECK-NEXT:    or %o0, 20, %i0
; CHECK-NEXT:    ret
; CHECK-NEXT:    restore
;
; UNOPT-LABEL: restore_or_imm:
; UNOPT:         .cfi_startproc
; UNOPT-NEXT:  ! %bb.0: ! %entry
; UNOPT-NEXT:    save %sp, -96, %sp
; UNOPT-NEXT:    .cfi_def_cfa_register %fp
; UNOPT-NEXT:    .cfi_window_save
; UNOPT-NEXT:    .cfi_register %o7, %i7
; UNOPT-NEXT:    call bar
; UNOPT-NEXT:    mov %i0, %o0
; UNOPT-NEXT:    or %o0, 20, %i0
; UNOPT-NEXT:    ret
; UNOPT-NEXT:    restore
entry:
  %0 = tail call i32 @bar(i32 %a) nounwind
  %1 = or i32 %0, 20
  ret i32 %1
}


define i32 @restore_sethi(i32 %a) {
; CHECK-LABEL: restore_sethi:
; CHECK:         .cfi_startproc
; CHECK-NEXT:  ! %bb.0: ! %entry
; CHECK-NEXT:    save %sp, -96, %sp
; CHECK-NEXT:    .cfi_def_cfa_register %fp
; CHECK-NEXT:    .cfi_window_save
; CHECK-NEXT:    .cfi_register %o7, %i7
; CHECK-NEXT:    call bar
; CHECK-NEXT:    mov %i0, %o0
; CHECK-NEXT:    cmp %o0, 0
; CHECK-NEXT:    bne .LBB10_2
; CHECK-NEXT:    nop
; CHECK-NEXT:  ! %bb.1: ! %entry
; CHECK-NEXT:    ret
; CHECK-NEXT:    restore %g0, %g0, %o0
; CHECK-NEXT:  .LBB10_2:
; CHECK-NEXT:    ret
; CHECK-NEXT:    restore %g0, 3072, %o0
;
; UNOPT-LABEL: restore_sethi:
; UNOPT:         .cfi_startproc
; UNOPT-NEXT:  ! %bb.0: ! %entry
; UNOPT-NEXT:    save %sp, -104, %sp
; UNOPT-NEXT:    .cfi_def_cfa_register %fp
; UNOPT-NEXT:    .cfi_window_save
; UNOPT-NEXT:    .cfi_register %o7, %i7
; UNOPT-NEXT:    call bar
; UNOPT-NEXT:    mov %i0, %o0
; UNOPT-NEXT:    mov %g0, %i0
; UNOPT-NEXT:    st %i0, [%fp+-8] ! 4-byte Folded Spill
; UNOPT-NEXT:    sethi 3, %i0
; UNOPT-NEXT:    cmp %o0, 0
; UNOPT-NEXT:    bne .LBB10_2
; UNOPT-NEXT:    st %i0, [%fp+-4]
; UNOPT-NEXT:  ! %bb.1: ! %entry
; UNOPT-NEXT:    ld [%fp+-8], %i0 ! 4-byte Folded Reload
; UNOPT-NEXT:    st %i0, [%fp+-4] ! 4-byte Folded Spill
; UNOPT-NEXT:  .LBB10_2: ! %entry
; UNOPT-NEXT:    ld [%fp+-4], %i0 ! 4-byte Folded Reload
; UNOPT-NEXT:    ret
; UNOPT-NEXT:    restore
entry:
  %0 = tail call i32 @bar(i32 %a) nounwind
  %1 = icmp ne i32 %0, 0
  %2 = select i1 %1, i32 3072, i32 0
  ret i32 %2
}

define i32 @restore_sethi_3bit(i32 %a) {
; CHECK-LABEL: restore_sethi_3bit:
; CHECK:         .cfi_startproc
; CHECK-NEXT:  ! %bb.0: ! %entry
; CHECK-NEXT:    save %sp, -96, %sp
; CHECK-NEXT:    .cfi_def_cfa_register %fp
; CHECK-NEXT:    .cfi_window_save
; CHECK-NEXT:    .cfi_register %o7, %i7
; CHECK-NEXT:    call bar
; CHECK-NEXT:    mov %i0, %o0
; CHECK-NEXT:    cmp %o0, 0
; CHECK-NEXT:    bne .LBB11_2
; CHECK-NEXT:    nop
; CHECK-NEXT:  ! %bb.1: ! %entry
; CHECK-NEXT:    ret
; CHECK-NEXT:    restore %g0, %g0, %o0
; CHECK-NEXT:  .LBB11_2:
; CHECK-NEXT:    sethi 6, %i0
; CHECK-NEXT:    ret
; CHECK-NEXT:    restore
;
; UNOPT-LABEL: restore_sethi_3bit:
; UNOPT:         .cfi_startproc
; UNOPT-NEXT:  ! %bb.0: ! %entry
; UNOPT-NEXT:    save %sp, -104, %sp
; UNOPT-NEXT:    .cfi_def_cfa_register %fp
; UNOPT-NEXT:    .cfi_window_save
; UNOPT-NEXT:    .cfi_register %o7, %i7
; UNOPT-NEXT:    call bar
; UNOPT-NEXT:    mov %i0, %o0
; UNOPT-NEXT:    mov %g0, %i0
; UNOPT-NEXT:    st %i0, [%fp+-8] ! 4-byte Folded Spill
; UNOPT-NEXT:    sethi 6, %i0
; UNOPT-NEXT:    cmp %o0, 0
; UNOPT-NEXT:    bne .LBB11_2
; UNOPT-NEXT:    st %i0, [%fp+-4]
; UNOPT-NEXT:  ! %bb.1: ! %entry
; UNOPT-NEXT:    ld [%fp+-8], %i0 ! 4-byte Folded Reload
; UNOPT-NEXT:    st %i0, [%fp+-4] ! 4-byte Folded Spill
; UNOPT-NEXT:  .LBB11_2: ! %entry
; UNOPT-NEXT:    ld [%fp+-4], %i0 ! 4-byte Folded Reload
; UNOPT-NEXT:    ret
; UNOPT-NEXT:    restore
entry:
  %0 = tail call i32 @bar(i32 %a) nounwind
  %1 = icmp ne i32 %0, 0
  %2 = select i1 %1, i32 6144, i32 0
  ret i32 %2
}

define i32 @restore_sethi_large(i32 %a) {
; CHECK-LABEL: restore_sethi_large:
; CHECK:         .cfi_startproc
; CHECK-NEXT:  ! %bb.0: ! %entry
; CHECK-NEXT:    save %sp, -96, %sp
; CHECK-NEXT:    .cfi_def_cfa_register %fp
; CHECK-NEXT:    .cfi_window_save
; CHECK-NEXT:    .cfi_register %o7, %i7
; CHECK-NEXT:    call bar
; CHECK-NEXT:    mov %i0, %o0
; CHECK-NEXT:    cmp %o0, 0
; CHECK-NEXT:    bne .LBB12_2
; CHECK-NEXT:    nop
; CHECK-NEXT:  ! %bb.1: ! %entry
; CHECK-NEXT:    ret
; CHECK-NEXT:    restore %g0, %g0, %o0
; CHECK-NEXT:  .LBB12_2:
; CHECK-NEXT:    sethi 4000, %i0
; CHECK-NEXT:    ret
; CHECK-NEXT:    restore
;
; UNOPT-LABEL: restore_sethi_large:
; UNOPT:         .cfi_startproc
; UNOPT-NEXT:  ! %bb.0: ! %entry
; UNOPT-NEXT:    save %sp, -104, %sp
; UNOPT-NEXT:    .cfi_def_cfa_register %fp
; UNOPT-NEXT:    .cfi_window_save
; UNOPT-NEXT:    .cfi_register %o7, %i7
; UNOPT-NEXT:    call bar
; UNOPT-NEXT:    mov %i0, %o0
; UNOPT-NEXT:    mov %g0, %i0
; UNOPT-NEXT:    st %i0, [%fp+-8] ! 4-byte Folded Spill
; UNOPT-NEXT:    sethi 4000, %i0
; UNOPT-NEXT:    cmp %o0, 0
; UNOPT-NEXT:    bne .LBB12_2
; UNOPT-NEXT:    st %i0, [%fp+-4]
; UNOPT-NEXT:  ! %bb.1: ! %entry
; UNOPT-NEXT:    ld [%fp+-8], %i0 ! 4-byte Folded Reload
; UNOPT-NEXT:    st %i0, [%fp+-4] ! 4-byte Folded Spill
; UNOPT-NEXT:  .LBB12_2: ! %entry
; UNOPT-NEXT:    ld [%fp+-4], %i0 ! 4-byte Folded Reload
; UNOPT-NEXT:    ret
; UNOPT-NEXT:    restore
entry:
  %0 = tail call i32 @bar(i32 %a) nounwind
  %1 = icmp ne i32 %0, 0
  %2 = select i1 %1, i32 4096000, i32 0
  ret i32 %2
}

define i32 @test_generic_inst(i32 %arg) #0 {
; CHECK-LABEL: test_generic_inst:
; CHECK:       ! %bb.0:
; CHECK-NEXT:    save %sp, -96, %sp
; CHECK-NEXT:    call bar
; CHECK-NEXT:    mov %i0, %o0
; CHECK-NEXT:    andcc %o0, 1, %g0
; CHECK-NEXT:    ! fake_use: $i0
; CHECK-NEXT:    bne .LBB13_2
; CHECK-NEXT:    nop
; CHECK-NEXT:  ! %bb.1: ! %true
; CHECK-NEXT:    call bar
; CHECK-NEXT:    nop
; CHECK-NEXT:    ret
; CHECK-NEXT:    restore %g0, %o0, %o0
; CHECK-NEXT:  .LBB13_2: ! %false
; CHECK-NEXT:    ret
; CHECK-NEXT:    restore %o0, 1, %o0
;
; UNOPT-LABEL: test_generic_inst:
; UNOPT:       ! %bb.0:
; UNOPT-NEXT:    save %sp, -104, %sp
; UNOPT-NEXT:    mov %i0, %i1
; UNOPT-NEXT:    call bar
; UNOPT-NEXT:    mov %i1, %o0
; UNOPT-NEXT:    mov %o0, %i0
; UNOPT-NEXT:    st %i0, [%fp+-4] ! 4-byte Folded Spill
; UNOPT-NEXT:    and %o0, 1, %i0
; UNOPT-NEXT:    ! fake_use: $i1
; UNOPT-NEXT:    cmp %i0, 0
; UNOPT-NEXT:    bne .LBB13_2
; UNOPT-NEXT:    nop
; UNOPT-NEXT:    ba .LBB13_1
; UNOPT-NEXT:    nop
; UNOPT-NEXT:  .LBB13_1: ! %true
; UNOPT-NEXT:    call bar
; UNOPT-NEXT:    ld [%fp+-4], %o0
; UNOPT-NEXT:    ba .LBB13_3
; UNOPT-NEXT:    st %o0, [%fp+-8]
; UNOPT-NEXT:  .LBB13_2: ! %false
; UNOPT-NEXT:    ld [%fp+-4], %i0 ! 4-byte Folded Reload
; UNOPT-NEXT:    add %i0, 1, %i0
; UNOPT-NEXT:    ba .LBB13_3
; UNOPT-NEXT:    st %i0, [%fp+-8]
; UNOPT-NEXT:  .LBB13_3: ! %cont
; UNOPT-NEXT:    ld [%fp+-8], %i0 ! 4-byte Folded Reload
; UNOPT-NEXT:    ret
; UNOPT-NEXT:    restore
  %bar1 = call i32 @bar(i32 %arg)
  %even = and i32 %bar1, 1
  %cmp = icmp eq i32 %even, 0
  ; This shouldn't get reordered into a delay slot
  call void (...) @llvm.fake.use(i32 %arg)
  br i1 %cmp, label %true, label %false
true:
  %bar2 = call i32 @bar(i32 %bar1)
  br label %cont

false:
  %inc = add nsw i32 %bar1, 1
  br label %cont

cont:
  %ret = phi i32 [ %bar2, %true ], [ %inc, %false ]
  ret i32 %ret
}

declare void @llvm.fake.use(...)
attributes #0 = { nounwind "disable-tail-calls"="true" }