llvm/test/CodeGen/AArch64/udiv-const-optimization.ll

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -mtriple=aarch64-unknown-linux | FileCheck %s

; Test optimization of 32-bit unsigned division by constants with 33-bit magic
; constants (IsAdd=true) on AArch64. The optimization uses the umulh instruction.

define i32 @udiv_by_7(i32 %x) nounwind {
; CHECK-LABEL: udiv_by_7:
; CHECK:       // %bb.0:
; CHECK-NEXT:    mov x8, #2684354560 // =0xa0000000
; CHECK-NEXT:    mov w9, w0
; CHECK-NEXT:    movk x8, #18724, lsl #32
; CHECK-NEXT:    movk x8, #9362, lsl #48
; CHECK-NEXT:    umulh x0, x9, x8
; CHECK-NEXT:    // kill: def $w0 killed $w0 killed $x0
; CHECK-NEXT:    ret
  %div = udiv i32 %x, 7
  ret i32 %div
}

define i32 @udiv_by_19(i32 %x) nounwind {
; CHECK-LABEL: udiv_by_19:
; CHECK:       // %bb.0:
; CHECK-NEXT:    mov x8, #1476395008 // =0x58000000
; CHECK-NEXT:    mov w9, w0
; CHECK-NEXT:    movk x8, #17246, lsl #32
; CHECK-NEXT:    movk x8, #3449, lsl #48
; CHECK-NEXT:    umulh x0, x9, x8
; CHECK-NEXT:    // kill: def $w0 killed $w0 killed $x0
; CHECK-NEXT:    ret
  %div = udiv i32 %x, 19
  ret i32 %div
}

define i32 @udiv_by_21(i32 %x) nounwind {
; CHECK-LABEL: udiv_by_21:
; CHECK:       // %bb.0:
; CHECK-NEXT:    mov x8, #939524096 // =0x38000000
; CHECK-NEXT:    mov w9, w0
; CHECK-NEXT:    movk x8, #49932, lsl #32
; CHECK-NEXT:    movk x8, #3120, lsl #48
; CHECK-NEXT:    umulh x0, x9, x8
; CHECK-NEXT:    // kill: def $w0 killed $w0 killed $x0
; CHECK-NEXT:    ret
  %div = udiv i32 %x, 21
  ret i32 %div
}

; Test non-optimized case
define i32 @udiv_by_3(i32 %x) nounwind {
; CHECK-LABEL: udiv_by_3:
; CHECK:       // %bb.0:
; CHECK-NEXT:    mov w8, #43691 // =0xaaab
; CHECK-NEXT:    movk w8, #43690, lsl #16
; CHECK-NEXT:    umull x8, w0, w8
; CHECK-NEXT:    lsr x0, x8, #33
; CHECK-NEXT:    // kill: def $w0 killed $w0 killed $x0
; CHECK-NEXT:    ret
  %div = udiv i32 %x, 3
  ret i32 %div
}

; Even divisors with 33-bit magic. On AArch64 these use umulh (i32->i64
; zero-extension is free).
define i32 @udiv_by_14(i32 %x) nounwind {
; CHECK-LABEL: udiv_by_14:
; CHECK:       // %bb.0:
; CHECK-NEXT:    mov x8, #1342177280 // =0x50000000
; CHECK-NEXT:    mov w9, w0
; CHECK-NEXT:    movk x8, #9362, lsl #32
; CHECK-NEXT:    movk x8, #4681, lsl #48
; CHECK-NEXT:    umulh x0, x9, x8
; CHECK-NEXT:    // kill: def $w0 killed $w0 killed $x0
; CHECK-NEXT:    ret
  %div = udiv i32 %x, 14
  ret i32 %div
}

; Vector udiv is unaffected by the free-zext scalar rewrite: <4 x i32> keeps
; the even-divisor rewrite (this function is unchanged by the follow-up commit).
define <4 x i32> @udiv_by_14_vec(<4 x i32> %x) nounwind {
; CHECK-LABEL: udiv_by_14_vec:
; CHECK:       // %bb.0:
; CHECK-NEXT:    mov w8, #9363 // =0x2493
; CHECK-NEXT:    ushr v0.4s, v0.4s, #1
; CHECK-NEXT:    movk w8, #37449, lsl #16
; CHECK-NEXT:    dup v1.4s, w8
; CHECK-NEXT:    umull2 v2.2d, v0.4s, v1.4s
; CHECK-NEXT:    umull v0.2d, v0.2s, v1.2s
; CHECK-NEXT:    uzp2 v0.4s, v0.4s, v2.4s
; CHECK-NEXT:    ushr v0.4s, v0.4s, #2
; CHECK-NEXT:    ret
  %div = udiv <4 x i32> %x, <i32 14, i32 14, i32 14, i32 14>
  ret <4 x i32> %div
}