; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
; RUN: %if x86-registered-target %{ opt -S --passes=slp-vectorizer -mtriple=x86_64-unknown-linux-gnu -slp-threshold=-50 < %s | FileCheck %s --check-prefixes=CHECK,X86 %}
; RUN: %if aarch64-registered-target %{ opt -S --passes=slp-vectorizer -mtriple=aarch64-unknown-linux-gnu -slp-threshold=-50 < %s | FileCheck %s --check-prefixes=CHECK,AARCH64 %}

; Don't care about the profitability with these tests, just want to demonstrate the ability
; to combine opcodes

define void @shl_add(ptr %p, ptr %s) {
; CHECK-LABEL: @shl_add(
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[TMP0:%.*]] = load <4 x i16>, ptr [[P:%.*]], align 2
; CHECK-NEXT:    [[TMP1:%.*]] = shl <4 x i16> [[TMP0]], <i16 3, i16 5, i16 0, i16 3>
; CHECK-NEXT:    store <4 x i16> [[TMP1]], ptr [[S:%.*]], align 2
; CHECK-NEXT:    ret void
;
entry:
  %p1 = getelementptr i16, ptr %p, i64 1
  %p2 = getelementptr i16, ptr %p, i64 2
  %p3 = getelementptr i16, ptr %p, i64 3

  %l0 = load i16, ptr %p
  %l1 = load i16, ptr %p1
  %l2 = load i16, ptr %p2
  %l3 = load i16, ptr %p3

  %shl0 = shl i16 %l0, 3
  %shl1 = shl i16 %l1, 5
  %shl2 = add i16 %l2, 0
  %shl3 = shl i16 %l3, 3

  %s1 = getelementptr i16, ptr %s, i64 1
  %s2 = getelementptr i16, ptr %s, i64 2
  %s3 = getelementptr i16, ptr %s, i64 3

  store i16 %shl0, ptr %s
  store i16 %shl1, ptr %s1
  store i16 %shl2, ptr %s2
  store i16 %shl3, ptr %s3
  ret void
}

define void @add_shl(ptr %p, ptr %s) {
; CHECK-LABEL: @add_shl(
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[TMP0:%.*]] = load <4 x i16>, ptr [[P:%.*]], align 2
; CHECK-NEXT:    [[TMP1:%.*]] = add <4 x i16> [[TMP0]], <i16 0, i16 5, i16 2, i16 3>
; CHECK-NEXT:    store <4 x i16> [[TMP1]], ptr [[S:%.*]], align 2
; CHECK-NEXT:    ret void
;
entry:
  %p1 = getelementptr i16, ptr %p, i64 1
  %p2 = getelementptr i16, ptr %p, i64 2
  %p3 = getelementptr i16, ptr %p, i64 3

  %l0 = load i16, ptr %p
  %l1 = load i16, ptr %p1
  %l2 = load i16, ptr %p2
  %l3 = load i16, ptr %p3

  %add0 = shl i16 %l0, 0
  %add1 = add i16 %l1, 5
  %add2 = add i16 %l2, 2
  %add3 = add i16 %l3, 3

  %s1 = getelementptr i16, ptr %s, i64 1
  %s2 = getelementptr i16, ptr %s, i64 2
  %s3 = getelementptr i16, ptr %s, i64 3

  store i16 %add0, ptr %s
  store i16 %add1, ptr %s1
  store i16 %add2, ptr %s2
  store i16 %add3, ptr %s3
  ret void
}

define void @sub_mul(ptr %p, ptr %s) {
; CHECK-LABEL: @sub_mul(
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[TMP0:%.*]] = load <4 x i16>, ptr [[P:%.*]], align 2
; CHECK-NEXT:    [[TMP1:%.*]] = mul <4 x i16> [[TMP0]], <i16 1, i16 5, i16 2, i16 3>
; CHECK-NEXT:    store <4 x i16> [[TMP1]], ptr [[S:%.*]], align 2
; CHECK-NEXT:    ret void
;
entry:
  %p1 = getelementptr i16, ptr %p, i64 1
  %p2 = getelementptr i16, ptr %p, i64 2
  %p3 = getelementptr i16, ptr %p, i64 3

  %l0 = load i16, ptr %p
  %l1 = load i16, ptr %p1
  %l2 = load i16, ptr %p2
  %l3 = load i16, ptr %p3

  %mul0 = sub i16 %l0, 0
  %mul1 = mul i16 %l1, 5
  %mul2 = mul i16 %l2, 2
  %mul3 = mul i16 %l3, 3

  %s1 = getelementptr i16, ptr %s, i64 1
  %s2 = getelementptr i16, ptr %s, i64 2
  %s3 = getelementptr i16, ptr %s, i64 3

  store i16 %mul0, ptr %s
  store i16 %mul1, ptr %s1
  store i16 %mul2, ptr %s2
  store i16 %mul3, ptr %s3
  ret void
}

define void @add_mul(ptr %p, ptr %s) {
; CHECK-LABEL: @add_mul(
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[TMP0:%.*]] = load <4 x i16>, ptr [[P:%.*]], align 2
; CHECK-NEXT:    [[TMP1:%.*]] = mul <4 x i16> [[TMP0]], <i16 1, i16 5, i16 2, i16 3>
; CHECK-NEXT:    store <4 x i16> [[TMP1]], ptr [[S:%.*]], align 2
; CHECK-NEXT:    ret void
;
entry:
  %p1 = getelementptr i16, ptr %p, i64 1
  %p2 = getelementptr i16, ptr %p, i64 2
  %p3 = getelementptr i16, ptr %p, i64 3

  %l0 = load i16, ptr %p
  %l1 = load i16, ptr %p1
  %l2 = load i16, ptr %p2
  %l3 = load i16, ptr %p3

  %mul0 = add i16 %l0, 0
  %mul1 = mul i16 %l1, 5
  %mul2 = mul i16 %l2, 2
  %mul3 = mul i16 %l3, 3

  %s1 = getelementptr i16, ptr %s, i64 1
  %s2 = getelementptr i16, ptr %s, i64 2
  %s3 = getelementptr i16, ptr %s, i64 3

  store i16 %mul0, ptr %s
  store i16 %mul1, ptr %s1
  store i16 %mul2, ptr %s2
  store i16 %mul3, ptr %s3
  ret void
}

define void @sub_and(ptr %p, ptr %s) {
; CHECK-LABEL: @sub_and(
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[TMP0:%.*]] = load <4 x i16>, ptr [[P:%.*]], align 2
; CHECK-NEXT:    [[TMP1:%.*]] = and <4 x i16> [[TMP0]], <i16 -1, i16 5, i16 2, i16 3>
; CHECK-NEXT:    store <4 x i16> [[TMP1]], ptr [[S:%.*]], align 2
; CHECK-NEXT:    ret void
;
entry:
  %p1 = getelementptr i16, ptr %p, i64 1
  %p2 = getelementptr i16, ptr %p, i64 2
  %p3 = getelementptr i16, ptr %p, i64 3

  %l0 = load i16, ptr %p
  %l1 = load i16, ptr %p1
  %l2 = load i16, ptr %p2
  %l3 = load i16, ptr %p3

  %mul0 = add i16 %l0, 0
  %mul1 = and i16 %l1, 5
  %mul2 = and i16 %l2, 2
  %mul3 = and i16 %l3, 3

  %s1 = getelementptr i16, ptr %s, i64 1
  %s2 = getelementptr i16, ptr %s, i64 2
  %s3 = getelementptr i16, ptr %s, i64 3

  store i16 %mul0, ptr %s
  store i16 %mul1, ptr %s1
  store i16 %mul2, ptr %s2
  store i16 %mul3, ptr %s3
  ret void
}

define void @shl_mul(ptr %p, ptr %s) {
; CHECK-LABEL: @shl_mul(
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[TMP0:%.*]] = load <4 x i16>, ptr [[P:%.*]], align 2
; CHECK-NEXT:    [[TMP1:%.*]] = mul <4 x i16> [[TMP0]], <i16 8, i16 5, i16 2, i16 3>
; CHECK-NEXT:    store <4 x i16> [[TMP1]], ptr [[S:%.*]], align 2
; CHECK-NEXT:    ret void
;
entry:
  %p1 = getelementptr i16, ptr %p, i64 1
  %p2 = getelementptr i16, ptr %p, i64 2
  %p3 = getelementptr i16, ptr %p, i64 3

  %l0 = load i16, ptr %p
  %l1 = load i16, ptr %p1
  %l2 = load i16, ptr %p2
  %l3 = load i16, ptr %p3

  %mul0 = shl i16 %l0, 3
  %mul1 = mul i16 %l1, 5
  %mul2 = mul i16 %l2, 2
  %mul3 = mul i16 %l3, 3

  %s1 = getelementptr i16, ptr %s, i64 1
  %s2 = getelementptr i16, ptr %s, i64 2
  %s3 = getelementptr i16, ptr %s, i64 3

  store i16 %mul0, ptr %s
  store i16 %mul1, ptr %s1
  store i16 %mul2, ptr %s2
  store i16 %mul3, ptr %s3
  ret void
}

define void @lshr_div(ptr %p, ptr %s) {
; CHECK-LABEL: @lshr_div(
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[TMP0:%.*]] = load <4 x i16>, ptr [[P:%.*]], align 2
; CHECK-NEXT:    [[TMP2:%.*]] = lshr <4 x i16> [[TMP0]], <i16 3, i16 0, i16 0, i16 0>
; CHECK-NEXT:    [[TMP1:%.*]] = udiv <4 x i16> [[TMP2]], <i16 1, i16 5, i16 2, i16 3>
; CHECK-NEXT:    store <4 x i16> [[TMP1]], ptr [[S:%.*]], align 2
; CHECK-NEXT:    ret void
;
entry:
  %p1 = getelementptr i16, ptr %p, i64 1
  %p2 = getelementptr i16, ptr %p, i64 2
  %p3 = getelementptr i16, ptr %p, i64 3

  %l0 = load i16, ptr %p
  %l1 = load i16, ptr %p1
  %l2 = load i16, ptr %p2
  %l3 = load i16, ptr %p3

  %div0 = lshr i16 %l0, 3
  %div1 = udiv i16 %l1, 5
  %div2 = udiv i16 %l2, 2
  %div3 = udiv i16 %l3, 3

  %s1 = getelementptr i16, ptr %s, i64 1
  %s2 = getelementptr i16, ptr %s, i64 2
  %s3 = getelementptr i16, ptr %s, i64 3

  store i16 %div0, ptr %s
  store i16 %div1, ptr %s1
  store i16 %div2, ptr %s2
  store i16 %div3, ptr %s3
  ret void
}

define void @div_lshr(ptr %p, ptr %s) {
; CHECK-LABEL: @div_lshr(
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[TMP0:%.*]] = load <4 x i16>, ptr [[P:%.*]], align 2
; CHECK-NEXT:    [[TMP2:%.*]] = udiv <4 x i16> [[TMP0]], <i16 5, i16 1, i16 2, i16 1>
; CHECK-NEXT:    [[TMP1:%.*]] = lshr <4 x i16> [[TMP2]], <i16 0, i16 3, i16 0, i16 9>
; CHECK-NEXT:    store <4 x i16> [[TMP1]], ptr [[S:%.*]], align 2
; CHECK-NEXT:    ret void
;
entry:
  %p1 = getelementptr i16, ptr %p, i64 1
  %p2 = getelementptr i16, ptr %p, i64 2
  %p3 = getelementptr i16, ptr %p, i64 3

  %l0 = load i16, ptr %p
  %l1 = load i16, ptr %p1
  %l2 = load i16, ptr %p2
  %l3 = load i16, ptr %p3

  %div0 = udiv i16 %l0, 5
  %div1 = lshr i16 %l1, 3
  %div2 = udiv i16 %l2, 2
  %div3 = lshr i16 %l3, 9

  %s1 = getelementptr i16, ptr %s, i64 1
  %s2 = getelementptr i16, ptr %s, i64 2
  %s3 = getelementptr i16, ptr %s, i64 3

  store i16 %div0, ptr %s
  store i16 %div1, ptr %s1
  store i16 %div2, ptr %s2
  store i16 %div3, ptr %s3
  ret void
}

define void @div_lshr_too_large(ptr %p, ptr %s) {
; CHECK-LABEL: @div_lshr_too_large(
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[TMP0:%.*]] = load <4 x i16>, ptr [[P:%.*]], align 2
; CHECK-NEXT:    [[TMP1:%.*]] = udiv <4 x i16> [[TMP0]], <i16 5, i16 1, i16 2, i16 1>
; CHECK-NEXT:    [[TMP2:%.*]] = lshr <4 x i16> [[TMP1]], <i16 0, i16 17, i16 0, i16 9>
; CHECK-NEXT:    store <4 x i16> [[TMP2]], ptr [[S:%.*]], align 2
; CHECK-NEXT:    ret void
;
entry:
  %p1 = getelementptr i16, ptr %p, i64 1
  %p2 = getelementptr i16, ptr %p, i64 2
  %p3 = getelementptr i16, ptr %p, i64 3

  %l0 = load i16, ptr %p
  %l1 = load i16, ptr %p1
  %l2 = load i16, ptr %p2
  %l3 = load i16, ptr %p3

  %div0 = udiv i16 %l0, 5
  %div1 = lshr i16 %l1, 17
  %div2 = udiv i16 %l2, 2
  %div3 = lshr i16 %l3, 9

  %s1 = getelementptr i16, ptr %s, i64 1
  %s2 = getelementptr i16, ptr %s, i64 2
  %s3 = getelementptr i16, ptr %s, i64 3

  store i16 %div0, ptr %s
  store i16 %div1, ptr %s1
  store i16 %div2, ptr %s2
  store i16 %div3, ptr %s3
  ret void
}

define void @lshr_div2(ptr %p, ptr %s) {
; X86-LABEL: @lshr_div2(
; X86-NEXT:  entry:
; X86-NEXT:    [[P1:%.*]] = getelementptr i16, ptr [[P:%.*]], i64 1
; X86-NEXT:    [[P3:%.*]] = getelementptr i16, ptr [[P]], i64 3
; X86-NEXT:    [[L0:%.*]] = load i16, ptr [[P]], align 2
; X86-NEXT:    [[L3:%.*]] = load i16, ptr [[P3]], align 2
; X86-NEXT:    [[DIV0:%.*]] = lshr i16 [[L0]], 3
; X86-NEXT:    [[DIV3:%.*]] = lshr i16 [[L3]], 5
; X86-NEXT:    [[S1:%.*]] = getelementptr i16, ptr [[S:%.*]], i64 1
; X86-NEXT:    [[S3:%.*]] = getelementptr i16, ptr [[S]], i64 3
; X86-NEXT:    [[TMP0:%.*]] = load <2 x i16>, ptr [[P1]], align 2
; X86-NEXT:    [[TMP1:%.*]] = udiv <2 x i16> [[TMP0]], <i16 8, i16 2>
; X86-NEXT:    store i16 [[DIV0]], ptr [[S]], align 2
; X86-NEXT:    store <2 x i16> [[TMP1]], ptr [[S1]], align 2
; X86-NEXT:    store i16 [[DIV3]], ptr [[S3]], align 2
; X86-NEXT:    ret void
;
; AARCH64-LABEL: @lshr_div2(
; AARCH64-NEXT:  entry:
; AARCH64-NEXT:    [[TMP0:%.*]] = load <4 x i16>, ptr [[P:%.*]], align 2
; AARCH64-NEXT:    [[TMP1:%.*]] = lshr <4 x i16> [[TMP0]], <i16 3, i16 0, i16 0, i16 5>
; AARCH64-NEXT:    [[TMP2:%.*]] = udiv <4 x i16> [[TMP1]], <i16 1, i16 8, i16 2, i16 1>
; AARCH64-NEXT:    store <4 x i16> [[TMP2]], ptr [[S:%.*]], align 2
; AARCH64-NEXT:    ret void
;
entry:
  %p1 = getelementptr i16, ptr %p, i64 1
  %p2 = getelementptr i16, ptr %p, i64 2
  %p3 = getelementptr i16, ptr %p, i64 3

  %l0 = load i16, ptr %p
  %l1 = load i16, ptr %p1
  %l2 = load i16, ptr %p2
  %l3 = load i16, ptr %p3

  %div0 = lshr i16 %l0, 3
  %div1 = udiv i16 %l1, 8
  %div2 = udiv i16 %l2, 2
  %div3 = lshr i16 %l3, 5

  %s1 = getelementptr i16, ptr %s, i64 1
  %s2 = getelementptr i16, ptr %s, i64 2
  %s3 = getelementptr i16, ptr %s, i64 3

  store i16 %div0, ptr %s
  store i16 %div1, ptr %s1
  store i16 %div2, ptr %s2
  store i16 %div3, ptr %s3
  ret void
}

define void @div2_lshr(ptr %p, ptr %s) {
; CHECK-LABEL: @div2_lshr(
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[TMP0:%.*]] = load <4 x i16>, ptr [[P:%.*]], align 2
; CHECK-NEXT:    [[TMP2:%.*]] = udiv <4 x i16> [[TMP0]], <i16 8, i16 1, i16 4, i16 1>
; CHECK-NEXT:    [[TMP1:%.*]] = lshr <4 x i16> [[TMP2]], <i16 0, i16 3, i16 0, i16 4>
; CHECK-NEXT:    store <4 x i16> [[TMP1]], ptr [[S:%.*]], align 2
; CHECK-NEXT:    ret void
;
entry:
  %p1 = getelementptr i16, ptr %p, i64 1
  %p2 = getelementptr i16, ptr %p, i64 2
  %p3 = getelementptr i16, ptr %p, i64 3

  %l0 = load i16, ptr %p
  %l1 = load i16, ptr %p1
  %l2 = load i16, ptr %p2
  %l3 = load i16, ptr %p3

  %div0 = udiv i16 %l0, 8
  %div1 = lshr i16 %l1, 3
  %div2 = udiv i16 %l2, 4
  %div3 = lshr i16 %l3, 4

  %s1 = getelementptr i16, ptr %s, i64 1
  %s2 = getelementptr i16, ptr %s, i64 2
  %s3 = getelementptr i16, ptr %s, i64 3

  store i16 %div0, ptr %s
  store i16 %div1, ptr %s1
  store i16 %div2, ptr %s2
  store i16 %div3, ptr %s3
  ret void
}

define void @add_sub(ptr %p, ptr %s) {
; CHECK-LABEL: @add_sub(
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[TMP0:%.*]] = load <4 x i16>, ptr [[P:%.*]], align 2
; CHECK-NEXT:    [[TMP1:%.*]] = add <4 x i16> [[TMP0]], <i16 3, i16 5, i16 2, i16 3>
; CHECK-NEXT:    store <4 x i16> [[TMP1]], ptr [[S:%.*]], align 2
; CHECK-NEXT:    ret void
;
entry:
  %p1 = getelementptr i16, ptr %p, i64 1
  %p2 = getelementptr i16, ptr %p, i64 2
  %p3 = getelementptr i16, ptr %p, i64 3

  %l0 = load i16, ptr %p
  %l1 = load i16, ptr %p1
  %l2 = load i16, ptr %p2
  %l3 = load i16, ptr %p3

  %add0 = sub i16 %l0, -3
  %add1 = add i16 %l1, 5
  %add2 = add i16 %l2, 2
  %add3 = add i16 %l3, 3

  %s1 = getelementptr i16, ptr %s, i64 1
  %s2 = getelementptr i16, ptr %s, i64 2
  %s3 = getelementptr i16, ptr %s, i64 3

  store i16 %add0, ptr %s
  store i16 %add1, ptr %s1
  store i16 %add2, ptr %s2
  store i16 %add3, ptr %s3
  ret void
}