; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
; RUN: llc < %s -mtriple=thumbv8.1m.main-none-eabihf -mattr=+mve.fp | FileCheck %s --check-prefix=CHECKLE
; RUN: llc < %s -mtriple=thumbebv8.1m.main-none-eabihf -mattr=+mve.fp | FileCheck %s --check-prefix=CHECKBE


define <8 x i8> @inserti8_first(ptr %p) {
; CHECKLE-LABEL: inserti8_first:
; CHECKLE:       @ %bb.0:
; CHECKLE-NEXT:    vldrb.u16 q0, [r0]
; CHECKLE-NEXT:    bx lr
;
; CHECKBE-LABEL: inserti8_first:
; CHECKBE:       @ %bb.0:
; CHECKBE-NEXT:    vldrb.u16 q1, [r0]
; CHECKBE-NEXT:    vrev64.16 q0, q1
; CHECKBE-NEXT:    bx lr
  %q = getelementptr inbounds i8, ptr %p, i32 1
  %l1 = load <8 x i8>, ptr %q
  %l2 = load i8, ptr %p
  %s = shufflevector <8 x i8> %l1, <8 x i8> undef, <8 x i32> <i32 undef, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6>
  %ins = insertelement <8 x i8> %s, i8 %l2, i32 0
  ret <8 x i8> %ins
}

define <8 x i8> @inserti8_last(ptr %p) {
; CHECKLE-LABEL: inserti8_last:
; CHECKLE:       @ %bb.0:
; CHECKLE-NEXT:    vldrb.u16 q0, [r0, #1]
; CHECKLE-NEXT:    bx lr
;
; CHECKBE-LABEL: inserti8_last:
; CHECKBE:       @ %bb.0:
; CHECKBE-NEXT:    vldrb.u16 q1, [r0, #1]
; CHECKBE-NEXT:    vrev64.16 q0, q1
; CHECKBE-NEXT:    bx lr
  %q = getelementptr inbounds i8, ptr %p, i32 8
  %l1 = load <8 x i8>, ptr %p
  %l2 = load i8, ptr %q
  %s = shufflevector <8 x i8> %l1, <8 x i8> undef, <8 x i32> <i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 undef>
  %ins = insertelement <8 x i8> %s, i8 %l2, i32 7
  ret <8 x i8> %ins
}

define <8 x i16> @inserti8_first_sext(ptr %p) {
; CHECKLE-LABEL: inserti8_first_sext:
; CHECKLE:       @ %bb.0:
; CHECKLE-NEXT:    vldrb.s16 q0, [r0]
; CHECKLE-NEXT:    bx lr
;
; CHECKBE-LABEL: inserti8_first_sext:
; CHECKBE:       @ %bb.0:
; CHECKBE-NEXT:    vldrb.s16 q1, [r0]
; CHECKBE-NEXT:    vrev64.16 q0, q1
; CHECKBE-NEXT:    bx lr
  %q = getelementptr inbounds i8, ptr %p, i32 1
  %l1 = load <8 x i8>, ptr %q
  %s1 = sext <8 x i8> %l1 to <8 x i16>
  %l2 = load i8, ptr %p
  %s2 = sext i8 %l2 to i16
  %s = shufflevector <8 x i16> %s1, <8 x i16> undef, <8 x i32> <i32 undef, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6>
  %ins = insertelement <8 x i16> %s, i16 %s2, i32 0
  ret <8 x i16> %ins
}

define <8 x i16> @inserti8_last_sext(ptr %p) {
; CHECKLE-LABEL: inserti8_last_sext:
; CHECKLE:       @ %bb.0:
; CHECKLE-NEXT:    vldrb.s16 q0, [r0, #1]
; CHECKLE-NEXT:    bx lr
;
; CHECKBE-LABEL: inserti8_last_sext:
; CHECKBE:       @ %bb.0:
; CHECKBE-NEXT:    vldrb.s16 q1, [r0, #1]
; CHECKBE-NEXT:    vrev64.16 q0, q1
; CHECKBE-NEXT:    bx lr
  %q = getelementptr inbounds i8, ptr %p, i32 8
  %l1 = load <8 x i8>, ptr %p
  %s1 = sext <8 x i8> %l1 to <8 x i16>
  %l2 = load i8, ptr %q
  %s2 = sext i8 %l2 to i16
  %s = shufflevector <8 x i16> %s1, <8 x i16> undef, <8 x i32> <i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 undef>
  %ins = insertelement <8 x i16> %s, i16 %s2, i32 7
  ret <8 x i16> %ins
}

define <8 x i16> @inserti8_first_zext(ptr %p) {
; CHECKLE-LABEL: inserti8_first_zext:
; CHECKLE:       @ %bb.0:
; CHECKLE-NEXT:    vldrb.u16 q0, [r0]
; CHECKLE-NEXT:    bx lr
;
; CHECKBE-LABEL: inserti8_first_zext:
; CHECKBE:       @ %bb.0:
; CHECKBE-NEXT:    vldrb.u16 q1, [r0]
; CHECKBE-NEXT:    vrev64.16 q0, q1
; CHECKBE-NEXT:    bx lr
  %q = getelementptr inbounds i8, ptr %p, i32 1
  %l1 = load <8 x i8>, ptr %q
  %s1 = zext <8 x i8> %l1 to <8 x i16>
  %l2 = load i8, ptr %p
  %s2 = zext i8 %l2 to i16
  %s = shufflevector <8 x i16> %s1, <8 x i16> undef, <8 x i32> <i32 undef, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6>
  %ins = insertelement <8 x i16> %s, i16 %s2, i32 0
  ret <8 x i16> %ins
}

define <8 x i16> @inserti8_last_zext(ptr %p) {
; CHECKLE-LABEL: inserti8_last_zext:
; CHECKLE:       @ %bb.0:
; CHECKLE-NEXT:    vldrb.u16 q0, [r0, #1]
; CHECKLE-NEXT:    bx lr
;
; CHECKBE-LABEL: inserti8_last_zext:
; CHECKBE:       @ %bb.0:
; CHECKBE-NEXT:    vldrb.u16 q1, [r0, #1]
; CHECKBE-NEXT:    vrev64.16 q0, q1
; CHECKBE-NEXT:    bx lr
  %q = getelementptr inbounds i8, ptr %p, i32 8
  %l1 = load <8 x i8>, ptr %p
  %s1 = zext <8 x i8> %l1 to <8 x i16>
  %l2 = load i8, ptr %q
  %s2 = zext i8 %l2 to i16
  %s = shufflevector <8 x i16> %s1, <8 x i16> undef, <8 x i32> <i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 undef>
  %ins = insertelement <8 x i16> %s, i16 %s2, i32 7
  ret <8 x i16> %ins
}

define <8 x i32> @inserti32_first(ptr %p) {
; CHECKLE-LABEL: inserti32_first:
; CHECKLE:       @ %bb.0:
; CHECKLE-NEXT:    vldrw.u32 q2, [r0, #20]
; CHECKLE-NEXT:    vldr s4, [r0, #16]
; CHECKLE-NEXT:    vldrw.u32 q0, [r0]
; CHECKLE-NEXT:    vmov.f32 s5, s8
; CHECKLE-NEXT:    vmov.f32 s6, s9
; CHECKLE-NEXT:    vmov.f32 s7, s10
; CHECKLE-NEXT:    bx lr
;
; CHECKBE-LABEL: inserti32_first:
; CHECKBE:       @ %bb.0:
; CHECKBE-NEXT:    vldrw.u32 q3, [r0, #20]
; CHECKBE-NEXT:    vldrb.u8 q1, [r0]
; CHECKBE-NEXT:    vldr s8, [r0, #16]
; CHECKBE-NEXT:    vmov.f32 s9, s12
; CHECKBE-NEXT:    vrev64.8 q0, q1
; CHECKBE-NEXT:    vmov.f32 s10, s13
; CHECKBE-NEXT:    vmov.f32 s11, s14
; CHECKBE-NEXT:    vrev64.32 q1, q2
; CHECKBE-NEXT:    bx lr
  %q = getelementptr inbounds i8, ptr %p, i32 4
  %l1 = load <8 x i32>, ptr %q
  %l2 = load i32, ptr %p
  %s = shufflevector <8 x i32> %l1, <8 x i32> undef, <8 x i32> <i32 undef, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6>
  %ins = insertelement <8 x i32> %s, i32 %l2, i32 0
  ret <8 x i32> %ins
}

define <8 x i32> @inserti32_last(ptr %p) {
; CHECKLE-LABEL: inserti32_last:
; CHECKLE:       @ %bb.0:
; CHECKLE-NEXT:    vldrw.u32 q2, [r0]
; CHECKLE-NEXT:    vldr s3, [r0, #16]
; CHECKLE-NEXT:    vldrw.u32 q1, [r0, #20]
; CHECKLE-NEXT:    vmov.f32 s0, s9
; CHECKLE-NEXT:    vmov.f32 s1, s10
; CHECKLE-NEXT:    vmov.f32 s2, s11
; CHECKLE-NEXT:    bx lr
;
; CHECKBE-LABEL: inserti32_last:
; CHECKBE:       @ %bb.0:
; CHECKBE-NEXT:    vldrw.u32 q3, [r0]
; CHECKBE-NEXT:    vldrb.u8 q0, [r0, #20]
; CHECKBE-NEXT:    vldr s11, [r0, #16]
; CHECKBE-NEXT:    vmov.f32 s8, s13
; CHECKBE-NEXT:    vrev64.8 q1, q0
; CHECKBE-NEXT:    vmov.f32 s9, s14
; CHECKBE-NEXT:    vmov.f32 s10, s15
; CHECKBE-NEXT:    vrev64.32 q0, q2
; CHECKBE-NEXT:    bx lr
  %q = getelementptr inbounds i8, ptr %p, i32 32
  %l1 = load <8 x i32>, ptr %p
  %l2 = load i32, ptr %q
  %s = shufflevector <8 x i32> %l1, <8 x i32> undef, <8 x i32> <i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 undef>
  %ins = insertelement <8 x i32> %s, i32 %l2, i32 7
  ret <8 x i32> %ins
}

define <8 x i32> @inserti32_first_multiuse(ptr %p) {
; CHECKLE-LABEL: inserti32_first_multiuse:
; CHECKLE:       @ %bb.0:
; CHECKLE-NEXT:    vldrw.u32 q0, [r0, #20]
; CHECKLE-NEXT:    vldrw.u32 q2, [r0, #4]
; CHECKLE-NEXT:    vmov.f32 s4, s11
; CHECKLE-NEXT:    vmov.f32 s5, s0
; CHECKLE-NEXT:    vmov.f32 s6, s1
; CHECKLE-NEXT:    vmov.f32 s7, s2
; CHECKLE-NEXT:    vadd.i32 q1, q0, q1
; CHECKLE-NEXT:    vldrw.u32 q0, [r0]
; CHECKLE-NEXT:    vadd.i32 q0, q2, q0
; CHECKLE-NEXT:    bx lr
;
; CHECKBE-LABEL: inserti32_first_multiuse:
; CHECKBE:       @ %bb.0:
; CHECKBE-NEXT:    vldrw.u32 q0, [r0, #20]
; CHECKBE-NEXT:    vldrw.u32 q2, [r0, #4]
; CHECKBE-NEXT:    vmov.f32 s4, s11
; CHECKBE-NEXT:    vmov.f32 s5, s0
; CHECKBE-NEXT:    vmov.f32 s6, s1
; CHECKBE-NEXT:    vmov.f32 s7, s2
; CHECKBE-NEXT:    vadd.i32 q0, q0, q1
; CHECKBE-NEXT:    vrev64.32 q1, q0
; CHECKBE-NEXT:    vldrw.u32 q0, [r0]
; CHECKBE-NEXT:    vadd.i32 q2, q2, q0
; CHECKBE-NEXT:    vrev64.32 q0, q2
; CHECKBE-NEXT:    bx lr
  %q = getelementptr inbounds i8, ptr %p, i32 4
  %l1 = load <8 x i32>, ptr %q
  %l2 = load i32, ptr %p
  %s = shufflevector <8 x i32> %l1, <8 x i32> undef, <8 x i32> <i32 undef, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6>
  %ins = insertelement <8 x i32> %s, i32 %l2, i32 0
  %a = add <8 x i32> %l1, %ins
  ret <8 x i32> %a
}

define <8 x i32> @inserti32_last_multiuse(ptr %p) {
; CHECKLE-LABEL: inserti32_last_multiuse:
; CHECKLE:       @ %bb.0:
; CHECKLE-NEXT:    vldrw.u32 q0, [r0]
; CHECKLE-NEXT:    vldrw.u32 q1, [r0, #16]
; CHECKLE-NEXT:    vmov.f32 s8, s1
; CHECKLE-NEXT:    vmov.f32 s9, s2
; CHECKLE-NEXT:    vmov.f32 s10, s3
; CHECKLE-NEXT:    vmov.f32 s11, s4
; CHECKLE-NEXT:    vadd.i32 q0, q0, q2
; CHECKLE-NEXT:    vldrw.u32 q2, [r0, #20]
; CHECKLE-NEXT:    vadd.i32 q1, q1, q2
; CHECKLE-NEXT:    bx lr
;
; CHECKBE-LABEL: inserti32_last_multiuse:
; CHECKBE:       @ %bb.0:
; CHECKBE-NEXT:    vldrw.u32 q0, [r0]
; CHECKBE-NEXT:    vldrw.u32 q1, [r0, #16]
; CHECKBE-NEXT:    vmov.f32 s8, s1
; CHECKBE-NEXT:    vmov.f32 s9, s2
; CHECKBE-NEXT:    vmov.f32 s10, s3
; CHECKBE-NEXT:    vmov.f32 s11, s4
; CHECKBE-NEXT:    vadd.i32 q2, q0, q2
; CHECKBE-NEXT:    vrev64.32 q0, q2
; CHECKBE-NEXT:    vldrw.u32 q2, [r0, #20]
; CHECKBE-NEXT:    vadd.i32 q2, q1, q2
; CHECKBE-NEXT:    vrev64.32 q1, q2
; CHECKBE-NEXT:    bx lr
  %q = getelementptr inbounds i8, ptr %p, i32 32
  %l1 = load <8 x i32>, ptr %p
  %l2 = load i32, ptr %q
  %s = shufflevector <8 x i32> %l1, <8 x i32> undef, <8 x i32> <i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 undef>
  %ins = insertelement <8 x i32> %s, i32 %l2, i32 7
  %a = add <8 x i32> %l1, %ins
  ret <8 x i32> %a
}

define <4 x float> @insertf32_first(ptr %p) {
; CHECKLE-LABEL: insertf32_first:
; CHECKLE:       @ %bb.0:
; CHECKLE-NEXT:    vldrw.u32 q0, [r0]
; CHECKLE-NEXT:    bx lr
;
; CHECKBE-LABEL: insertf32_first:
; CHECKBE:       @ %bb.0:
; CHECKBE-NEXT:    vldrb.u8 q1, [r0]
; CHECKBE-NEXT:    vrev64.8 q0, q1
; CHECKBE-NEXT:    bx lr
  %q = getelementptr inbounds i8, ptr %p, i32 4
  %l1 = load <4 x float>, ptr %q
  %l2 = load float, ptr %p
  %s = shufflevector <4 x float> %l1, <4 x float> undef, <4 x i32> <i32 undef, i32 0, i32 1, i32 2>
  %ins = insertelement <4 x float> %s, float %l2, i32 0
  ret <4 x float> %ins
}

define <4 x float> @insertf32_last(ptr %p) {
; CHECKLE-LABEL: insertf32_last:
; CHECKLE:       @ %bb.0:
; CHECKLE-NEXT:    vldrw.u32 q0, [r0, #4]
; CHECKLE-NEXT:    bx lr
;
; CHECKBE-LABEL: insertf32_last:
; CHECKBE:       @ %bb.0:
; CHECKBE-NEXT:    vldrb.u8 q1, [r0, #4]
; CHECKBE-NEXT:    vrev64.8 q0, q1
; CHECKBE-NEXT:    bx lr
  %q = getelementptr inbounds i8, ptr %p, i32 16
  %l1 = load <4 x float>, ptr %p
  %l2 = load float, ptr %q
  %s = shufflevector <4 x float> %l1, <4 x float> undef, <4 x i32> <i32 1, i32 2, i32 3, i32 undef>
  %ins = insertelement <4 x float> %s, float %l2, i32 3
  ret <4 x float> %ins
}