test/CodeGen/RISCV/rvp-ext-rv64.ll

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -mtriple=riscv64 -mattr=+experimental-p,+m,+zbb \
; RUN:   -verify-machineinstrs < %s | \
; RUN:   FileCheck %s

; Test basic add/sub operations for v4i16
define <4 x i16> @test_padd_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_padd_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    padd.h a0, a0, a1
; CHECK-NEXT:    ret
  %res = add <4 x i16> %a, %b
  ret <4 x i16> %res
}

define <4 x i16> @test_psub_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_psub_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psub.h a0, a0, a1
; CHECK-NEXT:    ret
  %res = sub <4 x i16> %a, %b
  ret <4 x i16> %res
}

; Test basic add/sub operations for v8i8
define <8 x i8> @test_padd_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_padd_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    padd.b a0, a0, a1
; CHECK-NEXT:    ret
  %res = add <8 x i8> %a, %b
  ret <8 x i8> %res
}

define <8 x i8> @test_psub_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_psub_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psub.b a0, a0, a1
; CHECK-NEXT:    ret
  %res = sub <8 x i8> %a, %b
  ret <8 x i8> %res
}

; Test bitwise operations for v4i16 (use scalar instructions)
define <4 x i16> @test_and_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_and_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    and a0, a0, a1
; CHECK-NEXT:    ret
  %res = and <4 x i16> %a, %b
  ret <4 x i16> %res
}

define <4 x i16> @test_or_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_or_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    or a0, a0, a1
; CHECK-NEXT:    ret
  %res = or <4 x i16> %a, %b
  ret <4 x i16> %res
}

define <4 x i16> @test_xor_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_xor_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    xor a0, a0, a1
; CHECK-NEXT:    ret
  %res = xor <4 x i16> %a, %b
  ret <4 x i16> %res
}

define <4 x i16> @test_andn_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_andn_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    andn a0, a0, a1
; CHECK-NEXT:    ret
  %not = xor <4 x i16> %b, splat (i16 -1)
  %res = and <4 x i16> %a, %not
  ret <4 x i16> %res
}

define <4 x i16> @test_orn_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_orn_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    orn a0, a0, a1
; CHECK-NEXT:    ret
  %not = xor <4 x i16> %b, splat (i16 -1)
  %res = or <4 x i16> %a, %not
  ret <4 x i16> %res
}

define <4 x i16> @test_xnor_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_xnor_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    xnor a0, a1, a0
; CHECK-NEXT:    ret
  %not = xor <4 x i16> %b, splat (i16 -1)
  %res = xor <4 x i16> %a, %not
  ret <4 x i16> %res
}

; Test bitwise operations for v8i8 (use scalar instructions)
define <8 x i8> @test_and_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_and_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    and a0, a0, a1
; CHECK-NEXT:    ret
  %res = and <8 x i8> %a, %b
  ret <8 x i8> %res
}

define <8 x i8> @test_or_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_or_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    or a0, a0, a1
; CHECK-NEXT:    ret
  %res = or <8 x i8> %a, %b
  ret <8 x i8> %res
}

define <8 x i8> @test_xor_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_xor_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    xor a0, a0, a1
; CHECK-NEXT:    ret
  %res = xor <8 x i8> %a, %b
  ret <8 x i8> %res
}

define <8 x i8> @test_andn_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_andn_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    andn a0, a0, a1
; CHECK-NEXT:    ret
  %not = xor <8 x i8> %b, splat (i8 -1)
  %res = and <8 x i8> %a, %not
  ret <8 x i8> %res
}

define <8 x i8> @test_orn_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_orn_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    orn a0, a0, a1
; CHECK-NEXT:    ret
  %not = xor <8 x i8> %b, splat (i8 -1)
  %res = or <8 x i8> %a, %not
  ret <8 x i8> %res
}

define <8 x i8> @test_xnor_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_xnor_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    xnor a0, a1, a0
; CHECK-NEXT:    ret
  %not = xor <8 x i8> %b, splat (i8 -1)
  %res = xor <8 x i8> %a, %not
  ret <8 x i8> %res
}

; Test bitwise operations for v2i32 (use scalar instructions)
define <2 x i32> @test_and_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_and_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    and a0, a0, a1
; CHECK-NEXT:    ret
  %res = and <2 x i32> %a, %b
  ret <2 x i32> %res
}

define <2 x i32> @test_or_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_or_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    or a0, a0, a1
; CHECK-NEXT:    ret
  %res = or <2 x i32> %a, %b
  ret <2 x i32> %res
}

define <2 x i32> @test_xor_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_xor_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    xor a0, a0, a1
; CHECK-NEXT:    ret
  %res = xor <2 x i32> %a, %b
  ret <2 x i32> %res
}

define <2 x i32> @test_andn_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_andn_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    andn a0, a0, a1
; CHECK-NEXT:    ret
  %not = xor <2 x i32> %b, splat (i32 -1)
  %res = and <2 x i32> %a, %not
  ret <2 x i32> %res
}

define <2 x i32> @test_orn_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_orn_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    orn a0, a0, a1
; CHECK-NEXT:    ret
  %not = xor <2 x i32> %b, splat (i32 -1)
  %res = or <2 x i32> %a, %not
  ret <2 x i32> %res
}

define <2 x i32> @test_xnor_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_xnor_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    xnor a0, a1, a0
; CHECK-NEXT:    ret
  %not = xor <2 x i32> %b, splat (i32 -1)
  %res = xor <2 x i32> %a, %not
  ret <2 x i32> %res
}

define <4 x i16> @test_not_h(<4 x i16> %a) {
; CHECK-LABEL: test_not_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    not a0, a0
; CHECK-NEXT:    ret
  %res = xor <4 x i16> %a, splat(i16 -1)
  ret <4 x i16> %res
}

define <8 x i8> @test_not_b(<8 x i8> %a) {
; CHECK-LABEL: test_not_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    not a0, a0
; CHECK-NEXT:    ret
  %res = xor <8 x i8> %a, splat(i8 -1)
  ret <8 x i8> %res
}

define <2 x i32> @test_not_w(<2 x i32> %a) {
; CHECK-LABEL: test_not_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    not a0, a0
; CHECK-NEXT:    ret
  %res = xor <2 x i32> %a, splat(i32 -1)
  ret <2 x i32> %res
}

; Test saturating add operations for v4i16
define <4 x i16> @test_psadd_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_psadd_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psadd.h a0, a0, a1
; CHECK-NEXT:    ret
  %res = call <4 x i16> @llvm.sadd.sat.v4i16(<4 x i16> %a, <4 x i16> %b)
  ret <4 x i16> %res
}

define <4 x i16> @test_psaddu_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_psaddu_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psaddu.h a0, a0, a1
; CHECK-NEXT:    ret
  %res = call <4 x i16> @llvm.uadd.sat.v4i16(<4 x i16> %a, <4 x i16> %b)
  ret <4 x i16> %res
}

; Test saturating sub operations for v4i16
define <4 x i16> @test_pssub_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_pssub_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pssub.h a0, a0, a1
; CHECK-NEXT:    ret
  %res = call <4 x i16> @llvm.ssub.sat.v4i16(<4 x i16> %a, <4 x i16> %b)
  ret <4 x i16> %res
}

define <4 x i16> @test_pssubu_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_pssubu_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pssubu.h a0, a0, a1
; CHECK-NEXT:    ret
  %res = call <4 x i16> @llvm.usub.sat.v4i16(<4 x i16> %a, <4 x i16> %b)
  ret <4 x i16> %res
}

; Test saturating add operations for v8i8
define <8 x i8> @test_psadd_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_psadd_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psadd.b a0, a0, a1
; CHECK-NEXT:    ret
  %res = call <8 x i8> @llvm.sadd.sat.v8i8(<8 x i8> %a, <8 x i8> %b)
  ret <8 x i8> %res
}

define <8 x i8> @test_psaddu_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_psaddu_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psaddu.b a0, a0, a1
; CHECK-NEXT:    ret
  %res = call <8 x i8> @llvm.uadd.sat.v8i8(<8 x i8> %a, <8 x i8> %b)
  ret <8 x i8> %res
}

; Test saturating sub operations for v8i8
define <8 x i8> @test_pssub_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_pssub_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pssub.b a0, a0, a1
; CHECK-NEXT:    ret
  %res = call <8 x i8> @llvm.ssub.sat.v8i8(<8 x i8> %a, <8 x i8> %b)
  ret <8 x i8> %res
}

define <8 x i8> @test_pssubu_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_pssubu_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pssubu.b a0, a0, a1
; CHECK-NEXT:    ret
  %res = call <8 x i8> @llvm.usub.sat.v8i8(<8 x i8> %a, <8 x i8> %b)
  ret <8 x i8> %res
}

; Test averaging floor signed operations for v4i16
; avgfloors pattern: (a + b) arithmetic shift right 1
define <4 x i16> @test_paadd_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_paadd_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    paadd.h a0, a0, a1
; CHECK-NEXT:    ret
  %ext.a = sext <4 x i16> %a to <4 x i32>
  %ext.b = sext <4 x i16> %b to <4 x i32>
  %add = add nsw <4 x i32> %ext.a, %ext.b
  %shift = ashr <4 x i32> %add, <i32 1, i32 1, i32 1, i32 1>
  %res = trunc <4 x i32> %shift to <4 x i16>
  ret <4 x i16> %res
}

; Test averaging floor unsigned operations for v4i16
; avgflooru pattern: (a & b) + ((a ^ b) >> 1)
define <4 x i16> @test_paaddu_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_paaddu_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    paaddu.h a0, a0, a1
; CHECK-NEXT:    ret
  %and = and <4 x i16> %a, %b
  %xor = xor <4 x i16> %a, %b
  %shift = lshr <4 x i16> %xor, <i16 1, i16 1, i16 1, i16 1>
  %res = add <4 x i16> %and, %shift
  ret <4 x i16> %res
}

; Test averaging floor signed operations for v8i8
define <8 x i8> @test_paadd_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_paadd_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    paadd.b a0, a0, a1
; CHECK-NEXT:    ret
  %ext.a = sext <8 x i8> %a to <8 x i16>
  %ext.b = sext <8 x i8> %b to <8 x i16>
  %add = add nsw <8 x i16> %ext.a, %ext.b
  %shift = ashr <8 x i16> %add, <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>
  %res = trunc <8 x i16> %shift to <8 x i8>
  ret <8 x i8> %res
}

; Test averaging floor unsigned operations for v8i8
define <8 x i8> @test_paaddu_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_paaddu_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    paaddu.b a0, a0, a1
; CHECK-NEXT:    ret
  %and = and <8 x i8> %a, %b
  %xor = xor <8 x i8> %a, %b
  %shift = lshr <8 x i8> %xor, <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>
  %res = add <8 x i8> %and, %shift
  ret <8 x i8> %res
}

define <2 x i32> @test_pabs_w(<2 x i32> %a) {
; CHECK-LABEL: test_pabs_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pneg.w a1, a0
; CHECK-NEXT:    pmax.w a0, a0, a1
; CHECK-NEXT:    ret
  %res = call <2 x i32> @llvm.abs.v2i32(<2 x i32> %a, i1 0)
  ret <2 x i32> %res
}

define <4 x i16> @test_pabs_h(<4 x i16> %a) {
; CHECK-LABEL: test_pabs_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pabs.h a0, a0
; CHECK-NEXT:    ret
  %res = call <4 x i16> @llvm.abs.v4i16(<4 x i16> %a, i1 0)
  ret <4 x i16> %res
}

define <8 x i8> @test_pabs_b(<8 x i8> %a) {
; CHECK-LABEL: test_pabs_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pabs.b a0, a0
; CHECK-NEXT:    ret
  %res = call <8 x i8> @llvm.abs.v8i8(<8 x i8> %a, i1 0)
  ret <8 x i8> %res
}

; Test absolute difference signed for v2i32
; abds pattern: sub(smax(a,b), smin(a,b))
define <2 x i32> @test_pdif_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_pdif_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmin.w a2, a0, a1
; CHECK-NEXT:    pmax.w a0, a0, a1
; CHECK-NEXT:    psub.w a0, a0, a2
; CHECK-NEXT:    ret
  %min = call <2 x i32> @llvm.smin.v2i32(<2 x i32> %a, <2 x i32> %b)
  %max = call <2 x i32> @llvm.smax.v2i32(<2 x i32> %a, <2 x i32> %b)
  %res = sub <2 x i32> %max, %min
  ret <2 x i32> %res
}

; Test absolute difference unsigned for v2i32
; abdu pattern: sub(umax(a,b), umin(a,b))
define <2 x i32> @test_pdifu_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_pdifu_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pminu.w a2, a0, a1
; CHECK-NEXT:    pmaxu.w a0, a0, a1
; CHECK-NEXT:    psub.w a0, a0, a2
; CHECK-NEXT:    ret
  %min = call <2 x i32> @llvm.umin.v2i32(<2 x i32> %a, <2 x i32> %b)
  %max = call <2 x i32> @llvm.umax.v2i32(<2 x i32> %a, <2 x i32> %b)
  %res = sub <2 x i32> %max, %min
  ret <2 x i32> %res
}

; Test absolute difference signed for v4i16
; abds pattern: sub(smax(a,b), smin(a,b))
define <4 x i16> @test_pdif_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_pdif_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pabd.h a0, a0, a1
; CHECK-NEXT:    ret
  %min = call <4 x i16> @llvm.smin.v4i16(<4 x i16> %a, <4 x i16> %b)
  %max = call <4 x i16> @llvm.smax.v4i16(<4 x i16> %a, <4 x i16> %b)
  %res = sub <4 x i16> %max, %min
  ret <4 x i16> %res
}

; Test absolute difference unsigned for v4i16
; abdu pattern: sub(umax(a,b), umin(a,b))
define <4 x i16> @test_pdifu_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_pdifu_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pabdu.h a0, a0, a1
; CHECK-NEXT:    ret
  %min = call <4 x i16> @llvm.umin.v4i16(<4 x i16> %a, <4 x i16> %b)
  %max = call <4 x i16> @llvm.umax.v4i16(<4 x i16> %a, <4 x i16> %b)
  %res = sub <4 x i16> %max, %min
  ret <4 x i16> %res
}

; Test absolute difference signed for v8i8
define <8 x i8> @test_pdif_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_pdif_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pabd.b a0, a0, a1
; CHECK-NEXT:    ret
  %min = call <8 x i8> @llvm.smin.v8i8(<8 x i8> %a, <8 x i8> %b)
  %max = call <8 x i8> @llvm.smax.v8i8(<8 x i8> %a, <8 x i8> %b)
  %res = sub <8 x i8> %max, %min
  ret <8 x i8> %res
}

; Test absolute difference unsigned for v8i8
define <8 x i8> @test_pdifu_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_pdifu_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pabdu.b a0, a0, a1
; CHECK-NEXT:    ret
  %min = call <8 x i8> @llvm.umin.v8i8(<8 x i8> %a, <8 x i8> %b)
  %max = call <8 x i8> @llvm.umax.v8i8(<8 x i8> %a, <8 x i8> %b)
  %res = sub <8 x i8> %max, %min
  ret <8 x i8> %res
}

; Test averaging floor subtraction signed for v4i16
; pasub pattern: (a - b) arithmetic shift right 1
define <4 x i16> @test_pasub_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_pasub_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pasub.h a0, a0, a1
; CHECK-NEXT:    ret
  %a_ext = sext <4 x i16> %a to <4 x i32>
  %b_ext = sext <4 x i16> %b to <4 x i32>
  %sub = sub <4 x i32> %a_ext, %b_ext
  %res = ashr <4 x i32> %sub, <i32 1, i32 1, i32 1, i32 1>
  %res_trunc = trunc <4 x i32> %res to <4 x i16>
  ret <4 x i16> %res_trunc
}

; Test averaging floor subtraction unsigned for v4i16
; pasubu pattern: (a - b) logical shift right 1
define <4 x i16> @test_pasubu_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_pasubu_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pasubu.h a0, a0, a1
; CHECK-NEXT:    ret
  %a_ext = zext <4 x i16> %a to <4 x i32>
  %b_ext = zext <4 x i16> %b to <4 x i32>
  %sub = sub <4 x i32> %a_ext, %b_ext
  %res = lshr <4 x i32> %sub, <i32 1, i32 1, i32 1, i32 1>
  %res_trunc = trunc <4 x i32> %res to <4 x i16>
  ret <4 x i16> %res_trunc
}

; Test averaging floor subtraction signed for v8i8
define <8 x i8> @test_pasub_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_pasub_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pasub.b a0, a0, a1
; CHECK-NEXT:    ret
  %a_ext = sext <8 x i8> %a to <8 x i16>
  %b_ext = sext <8 x i8> %b to <8 x i16>
  %sub = sub <8 x i16> %a_ext, %b_ext
  %res = ashr <8 x i16> %sub, <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>
  %res_trunc = trunc <8 x i16> %res to <8 x i8>
  ret <8 x i8> %res_trunc
}

; Test averaging floor subtraction unsigned for v8i8
define <8 x i8> @test_pasubu_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_pasubu_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pasubu.b a0, a0, a1
; CHECK-NEXT:    ret
  %a_ext = zext <8 x i8> %a to <8 x i16>
  %b_ext = zext <8 x i8> %b to <8 x i16>
  %sub = sub <8 x i16> %a_ext, %b_ext
  %res = lshr <8 x i16> %sub, <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>
  %res_trunc = trunc <8 x i16> %res to <8 x i8>
  ret <8 x i8> %res_trunc
}

; Test PLI (pack load immediate) for v4i16
define <4 x i16> @test_pli_h() {
; CHECK-LABEL: test_pli_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pli.h a0, 100
; CHECK-NEXT:    ret
  ret <4 x i16> splat (i16 100)
}

define <2 x i32> @test_pli_h_v2i32() {
; CHECK-LABEL: test_pli_h_v2i32:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pli.h a0, 100
; CHECK-NEXT:    ret
  ret <2 x i32> splat (i32 u0x640064)
}

; Test PLI for v8i8 with unsigned immediate
define <8 x i8> @test_pli_b() {
; CHECK-LABEL: test_pli_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pli.b a0, 64
; CHECK-NEXT:    ret
  ret <8 x i8> splat (i8 64)
}

define <4 x i16> @test_pli_b_v4i16() {
; CHECK-LABEL: test_pli_b_v4i16:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pli.b a0, 64
; CHECK-NEXT:    ret
  ret <4 x i16> splat (i16 u0x4040)
}

define <2 x i32> @test_pli_b_v2i32() {
; CHECK-LABEL: test_pli_b_v2i32:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pli.b a0, 64
; CHECK-NEXT:    ret
  ret <2 x i32> splat (i32 u0x40404040)
}

; Test PLI for v2i32 with signed immediate
define <2 x i32> @test_pli_w() {
; CHECK-LABEL: test_pli_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pli.w a0, -256
; CHECK-NEXT:    ret
  ret <2 x i32> splat (i32 -256)
}

define <4 x i16> @test_plui_h() {
; CHECK-LABEL: test_plui_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    plui.h a0, 100
; CHECK-NEXT:    ret
  ret <4 x i16> splat (i16 u0x1900)
}

define <4 x i16> @test_plui_h_negative() {
; CHECK-LABEL: test_plui_h_negative:
; CHECK:       # %bb.0:
; CHECK-NEXT:    plui.h a0, -412
; CHECK-NEXT:    ret
  ret <4 x i16> splat (i16 u0x9900)
}

define <2 x i32> @test_plui_h_v2i32() {
; CHECK-LABEL: test_plui_h_v2i32:
; CHECK:       # %bb.0:
; CHECK-NEXT:    plui.h a0, 100
; CHECK-NEXT:    ret
  ret <2 x i32> splat (i32 u0x19001900)
}

define <2 x i32> @test_plui_h_negative_v2i32() {
; CHECK-LABEL: test_plui_h_negative_v2i32:
; CHECK:       # %bb.0:
; CHECK-NEXT:    plui.h a0, -412
; CHECK-NEXT:    ret
  ret <2 x i32> splat (i32 u0x99009900)
}

define <2 x i32> @test_plui_w() {
; CHECK-LABEL: test_plui_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    plui.w a0, 75
; CHECK-NEXT:    ret
  ret <2 x i32> splat (i32 u0x12c00000)
}

define <2 x i32> @test_plui_w_negative() {
; CHECK-LABEL: test_plui_w_negative:
; CHECK:       # %bb.0:
; CHECK-NEXT:    plui.w a0, -218
; CHECK-NEXT:    ret
  ret <2 x i32> splat (i32 u0xc9800000)
}

define <8 x i8> @test_allones_v8i8() {
; CHECK-LABEL: test_allones_v8i8:
; CHECK:       # %bb.0:
; CHECK-NEXT:    li a0, -1
; CHECK-NEXT:    ret
  ret <8 x i8> splat (i8 -1)
}

define <4 x i16> @test_allones_v4i16() {
; CHECK-LABEL: test_allones_v4i16:
; CHECK:       # %bb.0:
; CHECK-NEXT:    li a0, -1
; CHECK-NEXT:    ret
  ret <4 x i16> splat (i16 -1)
}

define <2 x i32> @test_allones_v2i32() {
; CHECK-LABEL: test_allones_v2i32:
; CHECK:       # %bb.0:
; CHECK-NEXT:    li a0, -1
; CHECK-NEXT:    ret
  ret <2 x i32> splat (i32 -1)
}

define i16 @test_extract_vector_16(<4 x i16> %a) {
; CHECK-LABEL: test_extract_vector_16:
; CHECK:       # %bb.0:
; CHECK-NEXT:    ret
  %extracted = extractelement <4 x i16> %a, i32 0
  ret i16 %extracted
}

define i8 @test_extract_vector_8(<8 x i8> %a) {
; CHECK-LABEL: test_extract_vector_8:
; CHECK:       # %bb.0:
; CHECK-NEXT:    ret
  %extracted = extractelement <8 x i8> %a, i32 0
  ret i8 %extracted
}

define i32 @test_extract_vector_32(<2 x i32> %a) {
; CHECK-LABEL: test_extract_vector_32:
; CHECK:       # %bb.0:
; CHECK-NEXT:    ret
  %extracted = extractelement <2 x i32> %a, i32 0
  ret i32 %extracted
}

define i32 @test_extract_vector_32_elem1(<2 x i32> %a) {
; CHECK-LABEL: test_extract_vector_32_elem1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    srli a0, a0, 32
; CHECK-NEXT:    ret
  %extracted = extractelement <2 x i32> %a, i32 1
  ret i32 %extracted
}

define <4 x i16> @test_insert_vector_16(<4 x i16> %a, i16 %val) {
; CHECK-LABEL: test_insert_vector_16:
; CHECK:       # %bb.0:
; CHECK-NEXT:    lui a2, 16
; CHECK-NEXT:    addi a2, a2, -1
; CHECK-NEXT:    mvm a0, a1, a2
; CHECK-NEXT:    ret
  %res = insertelement <4 x i16> %a, i16 %val, i32 0
  ret <4 x i16> %res
}

define <4 x i16> @test_insert_vector_16_elem2(<4 x i16> %a, i16 %val) {
; CHECK-LABEL: test_insert_vector_16_elem2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    slli a1, a1, 32
; CHECK-NEXT:    lui a2, 65535
; CHECK-NEXT:    slli a2, a2, 20
; CHECK-NEXT:    mvm a0, a1, a2
; CHECK-NEXT:    ret
  %res = insertelement <4 x i16> %a, i16 %val, i32 2
  ret <4 x i16> %res
}

define <8 x i8> @test_insert_vector_8(<8 x i8> %a, i8 %val) {
; CHECK-LABEL: test_insert_vector_8:
; CHECK:       # %bb.0:
; CHECK-NEXT:    li a2, 255
; CHECK-NEXT:    mvm a0, a1, a2
; CHECK-NEXT:    ret
  %res = insertelement <8 x i8> %a, i8 %val, i32 0
  ret <8 x i8> %res
}

define <8 x i8> @test_insert_vector_8_elem3(<8 x i8> %a, i8 %val) {
; CHECK-LABEL: test_insert_vector_8_elem3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    slli a1, a1, 24
; CHECK-NEXT:    li a2, 255
; CHECK-NEXT:    slli a2, a2, 24
; CHECK-NEXT:    mvm a0, a1, a2
; CHECK-NEXT:    ret
  %res = insertelement <8 x i8> %a, i8 %val, i32 3
  ret <8 x i8> %res
}

define <2 x i32> @test_insert_vector_32(<2 x i32> %a, i32 %val) {
; CHECK-LABEL: test_insert_vector_32:
; CHECK:       # %bb.0:
; CHECK-NEXT:    srli a0, a0, 32
; CHECK-NEXT:    pack a0, a1, a0
; CHECK-NEXT:    ret
  %res = insertelement <2 x i32> %a, i32 %val, i32 0
  ret <2 x i32> %res
}

define <2 x i32> @test_insert_vector_32_elem1(<2 x i32> %a, i32 %val) {
; CHECK-LABEL: test_insert_vector_32_elem1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pack a0, a0, a1
; CHECK-NEXT:    ret
  %res = insertelement <2 x i32> %a, i32 %val, i32 1
  ret <2 x i32> %res
}

; Test basic add/sub operations for v2i32 (RV64 only)
define <2 x i32> @test_padd_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_padd_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    padd.w a0, a0, a1
; CHECK-NEXT:    ret
  %res = add <2 x i32> %a, %b
  ret <2 x i32> %res
}

define <2 x i32> @test_psub_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_psub_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psub.w a0, a0, a1
; CHECK-NEXT:    ret
  %res = sub <2 x i32> %a, %b
  ret <2 x i32> %res
}

; Test saturating add operations for v2i32 (RV64 only)
define <2 x i32> @test_psadd_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_psadd_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psadd.w a0, a0, a1
; CHECK-NEXT:    ret
  %res = call <2 x i32> @llvm.sadd.sat.v2i32(<2 x i32> %a, <2 x i32> %b)
  ret <2 x i32> %res
}

define <2 x i32> @test_psaddu_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_psaddu_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psaddu.w a0, a0, a1
; CHECK-NEXT:    ret
  %res = call <2 x i32> @llvm.uadd.sat.v2i32(<2 x i32> %a, <2 x i32> %b)
  ret <2 x i32> %res
}

; Test saturating sub operations for v2i32 (RV64 only)
define <2 x i32> @test_pssub_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_pssub_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pssub.w a0, a0, a1
; CHECK-NEXT:    ret
  %res = call <2 x i32> @llvm.ssub.sat.v2i32(<2 x i32> %a, <2 x i32> %b)
  ret <2 x i32> %res
}

define <2 x i32> @test_pssubu_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_pssubu_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pssubu.w a0, a0, a1
; CHECK-NEXT:    ret
  %res = call <2 x i32> @llvm.usub.sat.v2i32(<2 x i32> %a, <2 x i32> %b)
  ret <2 x i32> %res
}

; Test averaging floor signed operations for v2i32 (RV64 only)
; avgfloors pattern: (a + b) arithmetic shift right 1
define <2 x i32> @test_paadd_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_paadd_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    paadd.w a0, a0, a1
; CHECK-NEXT:    ret
  %ext.a = sext <2 x i32> %a to <2 x i64>
  %ext.b = sext <2 x i32> %b to <2 x i64>
  %add = add nsw <2 x i64> %ext.a, %ext.b
  %shift = ashr <2 x i64> %add, <i64 1, i64 1>
  %res = trunc <2 x i64> %shift to <2 x i32>
  ret <2 x i32> %res
}

; Test averaging floor unsigned operations for v2i32 (RV64 only)
; avgflooru pattern: (a & b) + ((a ^ b) >> 1)
define <2 x i32> @test_paaddu_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_paaddu_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    paaddu.w a0, a0, a1
; CHECK-NEXT:    ret
  %and = and <2 x i32> %a, %b
  %xor = xor <2 x i32> %a, %b
  %shift = lshr <2 x i32> %xor, <i32 1, i32 1>
  %res = add <2 x i32> %and, %shift
  ret <2 x i32> %res
}

; Test averaging floor subtraction signed for v2i32 (RV64 only)
; pasub pattern: (a - b) arithmetic shift right 1
define <2 x i32> @test_pasub_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_pasub_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pasub.w a0, a0, a1
; CHECK-NEXT:    ret
  %a_ext = sext <2 x i32> %a to <2 x i64>
  %b_ext = sext <2 x i32> %b to <2 x i64>
  %sub = sub <2 x i64> %a_ext, %b_ext
  %res = ashr <2 x i64> %sub, <i64 1, i64 1>
  %res_trunc = trunc <2 x i64> %res to <2 x i32>
  ret <2 x i32> %res_trunc
}

; Test averaging floor subtraction unsigned for v2i32 (RV64 only)
; pasubu pattern: (a - b) logical shift right 1
define <2 x i32> @test_pasubu_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_pasubu_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pasubu.w a0, a0, a1
; CHECK-NEXT:    ret
  %a_ext = zext <2 x i32> %a to <2 x i64>
  %b_ext = zext <2 x i32> %b to <2 x i64>
  %sub = sub <2 x i64> %a_ext, %b_ext
  %res = lshr <2 x i64> %sub, <i64 1, i64 1>
  %res_trunc = trunc <2 x i64> %res to <2 x i32>
  ret <2 x i32> %res_trunc
}

; Test for splat
define <2 x i32> @test_non_const_splat_i32(i32 %elt) {
; CHECK-LABEL: test_non_const_splat_i32:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmv.ws a0, a0
; CHECK-NEXT:    ret
  %insert = insertelement <2 x i32> poison, i32 %elt, i32 0
  %splat = shufflevector <2 x i32> %insert, <2 x i32> poison, <2 x i32> zeroinitializer
  ret <2 x i32> %splat
}

; Test add(vec, splat(scalar)) pattern
define <8 x i8> @test_padd_bs_splat_lhs(<8 x i8> %a, i8 %b) {
; CHECK-LABEL: test_padd_bs_splat_lhs:
; CHECK:       # %bb.0:
; CHECK-NEXT:    padd.bs a0, a0, a1
; CHECK-NEXT:    ret
  %insert = insertelement <8 x i8> poison, i8 %b, i32 0
  %splat = shufflevector <8 x i8> %insert, <8 x i8> poison, <8 x i32> zeroinitializer
  %res = add <8 x i8> %splat, %a
  ret <8 x i8> %res
}

define <8 x i8> @test_padd_bs_splat_rhs(<8 x i8> %a, i8 %b) {
; CHECK-LABEL: test_padd_bs_splat_rhs:
; CHECK:       # %bb.0:
; CHECK-NEXT:    padd.bs a0, a0, a1
; CHECK-NEXT:    ret
  %insert = insertelement <8 x i8> poison, i8 %b, i32 0
  %splat = shufflevector <8 x i8> %insert, <8 x i8> poison, <8 x i32> zeroinitializer
  %res = add <8 x i8> %a, %splat
  ret <8 x i8> %res
}

define <4 x i16> @test_padd_hs_splat_lhs(<4 x i16> %a, i16 %b) {
; CHECK-LABEL: test_padd_hs_splat_lhs:
; CHECK:       # %bb.0:
; CHECK-NEXT:    padd.hs a0, a0, a1
; CHECK-NEXT:    ret
  %insert = insertelement <4 x i16> poison, i16 %b, i32 0
  %splat = shufflevector <4 x i16> %insert, <4 x i16> poison, <4 x i32> zeroinitializer
  %res = add <4 x i16> %splat, %a
  ret <4 x i16> %res
}

define <4 x i16> @test_padd_hs_splat_rhs(<4 x i16> %a, i16 %b) {
; CHECK-LABEL: test_padd_hs_splat_rhs:
; CHECK:       # %bb.0:
; CHECK-NEXT:    padd.hs a0, a0, a1
; CHECK-NEXT:    ret
  %insert = insertelement <4 x i16> poison, i16 %b, i32 0
  %splat = shufflevector <4 x i16> %insert, <4 x i16> poison, <4 x i32> zeroinitializer
  %res = add <4 x i16> %a, %splat
  ret <4 x i16> %res
}

define <2 x i32> @test_padd_ws_splat_lhs(<2 x i32> %a, i32 %b) {
; CHECK-LABEL: test_padd_ws_splat_lhs:
; CHECK:       # %bb.0:
; CHECK-NEXT:    padd.ws a0, a0, a1
; CHECK-NEXT:    ret
  %insert = insertelement <2 x i32> poison, i32 %b, i32 0
  %splat = shufflevector <2 x i32> %insert, <2 x i32> poison, <2 x i32> zeroinitializer
  %res = add <2 x i32> %splat, %a
  ret <2 x i32> %res
}

define <2 x i32> @test_padd_ws_splat_rhs(<2 x i32> %a, i32 %b) {
; CHECK-LABEL: test_padd_ws_splat_rhs:
; CHECK:       # %bb.0:
; CHECK-NEXT:    padd.ws a0, a0, a1
; CHECK-NEXT:    ret
  %insert = insertelement <2 x i32> poison, i32 %b, i32 0
  %splat = shufflevector <2 x i32> %insert, <2 x i32> poison, <2 x i32> zeroinitializer
  %res = add <2 x i32> %a, %splat
  ret <2 x i32> %res
}

define <8 x i8> @test_build_vector_i8(i8 %a, i8 %b, i8 %c, i8 %d, i8 %e, i8 %f, i8 %g, i8 %h) {
; CHECK-LABEL: test_build_vector_i8:
; CHECK:       # %bb.0:
; CHECK-NEXT:    ppaire.b a6, a6, a7
; CHECK-NEXT:    ppaire.b a4, a4, a5
; CHECK-NEXT:    ppaire.b a2, a2, a3
; CHECK-NEXT:    ppaire.b a0, a0, a1
; CHECK-NEXT:    ppaire.h a1, a4, a6
; CHECK-NEXT:    ppaire.h a0, a0, a2
; CHECK-NEXT:    pack a0, a0, a1
; CHECK-NEXT:    ret
  %v0 = insertelement <8 x i8> poison, i8 %a, i32 0
  %v1 = insertelement <8 x i8> %v0, i8 %b, i32 1
  %v2 = insertelement <8 x i8> %v1, i8 %c, i32 2
  %v3 = insertelement <8 x i8> %v2, i8 %d, i32 3
  %v4 = insertelement <8 x i8> %v3, i8 %e, i32 4
  %v5 = insertelement <8 x i8> %v4, i8 %f, i32 5
  %v6 = insertelement <8 x i8> %v5, i8 %g, i32 6
  %v7 = insertelement <8 x i8> %v6, i8 %h, i32 7
  ret <8 x i8> %v7
}

define <4 x i16> @test_build_vector_i16(i16 %a, i16 %b, i16 %c, i16 %d) {
; CHECK-LABEL: test_build_vector_i16:
; CHECK:       # %bb.0:
; CHECK-NEXT:    ppaire.h a2, a2, a3
; CHECK-NEXT:    ppaire.h a0, a0, a1
; CHECK-NEXT:    pack a0, a0, a2
; CHECK-NEXT:    ret
  %v0 = insertelement <4 x i16> poison, i16 %a, i32 0
  %v1 = insertelement <4 x i16> %v0, i16 %b, i32 1
  %v2 = insertelement <4 x i16> %v1, i16 %c, i32 2
  %v3 = insertelement <4 x i16> %v2, i16 %d, i32 3
  ret <4 x i16> %v3
}

define <2 x i32> @test_build_vector_i32(i32 %a, i32 %b) {
; CHECK-LABEL: test_build_vector_i32:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pack a0, a0, a1
; CHECK-NEXT:    ret
  %v0 = insertelement <2 x i32> poison, i32 %a, i32 0
  %v1 = insertelement <2 x i32> %v0, i32 %b, i32 1
  ret <2 x i32> %v1
}

; Test logical shift left immediate for v4i16
define <4 x i16> @test_pslli_h(<4 x i16> %a) {
; CHECK-LABEL: test_pslli_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pslli.h a0, a0, 2
; CHECK-NEXT:    ret
  %res = shl <4 x i16> %a, splat(i16 2)
  ret <4 x i16> %res
}

; Test logical shift left immediate for v8i8
define <8 x i8> @test_pslli_b(<8 x i8> %a) {
; CHECK-LABEL: test_pslli_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pslli.b a0, a0, 2
; CHECK-NEXT:    ret
  %res = shl <8 x i8> %a, splat(i8 2)
  ret <8 x i8> %res
}

; Test logical shift left immediate for v2i32
define <2 x i32> @test_pslli_w(<2 x i32> %a) {
; CHECK-LABEL: test_pslli_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pslli.w a0, a0, 2
; CHECK-NEXT:    ret
  %res = shl <2 x i32> %a, splat(i32 2)
  ret <2 x i32> %res
}

; Test logical shift right immediate
define <2 x i32> @test_psrli_w(<2 x i32> %a) {
; CHECK-LABEL: test_psrli_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psrli.w a0, a0, 2
; CHECK-NEXT:    ret
  %res = lshr <2 x i32> %a, splat(i32 2)
  ret <2 x i32> %res
}

define <4 x i16> @test_psrli_h(<4 x i16> %a) {
; CHECK-LABEL: test_psrli_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psrli.h a0, a0, 2
; CHECK-NEXT:    ret
  %res = lshr <4 x i16> %a, splat(i16 2)
  ret <4 x i16> %res
}

define <8 x i8> @test_psrli_b(<8 x i8> %a) {
; CHECK-LABEL: test_psrli_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psrli.b a0, a0, 2
; CHECK-NEXT:    ret
  %res = lshr <8 x i8> %a, splat(i8 2)
  ret <8 x i8> %res
}

; Test arithmetic shift right immediate
define <2 x i32> @test_psrai_w(<2 x i32> %a) {
; CHECK-LABEL: test_psrai_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psrai.w a0, a0, 2
; CHECK-NEXT:    ret
  %res = ashr <2 x i32> %a, splat(i32 2)
  ret <2 x i32> %res
}

define <4 x i16> @test_psrai_h(<4 x i16> %a) {
; CHECK-LABEL: test_psrai_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psrai.h a0, a0, 2
; CHECK-NEXT:    ret
  %res = ashr <4 x i16> %a, splat(i16 2)
  ret <4 x i16> %res
}

define <8 x i8> @test_psrai_b(<8 x i8> %a) {
; CHECK-LABEL: test_psrai_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psrai.b a0, a0, 2
; CHECK-NEXT:    ret
  %res = ashr <8 x i8> %a, splat(i8 2)
  ret <8 x i8> %res
}

; Test saturating shift left arithmetic with splat immediate shift amount for v2i32
define <2 x i32> @test_psslai_w(<2 x i32> %a) {
; CHECK-LABEL: test_psslai_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psslai.w a0, a0, 2
; CHECK-NEXT:    ret
  %res = call <2 x i32> @llvm.sshl.sat.v2i32(<2 x i32> %a, <2 x i32> splat(i32 2))
  ret <2 x i32> %res
}

; Test saturating shift left arithmetic with splat immediate shift amount for v4i16
define <4 x i16> @test_psslai_h(<4 x i16> %a) {
; CHECK-LABEL: test_psslai_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psslai.h a0, a0, 2
; CHECK-NEXT:    ret
  %res = call <4 x i16> @llvm.sshl.sat.v4i16(<4 x i16> %a, <4 x i16> splat(i16 2))
  ret <4 x i16> %res
}

; Test saturating shift left arithmetic with splat immediate shift amount for v8i8
define <8 x i8> @test_psslai_b(<8 x i8> %a) {
; CHECK-LABEL: test_psslai_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmsltz.b a1, a0
; CHECK-NEXT:    pli.b a2, -128
; CHECK-NEXT:    pli.b a3, 127
; CHECK-NEXT:    merge a1, a3, a2
; CHECK-NEXT:    pslli.b a2, a0, 2
; CHECK-NEXT:    psrai.b a3, a2, 2
; CHECK-NEXT:    pmseq.b a0, a0, a3
; CHECK-NEXT:    merge a0, a1, a2
; CHECK-NEXT:    ret
  %res = call <8 x i8> @llvm.sshl.sat.v8i8(<8 x i8> %a, <8 x i8> splat(i8 2))
  ret <8 x i8> %res
}

; Test saturating shift left arithmetic with splat shift amount for v4i16
define <4 x i16> @test_pssla_hs(<4 x i16> %a, i16 %shamt) {
; CHECK-LABEL: test_pssla_hs:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmsltz.h a2, a0
; CHECK-NEXT:    lui a3, 8
; CHECK-NEXT:    plui.h a4, -512
; CHECK-NEXT:    addi a3, a3, -1
; CHECK-NEXT:    pmv.hs a3, a3
; CHECK-NEXT:    merge a2, a3, a4
; CHECK-NEXT:    psll.hs a3, a0, a1
; CHECK-NEXT:    psra.hs a1, a3, a1
; CHECK-NEXT:    pmseq.h a0, a0, a1
; CHECK-NEXT:    merge a0, a2, a3
; CHECK-NEXT:    ret
  %insert = insertelement <4 x i16> poison, i16 %shamt, i32 0
  %b = shufflevector <4 x i16> %insert, <4 x i16> poison, <4 x i32> zeroinitializer
  %res = call <4 x i16> @llvm.sshl.sat.v4i16(<4 x i16> %a, <4 x i16> %b)
  ret <4 x i16> %res
}

; Test saturating shift left arithmetic with splat shift amount for v2i32
define <2 x i32> @test_pssla_ws(<2 x i32> %a, i32 %shamt) {
; CHECK-LABEL: test_pssla_ws:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmsltz.w a2, a0
; CHECK-NEXT:    lui a3, 524288
; CHECK-NEXT:    plui.w a4, -512
; CHECK-NEXT:    addiw a3, a3, -1
; CHECK-NEXT:    pmv.ws a3, a3
; CHECK-NEXT:    merge a2, a3, a4
; CHECK-NEXT:    psll.ws a3, a0, a1
; CHECK-NEXT:    psra.ws a1, a3, a1
; CHECK-NEXT:    pmseq.w a0, a0, a1
; CHECK-NEXT:    merge a0, a2, a3
; CHECK-NEXT:    ret
  %insert = insertelement <2 x i32> poison, i32 %shamt, i32 0
  %b = shufflevector <2 x i32> %insert, <2 x i32> poison, <2 x i32> zeroinitializer
  %res = call <2 x i32> @llvm.sshl.sat.v2i32(<2 x i32> %a, <2 x i32> %b)
  ret <2 x i32> %res
}

; Test saturating shift left arithmetic with non-splat shift amount for v4i16
define <4 x i16> @test_pssla_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_pssla_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    srli a2, a1, 48
; CHECK-NEXT:    srli a3, a0, 48
; CHECK-NEXT:    srli a4, a1, 32
; CHECK-NEXT:    srli a5, a0, 32
; CHECK-NEXT:    sll a6, a0, a1
; CHECK-NEXT:    srli a7, a1, 16
; CHECK-NEXT:    srli t0, a0, 16
; CHECK-NEXT:    pmsltz.h t1, a0
; CHECK-NEXT:    sll a3, a3, a2
; CHECK-NEXT:    sll a5, a5, a4
; CHECK-NEXT:    sll t0, t0, a7
; CHECK-NEXT:    sext.h t2, a6
; CHECK-NEXT:    sra a1, t2, a1
; CHECK-NEXT:    ppaire.h t2, a5, a3
; CHECK-NEXT:    ppaire.h a6, a6, t0
; CHECK-NEXT:    pack a6, a6, t2
; CHECK-NEXT:    lui t2, 8
; CHECK-NEXT:    sext.h a3, a3
; CHECK-NEXT:    sra a2, a3, a2
; CHECK-NEXT:    plui.h a3, -512
; CHECK-NEXT:    addi t2, t2, -1
; CHECK-NEXT:    sext.h a5, a5
; CHECK-NEXT:    sext.h t0, t0
; CHECK-NEXT:    pmv.hs t2, t2
; CHECK-NEXT:    sra a4, a5, a4
; CHECK-NEXT:    sra a5, t0, a7
; CHECK-NEXT:    ppaire.h a2, a4, a2
; CHECK-NEXT:    ppaire.h a1, a1, a5
; CHECK-NEXT:    pack a1, a1, a2
; CHECK-NEXT:    pmseq.h a0, a0, a1
; CHECK-NEXT:    merge t1, t2, a3
; CHECK-NEXT:    merge a0, t1, a6
; CHECK-NEXT:    ret
  %res = call <4 x i16> @llvm.sshl.sat.v4i16(<4 x i16> %a, <4 x i16> %b)
  ret <4 x i16> %res
}

; Test saturating shift left arithmetic with non-splat shift amount for v2i32
define <2 x i32> @test_pssla_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_pssla_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    sllw a2, a0, a1
; CHECK-NEXT:    srli a3, a1, 32
; CHECK-NEXT:    srli a4, a0, 32
; CHECK-NEXT:    pmsltz.w a5, a0
; CHECK-NEXT:    sllw a4, a4, a3
; CHECK-NEXT:    sraw a1, a2, a1
; CHECK-NEXT:    pack a2, a2, a4
; CHECK-NEXT:    sraw a3, a4, a3
; CHECK-NEXT:    lui a4, 524288
; CHECK-NEXT:    pack a1, a1, a3
; CHECK-NEXT:    plui.w a3, -512
; CHECK-NEXT:    addiw a4, a4, -1
; CHECK-NEXT:    pmv.ws a4, a4
; CHECK-NEXT:    pmseq.w a0, a0, a1
; CHECK-NEXT:    merge a5, a4, a3
; CHECK-NEXT:    merge a0, a5, a2
; CHECK-NEXT:    ret
  %res = call <2 x i32> @llvm.sshl.sat.v2i32(<2 x i32> %a, <2 x i32> %b)
  ret <2 x i32> %res
}

; Test logical shift left(scalar shamt)
define <4 x i16> @test_psll_hs(<4 x i16> %a, i16 %shamt) {
; CHECK-LABEL: test_psll_hs:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psll.hs a0, a0, a1
; CHECK-NEXT:    ret
  %insert = insertelement <4 x i16> poison, i16 %shamt, i32 0
  %b = shufflevector <4 x i16> %insert, <4 x i16> poison, <4 x i32> zeroinitializer
  %res = shl <4 x i16> %a, %b
  ret <4 x i16> %res
}

define <4 x i16> @test_psll_hs_mask(<4 x i16> %a, i16 %shamt) {
; CHECK-LABEL: test_psll_hs_mask:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psll.hs a0, a0, a1
; CHECK-NEXT:    ret
  %masked = and i16 %shamt, 15
  %insert = insertelement <4 x i16> poison, i16 %masked, i32 0
  %b = shufflevector <4 x i16> %insert, <4 x i16> poison, <4 x i32> zeroinitializer
  %res = shl <4 x i16> %a, %b
  ret <4 x i16> %res
}

define <8 x i8> @test_psll_bs(<8 x i8> %a, i8 %shamt) {
; CHECK-LABEL: test_psll_bs:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psll.bs a0, a0, a1
; CHECK-NEXT:    ret
  %insert = insertelement <8 x i8> poison, i8 %shamt, i32 0
  %b = shufflevector <8 x i8> %insert, <8 x i8> poison, <8 x i32> zeroinitializer
  %res = shl <8 x i8> %a, %b
  ret <8 x i8> %res
}

define <8 x i8> @test_psll_bs_mask(<8 x i8> %a, i8 %shamt) {
; CHECK-LABEL: test_psll_bs_mask:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psll.bs a0, a0, a1
; CHECK-NEXT:    ret
  %masked = and i8 %shamt, 7
  %insert = insertelement <8 x i8> poison, i8 %masked, i32 0
  %b = shufflevector <8 x i8> %insert, <8 x i8> poison, <8 x i32> zeroinitializer
  %res = shl <8 x i8> %a, %b
  ret <8 x i8> %res
}

define <2 x i32> @test_psll_ws(<2 x i32> %a, i32 %shamt) {
; CHECK-LABEL: test_psll_ws:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psll.ws a0, a0, a1
; CHECK-NEXT:    ret
  %insert = insertelement <2 x i32> poison, i32 %shamt, i32 0
  %b = shufflevector <2 x i32> %insert, <2 x i32> poison, <2 x i32> zeroinitializer
  %res = shl <2 x i32> %a, %b
  ret <2 x i32> %res
}

define <2 x i32> @test_psll_ws_mask(<2 x i32> %a, i32 %shamt) {
; CHECK-LABEL: test_psll_ws_mask:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psll.ws a0, a0, a1
; CHECK-NEXT:    ret
  %masked = and i32 %shamt, 31
  %insert = insertelement <2 x i32> poison, i32 %masked, i32 0
  %b = shufflevector <2 x i32> %insert, <2 x i32> poison, <2 x i32> zeroinitializer
  %res = shl <2 x i32> %a, %b
  ret <2 x i32> %res
}

; Test logical shift left(vector shamt)
define <2 x i32> @test_psll_ws_vec_shamt(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_psll_ws_vec_shamt:
; CHECK:       # %bb.0:
; CHECK-NEXT:    sllw a2, a0, a1
; CHECK-NEXT:    srli a1, a1, 32
; CHECK-NEXT:    srli a0, a0, 32
; CHECK-NEXT:    sllw a0, a0, a1
; CHECK-NEXT:    pack a0, a2, a0
; CHECK-NEXT:    ret
  %res = shl <2 x i32> %a, %b
  ret <2 x i32> %res
}

; Test logical shift right(scalar shamt)
define <4 x i16> @test_psrl_hs(<4 x i16> %a, i16 %shamt) {
; CHECK-LABEL: test_psrl_hs:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psrl.hs a0, a0, a1
; CHECK-NEXT:    ret
  %insert = insertelement <4 x i16> poison, i16 %shamt, i32 0
  %b = shufflevector <4 x i16> %insert, <4 x i16> poison, <4 x i32> zeroinitializer
  %res = lshr <4 x i16> %a, %b
  ret <4 x i16> %res
}

define <4 x i16> @test_psrl_hs_mask(<4 x i16> %a, i16 %shamt) {
; CHECK-LABEL: test_psrl_hs_mask:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psrl.hs a0, a0, a1
; CHECK-NEXT:    ret
  %masked = and i16 %shamt, 15
  %insert = insertelement <4 x i16> poison, i16 %masked, i32 0
  %b = shufflevector <4 x i16> %insert, <4 x i16> poison, <4 x i32> zeroinitializer
  %res = lshr <4 x i16> %a, %b
  ret <4 x i16> %res
}

define <8 x i8> @test_psrl_bs(<8 x i8> %a, i8 %shamt) {
; CHECK-LABEL: test_psrl_bs:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psrl.bs a0, a0, a1
; CHECK-NEXT:    ret
  %insert = insertelement <8 x i8> poison, i8 %shamt, i32 0
  %b = shufflevector <8 x i8> %insert, <8 x i8> poison, <8 x i32> zeroinitializer
  %res = lshr <8 x i8> %a, %b
  ret <8 x i8> %res
}

define <8 x i8> @test_psrl_bs_mask(<8 x i8> %a, i8 %shamt) {
; CHECK-LABEL: test_psrl_bs_mask:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psrl.bs a0, a0, a1
; CHECK-NEXT:    ret
  %masked = and i8 %shamt, 7
  %insert = insertelement <8 x i8> poison, i8 %masked, i32 0
  %b = shufflevector <8 x i8> %insert, <8 x i8> poison, <8 x i32> zeroinitializer
  %res = lshr <8 x i8> %a, %b
  ret <8 x i8> %res
}

define <2 x i32> @test_psrl_ws(<2 x i32> %a, i32 %shamt) {
; CHECK-LABEL: test_psrl_ws:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psrl.ws a0, a0, a1
; CHECK-NEXT:    ret
  %insert = insertelement <2 x i32> poison, i32 %shamt, i32 0
  %b = shufflevector <2 x i32> %insert, <2 x i32> poison, <2 x i32> zeroinitializer
  %res = lshr <2 x i32> %a, %b
  ret <2 x i32> %res
}

define <2 x i32> @test_psrl_ws_mask(<2 x i32> %a, i32 %shamt) {
; CHECK-LABEL: test_psrl_ws_mask:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psrl.ws a0, a0, a1
; CHECK-NEXT:    ret
  %masked = and i32 %shamt, 31
  %insert = insertelement <2 x i32> poison, i32 %masked, i32 0
  %b = shufflevector <2 x i32> %insert, <2 x i32> poison, <2 x i32> zeroinitializer
  %res = lshr <2 x i32> %a, %b
  ret <2 x i32> %res
}

; Test arithmetic shift right(scalar shamt)
define <4 x i16> @test_psra_hs(<4 x i16> %a, i16 %shamt) {
; CHECK-LABEL: test_psra_hs:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psra.hs a0, a0, a1
; CHECK-NEXT:    ret
  %insert = insertelement <4 x i16> poison, i16 %shamt, i32 0
  %b = shufflevector <4 x i16> %insert, <4 x i16> poison, <4 x i32> zeroinitializer
  %res = ashr <4 x i16> %a, %b
  ret <4 x i16> %res
}

define <4 x i16> @test_psra_hs_mask(<4 x i16> %a, i16 %shamt) {
; CHECK-LABEL: test_psra_hs_mask:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psra.hs a0, a0, a1
; CHECK-NEXT:    ret
  %masked = and i16 %shamt, 15
  %insert = insertelement <4 x i16> poison, i16 %masked, i32 0
  %b = shufflevector <4 x i16> %insert, <4 x i16> poison, <4 x i32> zeroinitializer
  %res = ashr <4 x i16> %a, %b
  ret <4 x i16> %res
}

define <8 x i8> @test_psra_bs(<8 x i8> %a, i8 %shamt) {
; CHECK-LABEL: test_psra_bs:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psra.bs a0, a0, a1
; CHECK-NEXT:    ret
  %insert = insertelement <8 x i8> poison, i8 %shamt, i32 0
  %b = shufflevector <8 x i8> %insert, <8 x i8> poison, <8 x i32> zeroinitializer
  %res = ashr <8 x i8> %a, %b
  ret <8 x i8> %res
}

define <8 x i8> @test_psra_bs_mask(<8 x i8> %a, i8 %shamt) {
; CHECK-LABEL: test_psra_bs_mask:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psra.bs a0, a0, a1
; CHECK-NEXT:    ret
  %masked = and i8 %shamt, 7
  %insert = insertelement <8 x i8> poison, i8 %masked, i32 0
  %b = shufflevector <8 x i8> %insert, <8 x i8> poison, <8 x i32> zeroinitializer
  %res = ashr <8 x i8> %a, %b
  ret <8 x i8> %res
}

define <2 x i32> @test_psra_ws(<2 x i32> %a, i32 %shamt) {
; CHECK-LABEL: test_psra_ws:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psra.ws a0, a0, a1
; CHECK-NEXT:    ret
  %insert = insertelement <2 x i32> poison, i32 %shamt, i32 0
  %b = shufflevector <2 x i32> %insert, <2 x i32> poison, <2 x i32> zeroinitializer
  %res = ashr <2 x i32> %a, %b
  ret <2 x i32> %res
}

define <2 x i32> @test_psra_ws_mask(<2 x i32> %a, i32 %shamt) {
; CHECK-LABEL: test_psra_ws_mask:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psra.ws a0, a0, a1
; CHECK-NEXT:    ret
  %masked = and i32 %shamt, 31
  %insert = insertelement <2 x i32> poison, i32 %masked, i32 0
  %b = shufflevector <2 x i32> %insert, <2 x i32> poison, <2 x i32> zeroinitializer
  %res = ashr <2 x i32> %a, %b
  ret <2 x i32> %res
}

; Test logical shift right(vector shamt)
define <2 x i32> @test_psrl_ws_vec_shamt(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_psrl_ws_vec_shamt:
; CHECK:       # %bb.0:
; CHECK-NEXT:    srlw a2, a0, a1
; CHECK-NEXT:    srli a1, a1, 32
; CHECK-NEXT:    srli a0, a0, 32
; CHECK-NEXT:    srlw a0, a0, a1
; CHECK-NEXT:    pack a0, a2, a0
; CHECK-NEXT:    ret
  %res = lshr <2 x i32> %a, %b
  ret <2 x i32> %res
}

; Test arithmetic shift right(vector shamt)
define <2 x i32> @test_psra_ws_vec_shamt(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_psra_ws_vec_shamt:
; CHECK:       # %bb.0:
; CHECK-NEXT:    sraw a2, a0, a1
; CHECK-NEXT:    srli a1, a1, 32
; CHECK-NEXT:    srli a0, a0, 32
; CHECK-NEXT:    sraw a0, a0, a1
; CHECK-NEXT:    pack a0, a2, a0
; CHECK-NEXT:    ret
  %res = ashr <2 x i32> %a, %b
  ret <2 x i32> %res
}

; Test packed multiply high signed
define <4 x i16> @test_pmulh_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_pmulh_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmulh.h a0, a0, a1
; CHECK-NEXT:    ret
  %a_ext = sext <4 x i16> %a to <4 x i32>
  %b_ext = sext <4 x i16> %b to <4 x i32>
  %mul = mul <4 x i32> %a_ext, %b_ext
  %shift = lshr <4 x i32> %mul, <i32 16, i32 16, i32 16, i32 16>
  %res = trunc <4 x i32> %shift to <4 x i16>
  ret <4 x i16> %res
}

define <2 x i32> @test_pmulh_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_pmulh_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmulh.w a0, a0, a1
; CHECK-NEXT:    ret
  %a_ext = sext <2 x i32> %a to <2 x i64>
  %b_ext = sext <2 x i32> %b to <2 x i64>
  %mul = mul <2 x i64> %a_ext, %b_ext
  %shift = lshr <2 x i64> %mul, <i64 32, i64 32>
  %res = trunc <2 x i64> %shift to <2 x i32>
  ret <2 x i32> %res
}

; Test packed multiply high unsigned
define <4 x i16> @test_pmulhu_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_pmulhu_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmulhu.h a0, a0, a1
; CHECK-NEXT:    ret
  %a_ext = zext <4 x i16> %a to <4 x i32>
  %b_ext = zext <4 x i16> %b to <4 x i32>
  %mul = mul <4 x i32> %a_ext, %b_ext
  %shift = lshr <4 x i32> %mul, <i32 16, i32 16, i32 16, i32 16>
  %res = trunc <4 x i32> %shift to <4 x i16>
  ret <4 x i16> %res
}

define <2 x i32> @test_pmulhu_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_pmulhu_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmulhu.w a0, a0, a1
; CHECK-NEXT:    ret
  %a_ext = zext <2 x i32> %a to <2 x i64>
  %b_ext = zext <2 x i32> %b to <2 x i64>
  %mul = mul <2 x i64> %a_ext, %b_ext
  %shift = lshr <2 x i64> %mul, <i64 32, i64 32>
  %res = trunc <2 x i64> %shift to <2 x i32>
  ret <2 x i32> %res
}

; Test packed multiply high signed-unsigned
define <4 x i16> @test_pmulhsu_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_pmulhsu_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmulhsu.h a0, a0, a1
; CHECK-NEXT:    ret
  %a_ext = sext <4 x i16> %a to <4 x i32>
  %b_ext = zext <4 x i16> %b to <4 x i32>
  %mul = mul <4 x i32> %a_ext, %b_ext
  %shift = lshr <4 x i32> %mul, <i32 16, i32 16, i32 16, i32 16>
  %res = trunc <4 x i32> %shift to <4 x i16>
  ret <4 x i16> %res
}

define <4 x i16> @test_pmulhsu_h_commuted(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_pmulhsu_h_commuted:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmulhsu.h a0, a1, a0
; CHECK-NEXT:    ret
  %a_ext = zext <4 x i16> %a to <4 x i32>
  %b_ext = sext <4 x i16> %b to <4 x i32>
  %mul = mul <4 x i32> %a_ext, %b_ext
  %shift = lshr <4 x i32> %mul, <i32 16, i32 16, i32 16, i32 16>
  %res = trunc <4 x i32> %shift to <4 x i16>
  ret <4 x i16> %res
}

define <2 x i32> @test_pmulhsu_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_pmulhsu_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmulhsu.w a0, a0, a1
; CHECK-NEXT:    ret
  %a_ext = sext <2 x i32> %a to <2 x i64>
  %b_ext = zext <2 x i32> %b to <2 x i64>
  %mul = mul <2 x i64> %a_ext, %b_ext
  %shift = lshr <2 x i64> %mul, <i64 32, i64 32>
  %res = trunc <2 x i64> %shift to <2 x i32>
  ret <2 x i32> %res
}

define <2 x i32> @test_pmulhsu_w_commuted(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_pmulhsu_w_commuted:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmulhsu.w a0, a1, a0
; CHECK-NEXT:    ret
  %a_ext = zext <2 x i32> %a to <2 x i64>
  %b_ext = sext <2 x i32> %b to <2 x i64>
  %mul = mul <2 x i64> %a_ext, %b_ext
  %shift = lshr <2 x i64> %mul, <i64 32, i64 32>
  %res = trunc <2 x i64> %shift to <2 x i32>
  ret <2 x i32> %res
}

; Test packed multiply high rounding signed
define <4 x i16> @test_pmulhr_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_pmulhr_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmulhr.h a0, a0, a1
; CHECK-NEXT:    ret
  %a_ext = sext <4 x i16> %a to <4 x i32>
  %b_ext = sext <4 x i16> %b to <4 x i32>
  %mul = mul <4 x i32> %a_ext, %b_ext
  %add = add <4 x i32> %mul, <i32 32768, i32 32768, i32 32768, i32 32768>
  %shift = lshr <4 x i32> %add, <i32 16, i32 16, i32 16, i32 16>
  %res = trunc <4 x i32> %shift to <4 x i16>
  ret <4 x i16> %res
}

define <2 x i32> @test_pmulhr_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_pmulhr_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmulhr.w a0, a0, a1
; CHECK-NEXT:    ret
  %a_ext = sext <2 x i32> %a to <2 x i64>
  %b_ext = sext <2 x i32> %b to <2 x i64>
  %mul = mul <2 x i64> %a_ext, %b_ext
  %add = add <2 x i64> %mul, <i64 2147483648, i64 2147483648>
  %shift = lshr <2 x i64> %add, <i64 32, i64 32>
  %res = trunc <2 x i64> %shift to <2 x i32>
  ret <2 x i32> %res
}

; Test packed multiply high rounding unsigned
define <4 x i16> @test_pmulhru_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_pmulhru_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmulhru.h a0, a0, a1
; CHECK-NEXT:    ret
  %a_ext = zext <4 x i16> %a to <4 x i32>
  %b_ext = zext <4 x i16> %b to <4 x i32>
  %mul = mul <4 x i32> %a_ext, %b_ext
  %add = add <4 x i32> %mul, <i32 32768, i32 32768, i32 32768, i32 32768>
  %shift = lshr <4 x i32> %add, <i32 16, i32 16, i32 16, i32 16>
  %res = trunc <4 x i32> %shift to <4 x i16>
  ret <4 x i16> %res
}

define <2 x i32> @test_pmulhru_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_pmulhru_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmulhru.w a0, a0, a1
; CHECK-NEXT:    ret
  %a_ext = zext <2 x i32> %a to <2 x i64>
  %b_ext = zext <2 x i32> %b to <2 x i64>
  %mul = mul <2 x i64> %a_ext, %b_ext
  %add = add <2 x i64> %mul, <i64 2147483648, i64 2147483648>
  %shift = lshr <2 x i64> %add, <i64 32, i64 32>
  %res = trunc <2 x i64> %shift to <2 x i32>
  ret <2 x i32> %res
}

; Test packed multiply high rounding signed-unsigned
define <4 x i16> @test_pmulhrsu_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_pmulhrsu_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmulhrsu.h a0, a0, a1
; CHECK-NEXT:    ret
  %a_ext = sext <4 x i16> %a to <4 x i32>
  %b_ext = zext <4 x i16> %b to <4 x i32>
  %mul = mul <4 x i32> %a_ext, %b_ext
  %add = add <4 x i32> %mul, <i32 32768, i32 32768, i32 32768, i32 32768>
  %shift = lshr <4 x i32> %add, <i32 16, i32 16, i32 16, i32 16>
  %res = trunc <4 x i32> %shift to <4 x i16>
  ret <4 x i16> %res
}

define <4 x i16> @test_pmulhrsu_h_commuted(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_pmulhrsu_h_commuted:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmulhrsu.h a0, a1, a0
; CHECK-NEXT:    ret
  %a_ext = zext <4 x i16> %a to <4 x i32>
  %b_ext = sext <4 x i16> %b to <4 x i32>
  %mul = mul <4 x i32> %a_ext, %b_ext
  %add = add <4 x i32> %mul, <i32 32768, i32 32768, i32 32768, i32 32768>
  %shift = lshr <4 x i32> %add, <i32 16, i32 16, i32 16, i32 16>
  %res = trunc <4 x i32> %shift to <4 x i16>
  ret <4 x i16> %res
}

define <2 x i32> @test_pmulhrsu_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_pmulhrsu_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmulhrsu.w a0, a0, a1
; CHECK-NEXT:    ret
  %a_ext = sext <2 x i32> %a to <2 x i64>
  %b_ext = zext <2 x i32> %b to <2 x i64>
  %mul = mul <2 x i64> %a_ext, %b_ext
  %add = add <2 x i64> %mul, <i64 2147483648, i64 2147483648>
  %shift = lshr <2 x i64> %add, <i64 32, i64 32>
  %res = trunc <2 x i64> %shift to <2 x i32>
  ret <2 x i32> %res
}

define <2 x i32> @test_pmulhrsu_w_commuted(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_pmulhrsu_w_commuted:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmulhrsu.w a0, a1, a0
; CHECK-NEXT:    ret
  %a_ext = zext <2 x i32> %a to <2 x i64>
  %b_ext = sext <2 x i32> %b to <2 x i64>
  %mul = mul <2 x i64> %a_ext, %b_ext
  %add = add <2 x i64> %mul, <i64 2147483648, i64 2147483648>
  %shift = lshr <2 x i64> %add, <i64 32, i64 32>
  %res = trunc <2 x i64> %shift to <2 x i32>
  ret <2 x i32> %res
}

; Test packed multiply low for v4i16
define <4 x i16> @test_pmul_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_pmul_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmul.w.h11 a2, a0, a1
; CHECK-NEXT:    pmul.w.h00 a0, a0, a1
; CHECK-NEXT:    ppaire.h a0, a0, a2
; CHECK-NEXT:    ret
  %res = mul <4 x i16> %a, %b
  ret <4 x i16> %res
}

; Test packed multiply low for v8i8
define <8 x i8> @test_pmul_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_pmul_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmul.h.b11 a2, a0, a1
; CHECK-NEXT:    pmul.h.b00 a0, a0, a1
; CHECK-NEXT:    ppaire.b a0, a0, a2
; CHECK-NEXT:    ret
  %res = mul <8 x i8> %a, %b
  ret <8 x i8> %res
}

; Test packed multiply low for v2i32
define <2 x i32> @test_pmul_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_pmul_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    mul.w11 a2, a0, a1
; CHECK-NEXT:    mul.w00 a0, a0, a1
; CHECK-NEXT:    pack a0, a0, a2
; CHECK-NEXT:    ret
  %res = mul <2 x i32> %a, %b
  ret <2 x i32> %res
}

; Division and remainder tests
define <4 x i16> @test_psdiv_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_psdiv_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    srai a2, a1, 48
; CHECK-NEXT:    srai a3, a0, 48
; CHECK-NEXT:    slli a4, a1, 16
; CHECK-NEXT:    sext.h a5, a1
; CHECK-NEXT:    divw a2, a3, a2
; CHECK-NEXT:    sext.h a3, a0
; CHECK-NEXT:    divw a3, a3, a5
; CHECK-NEXT:    slli a5, a0, 16
; CHECK-NEXT:    srai a4, a4, 48
; CHECK-NEXT:    srai a5, a5, 48
; CHECK-NEXT:    divw a4, a5, a4
; CHECK-NEXT:    slli a1, a1, 32
; CHECK-NEXT:    slli a0, a0, 32
; CHECK-NEXT:    srai a1, a1, 48
; CHECK-NEXT:    srai a0, a0, 48
; CHECK-NEXT:    divw a0, a0, a1
; CHECK-NEXT:    ppaire.h a1, a4, a2
; CHECK-NEXT:    ppaire.h a0, a3, a0
; CHECK-NEXT:    pack a0, a0, a1
; CHECK-NEXT:    ret
  %res = sdiv <4 x i16> %a, %b
  ret <4 x i16> %res
}

define <8 x i8> @test_psdiv_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_psdiv_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    srai a2, a1, 56
; CHECK-NEXT:    srai a3, a0, 56
; CHECK-NEXT:    slli a4, a1, 8
; CHECK-NEXT:    slli a5, a0, 8
; CHECK-NEXT:    slli a6, a1, 16
; CHECK-NEXT:    slli a7, a0, 16
; CHECK-NEXT:    slli t0, a1, 24
; CHECK-NEXT:    sext.b t1, a1
; CHECK-NEXT:    divw a2, a3, a2
; CHECK-NEXT:    sext.b a3, a0
; CHECK-NEXT:    divw a3, a3, t1
; CHECK-NEXT:    slli t1, a0, 24
; CHECK-NEXT:    srai a4, a4, 56
; CHECK-NEXT:    srai a5, a5, 56
; CHECK-NEXT:    divw a4, a5, a4
; CHECK-NEXT:    slli a5, a1, 32
; CHECK-NEXT:    srai a6, a6, 56
; CHECK-NEXT:    srai a7, a7, 56
; CHECK-NEXT:    divw a6, a7, a6
; CHECK-NEXT:    slli a7, a0, 32
; CHECK-NEXT:    srai t0, t0, 56
; CHECK-NEXT:    srai t1, t1, 56
; CHECK-NEXT:    divw t0, t1, t0
; CHECK-NEXT:    slli t1, a1, 40
; CHECK-NEXT:    srai a5, a5, 56
; CHECK-NEXT:    srai a7, a7, 56
; CHECK-NEXT:    divw a5, a7, a5
; CHECK-NEXT:    slli a7, a0, 40
; CHECK-NEXT:    srai t1, t1, 56
; CHECK-NEXT:    srai a7, a7, 56
; CHECK-NEXT:    divw a7, a7, t1
; CHECK-NEXT:    slli a1, a1, 48
; CHECK-NEXT:    slli a0, a0, 48
; CHECK-NEXT:    srai a1, a1, 56
; CHECK-NEXT:    srai a0, a0, 56
; CHECK-NEXT:    divw a0, a0, a1
; CHECK-NEXT:    ppaire.b a1, a4, a2
; CHECK-NEXT:    ppaire.b a2, t0, a6
; CHECK-NEXT:    ppaire.b a4, a7, a5
; CHECK-NEXT:    ppaire.b a0, a3, a0
; CHECK-NEXT:    ppaire.h a1, a2, a1
; CHECK-NEXT:    ppaire.h a0, a0, a4
; CHECK-NEXT:    pack a0, a0, a1
; CHECK-NEXT:    ret
  %res = sdiv <8 x i8> %a, %b
  ret <8 x i8> %res
}

define <2 x i32> @test_psdiv_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_psdiv_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    divw a2, a0, a1
; CHECK-NEXT:    srli a1, a1, 32
; CHECK-NEXT:    srli a0, a0, 32
; CHECK-NEXT:    divw a0, a0, a1
; CHECK-NEXT:    pack a0, a2, a0
; CHECK-NEXT:    ret
  %res = sdiv <2 x i32> %a, %b
  ret <2 x i32> %res
}

define <4 x i16> @test_pudiv_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_pudiv_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    srliw a2, a1, 16
; CHECK-NEXT:    srliw a3, a0, 16
; CHECK-NEXT:    zext.h a4, a1
; CHECK-NEXT:    zext.h a5, a0
; CHECK-NEXT:    divuw a2, a3, a2
; CHECK-NEXT:    srli a3, a1, 48
; CHECK-NEXT:    divuw a4, a5, a4
; CHECK-NEXT:    srli a5, a0, 48
; CHECK-NEXT:    divuw a3, a5, a3
; CHECK-NEXT:    slli a1, a1, 16
; CHECK-NEXT:    slli a0, a0, 16
; CHECK-NEXT:    srli a1, a1, 48
; CHECK-NEXT:    srli a0, a0, 48
; CHECK-NEXT:    divuw a0, a0, a1
; CHECK-NEXT:    ppaire.h a1, a4, a2
; CHECK-NEXT:    ppaire.h a0, a0, a3
; CHECK-NEXT:    pack a0, a1, a0
; CHECK-NEXT:    ret
  %res = udiv <4 x i16> %a, %b
  ret <4 x i16> %res
}

define <8 x i8> @test_pudiv_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_pudiv_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    srli a2, a1, 56
; CHECK-NEXT:    srli a3, a0, 56
; CHECK-NEXT:    slli a4, a1, 8
; CHECK-NEXT:    slli a5, a0, 8
; CHECK-NEXT:    slli a6, a1, 16
; CHECK-NEXT:    slli a7, a0, 16
; CHECK-NEXT:    srliw t0, a1, 24
; CHECK-NEXT:    srliw t1, a0, 24
; CHECK-NEXT:    divuw a2, a3, a2
; CHECK-NEXT:    zext.b a3, a1
; CHECK-NEXT:    divuw t0, t1, t0
; CHECK-NEXT:    zext.b t1, a0
; CHECK-NEXT:    divuw a3, t1, a3
; CHECK-NEXT:    slli t1, a1, 24
; CHECK-NEXT:    srli a4, a4, 56
; CHECK-NEXT:    srli a5, a5, 56
; CHECK-NEXT:    divuw a4, a5, a4
; CHECK-NEXT:    slli a5, a0, 24
; CHECK-NEXT:    srli a6, a6, 56
; CHECK-NEXT:    srli a7, a7, 56
; CHECK-NEXT:    divuw a6, a7, a6
; CHECK-NEXT:    slli a7, a1, 40
; CHECK-NEXT:    srli t1, t1, 56
; CHECK-NEXT:    srli a5, a5, 56
; CHECK-NEXT:    divuw a5, a5, t1
; CHECK-NEXT:    slli t1, a0, 40
; CHECK-NEXT:    srli a7, a7, 56
; CHECK-NEXT:    srli t1, t1, 56
; CHECK-NEXT:    divuw a7, t1, a7
; CHECK-NEXT:    slli a1, a1, 48
; CHECK-NEXT:    slli a0, a0, 48
; CHECK-NEXT:    srli a1, a1, 56
; CHECK-NEXT:    srli a0, a0, 56
; CHECK-NEXT:    divuw a0, a0, a1
; CHECK-NEXT:    ppaire.b a1, a4, a2
; CHECK-NEXT:    ppaire.b a2, a5, a6
; CHECK-NEXT:    ppaire.b a4, a7, t0
; CHECK-NEXT:    ppaire.b a0, a3, a0
; CHECK-NEXT:    ppaire.h a1, a2, a1
; CHECK-NEXT:    ppaire.h a0, a0, a4
; CHECK-NEXT:    pack a0, a0, a1
; CHECK-NEXT:    ret
  %res = udiv <8 x i8> %a, %b
  ret <8 x i8> %res
}

define <2 x i32> @test_pudiv_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_pudiv_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    divuw a2, a0, a1
; CHECK-NEXT:    srli a1, a1, 32
; CHECK-NEXT:    srli a0, a0, 32
; CHECK-NEXT:    divuw a0, a0, a1
; CHECK-NEXT:    pack a0, a2, a0
; CHECK-NEXT:    ret
  %res = udiv <2 x i32> %a, %b
  ret <2 x i32> %res
}

define <4 x i16> @test_psrem_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_psrem_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    srai a2, a1, 48
; CHECK-NEXT:    srai a3, a0, 48
; CHECK-NEXT:    slli a4, a1, 16
; CHECK-NEXT:    sext.h a5, a1
; CHECK-NEXT:    remw a2, a3, a2
; CHECK-NEXT:    sext.h a3, a0
; CHECK-NEXT:    remw a3, a3, a5
; CHECK-NEXT:    slli a5, a0, 16
; CHECK-NEXT:    srai a4, a4, 48
; CHECK-NEXT:    srai a5, a5, 48
; CHECK-NEXT:    remw a4, a5, a4
; CHECK-NEXT:    slli a1, a1, 32
; CHECK-NEXT:    slli a0, a0, 32
; CHECK-NEXT:    srai a1, a1, 48
; CHECK-NEXT:    srai a0, a0, 48
; CHECK-NEXT:    remw a0, a0, a1
; CHECK-NEXT:    ppaire.h a1, a4, a2
; CHECK-NEXT:    ppaire.h a0, a3, a0
; CHECK-NEXT:    pack a0, a0, a1
; CHECK-NEXT:    ret
  %res = srem <4 x i16> %a, %b
  ret <4 x i16> %res
}

define <8 x i8> @test_psrem_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_psrem_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    srai a2, a1, 56
; CHECK-NEXT:    srai a3, a0, 56
; CHECK-NEXT:    slli a4, a1, 8
; CHECK-NEXT:    slli a5, a0, 8
; CHECK-NEXT:    slli a6, a1, 16
; CHECK-NEXT:    slli a7, a0, 16
; CHECK-NEXT:    slli t0, a1, 24
; CHECK-NEXT:    sext.b t1, a1
; CHECK-NEXT:    remw a2, a3, a2
; CHECK-NEXT:    sext.b a3, a0
; CHECK-NEXT:    remw a3, a3, t1
; CHECK-NEXT:    slli t1, a0, 24
; CHECK-NEXT:    srai a4, a4, 56
; CHECK-NEXT:    srai a5, a5, 56
; CHECK-NEXT:    remw a4, a5, a4
; CHECK-NEXT:    slli a5, a1, 32
; CHECK-NEXT:    srai a6, a6, 56
; CHECK-NEXT:    srai a7, a7, 56
; CHECK-NEXT:    remw a6, a7, a6
; CHECK-NEXT:    slli a7, a0, 32
; CHECK-NEXT:    srai t0, t0, 56
; CHECK-NEXT:    srai t1, t1, 56
; CHECK-NEXT:    remw t0, t1, t0
; CHECK-NEXT:    slli t1, a1, 40
; CHECK-NEXT:    srai a5, a5, 56
; CHECK-NEXT:    srai a7, a7, 56
; CHECK-NEXT:    remw a5, a7, a5
; CHECK-NEXT:    slli a7, a0, 40
; CHECK-NEXT:    srai t1, t1, 56
; CHECK-NEXT:    srai a7, a7, 56
; CHECK-NEXT:    remw a7, a7, t1
; CHECK-NEXT:    slli a1, a1, 48
; CHECK-NEXT:    slli a0, a0, 48
; CHECK-NEXT:    srai a1, a1, 56
; CHECK-NEXT:    srai a0, a0, 56
; CHECK-NEXT:    remw a0, a0, a1
; CHECK-NEXT:    ppaire.b a1, a4, a2
; CHECK-NEXT:    ppaire.b a2, t0, a6
; CHECK-NEXT:    ppaire.b a4, a7, a5
; CHECK-NEXT:    ppaire.b a0, a3, a0
; CHECK-NEXT:    ppaire.h a1, a2, a1
; CHECK-NEXT:    ppaire.h a0, a0, a4
; CHECK-NEXT:    pack a0, a0, a1
; CHECK-NEXT:    ret
  %res = srem <8 x i8> %a, %b
  ret <8 x i8> %res
}

define <2 x i32> @test_psrem_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_psrem_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    remw a2, a0, a1
; CHECK-NEXT:    srai a1, a1, 32
; CHECK-NEXT:    srai a0, a0, 32
; CHECK-NEXT:    remw a0, a0, a1
; CHECK-NEXT:    pack a0, a2, a0
; CHECK-NEXT:    ret
  %res = srem <2 x i32> %a, %b
  ret <2 x i32> %res
}

define <4 x i16> @test_purem_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_purem_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    srliw a2, a1, 16
; CHECK-NEXT:    srliw a3, a0, 16
; CHECK-NEXT:    zext.h a4, a1
; CHECK-NEXT:    zext.h a5, a0
; CHECK-NEXT:    remuw a2, a3, a2
; CHECK-NEXT:    srli a3, a1, 48
; CHECK-NEXT:    remuw a4, a5, a4
; CHECK-NEXT:    srli a5, a0, 48
; CHECK-NEXT:    remuw a3, a5, a3
; CHECK-NEXT:    slli a1, a1, 16
; CHECK-NEXT:    slli a0, a0, 16
; CHECK-NEXT:    srli a1, a1, 48
; CHECK-NEXT:    srli a0, a0, 48
; CHECK-NEXT:    remuw a0, a0, a1
; CHECK-NEXT:    ppaire.h a1, a4, a2
; CHECK-NEXT:    ppaire.h a0, a0, a3
; CHECK-NEXT:    pack a0, a1, a0
; CHECK-NEXT:    ret
  %res = urem <4 x i16> %a, %b
  ret <4 x i16> %res
}

define <8 x i8> @test_purem_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_purem_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    srli a2, a1, 56
; CHECK-NEXT:    srli a3, a0, 56
; CHECK-NEXT:    slli a4, a1, 8
; CHECK-NEXT:    slli a5, a0, 8
; CHECK-NEXT:    slli a6, a1, 16
; CHECK-NEXT:    slli a7, a0, 16
; CHECK-NEXT:    srliw t0, a1, 24
; CHECK-NEXT:    srliw t1, a0, 24
; CHECK-NEXT:    remuw a2, a3, a2
; CHECK-NEXT:    zext.b a3, a1
; CHECK-NEXT:    remuw t0, t1, t0
; CHECK-NEXT:    zext.b t1, a0
; CHECK-NEXT:    remuw a3, t1, a3
; CHECK-NEXT:    slli t1, a1, 24
; CHECK-NEXT:    srli a4, a4, 56
; CHECK-NEXT:    srli a5, a5, 56
; CHECK-NEXT:    remuw a4, a5, a4
; CHECK-NEXT:    slli a5, a0, 24
; CHECK-NEXT:    srli a6, a6, 56
; CHECK-NEXT:    srli a7, a7, 56
; CHECK-NEXT:    remuw a6, a7, a6
; CHECK-NEXT:    slli a7, a1, 40
; CHECK-NEXT:    srli t1, t1, 56
; CHECK-NEXT:    srli a5, a5, 56
; CHECK-NEXT:    remuw a5, a5, t1
; CHECK-NEXT:    slli t1, a0, 40
; CHECK-NEXT:    srli a7, a7, 56
; CHECK-NEXT:    srli t1, t1, 56
; CHECK-NEXT:    remuw a7, t1, a7
; CHECK-NEXT:    slli a1, a1, 48
; CHECK-NEXT:    slli a0, a0, 48
; CHECK-NEXT:    srli a1, a1, 56
; CHECK-NEXT:    srli a0, a0, 56
; CHECK-NEXT:    remuw a0, a0, a1
; CHECK-NEXT:    ppaire.b a1, a4, a2
; CHECK-NEXT:    ppaire.b a2, a5, a6
; CHECK-NEXT:    ppaire.b a4, a7, t0
; CHECK-NEXT:    ppaire.b a0, a3, a0
; CHECK-NEXT:    ppaire.h a1, a2, a1
; CHECK-NEXT:    ppaire.h a0, a0, a4
; CHECK-NEXT:    pack a0, a0, a1
; CHECK-NEXT:    ret
  %res = urem <8 x i8> %a, %b
  ret <8 x i8> %res
}

define <2 x i32> @test_purem_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_purem_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    remuw a2, a0, a1
; CHECK-NEXT:    srli a1, a1, 32
; CHECK-NEXT:    srli a0, a0, 32
; CHECK-NEXT:    remuw a0, a0, a1
; CHECK-NEXT:    pack a0, a2, a0
; CHECK-NEXT:    ret
  %res = urem <2 x i32> %a, %b
  ret <2 x i32> %res
}

; Comparison operations for v4i16
define <4 x i16> @test_eq_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_eq_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmseq.h a0, a0, a1
; CHECK-NEXT:    ret
  %cmp = icmp eq <4 x i16> %a, %b
  %res = sext <4 x i1> %cmp to <4 x i16>
  ret <4 x i16> %res
}

define <4 x i16> @test_ne_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_ne_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmseq.h a0, a0, a1
; CHECK-NEXT:    not a0, a0
; CHECK-NEXT:    ret
  %cmp = icmp ne <4 x i16> %a, %b
  %res = sext <4 x i1> %cmp to <4 x i16>
  ret <4 x i16> %res
}

define <4 x i16> @test_slt_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_slt_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmslt.h a0, a0, a1
; CHECK-NEXT:    ret
  %cmp = icmp slt <4 x i16> %a, %b
  %res = sext <4 x i1> %cmp to <4 x i16>
  ret <4 x i16> %res
}

define <4 x i16> @test_sle_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_sle_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmslt.h a0, a1, a0
; CHECK-NEXT:    not a0, a0
; CHECK-NEXT:    ret
  %cmp = icmp sle <4 x i16> %a, %b
  %res = sext <4 x i1> %cmp to <4 x i16>
  ret <4 x i16> %res
}

define <4 x i16> @test_sgt_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_sgt_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmslt.h a0, a1, a0
; CHECK-NEXT:    ret
  %cmp = icmp sgt <4 x i16> %a, %b
  %res = sext <4 x i1> %cmp to <4 x i16>
  ret <4 x i16> %res
}

define <4 x i16> @test_sge_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_sge_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmslt.h a0, a0, a1
; CHECK-NEXT:    not a0, a0
; CHECK-NEXT:    ret
  %cmp = icmp sge <4 x i16> %a, %b
  %res = sext <4 x i1> %cmp to <4 x i16>
  ret <4 x i16> %res
}

define <4 x i16> @test_ult_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_ult_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmsltu.h a0, a0, a1
; CHECK-NEXT:    ret
  %cmp = icmp ult <4 x i16> %a, %b
  %res = sext <4 x i1> %cmp to <4 x i16>
  ret <4 x i16> %res
}

define <4 x i16> @test_ule_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_ule_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmsltu.h a0, a1, a0
; CHECK-NEXT:    not a0, a0
; CHECK-NEXT:    ret
  %cmp = icmp ule <4 x i16> %a, %b
  %res = sext <4 x i1> %cmp to <4 x i16>
  ret <4 x i16> %res
}

define <4 x i16> @test_ugt_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_ugt_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmsltu.h a0, a1, a0
; CHECK-NEXT:    ret
  %cmp = icmp ugt <4 x i16> %a, %b
  %res = sext <4 x i1> %cmp to <4 x i16>
  ret <4 x i16> %res
}

define <4 x i16> @test_uge_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_uge_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmsltu.h a0, a0, a1
; CHECK-NEXT:    not a0, a0
; CHECK-NEXT:    ret
  %cmp = icmp uge <4 x i16> %a, %b
  %res = sext <4 x i1> %cmp to <4 x i16>
  ret <4 x i16> %res
}

; Comparison operations for v8i8
define <8 x i8> @test_eq_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_eq_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmseq.b a0, a0, a1
; CHECK-NEXT:    ret
  %cmp = icmp eq <8 x i8> %a, %b
  %res = sext <8 x i1> %cmp to <8 x i8>
  ret <8 x i8> %res
}

define <8 x i8> @test_ne_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_ne_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmseq.b a0, a0, a1
; CHECK-NEXT:    not a0, a0
; CHECK-NEXT:    ret
  %cmp = icmp ne <8 x i8> %a, %b
  %res = sext <8 x i1> %cmp to <8 x i8>
  ret <8 x i8> %res
}

define <8 x i8> @test_slt_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_slt_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmslt.b a0, a0, a1
; CHECK-NEXT:    ret
  %cmp = icmp slt <8 x i8> %a, %b
  %res = sext <8 x i1> %cmp to <8 x i8>
  ret <8 x i8> %res
}

define <8 x i8> @test_sle_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_sle_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmslt.b a0, a1, a0
; CHECK-NEXT:    not a0, a0
; CHECK-NEXT:    ret
  %cmp = icmp sle <8 x i8> %a, %b
  %res = sext <8 x i1> %cmp to <8 x i8>
  ret <8 x i8> %res
}

define <8 x i8> @test_sgt_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_sgt_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmslt.b a0, a1, a0
; CHECK-NEXT:    ret
  %cmp = icmp sgt <8 x i8> %a, %b
  %res = sext <8 x i1> %cmp to <8 x i8>
  ret <8 x i8> %res
}

define <8 x i8> @test_sge_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_sge_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmslt.b a0, a0, a1
; CHECK-NEXT:    not a0, a0
; CHECK-NEXT:    ret
  %cmp = icmp sge <8 x i8> %a, %b
  %res = sext <8 x i1> %cmp to <8 x i8>
  ret <8 x i8> %res
}

define <8 x i8> @test_ult_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_ult_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmsltu.b a0, a0, a1
; CHECK-NEXT:    ret
  %cmp = icmp ult <8 x i8> %a, %b
  %res = sext <8 x i1> %cmp to <8 x i8>
  ret <8 x i8> %res
}

define <8 x i8> @test_ule_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_ule_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmsltu.b a0, a1, a0
; CHECK-NEXT:    not a0, a0
; CHECK-NEXT:    ret
  %cmp = icmp ule <8 x i8> %a, %b
  %res = sext <8 x i1> %cmp to <8 x i8>
  ret <8 x i8> %res
}

define <8 x i8> @test_ugt_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_ugt_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmsltu.b a0, a1, a0
; CHECK-NEXT:    ret
  %cmp = icmp ugt <8 x i8> %a, %b
  %res = sext <8 x i1> %cmp to <8 x i8>
  ret <8 x i8> %res
}

define <8 x i8> @test_uge_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_uge_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmsltu.b a0, a0, a1
; CHECK-NEXT:    not a0, a0
; CHECK-NEXT:    ret
  %cmp = icmp uge <8 x i8> %a, %b
  %res = sext <8 x i1> %cmp to <8 x i8>
  ret <8 x i8> %res
}

; Comparison operations for v2i32
define <2 x i32> @test_eq_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_eq_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmseq.w a0, a0, a1
; CHECK-NEXT:    ret
  %cmp = icmp eq <2 x i32> %a, %b
  %res = sext <2 x i1> %cmp to <2 x i32>
  ret <2 x i32> %res
}

define <2 x i32> @test_ne_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_ne_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmseq.w a0, a0, a1
; CHECK-NEXT:    not a0, a0
; CHECK-NEXT:    ret
  %cmp = icmp ne <2 x i32> %a, %b
  %res = sext <2 x i1> %cmp to <2 x i32>
  ret <2 x i32> %res
}

define <2 x i32> @test_slt_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_slt_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmslt.w a0, a0, a1
; CHECK-NEXT:    ret
  %cmp = icmp slt <2 x i32> %a, %b
  %res = sext <2 x i1> %cmp to <2 x i32>
  ret <2 x i32> %res
}

define <2 x i32> @test_sle_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_sle_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmslt.w a0, a1, a0
; CHECK-NEXT:    not a0, a0
; CHECK-NEXT:    ret
  %cmp = icmp sle <2 x i32> %a, %b
  %res = sext <2 x i1> %cmp to <2 x i32>
  ret <2 x i32> %res
}

define <2 x i32> @test_sgt_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_sgt_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmslt.w a0, a1, a0
; CHECK-NEXT:    ret
  %cmp = icmp sgt <2 x i32> %a, %b
  %res = sext <2 x i1> %cmp to <2 x i32>
  ret <2 x i32> %res
}

define <2 x i32> @test_sge_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_sge_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmslt.w a0, a0, a1
; CHECK-NEXT:    not a0, a0
; CHECK-NEXT:    ret
  %cmp = icmp sge <2 x i32> %a, %b
  %res = sext <2 x i1> %cmp to <2 x i32>
  ret <2 x i32> %res
}

define <2 x i32> @test_ult_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_ult_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmsltu.w a0, a0, a1
; CHECK-NEXT:    ret
  %cmp = icmp ult <2 x i32> %a, %b
  %res = sext <2 x i1> %cmp to <2 x i32>
  ret <2 x i32> %res
}

define <2 x i32> @test_ule_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_ule_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmsltu.w a0, a1, a0
; CHECK-NEXT:    not a0, a0
; CHECK-NEXT:    ret
  %cmp = icmp ule <2 x i32> %a, %b
  %res = sext <2 x i1> %cmp to <2 x i32>
  ret <2 x i32> %res
}

define <2 x i32> @test_ugt_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_ugt_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmsltu.w a0, a1, a0
; CHECK-NEXT:    ret
  %cmp = icmp ugt <2 x i32> %a, %b
  %res = sext <2 x i1> %cmp to <2 x i32>
  ret <2 x i32> %res
}

define <2 x i32> @test_uge_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_uge_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmsltu.w a0, a0, a1
; CHECK-NEXT:    not a0, a0
; CHECK-NEXT:    ret
  %cmp = icmp uge <2 x i32> %a, %b
  %res = sext <2 x i1> %cmp to <2 x i32>
  ret <2 x i32> %res
}

; Test 8/16/32-bit [s|u]min/[s|u]max
define <4 x i16> @test_smin_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_smin_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmin.h a0, a0, a1
; CHECK-NEXT:    ret
  %min = call <4 x i16> @llvm.smin.v2i16(<4 x i16> %a, <4 x i16> %b)
  ret <4 x i16> %min
}

define <4 x i16> @test_umin_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_umin_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pminu.h a0, a0, a1
; CHECK-NEXT:    ret
  %min = call <4 x i16> @llvm.umin.v2i16(<4 x i16> %a, <4 x i16> %b)
  ret <4 x i16> %min
}

define <8 x i8> @test_smin_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_smin_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmin.b a0, a0, a1
; CHECK-NEXT:    ret
  %min = call <8 x i8> @llvm.smin.v4i8(<8 x i8> %a, <8 x i8> %b)
  ret <8 x i8> %min
}

define <8 x i8> @test_umin_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_umin_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pminu.b a0, a0, a1
; CHECK-NEXT:    ret
  %min = call <8 x i8> @llvm.umin.v4i8(<8 x i8> %a, <8 x i8> %b)
  ret <8 x i8> %min
}

define <2 x i32> @test_smin_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_smin_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmin.w a0, a0, a1
; CHECK-NEXT:    ret
  %min = call <2 x i32> @llvm.smin.v2i32(<2 x i32> %a, <2 x i32> %b)
  ret <2 x i32> %min
}

define <2 x i32> @test_umin_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_umin_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pminu.w a0, a0, a1
; CHECK-NEXT:    ret
  %min = call <2 x i32> @llvm.umin.v2i32(<2 x i32> %a, <2 x i32> %b)
  ret <2 x i32> %min
}

define <4 x i16> @test_smax_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_smax_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmax.h a0, a0, a1
; CHECK-NEXT:    ret
  %max = call <4 x i16> @llvm.smax.v2i16(<4 x i16> %a, <4 x i16> %b)
  ret <4 x i16> %max
}

define <4 x i16> @test_umax_h(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_umax_h:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmaxu.h a0, a0, a1
; CHECK-NEXT:    ret
  %max = call <4 x i16> @llvm.umax.v2i16(<4 x i16> %a, <4 x i16> %b)
  ret <4 x i16> %max
}

define <8 x i8> @test_smax_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_smax_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmax.b a0, a0, a1
; CHECK-NEXT:    ret
  %max = call <8 x i8> @llvm.smax.v4i8(<8 x i8> %a, <8 x i8> %b)
  ret <8 x i8> %max
}

define <8 x i8> @test_umax_b(<8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_umax_b:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmaxu.b a0, a0, a1
; CHECK-NEXT:    ret
  %max = call <8 x i8> @llvm.umax.v4i8(<8 x i8> %a, <8 x i8> %b)
  ret <8 x i8> %max
}

define <2 x i32> @test_smax_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_smax_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmax.w a0, a0, a1
; CHECK-NEXT:    ret
  %max = call <2 x i32> @llvm.smax.v2i32(<2 x i32> %a, <2 x i32> %b)
  ret <2 x i32> %max
}

define <2 x i32> @test_umax_w(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_umax_w:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmaxu.w a0, a0, a1
; CHECK-NEXT:    ret
  %max = call <2 x i32> @llvm.umax.v2i32(<2 x i32> %a, <2 x i32> %b)
  ret <2 x i32> %max
}

; Test select operations
define <4 x i16> @test_select_v4i16(i1 %cond, <4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: test_select_v4i16:
; CHECK:       # %bb.0:
; CHECK-NEXT:    andi a3, a0, 1
; CHECK-NEXT:    mv a0, a1
; CHECK-NEXT:    bnez a3, .LBB205_2
; CHECK-NEXT:  # %bb.1:
; CHECK-NEXT:    mv a0, a2
; CHECK-NEXT:  .LBB205_2:
; CHECK-NEXT:    ret
  %res = select i1 %cond, <4 x i16> %a, <4 x i16> %b
  ret <4 x i16> %res
}

define <8 x i8> @test_select_v8i8(i1 %cond, <8 x i8> %a, <8 x i8> %b) {
; CHECK-LABEL: test_select_v8i8:
; CHECK:       # %bb.0:
; CHECK-NEXT:    andi a3, a0, 1
; CHECK-NEXT:    mv a0, a1
; CHECK-NEXT:    bnez a3, .LBB206_2
; CHECK-NEXT:  # %bb.1:
; CHECK-NEXT:    mv a0, a2
; CHECK-NEXT:  .LBB206_2:
; CHECK-NEXT:    ret
  %res = select i1 %cond, <8 x i8> %a, <8 x i8> %b
  ret <8 x i8> %res
}

define <2 x i32> @test_select_v2i32(i1 %cond, <2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: test_select_v2i32:
; CHECK:       # %bb.0:
; CHECK-NEXT:    andi a3, a0, 1
; CHECK-NEXT:    mv a0, a1
; CHECK-NEXT:    bnez a3, .LBB207_2
; CHECK-NEXT:  # %bb.1:
; CHECK-NEXT:    mv a0, a2
; CHECK-NEXT:  .LBB207_2:
; CHECK-NEXT:    ret
  %res = select i1 %cond, <2 x i32> %a, <2 x i32> %b
  ret <2 x i32> %res
}

; Test vselect operations
define <4 x i16> @test_vselect_v4i16(<4 x i16> %a, <4 x i16> %b, <4 x i16> %c) {
; CHECK-LABEL: test_vselect_v4i16:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmseq.h a0, a0, a1
; CHECK-NEXT:    merge a0, a1, a2
; CHECK-NEXT:    ret
  %mask = icmp eq <4 x i16> %a, %b
  %res = select <4 x i1> %mask, <4 x i16> %c, <4 x i16> %b
  ret <4 x i16> %res
}

define <8 x i8> @test_vselect_v8i8(<8 x i8> %a, <8 x i8> %b, <8 x i8> %c) {
; CHECK-LABEL: test_vselect_v8i8:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmsltu.b a0, a0, a1
; CHECK-NEXT:    merge a0, a1, a2
; CHECK-NEXT:    ret
  %mask = icmp ult <8 x i8> %a, %b
  %res = select <8 x i1> %mask, <8 x i8> %c, <8 x i8> %b
  ret <8 x i8> %res
}

define <2 x i32> @test_vselect_v2i32(<2 x i32> %a, <2 x i32> %b, <2 x i32> %c) {
; CHECK-LABEL: test_vselect_v2i32:
; CHECK:       # %bb.0:
; CHECK-NEXT:    pmslt.w a0, a1, a0
; CHECK-NEXT:    merge a0, a1, a2
; CHECK-NEXT:    ret
  %mask = icmp sgt <2 x i32> %a, %b
  %res = select <2 x i1> %mask, <2 x i32> %c, <2 x i32> %b
  ret <2 x i32> %res
}

define <4 x i16> @test_bswap_v4i16(<4 x i16> %a) {
; CHECK-LABEL: test_bswap_v4i16:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psrli.h a1, a0, 8
; CHECK-NEXT:    pslli.h a0, a0, 8
; CHECK-NEXT:    or a0, a0, a1
; CHECK-NEXT:    ret
  %res = call <4 x i16> @llvm.bswap.v4i16(<4 x i16> %a)
  ret <4 x i16> %res
}

define <2 x i32> @test_bswap_v2i32(<2 x i32> %a) {
; CHECK-LABEL: test_bswap_v2i32:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psrli.w a1, a0, 8
; CHECK-NEXT:    lui a2, 16
; CHECK-NEXT:    psrli.w a3, a0, 24
; CHECK-NEXT:    addi a2, a2, -256
; CHECK-NEXT:    pmv.ws a2, a2
; CHECK-NEXT:    and a1, a1, a2
; CHECK-NEXT:    and a2, a0, a2
; CHECK-NEXT:    or a1, a1, a3
; CHECK-NEXT:    pslli.w a2, a2, 8
; CHECK-NEXT:    pslli.w a0, a0, 24
; CHECK-NEXT:    or a0, a0, a2
; CHECK-NEXT:    or a0, a0, a1
; CHECK-NEXT:    ret
  %res = call <2 x i32> @llvm.bswap.v2i32(<2 x i32> %a)
  ret <2 x i32> %res
}

define <8 x i8> @test_bitreverse_v8i8(<8 x i8> %a) {
; CHECK-LABEL: test_bitreverse_v8i8:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psrli.b a1, a0, 4
; CHECK-NEXT:    pli.b a2, 15
; CHECK-NEXT:    and a1, a1, a2
; CHECK-NEXT:    and a0, a0, a2
; CHECK-NEXT:    pli.b a2, 51
; CHECK-NEXT:    pslli.b a0, a0, 4
; CHECK-NEXT:    or a0, a1, a0
; CHECK-NEXT:    psrli.b a1, a0, 2
; CHECK-NEXT:    and a0, a0, a2
; CHECK-NEXT:    and a1, a1, a2
; CHECK-NEXT:    pli.b a2, 85
; CHECK-NEXT:    pslli.b a0, a0, 2
; CHECK-NEXT:    or a0, a1, a0
; CHECK-NEXT:    psrli.b a1, a0, 1
; CHECK-NEXT:    and a0, a0, a2
; CHECK-NEXT:    and a1, a1, a2
; CHECK-NEXT:    pslli.b a0, a0, 1
; CHECK-NEXT:    or a0, a1, a0
; CHECK-NEXT:    ret
  %res = call <8 x i8> @llvm.bitreverse.v8i8(<8 x i8> %a)
  ret <8 x i8> %res
}

define <4 x i16> @test_bitreverse_v4i16(<4 x i16> %a) {
; CHECK-LABEL: test_bitreverse_v4i16:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psrli.h a1, a0, 8
; CHECK-NEXT:    pslli.h a0, a0, 8
; CHECK-NEXT:    pli.b a2, 15
; CHECK-NEXT:    or a0, a0, a1
; CHECK-NEXT:    psrli.h a1, a0, 4
; CHECK-NEXT:    and a0, a0, a2
; CHECK-NEXT:    and a1, a1, a2
; CHECK-NEXT:    pli.b a2, 51
; CHECK-NEXT:    pslli.h a0, a0, 4
; CHECK-NEXT:    or a0, a1, a0
; CHECK-NEXT:    psrli.h a1, a0, 2
; CHECK-NEXT:    and a0, a0, a2
; CHECK-NEXT:    and a1, a1, a2
; CHECK-NEXT:    pli.b a2, 85
; CHECK-NEXT:    pslli.h a0, a0, 2
; CHECK-NEXT:    or a0, a1, a0
; CHECK-NEXT:    psrli.h a1, a0, 1
; CHECK-NEXT:    and a0, a0, a2
; CHECK-NEXT:    and a1, a1, a2
; CHECK-NEXT:    pslli.h a0, a0, 1
; CHECK-NEXT:    or a0, a1, a0
; CHECK-NEXT:    ret
  %res = call <4 x i16> @llvm.bitreverse.v4i16(<4 x i16> %a)
  ret <4 x i16> %res
}

define <2 x i32> @test_bitreverse_v2i32(<2 x i32> %a) {
; CHECK-LABEL: test_bitreverse_v2i32:
; CHECK:       # %bb.0:
; CHECK-NEXT:    psrli.w a1, a0, 8
; CHECK-NEXT:    lui a2, 16
; CHECK-NEXT:    psrli.w a3, a0, 24
; CHECK-NEXT:    addi a2, a2, -256
; CHECK-NEXT:    pmv.ws a2, a2
; CHECK-NEXT:    and a1, a1, a2
; CHECK-NEXT:    and a2, a0, a2
; CHECK-NEXT:    pslli.w a0, a0, 24
; CHECK-NEXT:    or a1, a1, a3
; CHECK-NEXT:    pli.b a3, 15
; CHECK-NEXT:    pslli.w a2, a2, 8
; CHECK-NEXT:    or a0, a0, a2
; CHECK-NEXT:    pli.b a2, 51
; CHECK-NEXT:    or a0, a0, a1
; CHECK-NEXT:    psrli.w a1, a0, 4
; CHECK-NEXT:    and a0, a0, a3
; CHECK-NEXT:    and a1, a1, a3
; CHECK-NEXT:    pli.b a3, 85
; CHECK-NEXT:    pslli.w a0, a0, 4
; CHECK-NEXT:    or a0, a1, a0
; CHECK-NEXT:    psrli.w a1, a0, 2
; CHECK-NEXT:    and a0, a0, a2
; CHECK-NEXT:    and a1, a1, a2
; CHECK-NEXT:    pslli.w a0, a0, 2
; CHECK-NEXT:    or a0, a1, a0
; CHECK-NEXT:    psrli.w a1, a0, 1
; CHECK-NEXT:    and a0, a0, a3
; CHECK-NEXT:    and a1, a1, a3
; CHECK-NEXT:    pslli.w a0, a0, 1
; CHECK-NEXT:    or a0, a1, a0
; CHECK-NEXT:    ret
  %res = call <2 x i32> @llvm.bitreverse.v2i32(<2 x i32> %a)
  ret <2 x i32> %res
}