; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -fast-isel -mtriple=x86_64-unknown-unknown -mattr=+kl,+widekl | FileCheck %s

; NOTE: This should use IR equivalent to what is generated by clang/test/CodeGen/X86/keylocker-builtins.c

define void @test_loadiwkey(i32 %ctl, <2 x i64> %intkey, <2 x i64> %enkey_lo, <2 x i64> %enkey_hi) {
; CHECK-LABEL: test_loadiwkey:
; CHECK:       # %bb.0: # %entry
; CHECK-NEXT:    movl %edi, %eax
; CHECK-NEXT:    loadiwkey %xmm2, %xmm1
; CHECK-NEXT:    retq
entry:
  tail call void @llvm.x86.loadiwkey(<2 x i64> %intkey, <2 x i64> %enkey_lo, <2 x i64> %enkey_hi, i32 %ctl)
  ret void
}

define i32 @test_encodekey128_u32(i32 %htype, <2 x i64> %key, ptr nocapture %h) {
; CHECK-LABEL: test_encodekey128_u32:
; CHECK:       # %bb.0: # %entry
; CHECK-NEXT:    encodekey128 %edi, %eax
; CHECK-NEXT:    movups %xmm0, (%rsi)
; CHECK-NEXT:    movups %xmm1, 16(%rsi)
; CHECK-NEXT:    movups %xmm2, 32(%rsi)
; CHECK-NEXT:    movups %xmm4, 48(%rsi)
; CHECK-NEXT:    movups %xmm5, 64(%rsi)
; CHECK-NEXT:    movups %xmm6, 80(%rsi)
; CHECK-NEXT:    retq
entry:
  %0 = tail call { i32, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } @llvm.x86.encodekey128(i32 %htype, <2 x i64> %key)
  %1 = extractvalue { i32, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %0, 1
  store <2 x i64> %1, ptr %h, align 1
  %2 = extractvalue { i32, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %0, 2
  %3 = getelementptr i8, ptr %h, i64 16
  store <2 x i64> %2, ptr %3, align 1
  %4 = extractvalue { i32, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %0, 3
  %5 = getelementptr i8, ptr %h, i64 32
  store <2 x i64> %4, ptr %5, align 1
  %6 = extractvalue { i32, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %0, 4
  %7 = getelementptr i8, ptr %h, i64 48
  store <2 x i64> %6, ptr %7, align 1
  %8 = extractvalue { i32, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %0, 5
  %9 = getelementptr i8, ptr %h, i64 64
  store <2 x i64> %8, ptr %9, align 1
  %10 = extractvalue { i32, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %0, 6
  %11 = getelementptr i8, ptr %h, i64 80
  store <2 x i64> %10, ptr %11, align 1
  %12 = extractvalue { i32, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %0, 0
  ret i32 %12
}

define i32 @test_encodekey256_u32(i32 %htype, <2 x i64> %key_lo, <2 x i64> %key_hi, ptr nocapture %h) {
; CHECK-LABEL: test_encodekey256_u32:
; CHECK:       # %bb.0: # %entry
; CHECK-NEXT:    encodekey256 %edi, %eax
; CHECK-NEXT:    movups %xmm0, (%rsi)
; CHECK-NEXT:    movups %xmm1, 16(%rsi)
; CHECK-NEXT:    movups %xmm2, 32(%rsi)
; CHECK-NEXT:    movups %xmm3, 48(%rsi)
; CHECK-NEXT:    movups %xmm4, 64(%rsi)
; CHECK-NEXT:    movups %xmm5, 80(%rsi)
; CHECK-NEXT:    movups %xmm6, 96(%rsi)
; CHECK-NEXT:    retq
entry:
  %0 = tail call { i32, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } @llvm.x86.encodekey256(i32 %htype, <2 x i64> %key_lo, <2 x i64> %key_hi)
  %1 = extractvalue { i32, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %0, 1
  store <2 x i64> %1, ptr %h, align 1
  %2 = extractvalue { i32, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %0, 2
  %3 = getelementptr i8, ptr %h, i64 16
  store <2 x i64> %2, ptr %3, align 1
  %4 = extractvalue { i32, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %0, 3
  %5 = getelementptr i8, ptr %h, i64 32
  store <2 x i64> %4, ptr %5, align 1
  %6 = extractvalue { i32, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %0, 4
  %7 = getelementptr i8, ptr %h, i64 48
  store <2 x i64> %6, ptr %7, align 1
  %8 = extractvalue { i32, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %0, 5
  %9 = getelementptr i8, ptr %h, i64 64
  store <2 x i64> %8, ptr %9, align 1
  %10 = extractvalue { i32, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %0, 6
  %11 = getelementptr i8, ptr %h, i64 80
  store <2 x i64> %10, ptr %11, align 1
  %12 = extractvalue { i32, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %0, 7
  %13 = getelementptr i8, ptr %h, i64 96
  store <2 x i64> %12, ptr %13, align 1
  %14 = extractvalue { i32, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %0, 0
  ret i32 %14
}

define zeroext i8 @test_mm_aesenc256kl_u8(ptr %odata, <2 x i64> %idata, ptr %h) {
; CHECK-LABEL: test_mm_aesenc256kl_u8:
; CHECK:       # %bb.0: # %entry
; CHECK-NEXT:    aesenc256kl (%rsi), %xmm0
; CHECK-NEXT:    sete %al
; CHECK-NEXT:    movaps %xmm0, (%rdi)
; CHECK-NEXT:    retq
entry:
  %0 = tail call { i8, <2 x i64> } @llvm.x86.aesenc256kl(<2 x i64> %idata, ptr %h) #1
  %1 = extractvalue { i8, <2 x i64> } %0, 1
  store <2 x i64> %1, ptr %odata, align 16
  %2 = extractvalue { i8, <2 x i64> } %0, 0
  ret i8 %2
}

define zeroext i8 @test_mm_aesdec256kl_u8(ptr %odata, <2 x i64> %idata, ptr %h) {
; CHECK-LABEL: test_mm_aesdec256kl_u8:
; CHECK:       # %bb.0: # %entry
; CHECK-NEXT:    aesdec256kl (%rsi), %xmm0
; CHECK-NEXT:    sete %al
; CHECK-NEXT:    movaps %xmm0, (%rdi)
; CHECK-NEXT:    retq
entry:
  %0 = tail call { i8, <2 x i64> } @llvm.x86.aesdec256kl(<2 x i64> %idata, ptr %h) #1
  %1 = extractvalue { i8, <2 x i64> } %0, 1
  store <2 x i64> %1, ptr %odata, align 16
  %2 = extractvalue { i8, <2 x i64> } %0, 0
  ret i8 %2
}

define zeroext i8 @test_mm_aesenc128kl_u8(ptr %odata, <2 x i64> %idata, ptr %h) {
; CHECK-LABEL: test_mm_aesenc128kl_u8:
; CHECK:       # %bb.0: # %entry
; CHECK-NEXT:    aesenc128kl (%rsi), %xmm0
; CHECK-NEXT:    sete %al
; CHECK-NEXT:    movaps %xmm0, (%rdi)
; CHECK-NEXT:    retq
entry:
  %0 = tail call { i8, <2 x i64> } @llvm.x86.aesenc128kl(<2 x i64> %idata, ptr %h) #1
  %1 = extractvalue { i8, <2 x i64> } %0, 1
  store <2 x i64> %1, ptr %odata, align 16
  %2 = extractvalue { i8, <2 x i64> } %0, 0
  ret i8 %2
}

define zeroext i8 @test_mm_aesdec128kl_u8(ptr %odata, <2 x i64> %idata, ptr %h) {
; CHECK-LABEL: test_mm_aesdec128kl_u8:
; CHECK:       # %bb.0: # %entry
; CHECK-NEXT:    aesdec128kl (%rsi), %xmm0
; CHECK-NEXT:    sete %al
; CHECK-NEXT:    movaps %xmm0, (%rdi)
; CHECK-NEXT:    retq
entry:
  %0 = tail call { i8, <2 x i64> } @llvm.x86.aesdec128kl(<2 x i64> %idata, ptr %h) #1
  %1 = extractvalue { i8, <2 x i64> } %0, 1
  store <2 x i64> %1, ptr %odata, align 16
  %2 = extractvalue { i8, <2 x i64> } %0, 0
  ret i8 %2
}

define zeroext i8 @test__mm_aesencwide128kl_u8(ptr %odata, ptr %idata, ptr %h) {
; CHECK-LABEL: test__mm_aesencwide128kl_u8:
; CHECK:       # %bb.0: # %entry
; CHECK-NEXT:    movaps (%rsi), %xmm0
; CHECK-NEXT:    movaps 16(%rsi), %xmm1
; CHECK-NEXT:    movaps 32(%rsi), %xmm2
; CHECK-NEXT:    movaps 48(%rsi), %xmm3
; CHECK-NEXT:    movaps 64(%rsi), %xmm4
; CHECK-NEXT:    movaps 80(%rsi), %xmm5
; CHECK-NEXT:    movaps 96(%rsi), %xmm6
; CHECK-NEXT:    movaps 112(%rsi), %xmm7
; CHECK-NEXT:    aesencwide128kl (%rdx)
; CHECK-NEXT:    sete %al
; CHECK-NEXT:    movaps %xmm0, (%rdi)
; CHECK-NEXT:    movaps %xmm1, 16(%rdi)
; CHECK-NEXT:    movaps %xmm2, 32(%rdi)
; CHECK-NEXT:    movaps %xmm3, 48(%rdi)
; CHECK-NEXT:    movaps %xmm4, 64(%rdi)
; CHECK-NEXT:    movaps %xmm5, 80(%rdi)
; CHECK-NEXT:    movaps %xmm6, 96(%rdi)
; CHECK-NEXT:    movaps %xmm7, 112(%rdi)
; CHECK-NEXT:    retq
entry:
  %0 = load <2 x i64>, ptr %idata, align 16
  %1 = getelementptr <2 x i64>, ptr %idata, i64 1
  %2 = load <2 x i64>, ptr %1, align 16
  %3 = getelementptr <2 x i64>, ptr %idata, i64 2
  %4 = load <2 x i64>, ptr %3, align 16
  %5 = getelementptr <2 x i64>, ptr %idata, i64 3
  %6 = load <2 x i64>, ptr %5, align 16
  %7 = getelementptr <2 x i64>, ptr %idata, i64 4
  %8 = load <2 x i64>, ptr %7, align 16
  %9 = getelementptr <2 x i64>, ptr %idata, i64 5
  %10 = load <2 x i64>, ptr %9, align 16
  %11 = getelementptr <2 x i64>, ptr %idata, i64 6
  %12 = load <2 x i64>, ptr %11, align 16
  %13 = getelementptr <2 x i64>, ptr %idata, i64 7
  %14 = load <2 x i64>, ptr %13, align 16
  %15 = tail call { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } @llvm.x86.aesencwide128kl(ptr %h, <2 x i64> %0, <2 x i64> %2, <2 x i64> %4, <2 x i64> %6, <2 x i64> %8, <2 x i64> %10, <2 x i64> %12, <2 x i64> %14) #1
  %16 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 1
  store <2 x i64> %16, ptr %odata, align 16
  %17 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 2
  %18 = getelementptr <2 x i64>, ptr %odata, i64 1
  store <2 x i64> %17, ptr %18, align 16
  %19 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 3
  %20 = getelementptr <2 x i64>, ptr %odata, i64 2
  store <2 x i64> %19, ptr %20, align 16
  %21 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 4
  %22 = getelementptr <2 x i64>, ptr %odata, i64 3
  store <2 x i64> %21, ptr %22, align 16
  %23 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 5
  %24 = getelementptr <2 x i64>, ptr %odata, i64 4
  store <2 x i64> %23, ptr %24, align 16
  %25 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 6
  %26 = getelementptr <2 x i64>, ptr %odata, i64 5
  store <2 x i64> %25, ptr %26, align 16
  %27 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 7
  %28 = getelementptr <2 x i64>, ptr %odata, i64 6
  store <2 x i64> %27, ptr %28, align 16
  %29 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 8
  %30 = getelementptr <2 x i64>, ptr %odata, i64 7
  store <2 x i64> %29, ptr %30, align 16
  %31 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 0
  ret i8 %31
}

define zeroext i8 @test__mm_aesdecwide128kl_u8(ptr %odata, ptr %idata, ptr %h) {
; CHECK-LABEL: test__mm_aesdecwide128kl_u8:
; CHECK:       # %bb.0: # %entry
; CHECK-NEXT:    movaps (%rsi), %xmm0
; CHECK-NEXT:    movaps 16(%rsi), %xmm1
; CHECK-NEXT:    movaps 32(%rsi), %xmm2
; CHECK-NEXT:    movaps 48(%rsi), %xmm3
; CHECK-NEXT:    movaps 64(%rsi), %xmm4
; CHECK-NEXT:    movaps 80(%rsi), %xmm5
; CHECK-NEXT:    movaps 96(%rsi), %xmm6
; CHECK-NEXT:    movaps 112(%rsi), %xmm7
; CHECK-NEXT:    aesdecwide128kl (%rdx)
; CHECK-NEXT:    sete %al
; CHECK-NEXT:    movaps %xmm0, (%rdi)
; CHECK-NEXT:    movaps %xmm1, 16(%rdi)
; CHECK-NEXT:    movaps %xmm2, 32(%rdi)
; CHECK-NEXT:    movaps %xmm3, 48(%rdi)
; CHECK-NEXT:    movaps %xmm4, 64(%rdi)
; CHECK-NEXT:    movaps %xmm5, 80(%rdi)
; CHECK-NEXT:    movaps %xmm6, 96(%rdi)
; CHECK-NEXT:    movaps %xmm7, 112(%rdi)
; CHECK-NEXT:    retq
entry:
  %0 = load <2 x i64>, ptr %idata, align 16
  %1 = getelementptr <2 x i64>, ptr %idata, i64 1
  %2 = load <2 x i64>, ptr %1, align 16
  %3 = getelementptr <2 x i64>, ptr %idata, i64 2
  %4 = load <2 x i64>, ptr %3, align 16
  %5 = getelementptr <2 x i64>, ptr %idata, i64 3
  %6 = load <2 x i64>, ptr %5, align 16
  %7 = getelementptr <2 x i64>, ptr %idata, i64 4
  %8 = load <2 x i64>, ptr %7, align 16
  %9 = getelementptr <2 x i64>, ptr %idata, i64 5
  %10 = load <2 x i64>, ptr %9, align 16
  %11 = getelementptr <2 x i64>, ptr %idata, i64 6
  %12 = load <2 x i64>, ptr %11, align 16
  %13 = getelementptr <2 x i64>, ptr %idata, i64 7
  %14 = load <2 x i64>, ptr %13, align 16
  %15 = tail call { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } @llvm.x86.aesdecwide128kl(ptr %h, <2 x i64> %0, <2 x i64> %2, <2 x i64> %4, <2 x i64> %6, <2 x i64> %8, <2 x i64> %10, <2 x i64> %12, <2 x i64> %14) #1
  %16 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 1
  store <2 x i64> %16, ptr %odata, align 16
  %17 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 2
  %18 = getelementptr <2 x i64>, ptr %odata, i64 1
  store <2 x i64> %17, ptr %18, align 16
  %19 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 3
  %20 = getelementptr <2 x i64>, ptr %odata, i64 2
  store <2 x i64> %19, ptr %20, align 16
  %21 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 4
  %22 = getelementptr <2 x i64>, ptr %odata, i64 3
  store <2 x i64> %21, ptr %22, align 16
  %23 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 5
  %24 = getelementptr <2 x i64>, ptr %odata, i64 4
  store <2 x i64> %23, ptr %24, align 16
  %25 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 6
  %26 = getelementptr <2 x i64>, ptr %odata, i64 5
  store <2 x i64> %25, ptr %26, align 16
  %27 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 7
  %28 = getelementptr <2 x i64>, ptr %odata, i64 6
  store <2 x i64> %27, ptr %28, align 16
  %29 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 8
  %30 = getelementptr <2 x i64>, ptr %odata, i64 7
  store <2 x i64> %29, ptr %30, align 16
  %31 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 0
  ret i8 %31
}

define zeroext i8 @test__mm_aesencwide256kl_u8(ptr %odata, ptr %idata, ptr %h) {
; CHECK-LABEL: test__mm_aesencwide256kl_u8:
; CHECK:       # %bb.0: # %entry
; CHECK-NEXT:    movaps (%rsi), %xmm0
; CHECK-NEXT:    movaps 16(%rsi), %xmm1
; CHECK-NEXT:    movaps 32(%rsi), %xmm2
; CHECK-NEXT:    movaps 48(%rsi), %xmm3
; CHECK-NEXT:    movaps 64(%rsi), %xmm4
; CHECK-NEXT:    movaps 80(%rsi), %xmm5
; CHECK-NEXT:    movaps 96(%rsi), %xmm6
; CHECK-NEXT:    movaps 112(%rsi), %xmm7
; CHECK-NEXT:    aesencwide256kl (%rdx)
; CHECK-NEXT:    sete %al
; CHECK-NEXT:    movaps %xmm0, (%rdi)
; CHECK-NEXT:    movaps %xmm1, 16(%rdi)
; CHECK-NEXT:    movaps %xmm2, 32(%rdi)
; CHECK-NEXT:    movaps %xmm3, 48(%rdi)
; CHECK-NEXT:    movaps %xmm4, 64(%rdi)
; CHECK-NEXT:    movaps %xmm5, 80(%rdi)
; CHECK-NEXT:    movaps %xmm6, 96(%rdi)
; CHECK-NEXT:    movaps %xmm7, 112(%rdi)
; CHECK-NEXT:    retq
entry:
  %0 = load <2 x i64>, ptr %idata, align 16
  %1 = getelementptr <2 x i64>, ptr %idata, i64 1
  %2 = load <2 x i64>, ptr %1, align 16
  %3 = getelementptr <2 x i64>, ptr %idata, i64 2
  %4 = load <2 x i64>, ptr %3, align 16
  %5 = getelementptr <2 x i64>, ptr %idata, i64 3
  %6 = load <2 x i64>, ptr %5, align 16
  %7 = getelementptr <2 x i64>, ptr %idata, i64 4
  %8 = load <2 x i64>, ptr %7, align 16
  %9 = getelementptr <2 x i64>, ptr %idata, i64 5
  %10 = load <2 x i64>, ptr %9, align 16
  %11 = getelementptr <2 x i64>, ptr %idata, i64 6
  %12 = load <2 x i64>, ptr %11, align 16
  %13 = getelementptr <2 x i64>, ptr %idata, i64 7
  %14 = load <2 x i64>, ptr %13, align 16
  %15 = tail call { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } @llvm.x86.aesencwide256kl(ptr %h, <2 x i64> %0, <2 x i64> %2, <2 x i64> %4, <2 x i64> %6, <2 x i64> %8, <2 x i64> %10, <2 x i64> %12, <2 x i64> %14) #1
  %16 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 1
  store <2 x i64> %16, ptr %odata, align 16
  %17 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 2
  %18 = getelementptr <2 x i64>, ptr %odata, i64 1
  store <2 x i64> %17, ptr %18, align 16
  %19 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 3
  %20 = getelementptr <2 x i64>, ptr %odata, i64 2
  store <2 x i64> %19, ptr %20, align 16
  %21 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 4
  %22 = getelementptr <2 x i64>, ptr %odata, i64 3
  store <2 x i64> %21, ptr %22, align 16
  %23 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 5
  %24 = getelementptr <2 x i64>, ptr %odata, i64 4
  store <2 x i64> %23, ptr %24, align 16
  %25 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 6
  %26 = getelementptr <2 x i64>, ptr %odata, i64 5
  store <2 x i64> %25, ptr %26, align 16
  %27 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 7
  %28 = getelementptr <2 x i64>, ptr %odata, i64 6
  store <2 x i64> %27, ptr %28, align 16
  %29 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 8
  %30 = getelementptr <2 x i64>, ptr %odata, i64 7
  store <2 x i64> %29, ptr %30, align 16
  %31 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 0
  ret i8 %31
}

define zeroext i8 @test__mm_aesdecwide256kl_u8(ptr %odata, ptr %idata, ptr %h) {
; CHECK-LABEL: test__mm_aesdecwide256kl_u8:
; CHECK:       # %bb.0: # %entry
; CHECK-NEXT:    movaps (%rsi), %xmm0
; CHECK-NEXT:    movaps 16(%rsi), %xmm1
; CHECK-NEXT:    movaps 32(%rsi), %xmm2
; CHECK-NEXT:    movaps 48(%rsi), %xmm3
; CHECK-NEXT:    movaps 64(%rsi), %xmm4
; CHECK-NEXT:    movaps 80(%rsi), %xmm5
; CHECK-NEXT:    movaps 96(%rsi), %xmm6
; CHECK-NEXT:    movaps 112(%rsi), %xmm7
; CHECK-NEXT:    aesdecwide256kl (%rdx)
; CHECK-NEXT:    sete %al
; CHECK-NEXT:    movaps %xmm0, (%rdi)
; CHECK-NEXT:    movaps %xmm1, 16(%rdi)
; CHECK-NEXT:    movaps %xmm2, 32(%rdi)
; CHECK-NEXT:    movaps %xmm3, 48(%rdi)
; CHECK-NEXT:    movaps %xmm4, 64(%rdi)
; CHECK-NEXT:    movaps %xmm5, 80(%rdi)
; CHECK-NEXT:    movaps %xmm6, 96(%rdi)
; CHECK-NEXT:    movaps %xmm7, 112(%rdi)
; CHECK-NEXT:    retq
entry:
  %0 = load <2 x i64>, ptr %idata, align 16
  %1 = getelementptr <2 x i64>, ptr %idata, i64 1
  %2 = load <2 x i64>, ptr %1, align 16
  %3 = getelementptr <2 x i64>, ptr %idata, i64 2
  %4 = load <2 x i64>, ptr %3, align 16
  %5 = getelementptr <2 x i64>, ptr %idata, i64 3
  %6 = load <2 x i64>, ptr %5, align 16
  %7 = getelementptr <2 x i64>, ptr %idata, i64 4
  %8 = load <2 x i64>, ptr %7, align 16
  %9 = getelementptr <2 x i64>, ptr %idata, i64 5
  %10 = load <2 x i64>, ptr %9, align 16
  %11 = getelementptr <2 x i64>, ptr %idata, i64 6
  %12 = load <2 x i64>, ptr %11, align 16
  %13 = getelementptr <2 x i64>, ptr %idata, i64 7
  %14 = load <2 x i64>, ptr %13, align 16
  %15 = tail call { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } @llvm.x86.aesdecwide256kl(ptr %h, <2 x i64> %0, <2 x i64> %2, <2 x i64> %4, <2 x i64> %6, <2 x i64> %8, <2 x i64> %10, <2 x i64> %12, <2 x i64> %14) #1
  %16 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 1
  store <2 x i64> %16, ptr %odata, align 16
  %17 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 2
  %18 = getelementptr <2 x i64>, ptr %odata, i64 1
  store <2 x i64> %17, ptr %18, align 16
  %19 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 3
  %20 = getelementptr <2 x i64>, ptr %odata, i64 2
  store <2 x i64> %19, ptr %20, align 16
  %21 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 4
  %22 = getelementptr <2 x i64>, ptr %odata, i64 3
  store <2 x i64> %21, ptr %22, align 16
  %23 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 5
  %24 = getelementptr <2 x i64>, ptr %odata, i64 4
  store <2 x i64> %23, ptr %24, align 16
  %25 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 6
  %26 = getelementptr <2 x i64>, ptr %odata, i64 5
  store <2 x i64> %25, ptr %26, align 16
  %27 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 7
  %28 = getelementptr <2 x i64>, ptr %odata, i64 6
  store <2 x i64> %27, ptr %28, align 16
  %29 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 8
  %30 = getelementptr <2 x i64>, ptr %odata, i64 7
  store <2 x i64> %29, ptr %30, align 16
  %31 = extractvalue { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } %15, 0
  ret i8 %31
}

declare void @llvm.x86.loadiwkey(<2 x i64>, <2 x i64>, <2 x i64>, i32)
declare { i32, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } @llvm.x86.encodekey128(i32, <2 x i64>)
declare { i32, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } @llvm.x86.encodekey256(i32, <2 x i64>, <2 x i64>)
declare { i8, <2 x i64> } @llvm.x86.aesenc256kl(<2 x i64>, ptr)
declare { i8, <2 x i64> } @llvm.x86.aesdec256kl(<2 x i64>, ptr)
declare { i8, <2 x i64> } @llvm.x86.aesenc128kl(<2 x i64>, ptr)
declare { i8, <2 x i64> } @llvm.x86.aesdec128kl(<2 x i64>, ptr)
declare { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } @llvm.x86.aesencwide128kl(ptr, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>)
declare { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } @llvm.x86.aesdecwide128kl(ptr, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>)
declare { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } @llvm.x86.aesencwide256kl(ptr, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>)
declare { i8, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } @llvm.x86.aesdecwide256kl(ptr, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>)