; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
; RUN: llc --mtriple=loongarch64 -O3 < %s | FileCheck %s --check-prefix=DEFAULT
; RUN: llc --mtriple=loongarch64 -O3 -loongarch-disable-reloc-sched < %s | FileCheck %s --check-prefix=NOSCHED

@g = external global i32

define i32 @boundary_test(i32 %a, i32 %b) {
; DEFAULT-LABEL: boundary_test:
; DEFAULT:       # %bb.0: # %entry
; DEFAULT-NEXT:    pcalau12i $a2, %got_pc_hi20(g)
; DEFAULT-NEXT:    ld.d $a2, $a2, %got_pc_lo12(g)
; DEFAULT-NEXT:    mul.d $a3, $a0, $a1
; DEFAULT-NEXT:    ld.w $a2, $a2, 0
; DEFAULT-NEXT:    addi.w $a1, $a1, 0
; DEFAULT-NEXT:    addi.w $a0, $a0, 0
; DEFAULT-NEXT:    div.w $a0, $a0, $a1
; DEFAULT-NEXT:    add.d $a0, $a3, $a0
; DEFAULT-NEXT:    add.w $a0, $a0, $a2
; DEFAULT-NEXT:    ret
;
; NOSCHED-LABEL: boundary_test:
; NOSCHED:       # %bb.0: # %entry
; NOSCHED-NEXT:    mul.d $a2, $a0, $a1
; NOSCHED-NEXT:    pcalau12i $a3, %got_pc_hi20(g)
; NOSCHED-NEXT:    ld.d $a3, $a3, %got_pc_lo12(g)
; NOSCHED-NEXT:    ld.w $a3, $a3, 0
; NOSCHED-NEXT:    addi.w $a1, $a1, 0
; NOSCHED-NEXT:    addi.w $a0, $a0, 0
; NOSCHED-NEXT:    div.w $a0, $a0, $a1
; NOSCHED-NEXT:    add.d $a0, $a2, $a0
; NOSCHED-NEXT:    add.w $a0, $a0, $a3
; NOSCHED-NEXT:    ret


entry:
  ; Mul is fast, but without scheduling boundary it might be scheduled after the long latency load address setup
  %x = mul i32 %a, %b

  ; Load global (pcalau12i) - this has target flags so it becomes a boundary with -loongarch-disable-reloc-sched
  %v = load i32, ptr @g

  ; Div is slow
  %y = sdiv i32 %a, %b

  %s1 = add i32 %x, %y
  %s2 = add i32 %s1, %v
  ret i32 %s2
}