[Clang][avx512][Builtin] Adding Builtin to clang.

[m_zuckerman]

Summary

builtin_ia32_ptestmb512
builtin_ia32_ptestmw512
builtin_ia32_ptestnmb512
builtin_ia32_ptestnmw512
builtin_ia32_storedquhi512_mask
builtin_ia32_storedquqi512_mask
builtin_ia32_rangesd128_round_mask
builtin_ia32_rangess128_round_mask
builtin_ia32_reducesd_mask
builtin_ia32_reducess_mask
builtin_ia32_pslld512_mask
builtin_ia32_psllq512_mask
builtin_ia32_psllv16si_mask
builtin_ia32_psllv8di_mask
builtin_ia32_psrad512_mask
builtin_ia32_psraq512_mask
builtin_ia32_psrav16si_mask
builtin_ia32_psrav8di_mask
builtin_ia32_psrld512_mask
builtin_ia32_psrlq512_mask
builtin_ia32_psrlv16si_mask
builtin_ia32_psrlv8di_mask
builtin_ia32_pternlogd512_mask
builtin_ia32_pternlogd512_maskz
builtin_ia32_pternlogq512_mask
builtin_ia32_pternlogq512_maskz
builtin_ia32_ptestnmd512
builtin_ia32_ptestnmq512
builtin_ia32_punpckhdq512_mask
builtin_ia32_punpckhqdq512_mask
builtin_ia32_punpckldq512_mask
builtin_ia32_punpcklqdq512_mask
builtin_ia32_rcp14sd
builtin_ia32_rcp14ss
builtin_ia32_rndscalesd_round_mask
builtin_ia32_rndscaless_round_mask
builtin_ia32_rsqrt14sd
builtin_ia32_rsqrt14ss
builtin_ia32_scalefpd512_mask
builtin_ia32_scalefps512_mask
builtin_ia32_scalefsd_round_mask
builtin_ia32_scalefss_round_mask
builtin_ia32_shuf_f32x4_mask
builtin_ia32_shuf_f64x2_mask
builtin_ia32_shuf_i32x4_mask
builtin_ia32_shuf_i64x2_mask
builtin_ia32_shufpd512_mask
builtin_ia32_shufps512_mask
builtin_ia32_sqrtsd_round_mask
builtin_ia32_sqrtss_round_mask
builtin_ia32_unpckhpd512_mask
builtin_ia32_unpckhps512_mask
builtin_ia32_unpcklpd512_mask
builtin_ia32_unpcklps512_mask
builtin_ia32_vcvtsd2si64
builtin_ia32_vcvtsd2si32
builtin_ia32_vcvtsd2usi32
builtin_ia32_vcvtsd2usi64
builtin_ia32_vcvtss2si32
builtin_ia32_vcvtss2si64
builtin_ia32_vcvtss2usi32
builtin_ia32_vcvtss2usi64
builtin_ia32_vcvttsd2si32
builtin_ia32_vcvttsd2si64
builtin_ia32_vcvttsd2usi32
builtin_ia32_vcvttsd2usi64
builtin_ia32_vcvttss2si32
builtin_ia32_vcvttss2si64
builtin_ia32_vcvttss2usi32
builtin_ia32_vcvttss2usi64
builtin_ia32_vpermi2vard512_mask
builtin_ia32_vpermi2varpd512_mask
builtin_ia32_vpermi2varps512_mask
builtin_ia32_vpermi2varq512_mask
builtin_ia32_vpermilpd512_mask
builtin_ia32_vpermilps512_mask
builtin_ia32_vpermilvarpd512_mask
builtin_ia32_vpermilvarps512_mask
builtin_ia32_vpermt2vard512_maskz
builtin_ia32_vpermt2varpd512_maskz
builtin_ia32_vpermt2varps512_maskz
builtin_ia32_vpermt2varq512_maskz
builtin_ia32_ptestmb128
builtin_ia32_ptestmb256
builtin_ia32_ptestmw128
builtin_ia32_ptestmw256
builtin_ia32_ptestnmb128
builtin_ia32_ptestnmb256
builtin_ia32_ptestnmw128
builtin_ia32_ptestnmw256
builtin_ia32_storedquhi128_mask
builtin_ia32_storedquhi256_mask
builtin_ia32_storedquqi128_mask
builtin_ia32_storedquqi256_mask
builtin_ia32_pternlogd128_mask
builtin_ia32_pternlogd128_maskz
builtin_ia32_pternlogd256_mask
builtin_ia32_pternlogd256_maskz
builtin_ia32_pternlogq128_mask
builtin_ia32_pternlogq128_maskz
builtin_ia32_pternlogq256_mask
builtin_ia32_pternlogq256_maskz
builtin_ia32_ptestmd128
builtin_ia32_ptestmd256
builtin_ia32_ptestmq128
builtin_ia32_ptestmq256
builtin_ia32_ptestnmd128
builtin_ia32_ptestnmd256
builtin_ia32_ptestnmq128
builtin_ia32_ptestnmq256
builtin_ia32_punpckhdq128_mask
builtin_ia32_punpckhdq256_mask
builtin_ia32_punpckhqdq128_mask
builtin_ia32_punpckhqdq256_mask
builtin_ia32_punpckldq128_mask
builtin_ia32_punpckldq256_mask
builtin_ia32_punpcklqdq128_mask
builtin_ia32_punpcklqdq256_mask
builtin_ia32_rcp14pd128_mask
builtin_ia32_rcp14pd256_mask
builtin_ia32_rcp14ps128_mask
builtin_ia32_rcp14ps256_mask
builtin_ia32_rsqrt14pd128_mask
builtin_ia32_rsqrt14pd256_mask
builtin_ia32_rsqrt14ps128_mask
builtin_ia32_rsqrt14ps256_mask
builtin_ia32_shuf_f32x4_256_mask
builtin_ia32_shuf_f64x2_256_mask
builtin_ia32_shuf_i32x4_256_mask
builtin_ia32_shuf_i64x2_256_mask
builtin_ia32_shufpd128_mask
builtin_ia32_shufpd256_mask
builtin_ia32_shufps128_mask
builtin_ia32_shufps256_mask
builtin_ia32_storeapd128_mask
builtin_ia32_storeapd256_mask
builtin_ia32_storeaps128_mask
builtin_ia32_storeaps256_mask
builtin_ia32_storedqudi128_mask
builtin_ia32_storedqudi256_mask
builtin_ia32_storedqusi128_mask
builtin_ia32_storedqusi256_mask
builtin_ia32_storeupd128_mask
builtin_ia32_storeupd256_mask
builtin_ia32_storeups128_mask
builtin_ia32_storeups256_mask
builtin_ia32_unpckhpd128_mask
builtin_ia32_unpckhpd256_mask
builtin_ia32_unpckhps128_mask
builtin_ia32_unpckhps256_mask
builtin_ia32_unpcklpd128_mask
builtin_ia32_unpcklpd256_mask
builtin_ia32_unpcklps128_mask
builtin_ia32_unpcklps256_mask
builtin_ia32_vcvtph2ps_mask
builtin_ia32_vcvtph2ps256_mask
builtin_ia32_vcvtps2ph_mask
builtin_ia32_vcvtps2ph256_mask
builtin_ia32_vpconflictdi_128_mask
builtin_ia32_vpconflictdi_256_mask
builtin_ia32_vpconflictsi_128_mask
builtin_ia32_vpconflictsi_256_mask
builtin_ia32_vpermilpd_mask
builtin_ia32_vpermilpd256_mask
builtin_ia32_vpermilps_mask
builtin_ia32_vpermilps256_mask
builtin_ia32_vpermilvarpd_mask
builtin_ia32_vpermilvarpd256_mask
builtin_ia32_vpermilvarps_mask
builtin_ia32_vpermilvarps256_mask
builtin_ia32_vplzcntd_128_mask
builtin_ia32_vplzcntd_256_mask
builtin_ia32_vplzcntq_128_mask
builtin_ia32_vplzcntq_256_mask
builtin_ia32_pshufd512_mask
builtin_ia32_pshufd128_mask
builtin_ia32_pshufd256_mask

[m_zuckerman]

ping

test/CodeGen/avx512f-builtins.c

1543

Delete extra line

[m_zuckerman]

test/CodeGen/avx512bw-builtins.c

745

delete extra line

750

delete extra line

[m_zuckerman]

ping

test/CodeGen/avx512dq-builtins.c

745 ↗

(On Diff #51552)

add this line

[m_zuckerman]

ping

include/clang/Basic/BuiltinsX86.def

1926

Delete this built, will be in another commit

1927–1928

There is no intrinsics use it.

lib/Headers/avx512fintrin.h

3796–3804

Delete this intrinsics . Will be in another commit

[AsafBadouh]

include/clang/Basic/BuiltinsX86.def

1758

missing implementation and tests for all "avx512dq"

1784

missing feature flags

1788

missing implementation

1792

missing feature flags

1800

missing feature flags

lib/Headers/avx512fintrin.h

4344

mask|z versions are missing for "roundscale_[round]_{sd|ss}"

4359

(v2df) _mm_setzero_ps()
--->
(v4sf) _mm_setzero_ps(),

4365

(v2df) _mm_setzero_ps()
--->
(v4sf) _mm_setzero_ps(),

test/CodeGen/avx512f-builtins.c

2416

It should it be @llvm.x86.avx512.mask.psllv.d

2434

same here, @llvm.x86.avx512.mask.psllv.q

2488

.mask.psrav.d

2512

.mask.psrav.q

2560

// CHECK: @llvm.x86.avx512.mask.psrlv.d

2578

// CHECK: @llvm.x86.avx512.mask.psrlv.q

[m_zuckerman]

include/clang/Basic/BuiltinsX86.def

1784

Missing implementation and test

1785

Missing implementation and test

1789

missing implementation

1907–1913

Missing implementation and test

[m_zuckerman]

test/CodeGen/avx512vlbw-builtins.c

1241

delete extra line

include/clang/Basic/BuiltinsX86.def

 TARGET_BUILTIN(__builtin_ia32_vpermi2varqi128_mask, "V16cV16cV16cV16cUs","","avx512vbmi,avx512vl")
 TARGET_BUILTIN(__builtin_ia32_vpermi2varqi256_mask, "V32cV32cV32cV32cUi","","avx512vbmi,avx512vl")
 TARGET_BUILTIN(__builtin_ia32_vpermt2varqi128_mask, "V16cV16cV16cV16cUs","","avx512vbmi,avx512vl")
 TARGET_BUILTIN(__builtin_ia32_vpermt2varqi128_maskz, "V16cV16cV16cV16cUs","","avx512vbmi,avx512vl")
 TARGET_BUILTIN(__builtin_ia32_vpermt2varqi256_mask, "V32cV32cV32cV32cUi","","avx512vbmi,avx512vl")
 TARGET_BUILTIN(__builtin_ia32_vpermt2varqi256_maskz, "V32cV32cV32cV32cUi","","avx512vbmi,avx512vl")
 TARGET_BUILTIN(__builtin_ia32_vcomisd, "iV2dV2dIiIi","","avx512f")
 TARGET_BUILTIN(__builtin_ia32_vcomiss, "iV4fV4fIiIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_ptestmb512, "ULLiV64cV64cULLi","","avx512bw")
+TARGET_BUILTIN(__builtin_ia32_ptestmw512, "UiV32sV32sUi","","avx512bw")
+TARGET_BUILTIN(__builtin_ia32_ptestnmb512, "ULLiV64cV64cULLi","","avx512bw")
+TARGET_BUILTIN(__builtin_ia32_ptestnmw512, "UiV32sV32sUi","","avx512bw")
+TARGET_BUILTIN(__builtin_ia32_storedquhi512_mask, "vV32s*V32sUi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_storedquqi512_mask, "vV64c*V64cULLi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_pslld512_mask, "V16iV16iV4iV16iUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_psllq512_mask, "V8LLiV8LLiV2LLiV8LLiUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_psllv16si_mask, "V16iV16iV16iV16iUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_psllv8di_mask, "V8LLiV8LLiV8LLiV8LLiUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_psrad512_mask, "V16iV16iV4iV16iUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_psraq512_mask, "V8LLiV8LLiV2LLiV8LLiUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_psrav16si_mask, "V16iV16iV16iV16iUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_psrav8di_mask, "V8LLiV8LLiV8LLiV8LLiUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_psrld512_mask, "V16iV16iV4iV16iUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_psrlq512_mask, "V8LLiV8LLiV2LLiV8LLiUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_psrlv16si_mask, "V16iV16iV16iV16iUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_psrlv8di_mask, "V8LLiV8LLiV8LLiV8LLiUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_pternlogd512_mask, "V16iV16iV16iV16iIiUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_pternlogd512_maskz, "V16iV16iV16iV16iIiUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_pternlogq512_mask, "V8LLiV8LLiV8LLiV8LLiIiUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_pternlogq512_maskz, "V8LLiV8LLiV8LLiV8LLiIiUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_ptestnmd512, "UsV16iV16iUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_ptestnmq512, "UcV8LLiV8LLiUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_punpckhdq512_mask, "V16iV16iV16iV16iUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_punpckhqdq512_mask, "V8LLiV8LLiV8LLiV8LLiUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_punpckldq512_mask, "V16iV16iV16iV16iUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_punpcklqdq512_mask, "V8LLiV8LLiV8LLiV8LLiUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_rndscalesd_round_mask, "V2dV2dV2dV2dUcIiIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_rndscaless_round_mask, "V4fV4fV4fV4fUcIiIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_scalefpd512_mask, "V8dV8dV8dV8dUcIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_scalefps512_mask, "V16fV16fV16fV16fUsIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_scalefsd_round_mask, "V2dV2dV2dV2dUcIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_scalefss_round_mask, "V4fV4fV4fV4fUcIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_shuf_f32x4_mask, "V16fV16fV16fIiV16fUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_shuf_f64x2_mask, "V8dV8dV8dIiV8dUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_shuf_i32x4_mask, "V16iV16iV16iIiV16iUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_shuf_i64x2_mask, "V8LLiV8LLiV8LLiIiV8LLiUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_shufpd512_mask, "V8dV8dV8dIiV8dUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_shufps512_mask, "V16fV16fV16fIiV16fUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_sqrtsd_round_mask, "V2dV2dV2dV2dUcIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_sqrtss_round_mask, "V4fV4fV4fV4fUcIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_unpckhpd512_mask, "V8dV8dV8dV8dUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_unpckhps512_mask, "V16fV16fV16fV16fUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_unpcklpd512_mask, "V8dV8dV8dV8dUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_unpcklps512_mask, "V16fV16fV16fV16fUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvtsd2si64, "LLiV2dIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvtsd2si32, "iV2dIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvtsd2usi32, "UiV2dIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvtsd2usi64, "ULLiV2dIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvtss2si32, "iV4fIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvtss2si64, "LLiV4fIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvtss2usi32, "UiV4fIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvtss2usi64, "ULLiV4fIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvttsd2si32, "iV2dIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvttsd2si64, "LLiV2dIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvttsd2usi32, "UiV2dIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvttsd2usi64, "ULLiV2dIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvttss2si32, "iV4fIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvttss2si64, "LLiV4fIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvttss2usi32, "UiV4fIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvttss2usi64, "ULLiV4fIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpermi2vard512_mask, "V16iV16iV16iV16iUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpermi2varpd512_mask, "V8dV8dV8LLiV8dUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpermi2varps512_mask, "V16fV16fV16iV16fUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpermi2varq512_mask, "V8LLiV8LLiV8LLiV8LLiUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpermilpd512_mask, "V8dV8dIiV8dUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpermilps512_mask, "V16fV16fIiV16fUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpermilvarpd512_mask, "V8dV8dV8LLiV8dUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpermilvarps512_mask, "V16fV16fV16iV16fUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpermt2vard512_maskz, "V16iV16iV16iV16iUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varpd512_maskz, "V8dV8LLiV8dV8dUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varps512_maskz, "V16fV16iV16fV16fUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varq512_maskz, "V8LLiV8LLiV8LLiV8LLiUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_ptestmb128, "UsV16cV16cUs","","avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_ptestmb256, "UiV32cV32cUi","","avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_ptestmw128, "UcV8sV8sUc","","avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_ptestmw256, "UsV16sV16sUs","","avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_ptestnmb128, "UsV16cV16cUs","","avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_ptestnmb256, "UiV32cV32cUi","","avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_ptestnmw128, "UcV8sV8sUc","","avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_ptestnmw256, "UsV16sV16sUs","","avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_storedquhi128_mask, "vV8s*V8sUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_storedquhi256_mask, "vV16s*V16sUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_storedquqi128_mask, "vV16c*V16cUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_storedquqi256_mask, "vV32c*V32cUi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_pternlogd128_mask, "V4iV4iV4iV4iIiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pternlogd128_maskz, "V4iV4iV4iV4iIiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pternlogd256_mask, "V8iV8iV8iV8iIiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pternlogd256_maskz, "V8iV8iV8iV8iIiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pternlogq128_mask, "V2LLiV2LLiV2LLiV2LLiIiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pternlogq128_maskz, "V2LLiV2LLiV2LLiV2LLiIiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pternlogq256_mask, "V4LLiV4LLiV4LLiV4LLiIiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pternlogq256_maskz, "V4LLiV4LLiV4LLiV4LLiIiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_ptestmd128, "UcV4iV4iUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_ptestmd256, "UcV8iV8iUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_ptestmq128, "UcV2LLiV2LLiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_ptestmq256, "UcV4LLiV4LLiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_ptestnmd128, "UcV4iV4iUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_ptestnmd256, "UcV8iV8iUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_ptestnmq128, "UcV2LLiV2LLiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_ptestnmq256, "UcV4LLiV4LLiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_punpckhdq128_mask, "V4iV4iV4iV4iUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_punpckhdq256_mask, "V8iV8iV8iV8iUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_punpckhqdq128_mask, "V2LLiV2LLiV2LLiV2LLiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_punpckhqdq256_mask, "V4LLiV4LLiV4LLiV4LLiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_punpckldq128_mask, "V4iV4iV4iV4iUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_punpckldq256_mask, "V8iV8iV8iV8iUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_punpcklqdq128_mask, "V2LLiV2LLiV2LLiV2LLiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_punpcklqdq256_mask, "V4LLiV4LLiV4LLiV4LLiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_rcp14pd128_mask, "V2dV2dV2dUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_rcp14pd256_mask, "V4dV4dV4dUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_rcp14ps128_mask, "V4fV4fV4fUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_rcp14ps256_mask, "V8fV8fV8fUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_rsqrt14pd128_mask, "V2dV2dV2dUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_rsqrt14pd256_mask, "V4dV4dV4dUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_rsqrt14ps128_mask, "V4fV4fV4fUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_rsqrt14ps256_mask, "V8fV8fV8fUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_shuf_f32x4_256_mask, "V8fV8fV8fIiV8fUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_shuf_f64x2_256_mask, "V4dV4dV4dIiV4dUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_shuf_i32x4_256_mask, "V8iV8iV8iIiV8iUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_shuf_i64x2_256_mask, "V4LLiV4LLiV4LLiIiV4LLiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_shufpd128_mask, "V2dV2dV2dIiV2dUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_shufpd256_mask, "V4dV4dV4dIiV4dUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_shufps128_mask, "V4fV4fV4fIiV4fUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_shufps256_mask, "V8fV8fV8fIiV8fUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_storeapd128_mask, "vV2d*V2dUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_storeapd256_mask, "vV4d*V4dUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_storeaps128_mask, "vV4f*V4fUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_storeaps256_mask, "vV8f*V8fUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_storedqudi128_mask, "vV2LLi*V2LLiUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_storedqudi256_mask, "vV4LLi*V4LLiUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_storedqusi128_mask, "vV4i*V4iUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_storedqusi256_mask, "vV8i*V8iUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_storeupd128_mask, "vV2d*V2dUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_storeupd256_mask, "vV4d*V4dUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_storeups128_mask, "vV4f*V4fUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_storeups256_mask, "vV8f*V8fUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_unpckhpd128_mask, "V2dV2dV2dV2dUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_unpckhpd256_mask, "V4dV4dV4dV4dUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_unpckhps128_mask, "V4fV4fV4fV4fUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_unpckhps256_mask, "V8fV8fV8fV8fUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_unpcklpd128_mask, "V2dV2dV2dV2dUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_unpcklpd256_mask, "V4dV4dV4dV4dUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_unpcklps128_mask, "V4fV4fV4fV4fUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_unpcklps256_mask, "V8fV8fV8fV8fUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermilpd_mask, "V2dV2dIiV2dUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermilpd256_mask, "V4dV4dIiV4dUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermilps_mask, "V4fV4fIiV4fUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermilps256_mask, "V8fV8fIiV8fUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermilvarpd_mask, "V2dV2dV2LLiV2dUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermilvarpd256_mask, "V4dV4dV4LLiV4dUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermilvarps_mask, "V4fV4fV4iV4fUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermilvarps256_mask, "V8fV8fV8iV8fUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vplzcntd_128_mask, "V4iV4iV4iUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vplzcntd_256_mask, "V8iV8iV8iUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vplzcntq_128_mask, "V2LLiV2LLiV2LLiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vplzcntq_256_mask, "V4LLiV4LLiV4LLiUc","","avx512vl")
 
 #undef BUILTIN
 #undef TARGET_BUILTIN

lib/Headers/avx512bwintrin.h

 _mm512_maskz_set1_epi8 (__mmask64 __M, char __A)
 {
   return (__m512i) __builtin_ia32_pbroadcastb512_gpr_mask (__A,
                  (__v64qi)
                  _mm512_setzero_qi(),
                  __M);
 }
 
+static __inline__ __mmask64 __DEFAULT_FN_ATTRS
+_mm512_test_epi8_mask (__m512i __A, __m512i __B)
+{
+  return (__mmask64) __builtin_ia32_ptestmb512 ((__v64qi) __A,
+            (__v64qi) __B,
+            (__mmask64) -1);
+}
+
+static __inline__ __mmask64 __DEFAULT_FN_ATTRS
+_mm512_mask_test_epi8_mask (__mmask64 __U, __m512i __A, __m512i __B)
+{
+  return (__mmask64) __builtin_ia32_ptestmb512 ((__v64qi) __A,
+            (__v64qi) __B, __U);
+}
+
+static __inline__ __mmask32 __DEFAULT_FN_ATTRS
+_mm512_test_epi16_mask (__m512i __A, __m512i __B)
+{
+  return (__mmask32) __builtin_ia32_ptestmw512 ((__v32hi) __A,
+            (__v32hi) __B,
+            (__mmask32) -1);
+}
+
+static __inline__ __mmask32 __DEFAULT_FN_ATTRS
+_mm512_mask_test_epi16_mask (__mmask32 __U, __m512i __A, __m512i __B)
+{
+  return (__mmask32) __builtin_ia32_ptestmw512 ((__v32hi) __A,
+            (__v32hi) __B, __U);
+}
+
+static __inline__ __mmask64 __DEFAULT_FN_ATTRS
+_mm512_testn_epi8_mask (__m512i __A, __m512i __B)
+{
+  return (__mmask64) __builtin_ia32_ptestnmb512 ((__v64qi) __A,
+             (__v64qi) __B,
+             (__mmask64) -1);
+}
+
+static __inline__ __mmask64 __DEFAULT_FN_ATTRS
+_mm512_mask_testn_epi8_mask (__mmask64 __U, __m512i __A, __m512i __B)
+{
+  return (__mmask64) __builtin_ia32_ptestnmb512 ((__v64qi) __A,
+             (__v64qi) __B, __U);
+}
+
+static __inline__ __mmask32 __DEFAULT_FN_ATTRS
+_mm512_testn_epi16_mask (__m512i __A, __m512i __B)
+{
+  return (__mmask32) __builtin_ia32_ptestnmw512 ((__v32hi) __A,
+             (__v32hi) __B,
+             (__mmask32) -1);
+}
+
+static __inline__ __mmask32 __DEFAULT_FN_ATTRS
+_mm512_mask_testn_epi16_mask (__mmask32 __U, __m512i __A, __m512i __B)
+{
+  return (__mmask32) __builtin_ia32_ptestnmw512 ((__v32hi) __A,
+             (__v32hi) __B, __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm512_mask_storeu_epi16 (void *__P, __mmask32 __U, __m512i __A)
+{
+  __builtin_ia32_storedquhi512_mask ((__v32hi *) __P,
+             (__v32hi) __A,
+             (__mmask32) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm512_mask_storeu_epi8 (void *__P, __mmask64 __U, __m512i __A)
+{
+  __builtin_ia32_storedquqi512_mask ((__v64qi *) __P,
+             (__v64qi) __A,
+             (__mmask64) __U);
+}
 #undef __DEFAULT_FN_ATTRS
 
 #endif

lib/Headers/avx512fintrin.h

 
 typedef float __m512 __attribute__((__vector_size__(64)));
 typedef double __m512d __attribute__((__vector_size__(64)));
 typedef long long __m512i __attribute__((__vector_size__(64)));
 
 typedef unsigned char __mmask8;
 typedef unsigned short __mmask16;
 
+typedef enum
+{
+  _MM_PERM_AAAA = 0x00, _MM_PERM_AAAB = 0x01, _MM_PERM_AAAC = 0x02,
+  _MM_PERM_AAAD = 0x03, _MM_PERM_AABA = 0x04, _MM_PERM_AABB = 0x05,
+  _MM_PERM_AABC = 0x06, _MM_PERM_AABD = 0x07, _MM_PERM_AACA = 0x08,
+  _MM_PERM_AACB = 0x09, _MM_PERM_AACC = 0x0A, _MM_PERM_AACD = 0x0B,
+  _MM_PERM_AADA = 0x0C, _MM_PERM_AADB = 0x0D, _MM_PERM_AADC = 0x0E,
+  _MM_PERM_AADD = 0x0F, _MM_PERM_ABAA = 0x10, _MM_PERM_ABAB = 0x11,
+  _MM_PERM_ABAC = 0x12, _MM_PERM_ABAD = 0x13, _MM_PERM_ABBA = 0x14,
+  _MM_PERM_ABBB = 0x15, _MM_PERM_ABBC = 0x16, _MM_PERM_ABBD = 0x17,
+  _MM_PERM_ABCA = 0x18, _MM_PERM_ABCB = 0x19, _MM_PERM_ABCC = 0x1A,
+  _MM_PERM_ABCD = 0x1B, _MM_PERM_ABDA = 0x1C, _MM_PERM_ABDB = 0x1D,
+  _MM_PERM_ABDC = 0x1E, _MM_PERM_ABDD = 0x1F, _MM_PERM_ACAA = 0x20,
+  _MM_PERM_ACAB = 0x21, _MM_PERM_ACAC = 0x22, _MM_PERM_ACAD = 0x23,
+  _MM_PERM_ACBA = 0x24, _MM_PERM_ACBB = 0x25, _MM_PERM_ACBC = 0x26,
+  _MM_PERM_ACBD = 0x27, _MM_PERM_ACCA = 0x28, _MM_PERM_ACCB = 0x29,
+  _MM_PERM_ACCC = 0x2A, _MM_PERM_ACCD = 0x2B, _MM_PERM_ACDA = 0x2C,
+  _MM_PERM_ACDB = 0x2D, _MM_PERM_ACDC = 0x2E, _MM_PERM_ACDD = 0x2F,
+  _MM_PERM_ADAA = 0x30, _MM_PERM_ADAB = 0x31, _MM_PERM_ADAC = 0x32,
+  _MM_PERM_ADAD = 0x33, _MM_PERM_ADBA = 0x34, _MM_PERM_ADBB = 0x35,
+  _MM_PERM_ADBC = 0x36, _MM_PERM_ADBD = 0x37, _MM_PERM_ADCA = 0x38,
+  _MM_PERM_ADCB = 0x39, _MM_PERM_ADCC = 0x3A, _MM_PERM_ADCD = 0x3B,
+  _MM_PERM_ADDA = 0x3C, _MM_PERM_ADDB = 0x3D, _MM_PERM_ADDC = 0x3E,
+  _MM_PERM_ADDD = 0x3F, _MM_PERM_BAAA = 0x40, _MM_PERM_BAAB = 0x41,
+  _MM_PERM_BAAC = 0x42, _MM_PERM_BAAD = 0x43, _MM_PERM_BABA = 0x44,
+  _MM_PERM_BABB = 0x45, _MM_PERM_BABC = 0x46, _MM_PERM_BABD = 0x47,
+  _MM_PERM_BACA = 0x48, _MM_PERM_BACB = 0x49, _MM_PERM_BACC = 0x4A,
+  _MM_PERM_BACD = 0x4B, _MM_PERM_BADA = 0x4C, _MM_PERM_BADB = 0x4D,
+  _MM_PERM_BADC = 0x4E, _MM_PERM_BADD = 0x4F, _MM_PERM_BBAA = 0x50,
+  _MM_PERM_BBAB = 0x51, _MM_PERM_BBAC = 0x52, _MM_PERM_BBAD = 0x53,
+  _MM_PERM_BBBA = 0x54, _MM_PERM_BBBB = 0x55, _MM_PERM_BBBC = 0x56,
+  _MM_PERM_BBBD = 0x57, _MM_PERM_BBCA = 0x58, _MM_PERM_BBCB = 0x59,
+  _MM_PERM_BBCC = 0x5A, _MM_PERM_BBCD = 0x5B, _MM_PERM_BBDA = 0x5C,
+  _MM_PERM_BBDB = 0x5D, _MM_PERM_BBDC = 0x5E, _MM_PERM_BBDD = 0x5F,
+  _MM_PERM_BCAA = 0x60, _MM_PERM_BCAB = 0x61, _MM_PERM_BCAC = 0x62,
+  _MM_PERM_BCAD = 0x63, _MM_PERM_BCBA = 0x64, _MM_PERM_BCBB = 0x65,
+  _MM_PERM_BCBC = 0x66, _MM_PERM_BCBD = 0x67, _MM_PERM_BCCA = 0x68,
+  _MM_PERM_BCCB = 0x69, _MM_PERM_BCCC = 0x6A, _MM_PERM_BCCD = 0x6B,
+  _MM_PERM_BCDA = 0x6C, _MM_PERM_BCDB = 0x6D, _MM_PERM_BCDC = 0x6E,
+  _MM_PERM_BCDD = 0x6F, _MM_PERM_BDAA = 0x70, _MM_PERM_BDAB = 0x71,
+  _MM_PERM_BDAC = 0x72, _MM_PERM_BDAD = 0x73, _MM_PERM_BDBA = 0x74,
+  _MM_PERM_BDBB = 0x75, _MM_PERM_BDBC = 0x76, _MM_PERM_BDBD = 0x77,
+  _MM_PERM_BDCA = 0x78, _MM_PERM_BDCB = 0x79, _MM_PERM_BDCC = 0x7A,
+  _MM_PERM_BDCD = 0x7B, _MM_PERM_BDDA = 0x7C, _MM_PERM_BDDB = 0x7D,
+  _MM_PERM_BDDC = 0x7E, _MM_PERM_BDDD = 0x7F, _MM_PERM_CAAA = 0x80,
+  _MM_PERM_CAAB = 0x81, _MM_PERM_CAAC = 0x82, _MM_PERM_CAAD = 0x83,
+  _MM_PERM_CABA = 0x84, _MM_PERM_CABB = 0x85, _MM_PERM_CABC = 0x86,
+  _MM_PERM_CABD = 0x87, _MM_PERM_CACA = 0x88, _MM_PERM_CACB = 0x89,
+  _MM_PERM_CACC = 0x8A, _MM_PERM_CACD = 0x8B, _MM_PERM_CADA = 0x8C,
+  _MM_PERM_CADB = 0x8D, _MM_PERM_CADC = 0x8E, _MM_PERM_CADD = 0x8F,
+  _MM_PERM_CBAA = 0x90, _MM_PERM_CBAB = 0x91, _MM_PERM_CBAC = 0x92,
+  _MM_PERM_CBAD = 0x93, _MM_PERM_CBBA = 0x94, _MM_PERM_CBBB = 0x95,
+  _MM_PERM_CBBC = 0x96, _MM_PERM_CBBD = 0x97, _MM_PERM_CBCA = 0x98,
+  _MM_PERM_CBCB = 0x99, _MM_PERM_CBCC = 0x9A, _MM_PERM_CBCD = 0x9B,
+  _MM_PERM_CBDA = 0x9C, _MM_PERM_CBDB = 0x9D, _MM_PERM_CBDC = 0x9E,
+  _MM_PERM_CBDD = 0x9F, _MM_PERM_CCAA = 0xA0, _MM_PERM_CCAB = 0xA1,
+  _MM_PERM_CCAC = 0xA2, _MM_PERM_CCAD = 0xA3, _MM_PERM_CCBA = 0xA4,
+  _MM_PERM_CCBB = 0xA5, _MM_PERM_CCBC = 0xA6, _MM_PERM_CCBD = 0xA7,
+  _MM_PERM_CCCA = 0xA8, _MM_PERM_CCCB = 0xA9, _MM_PERM_CCCC = 0xAA,
+  _MM_PERM_CCCD = 0xAB, _MM_PERM_CCDA = 0xAC, _MM_PERM_CCDB = 0xAD,
+  _MM_PERM_CCDC = 0xAE, _MM_PERM_CCDD = 0xAF, _MM_PERM_CDAA = 0xB0,
+  _MM_PERM_CDAB = 0xB1, _MM_PERM_CDAC = 0xB2, _MM_PERM_CDAD = 0xB3,
+  _MM_PERM_CDBA = 0xB4, _MM_PERM_CDBB = 0xB5, _MM_PERM_CDBC = 0xB6,
+  _MM_PERM_CDBD = 0xB7, _MM_PERM_CDCA = 0xB8, _MM_PERM_CDCB = 0xB9,
+  _MM_PERM_CDCC = 0xBA, _MM_PERM_CDCD = 0xBB, _MM_PERM_CDDA = 0xBC,
+  _MM_PERM_CDDB = 0xBD, _MM_PERM_CDDC = 0xBE, _MM_PERM_CDDD = 0xBF,
+  _MM_PERM_DAAA = 0xC0, _MM_PERM_DAAB = 0xC1, _MM_PERM_DAAC = 0xC2,
+  _MM_PERM_DAAD = 0xC3, _MM_PERM_DABA = 0xC4, _MM_PERM_DABB = 0xC5,
+  _MM_PERM_DABC = 0xC6, _MM_PERM_DABD = 0xC7, _MM_PERM_DACA = 0xC8,
+  _MM_PERM_DACB = 0xC9, _MM_PERM_DACC = 0xCA, _MM_PERM_DACD = 0xCB,
+  _MM_PERM_DADA = 0xCC, _MM_PERM_DADB = 0xCD, _MM_PERM_DADC = 0xCE,
+  _MM_PERM_DADD = 0xCF, _MM_PERM_DBAA = 0xD0, _MM_PERM_DBAB = 0xD1,
+  _MM_PERM_DBAC = 0xD2, _MM_PERM_DBAD = 0xD3, _MM_PERM_DBBA = 0xD4,
+  _MM_PERM_DBBB = 0xD5, _MM_PERM_DBBC = 0xD6, _MM_PERM_DBBD = 0xD7,
+  _MM_PERM_DBCA = 0xD8, _MM_PERM_DBCB = 0xD9, _MM_PERM_DBCC = 0xDA,
+  _MM_PERM_DBCD = 0xDB, _MM_PERM_DBDA = 0xDC, _MM_PERM_DBDB = 0xDD,
+  _MM_PERM_DBDC = 0xDE, _MM_PERM_DBDD = 0xDF, _MM_PERM_DCAA = 0xE0,
+  _MM_PERM_DCAB = 0xE1, _MM_PERM_DCAC = 0xE2, _MM_PERM_DCAD = 0xE3,
+  _MM_PERM_DCBA = 0xE4, _MM_PERM_DCBB = 0xE5, _MM_PERM_DCBC = 0xE6,
+  _MM_PERM_DCBD = 0xE7, _MM_PERM_DCCA = 0xE8, _MM_PERM_DCCB = 0xE9,
+  _MM_PERM_DCCC = 0xEA, _MM_PERM_DCCD = 0xEB, _MM_PERM_DCDA = 0xEC,
+  _MM_PERM_DCDB = 0xED, _MM_PERM_DCDC = 0xEE, _MM_PERM_DCDD = 0xEF,
+  _MM_PERM_DDAA = 0xF0, _MM_PERM_DDAB = 0xF1, _MM_PERM_DDAC = 0xF2,
+  _MM_PERM_DDAD = 0xF3, _MM_PERM_DDBA = 0xF4, _MM_PERM_DDBB = 0xF5,
+  _MM_PERM_DDBC = 0xF6, _MM_PERM_DDBD = 0xF7, _MM_PERM_DDCA = 0xF8,
+  _MM_PERM_DDCB = 0xF9, _MM_PERM_DDCC = 0xFA, _MM_PERM_DDCD = 0xFB,
+  _MM_PERM_DDDA = 0xFC, _MM_PERM_DDDB = 0xFD, _MM_PERM_DDDC = 0xFE,
+  _MM_PERM_DDDD = 0xFF
+} _MM_PERM_ENUM;
+
 /* Rounding mode macros.  */
 #define _MM_FROUND_TO_NEAREST_INT   0x00
 #define _MM_FROUND_TO_NEG_INF       0x01
 #define _MM_FROUND_TO_POS_INF       0x02
 #define _MM_FROUND_TO_ZERO          0x03
 #define _MM_FROUND_CUR_DIRECTION    0x04
 
 /* Define the default attributes for the functions in this file. */

 #define _mm_comi_round_sd(__A, __B, __P, __R) __extension__ ({\
 __builtin_ia32_vcomisd ((__v2df) (__A), (__v2df) (__B), ( __P), (  __R));\
 })
 
 #define _mm_comi_round_ss( __A,  __B, __P, __R) __extension__ ({\
 __builtin_ia32_vcomiss ((__v4sf) (__A), (__v4sf) (__B), ( __P), ( __R));\
 })
 
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_sll_epi32 (__m512i __A, __m128i __B)
+{
+  return (__m512i) __builtin_ia32_pslld512_mask ((__v16si) __A,
+             (__v4si) __B,
+             (__v16si)
+             _mm512_setzero_si512 (),
+             (__mmask16) -1);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_mask_sll_epi32 (__m512i __W, __mmask16 __U, __m512i __A, __m128i __B)
+{
+  return (__m512i) __builtin_ia32_pslld512_mask ((__v16si) __A,
+             (__v4si) __B,
+             (__v16si) __W,
+             (__mmask16) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_maskz_sll_epi32 (__mmask16 __U, __m512i __A, __m128i __B)
+{
+  return (__m512i) __builtin_ia32_pslld512_mask ((__v16si) __A,
+             (__v4si) __B,
+             (__v16si)
+             _mm512_setzero_si512 (),
+             (__mmask16) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_sll_epi64 (__m512i __A, __m128i __B)
+{
+  return (__m512i) __builtin_ia32_psllq512_mask ((__v8di) __A,
+             (__v2di) __B,
+             (__v8di)
+             _mm512_setzero_si512 (),
+             (__mmask8) -1);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_mask_sll_epi64 (__m512i __W, __mmask8 __U, __m512i __A, __m128i __B)
+{
+  return (__m512i) __builtin_ia32_psllq512_mask ((__v8di) __A,
+             (__v2di) __B,
+             (__v8di) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_maskz_sll_epi64 (__mmask8 __U, __m512i __A, __m128i __B)
+{
+  return (__m512i) __builtin_ia32_psllq512_mask ((__v8di) __A,
+             (__v2di) __B,
+             (__v8di)
+             _mm512_setzero_si512 (),
+             (__mmask8) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_sllv_epi32 (__m512i __X, __m512i __Y)
+{
+  return (__m512i) __builtin_ia32_psllv16si_mask ((__v16si) __X,
+              (__v16si) __Y,
+              (__v16si)
+              _mm512_setzero_si512 (),
+              (__mmask16) -1);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_mask_sllv_epi32 (__m512i __W, __mmask16 __U, __m512i __X, __m512i __Y)
+{
+  return (__m512i) __builtin_ia32_psllv16si_mask ((__v16si) __X,
+              (__v16si) __Y,
+              (__v16si) __W,
+              (__mmask16) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_maskz_sllv_epi32 (__mmask16 __U, __m512i __X, __m512i __Y)
+{
+  return (__m512i) __builtin_ia32_psllv16si_mask ((__v16si) __X,
+              (__v16si) __Y,
+              (__v16si)
+              _mm512_setzero_si512 (),
+              (__mmask16) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_sllv_epi64 (__m512i __X, __m512i __Y)
+{
+  return (__m512i) __builtin_ia32_psllv8di_mask ((__v8di) __X,
+             (__v8di) __Y,
+             (__v8di)
+             _mm512_undefined_pd (),
+             (__mmask8) -1);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_mask_sllv_epi64 (__m512i __W, __mmask8 __U, __m512i __X, __m512i __Y)
+{
+  return (__m512i) __builtin_ia32_psllv8di_mask ((__v8di) __X,
+             (__v8di) __Y,
+             (__v8di) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_maskz_sllv_epi64 (__mmask8 __U, __m512i __X, __m512i __Y)
+{
+  return (__m512i) __builtin_ia32_psllv8di_mask ((__v8di) __X,
+             (__v8di) __Y,
+             (__v8di)
+             _mm512_setzero_si512 (),
+             (__mmask8) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_sra_epi32 (__m512i __A, __m128i __B)
+{
+  return (__m512i) __builtin_ia32_psrad512_mask ((__v16si) __A,
+             (__v4si) __B,
+             (__v16si)
+             _mm512_setzero_si512 (),
+             (__mmask16) -1);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_mask_sra_epi32 (__m512i __W, __mmask16 __U, __m512i __A, __m128i __B)
+{
+  return (__m512i) __builtin_ia32_psrad512_mask ((__v16si) __A,
+             (__v4si) __B,
+             (__v16si) __W,
+             (__mmask16) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_maskz_sra_epi32 (__mmask16 __U, __m512i __A, __m128i __B)
+{
+  return (__m512i) __builtin_ia32_psrad512_mask ((__v16si) __A,
+             (__v4si) __B,
+             (__v16si)
+             _mm512_setzero_si512 (),
+             (__mmask16) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_sra_epi64 (__m512i __A, __m128i __B)
+{
+  return (__m512i) __builtin_ia32_psraq512_mask ((__v8di) __A,
+             (__v2di) __B,
+             (__v8di)
+             _mm512_setzero_si512 (),
+             (__mmask8) -1);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_mask_sra_epi64 (__m512i __W, __mmask8 __U, __m512i __A, __m128i __B)
+{
+  return (__m512i) __builtin_ia32_psraq512_mask ((__v8di) __A,
+             (__v2di) __B,
+             (__v8di) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_maskz_sra_epi64 (__mmask8 __U, __m512i __A, __m128i __B)
+{
+  return (__m512i) __builtin_ia32_psraq512_mask ((__v8di) __A,
+             (__v2di) __B,
+             (__v8di)
+             _mm512_setzero_si512 (),
+             (__mmask8) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_srav_epi32 (__m512i __X, __m512i __Y)
+{
+  return (__m512i) __builtin_ia32_psrav16si_mask ((__v16si) __X,
+              (__v16si) __Y,
+              (__v16si)
+              _mm512_setzero_si512 (),
+              (__mmask16) -1);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_mask_srav_epi32 (__m512i __W, __mmask16 __U, __m512i __X, __m512i __Y)
+{
+  return (__m512i) __builtin_ia32_psrav16si_mask ((__v16si) __X,
+              (__v16si) __Y,
+              (__v16si) __W,
+              (__mmask16) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_maskz_srav_epi32 (__mmask16 __U, __m512i __X, __m512i __Y)
+{
+  return (__m512i) __builtin_ia32_psrav16si_mask ((__v16si) __X,
+              (__v16si) __Y,
+              (__v16si)
+              _mm512_setzero_si512 (),
+              (__mmask16) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_srav_epi64 (__m512i __X, __m512i __Y)
+{
+  return (__m512i) __builtin_ia32_psrav8di_mask ((__v8di) __X,
+             (__v8di) __Y,
+             (__v8di)
+             _mm512_setzero_si512 (),
+             (__mmask8) -1);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_mask_srav_epi64 (__m512i __W, __mmask8 __U, __m512i __X, __m512i __Y)
+{
+  return (__m512i) __builtin_ia32_psrav8di_mask ((__v8di) __X,
+             (__v8di) __Y,
+             (__v8di) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_maskz_srav_epi64 (__mmask8 __U, __m512i __X, __m512i __Y)
+{
+  return (__m512i) __builtin_ia32_psrav8di_mask ((__v8di) __X,
+             (__v8di) __Y,
+             (__v8di)
+             _mm512_setzero_si512 (),
+             (__mmask8) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_srl_epi32 (__m512i __A, __m128i __B)
+{
+  return (__m512i) __builtin_ia32_psrld512_mask ((__v16si) __A,
+             (__v4si) __B,
+             (__v16si)
+             _mm512_setzero_si512 (),
+             (__mmask16) -1);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_mask_srl_epi32 (__m512i __W, __mmask16 __U, __m512i __A, __m128i __B)
+{
+  return (__m512i) __builtin_ia32_psrld512_mask ((__v16si) __A,
+             (__v4si) __B,
+             (__v16si) __W,
+             (__mmask16) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_maskz_srl_epi32 (__mmask16 __U, __m512i __A, __m128i __B)
+{
+  return (__m512i) __builtin_ia32_psrld512_mask ((__v16si) __A,
+             (__v4si) __B,
+             (__v16si)
+             _mm512_setzero_si512 (),
+             (__mmask16) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_srl_epi64 (__m512i __A, __m128i __B)
+{
+  return (__m512i) __builtin_ia32_psrlq512_mask ((__v8di) __A,
+             (__v2di) __B,
+             (__v8di)
+             _mm512_setzero_si512 (),
+             (__mmask8) -1);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_mask_srl_epi64 (__m512i __W, __mmask8 __U, __m512i __A, __m128i __B)
+{
+  return (__m512i) __builtin_ia32_psrlq512_mask ((__v8di) __A,
+             (__v2di) __B,
+             (__v8di) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_maskz_srl_epi64 (__mmask8 __U, __m512i __A, __m128i __B)
+{
+  return (__m512i) __builtin_ia32_psrlq512_mask ((__v8di) __A,
+             (__v2di) __B,
+             (__v8di)
+             _mm512_setzero_si512 (),
+             (__mmask8) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_srlv_epi32 (__m512i __X, __m512i __Y)
+{
+  return (__m512i) __builtin_ia32_psrlv16si_mask ((__v16si) __X,
+              (__v16si) __Y,
+              (__v16si)
+              _mm512_setzero_si512 (),
+              (__mmask16) -1);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_mask_srlv_epi32 (__m512i __W, __mmask16 __U, __m512i __X, __m512i __Y)
+{
+  return (__m512i) __builtin_ia32_psrlv16si_mask ((__v16si) __X,
+              (__v16si) __Y,
+              (__v16si) __W,
+              (__mmask16) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_maskz_srlv_epi32 (__mmask16 __U, __m512i __X, __m512i __Y)
+{
+  return (__m512i) __builtin_ia32_psrlv16si_mask ((__v16si) __X,
+              (__v16si) __Y,
+              (__v16si)
+              _mm512_setzero_si512 (),
+              (__mmask16) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_srlv_epi64 (__m512i __X, __m512i __Y)
+{
+  return (__m512i) __builtin_ia32_psrlv8di_mask ((__v8di) __X,
+             (__v8di) __Y,
+             (__v8di)
+             _mm512_setzero_si512 (),
+             (__mmask8) -1);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_mask_srlv_epi64 (__m512i __W, __mmask8 __U, __m512i __X, __m512i __Y)
+{
+  return (__m512i) __builtin_ia32_psrlv8di_mask ((__v8di) __X,
+             (__v8di) __Y,
+             (__v8di) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_maskz_srlv_epi64 (__mmask8 __U, __m512i __X, __m512i __Y)
+{
+  return (__m512i) __builtin_ia32_psrlv8di_mask ((__v8di) __X,
+             (__v8di) __Y,
+             (__v8di)
+             _mm512_setzero_si512 (),
+             (__mmask8) __U);
+}
+
+#define _mm512_ternarylogic_epi32( __A, __B, __C, imm) __extension__ ({ \
+__builtin_ia32_pternlogd512_mask ((__v16si)( __A),\
+                 (__v16si)( __B),\
+                 (__v16si)( __C),\
+                ( imm), (__mmask16) -1);\
+})
+
+#define _mm512_mask_ternarylogic_epi32( __A, __U, __B, __C, imm) __extension__ ({ \
+__builtin_ia32_pternlogd512_mask ((__v16si)( __A),\
+                 (__v16si)( __B),\
+                 (__v16si)( __C),\
+                ( imm), (__mmask16)( __U));\
+})
+
+#define _mm512_maskz_ternarylogic_epi32( __U, __A, __B, __C, imm) __extension__ ({ \
+__builtin_ia32_pternlogd512_maskz ((__v16si)( __A),\
+                  (__v16si)( __B),\
+                  (__v16si)( __C),\
+                 ( imm), (__mmask16)( __U));\
+})
+
+#define _mm512_ternarylogic_epi64( __A, __B, __C, imm) __extension__ ({ \
+__builtin_ia32_pternlogq512_mask ((__v8di)( __A),\
+                 (__v8di)( __B),\
+                 (__v8di)( __C),( imm),\
+                 (__mmask8) -1);\
+})
+
+#define _mm512_mask_ternarylogic_epi64( __A, __U, __B, __C, imm) __extension__ ({ \
+__builtin_ia32_pternlogq512_mask ((__v8di)( __A),\
+                 (__v8di)( __B),\
+                 (__v8di)( __C),( imm),\
+                 (__mmask8)( __U));\
+})
+
+#define _mm512_maskz_ternarylogic_epi64( __U, __A, __B, __C, imm) __extension__ ({ \
+__builtin_ia32_pternlogq512_maskz ((__v8di)( __A),\
+                  (__v8di)( __B),\
+                  (__v8di)( __C),\
+                 ( imm), (__mmask8)( __U));\
+})
+
+static __inline__ __mmask16 __DEFAULT_FN_ATTRS
+_mm512_testn_epi32_mask (__m512i __A, __m512i __B)
+{
+  return (__mmask16) __builtin_ia32_ptestnmd512 ((__v16si) __A,
+             (__v16si) __B,
+             (__mmask16) -1);
+}
+
+static __inline__ __mmask16 __DEFAULT_FN_ATTRS
+_mm512_mask_testn_epi32_mask (__mmask16 __U, __m512i __A, __m512i __B)
+{
+  return (__mmask16) __builtin_ia32_ptestnmd512 ((__v16si) __A,
+             (__v16si) __B, __U);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm512_testn_epi64_mask (__m512i __A, __m512i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestnmq512 ((__v8di) __A,
+            (__v8di) __B,
+            (__mmask8) -1);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm512_mask_testn_epi64_mask (__mmask8 __U, __m512i __A, __m512i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestnmq512 ((__v8di) __A,
+            (__v8di) __B, __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_unpackhi_epi32 (__m512i __A, __m512i __B)
+{
+  return (__m512i) __builtin_ia32_punpckhdq512_mask ((__v16si) __A,
+                 (__v16si) __B,
+                 (__v16si)
+                 _mm512_setzero_si512 (),
+                 (__mmask16) -1);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_mask_unpackhi_epi32 (__m512i __W, __mmask16 __U, __m512i __A,
+          __m512i __B)
+{
+  return (__m512i) __builtin_ia32_punpckhdq512_mask ((__v16si) __A,
+                 (__v16si) __B,
+                 (__v16si) __W,
+                 (__mmask16) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_maskz_unpackhi_epi32 (__mmask16 __U, __m512i __A, __m512i __B)
+{
+  return (__m512i) __builtin_ia32_punpckhdq512_mask ((__v16si) __A,
+                 (__v16si) __B,
+                 (__v16si)
+                 _mm512_setzero_si512 (),
+                 (__mmask16) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_unpackhi_epi64 (__m512i __A, __m512i __B)
+{
+  return (__m512i) __builtin_ia32_punpckhqdq512_mask ((__v8di) __A,
+                  (__v8di) __B,
+                  (__v8di)
+                  _mm512_setzero_si512 (),
+                  (__mmask8) -1);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_mask_unpackhi_epi64 (__m512i __W, __mmask8 __U, __m512i __A, __m512i __B)
+{
+  return (__m512i) __builtin_ia32_punpckhqdq512_mask ((__v8di) __A,
+                  (__v8di) __B,
+                  (__v8di) __W,
+                  (__mmask8) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_maskz_unpackhi_epi64 (__mmask8 __U, __m512i __A, __m512i __B)
+{
+  return (__m512i) __builtin_ia32_punpckhqdq512_mask ((__v8di) __A,
+                  (__v8di) __B,
+                  (__v8di)
+                  _mm512_setzero_si512 (),
+                  (__mmask8) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_unpacklo_epi32 (__m512i __A, __m512i __B)
+{
+  return (__m512i) __builtin_ia32_punpckldq512_mask ((__v16si) __A,
+                 (__v16si) __B,
+                 (__v16si)
+                 _mm512_setzero_si512 (),
+                 (__mmask16) -1);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_mask_unpacklo_epi32 (__m512i __W, __mmask16 __U, __m512i __A,
+          __m512i __B)
+{
+  return (__m512i) __builtin_ia32_punpckldq512_mask ((__v16si) __A,
+                 (__v16si) __B,
+                 (__v16si) __W,
+                 (__mmask16) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_maskz_unpacklo_epi32 (__mmask16 __U, __m512i __A, __m512i __B)
+{
+  return (__m512i) __builtin_ia32_punpckldq512_mask ((__v16si) __A,
+                 (__v16si) __B,
+                 (__v16si)
+                 _mm512_setzero_si512 (),
+                 (__mmask16) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_unpacklo_epi64 (__m512i __A, __m512i __B)
+{
+  return (__m512i) __builtin_ia32_punpcklqdq512_mask ((__v8di) __A,
+                  (__v8di) __B,
+                  (__v8di)
+                  _mm512_setzero_si512 (),
+                  (__mmask8) -1);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_mask_unpacklo_epi64 (__m512i __W, __mmask8 __U, __m512i __A, __m512i __B)
+{
+  return (__m512i) __builtin_ia32_punpcklqdq512_mask ((__v8di) __A,
+                  (__v8di) __B,
+                  (__v8di) __W,
+                  (__mmask8) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_maskz_unpacklo_epi64 (__mmask8 __U, __m512i __A, __m512i __B)
+{
+  return (__m512i) __builtin_ia32_punpcklqdq512_mask ((__v8di) __A,
+                  (__v8di) __B,
+                  (__v8di)
+                  _mm512_setzero_si512 (),
+                  (__mmask8) __U);
+}
+
+#define _mm_roundscale_round_sd( __A, __B, __imm, __R) __extension__ ({ \
+__builtin_ia32_rndscalesd_round_mask ((__v2df)( __A),\
+                (__v2df)( __B), (__v2df) _mm_setzero_pd(),\
+                (__mmask8) -1,( __imm),( __R));\
+})
+
+#define _mm_roundscale_sd( __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_rndscalesd_round_mask ((__v2df)( __A),\
+                (__v2df)( __B), (__v2df) _mm_setzero_pd(),\
+                (__mmask8) -1, ( __imm),\
+               _MM_FROUND_CUR_DIRECTION);\
+})
+
+#define _mm_mask_roundscale_sd( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_rndscalesd_round_mask ( (__v2df)( __A),\
+                (__v2df)( __B),\
+                (__v2df)( __W),\
+                (__mmask8)( __U),\
+                (__imm),\
+                _MM_FROUND_CUR_DIRECTION);\
+})
+
+#define _mm_mask_roundscale_round_sd( __W, __U, __A, __B, __I, __R) __extension__ ({ \
+__builtin_ia32_rndscalesd_round_mask ( (__v2df)( __A),\
+                (__v2df)( __B),\
+                (__v2df)( __W),\
+                (__mmask8)( __U),\
+                __I,\
+                __R);\
+})
+
+#define _mm_maskz_roundscale_sd( __U, __A, __B, __I) __extension__ ({ \
+__builtin_ia32_rndscalesd_round_mask ( (__v2df)( __A),\
+                 (__v2df)( __B),\
+                (__v2df) _mm_setzero_pd (),\
+                (__mmask8)( __U),\
+                __I,\
+                _MM_FROUND_CUR_DIRECTION);\
+})
+
+#define _mm_maskz_roundscale_round_sd( __U, __A, __B, __I, __R) __extension__ ({ \
+__builtin_ia32_rndscalesd_round_mask ( (__v2df)( __A),\
+                 (__v2df)( __B),\
+                (__v2df) _mm_setzero_pd (),\
+                (__mmask8)( __U),\
+                __I,\
+                __R);\
+})
+
+#define _mm_roundscale_round_ss( __A, __B, __imm, __R) __extension__ ({ \
+__builtin_ia32_rndscaless_round_mask ((__v4sf)( __A),\
+               (__v4sf)( __B),  (__v4sf) _mm_setzero_ps(),\
+                (__mmask8) -1, __imm, __R);\
+})
+
+#define _mm_roundscale_ss( __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_rndscaless_round_mask ((__v4sf)( __A),\
+               (__v4sf)( __B), (__v4sf) _mm_setzero_ps(),\
+                (__mmask8) -1, ( __imm),\
+               _MM_FROUND_CUR_DIRECTION);\
+})
+
+#define _mm_mask_roundscale_ss( __W, __U, __A, __B, __I) __extension__ ({ \
+__builtin_ia32_rndscaless_round_mask ( (__v4sf)(( __)A),\
+                 (__v4sf)( __B),\
+                 (__v4sf)( __W),\
+                 (__mmask8)( __U),\
+                 __I,\
+                 _MM_FROUND_CUR_DIRECTION);\
+})
+
+#define _mm_mask_roundscale_round_ss( __W, __U, __A, __B, __I, __R) __extension__ ({ \
+__builtin_ia32_rndscaless_round_mask ( (__v4sf)( __A),\
+                (__v4sf)( __B),\
+                (__v4sf)( __W),\
+                (__mmask8)( __U),\
+                __I,\
+                __R);\
+})
+
+#define _mm_maskz_roundscale_ss( __U, __A, __B, __I) __extension__ ({ \
+__builtin_ia32_rndscaless_round_mask ( (__v4sf)( __A),\
+                 (__v4sf)( __B),\
+                (__v4sf) _mm_setzero_ps (),\
+                (__mmask8)( __U),\
+                __I,\
+                _MM_FROUND_CUR_DIRECTION);\
+})
+
+#define _mm_maskz_roundscale_ss( __U, __A, __B, __I, __R) __extension__ ({ \
+__builtin_ia32_rndscaless_round_mask ( (__v4sf)( __A),\
+                 (__v4sf)( __B),\
+                (__v4sf) _mm_setzero_ps (),\
+                (__mmask8)( __U),\
+                __I,\
+                __R);\
+})
+
+
+#define _mm512_scalef_round_pd( __A, __B, __R) __extension__ ({ \
+__builtin_ia32_scalefpd512_mask ((__v8df)( __A),\
+                (__v8df)( __B),\
+                (__v8df)\
+                _mm512_undefined_pd (),\
+                (__mmask8) -1,( __R));\
+})
+
+#define _mm512_mask_scalef_round_pd( __W, __U, __A, __B, __R) __extension__ ({ \
+__builtin_ia32_scalefpd512_mask ((__v8df)( __A),\
+                (__v8df)( __B),\
+                (__v8df)( __W),\
+                (__mmask8)( __U),( __R));\
+})
+
+#define _mm512_maskz_scalef_round_pd( __U, __A, __B, __R) __extension__ ({ \
+__builtin_ia32_scalefpd512_mask ((__v8df)( __A),\
+                (__v8df)( __B),\
+                (__v8df)\
+                _mm512_setzero_pd (),\
+                (__mmask8)( __U),( __R));\
+})
+
+static __inline__ __m512d __DEFAULT_FN_ATTRS
+_mm512_scalef_pd (__m512d __A, __m512d __B)
+{
+  return (__m512d) __builtin_ia32_scalefpd512_mask ((__v8df) __A,
+                (__v8df) __B,
+                (__v8df)
+                _mm512_undefined_pd (),
+                (__mmask8) -1,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+static __inline__ __m512d __DEFAULT_FN_ATTRS
+_mm512_mask_scalef_pd (__m512d __W, __mmask8 __U, __m512d __A, __m512d __B)
+{
+  return (__m512d) __builtin_ia32_scalefpd512_mask ((__v8df) __A,
+                (__v8df) __B,
+                (__v8df) __W,
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+static __inline__ __m512d __DEFAULT_FN_ATTRS
+_mm512_maskz_scalef_pd (__mmask8 __U, __m512d __A, __m512d __B)
+{
+  return (__m512d) __builtin_ia32_scalefpd512_mask ((__v8df) __A,
+                (__v8df) __B,
+                (__v8df)
+                _mm512_setzero_pd (),
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm512_scalef_round_ps( __A, __B, __R) __extension__ ({ \
+__builtin_ia32_scalefps512_mask ((__v16sf)( __A),\
+               (__v16sf)( __B),\
+               (__v16sf)\
+               _mm512_undefined_ps (),\
+               (__mmask16) -1,( __R));\
+})
+
+#define _mm512_mask_scalef_round_ps( __W, __U, __A, __B, __R) __extension__ ({ \
+__builtin_ia32_scalefps512_mask ((__v16sf)( __A),\
+               (__v16sf)( __B),\
+               (__v16sf)( __W),\
+               (__mmask16)( __U),( __R));\
+})
+
+#define _mm512_maskz_scalef_round_ps( __U, __A, __B, __R) __extension__ ({ \
+__builtin_ia32_scalefps512_mask ((__v16sf)( __A),\
+               (__v16sf)( __B),\
+               (__v16sf)\
+               _mm512_setzero_ps (),\
+               (__mmask16)( __U),( __R));\
+})
+
+static __inline__ __m512 __DEFAULT_FN_ATTRS
+_mm512_scalef_ps (__m512 __A, __m512 __B)
+{
+  return (__m512) __builtin_ia32_scalefps512_mask ((__v16sf) __A,
+               (__v16sf) __B,
+               (__v16sf)
+               _mm512_undefined_ps (),
+               (__mmask16) -1,
+               _MM_FROUND_CUR_DIRECTION);
+}
+
+static __inline__ __m512 __DEFAULT_FN_ATTRS
+_mm512_mask_scalef_ps (__m512 __W, __mmask16 __U, __m512 __A, __m512 __B)
+{
+  return (__m512) __builtin_ia32_scalefps512_mask ((__v16sf) __A,
+               (__v16sf) __B,
+               (__v16sf) __W,
+               (__mmask16) __U,
+               _MM_FROUND_CUR_DIRECTION);
+}
+
+static __inline__ __m512 __DEFAULT_FN_ATTRS
+_mm512_maskz_scalef_ps (__mmask16 __U, __m512 __A, __m512 __B)
+{
+  return (__m512) __builtin_ia32_scalefps512_mask ((__v16sf) __A,
+               (__v16sf) __B,
+               (__v16sf)
+               _mm512_setzero_ps (),
+               (__mmask16) __U,
+               _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_scalef_round_sd( __A, __B, __R) __extension__ ({ \
+__builtin_ia32_scalefsd_round_mask ((__v2df)( __A),\
+              (__v2df)( __B), (__v2df) _mm_setzero_pd(),\
+              (__mmask8) -1,\
+              ( __R));\
+})
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_scalef_sd (__m128d __A, __m128d __B)
+{
+  return (__m128d) __builtin_ia32_scalefsd_round_mask ((__v2df) __A,
+              (__v2df)( __B), (__v2df) _mm_setzero_pd(),
+              (__mmask8) -1,
+              _MM_FROUND_CUR_DIRECTION);
+}
+
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_mask_scalef_sd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) 
+{
+ return (__m128d) __builtin_ia32_scalefsd_round_mask ( (__v2df) __A,
+                 (__v2df) __B,
+                (__v2df) __W,
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+
+#define _mm_mask_scalef_round_sd( __W, __U, __A, __B, __R) __extension__ ({ \
+__builtin_ia32_scalefsd_round_mask ((__v2df)( __A),\
+              (__v2df)( __B), (__v2df) __W,\
+              (__mmask8) __U,\
+              ( __R));\
+})
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_scalef_sd (__mmask8 __U, __m128d __A, __m128d __B) 
+{
+ return (__m128d) __builtin_ia32_scalefsd_round_mask ( (__v2df) __A,
+                 (__v2df) __B,
+                (__v2df) _mm_setzero_pd (),
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_maskz_scalef_round_sd( __U, __A, __B, __R) __extension__ ({ \
+__builtin_ia32_scalefsd_round_mask ((__v2df)( __A),\
+              (__v2df)( __B), (__v2df) _mm_setzero_pd (),\
+              (__mmask8) __U,\
+              ( __R));\
+})
+
+#define _mm_scalef_round_ss( __A, __B, __R) __extension__ ({ \
+__builtin_ia32_scalefss_round_mask ((__v4sf)( __A),\
+             (__v4sf)( __B), (__v4sf) _mm_setzero_ps(),\
+             (__mmask8) -1,\
+             ( __R));\
+})
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_scalef_ss (__m128 __A, __m128 __B)
+{
+  return (__m128) __builtin_ia32_scalefss_round_mask ((__v4sf) __A,
+             (__v4sf)( __B), (__v4sf) _mm_setzero_ps(),
+             (__mmask8) -1,
+             _MM_FROUND_CUR_DIRECTION);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_scalef_ss (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) 
+{
+ return (__m128) __builtin_ia32_scalefss_round_mask ( (__v4sf) __A,
+                (__v4sf) __B,
+                (__v4sf) __W,
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_mask_scalef_round_ss( __W, __U, __A, __B, __R) __extension__ ({ \
+__builtin_ia32_scalefss_round_mask ((__v4sf)( __A),\
+             (__v4sf)( __B), (__v4sf) __W,\
+             (__mmask8) __U,\
+             ( __R));\
+})
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_scalef_ss (__mmask8 __U, __m128 __A, __m128 __B) 
+{
+ return (__m128) __builtin_ia32_scalefss_round_mask ( (__v4sf) __A,
+                 (__v4sf) __B,
+                (__v4sf) _mm_setzero_ps (),
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_maskz_scalef_round_ss( __U, __A, __B, __R) __extension__ ({ \
+__builtin_ia32_scalefss_round_mask ((__v4sf)( __A),\
+             (__v4sf)( __B), (__v4sf) _mm_setzero_ps(),\
+             (__mmask8) __U,\
+             _MM_FROUND_CUR_DIRECTION);\
+})
+
+#define _mm512_i64scatter_ps( __addr,  __index,  __v1,  __scale) __extension__ ({ \
+__builtin_ia32_scatterdiv16sf ((__addr), (__mmask8) 0xFF,\
+         (__v8di) (__index), (__v8sf) (__v1), (__scale));\
+})
+
+#define _mm512_mask_i64scatter_ps( __addr, __mask, __index, __v1, __scale) __extension__ ({ \
+__builtin_ia32_scatterdiv16sf(__addr,( __mask), (__v8di)( __index),\
+         (__v8sf)( __v1),( __scale));\
+})
+
+#define _mm512_i64scatter_epi32( __addr, __index, __v1, __scale) __extension__ ({ \
+__builtin_ia32_scatterdiv16si (__addr, (__mmask8) 0xFF,\
+         (__v8di)( __index), (__v8si)( __v1),( __scale));\
+})
+
+#define _mm512_mask_i64scatter_epi32( __addr, __mask, __index, __v1, __scale) __extension__ ({ \
+__builtin_ia32_scatterdiv16si (__addr,( __mask), (__v8di)( __index),\
+         (__v8si)( __v1),( __scale));\
+})
+
+#define _mm512_i64scatter_pd( __addr, __index, __v1, __scale) __extension__ ({ \
+__builtin_ia32_scatterdiv8df (__addr, (__mmask8) 0xFF,\
+        (__v8di)( __index), (__v8df)( __v1),( __scale));\
+})
+
+#define _mm512_mask_i64scatter_pd( __addr, __mask, __index, __v1, __scale) __extension__ ({ \
+__builtin_ia32_scatterdiv8df (__addr,( __mask), (__v8di)( __index),\
+        (__v8df)( __v1),( __scale));\
+})
+
+#define _mm512_i64scatter_epi64( __addr, __mask,__v1, __scale) __extension__ ({ \
+__builtin_ia32_scatterdiv8di (__addr, (__mmask8) 0xFF,\
+        (__v8di) __index, (__v8di)( __v1),( __scale));\
+})
+
+#define _mm512_mask_i64scatter_epi64( __addr, __mask,__index, __v1, __scale) __extension__ ({ \
+__builtin_ia32_scatterdiv8di (__addr, __mask, (__v8di)( __index),\
+        (__v8di)( __v1),( __scale));\
+})
+
+#define _mm512_i32scatter_ps( __addr, __index, __v1, __scale) __extension__ ({ \
+__builtin_ia32_scattersiv16sf (__addr, (__mmask16) 0xFFFF,\
+         (__v16si)( __index), (__v16sf)( __v1),( __scale));\
+})
+
+#define _mm512_mask_i32scatter_ps( __addr, __mask, __index, __v1, __scale) __extension__ ({ \
+__builtin_ia32_scattersiv16sf (__addr,( __mask), (__v16si)( __index),\
+         (__v16sf)( __v1),( __scale));\
+})
+
+#define _mm512_i32scatter_epi32( __addr, __index, __v1, __scale) __extension__ ({ \
+__builtin_ia32_scattersiv16si (__addr, (__mmask16) 0xFFFF,\
+         (__v16si)( __index), (__v16si)( __v1),( __scale));\
+})
+
+#define _mm512_mask_i32scatter_epi32( __addr, __mask, __index, __v1, __scale) __extension__ ({ \
+__builtin_ia32_scattersiv16si (__addr,( __mask), (__v16si)( __index),\
+         (__v16si)( __v1),( __scale));\
+})
+
+#define _mm512_i32scatter_pd( __addr, __index, __v1, __scale) __extension__ ({ \
+__builtin_ia32_scattersiv8df (__addr, (__mmask8) 0xFF,\
+        (__v8si)( __index), (__v8df)( __v1),( __scale));\
+})
+
+#define _mm512_mask_i32scatter_pd( __addr, __mask, __index, __v1, __scale) __extension__ ({ \
+__builtin_ia32_scattersiv8df (__addr,( __mask), (__v8si)( __index),\
+        (__v8df)( __v1),( __scale));\
+})
+
+#define _mm512_i32scatter_epi64( __addr, __mask,__v1, __scale) __extension__ ({ \
+__builtin_ia32_scattersiv8di (__addr, (__mmask8) 0xFF,\
+        (__v8si) __index, (__v8di)( __v1),( __scale));\
+})
+
+#define _mm512_mask_i32scatter_epi64( __addr, __mask,__index, __v1, __scale) __extension__ ({ \
+__builtin_ia32_scattersiv8di (__addr, __mask, (__v8si)( __index),\
+        (__v8di)( __v1),( __scale));\
+})
+
+#define _mm512_shuffle_f32x4( __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_f32x4_mask ((__v16sf)( __A),\
+              (__v16sf)( __B),( __imm),\
+              (__v16sf)\
+              _mm512_undefined_ps (),\
+              (__mmask16) -1);\
+})
+
+#define _mm512_mask_shuffle_f32x4( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_f32x4_mask ((__v16sf)( __A),\
+              (__v16sf)( __B),( __imm),\
+              (__v16sf)( __W),\
+              (__mmask16)( __U));\
+})
+
+#define _mm512_maskz_shuffle_f32x4( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_f32x4_mask ((__v16sf)( __A),\
+              (__v16sf)( __B),( __imm),\
+              (__v16sf)\
+              _mm512_setzero_ps (),\
+              (__mmask16)( __U));\
+})
+
+#define _mm512_shuffle_f64x2( __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_f64x2_mask ((__v8df)( __A),\
+               (__v8df)( __B),( __imm),\
+               (__v8df)\
+               _mm512_undefined_pd (),\
+               (__mmask8) -1);\
+})
+
+#define _mm512_mask_shuffle_f64x2( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_f64x2_mask ((__v8df)( __A),\
+               (__v8df)( __B),( __imm),\
+               (__v8df)( __W),\
+               (__mmask8)( __U));\
+})
+
+#define _mm512_maskz_shuffle_f64x2( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_f64x2_mask ((__v8df)( __A),\
+               (__v8df)( __B),( __imm),\
+               (__v8df)\
+               _mm512_setzero_pd (),\
+               (__mmask8)( __U));\
+})
+
+#define _mm512_shuffle_i32x4( __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_i32x4_mask ((__v16si)( __A),\
+               (__v16si)( __B),\
+              ( __imm),\
+               (__v16si)\
+               _mm512_setzero_si512 (),\
+               (__mmask16) -1);\
+})
+
+#define _mm512_mask_shuffle_i32x4( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_i32x4_mask ((__v16si)( __A),\
+               (__v16si)( __B),\
+              ( __imm),\
+               (__v16si)( __W),\
+               (__mmask16)( __U));\
+})
+
+#define _mm512_maskz_shuffle_i32x4( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_i32x4_mask ((__v16si)( __A),\
+               (__v16si)( __B),\
+              ( __imm),\
+               (__v16si)\
+               _mm512_setzero_si512 (),\
+               (__mmask16)( __U));\
+})
+
+#define _mm512_shuffle_i64x2( __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_i64x2_mask ((__v8di)( __A),\
+               (__v8di)( __B),( __imm),\
+               (__v8di)\
+               _mm512_setzero_si512 (),\
+               (__mmask8) -1);\
+})
+
+#define _mm512_mask_shuffle_i64x2( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_i64x2_mask ((__v8di)( __A),\
+               (__v8di)( __B),( __imm),\
+               (__v8di)( __W),\
+               (__mmask8)( __U));\
+})
+
+#define _mm512_maskz_shuffle_i64x2( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_i64x2_mask ((__v8di)( __A),\
+               (__v8di)( __B),( __imm),\
+               (__v8di)\
+               _mm512_setzero_si512 (),\
+               (__mmask8)( __U));\
+})
+
+#define _mm512_shuffle_pd( __M, __V, __imm) __extension__ ({ \
+__builtin_ia32_shufpd512_mask ((__v8df)( __M),\
+              (__v8df)( __V),( __imm),\
+              (__v8df)\
+              _mm512_undefined_pd (),\
+              (__mmask8) -1);\
+})
+
+#define _mm512_mask_shuffle_pd( __W, __U, __M, __V, __imm) __extension__ ({ \
+__builtin_ia32_shufpd512_mask ((__v8df)( __M),\
+              (__v8df)( __V),( __imm),\
+              (__v8df)( __W),\
+              (__mmask8)( __U));\
+})
+
+#define _mm512_maskz_shuffle_pd( __U, __M, __V, __imm) __extension__ ({ \
+__builtin_ia32_shufpd512_mask ((__v8df)( __M),\
+              (__v8df)( __V),( __imm),\
+              (__v8df)\
+              _mm512_setzero_pd (),\
+              (__mmask8)( __U));\
+})
+
+#define _mm512_shuffle_ps( __M, __V, __imm) __extension__ ({ \
+__builtin_ia32_shufps512_mask ((__v16sf)( __M),\
+             (__v16sf)( __V),( __imm),\
+             (__v16sf)\
+             _mm512_undefined_ps (),\
+             (__mmask16) -1);\
+})
+
+#define _mm512_mask_shuffle_ps( __W, __U, __M, __V, __imm) __extension__ ({ \
+__builtin_ia32_shufps512_mask ((__v16sf)( __M),\
+             (__v16sf)( __V),( __imm),\
+             (__v16sf)( __W),\
+             (__mmask16)( __U));\
+})
+
+#define _mm512_maskz_shuffle_ps( __U, __M, __V, __imm) __extension__ ({ \
+__builtin_ia32_shufps512_mask ((__v16sf)( __M),\
+             (__v16sf)( __V),( __imm),\
+             (__v16sf)\
+             _mm512_setzero_ps (),\
+             (__mmask16)( __U));\
+})
+
+#define _mm_sqrt_round_sd( __A, __B, __R) __extension__ ({ \
+__builtin_ia32_sqrtsd_round_mask ((__v2df)( __B),\
+            (__v2df)( __A),(__v2df) _mm_setzero_pd(),\
+            (__mmask8) -1,\
+           ( __R));\
+})
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_mask_sqrt_sd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) 
+{
+ return (__m128d) __builtin_ia32_sqrtsd_round_mask ( (__v2df) __B,
+                 (__v2df) __A,
+                (__v2df) __W,
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_mask_sqrt_round_sd( __W, __U, __A, __B, __R) __extension__ ({ \
+__builtin_ia32_sqrtsd_round_mask ((__v2df)( __B),\
+            (__v2df)( __A),(__v2df) __W,\
+            (__mmask8) __U,\
+            ( __R));\
+})
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_sqrt_sd (__mmask8 __U, __m128d __A, __m128d __B) 
+{
+ return (__m128d) __builtin_ia32_sqrtsd_round_mask ( (__v2df) __B,
+                 (__v2df) __A,
+                (__v2df) _mm_setzero_pd (),
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_maskz_sqrt_round_sd( __U, __A, __B, __R) __extension__ ({ \
+__builtin_ia32_sqrtsd_round_mask ((__v2df)( __B),\
+            (__v2df)( __A),(__v2df) _mm_setzero_pd(),\
+            (__mmask8) __U,\
+            ( __R));\
+})
+
+#define _mm_sqrt_round_ss( __A, __B, __R) __extension__ ({ \
+__builtin_ia32_sqrtss_round_mask ((__v4sf)( __B),\
+                 (__v4sf)( __A),(__v4sf) _mm_setzero_ps(),\
+                 (__mmask8) -1,\
+                 ( __R));\
+})
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_sqrt_ss (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) 
+{
+ return (__m128) __builtin_ia32_sqrtss_round_mask ( (__v4sf) __B,
+                 (__v4sf) __A,
+                (__v4sf) __W,
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_mask_sqrt_round_ss( __W, __U, __A, __B, __R) __extension__ ({ \
+__builtin_ia32_sqrtss_round_mask ((__v4sf)( __B),\
+                 (__v4sf)( __A),(__v4sf) __W,\
+                 (__mmask8) __U,\
+                 ( __R));\
+})
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_sqrt_ss (__mmask8 __U, __m128 __A, __m128 __B) 
+{
+ return (__m128) __builtin_ia32_sqrtss_round_mask ( (__v4sf) __A,
+                 (__v4sf) __B,
+                (__v4sf) _mm_setzero_ps (),
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_maskz_sqrt_round_ss( __U, __A, __B, __R) __extension__ ({ \
+__builtin_ia32_sqrtss_round_mask ((__v4sf)( __B),\
+                 (__v4sf)( __A),(__v4sf) _mm_setzero_ps(),\
+                 (__mmask8) __U,\
+                 _MM_FROUND_CUR_DIRECTION);\
+})
+
+static __inline__ __m512d __DEFAULT_FN_ATTRS
+_mm512_mask_unpackhi_pd (__m512d __W, __mmask8 __U, __m512d __A, __m512d __B)
+{
+  return (__m512d) __builtin_ia32_unpckhpd512_mask ((__v8df) __A,
+                (__v8df) __B,
+                (__v8df) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m512d __DEFAULT_FN_ATTRS
+_mm512_maskz_unpackhi_pd (__mmask8 __U, __m512d __A, __m512d __B)
+{
+  return (__m512d) __builtin_ia32_unpckhpd512_mask ((__v8df) __A,
+                (__v8df) __B,
+                (__v8df)
+                _mm512_setzero_pd (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m512 __DEFAULT_FN_ATTRS
+_mm512_mask_unpackhi_ps (__m512 __W, __mmask16 __U, __m512 __A, __m512 __B)
+{
+  return (__m512) __builtin_ia32_unpckhps512_mask ((__v16sf) __A,
+               (__v16sf) __B,
+               (__v16sf) __W,
+               (__mmask16) __U);
+}
+
+static __inline__ __m512 __DEFAULT_FN_ATTRS
+_mm512_maskz_unpackhi_ps (__mmask16 __U, __m512 __A, __m512 __B)
+{
+  return (__m512) __builtin_ia32_unpckhps512_mask ((__v16sf) __A,
+               (__v16sf) __B,
+               (__v16sf)
+               _mm512_setzero_ps (),
+               (__mmask16) __U);
+}
+
+static __inline__ __m512d __DEFAULT_FN_ATTRS
+_mm512_mask_unpacklo_pd (__m512d __W, __mmask8 __U, __m512d __A, __m512d __B)
+{
+  return (__m512d) __builtin_ia32_unpcklpd512_mask ((__v8df) __A,
+                (__v8df) __B,
+                (__v8df) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m512d __DEFAULT_FN_ATTRS
+_mm512_maskz_unpacklo_pd (__mmask8 __U, __m512d __A, __m512d __B)
+{
+  return (__m512d) __builtin_ia32_unpcklpd512_mask ((__v8df) __A,
+                (__v8df) __B,
+                (__v8df)
+                _mm512_setzero_pd (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m512 __DEFAULT_FN_ATTRS
+_mm512_mask_unpacklo_ps (__m512 __W, __mmask16 __U, __m512 __A, __m512 __B)
+{
+  return (__m512) __builtin_ia32_unpcklps512_mask ((__v16sf) __A,
+               (__v16sf) __B,
+               (__v16sf) __W,
+               (__mmask16) __U);
+}
+
+static __inline__ __m512 __DEFAULT_FN_ATTRS
+_mm512_maskz_unpacklo_ps (__mmask16 __U, __m512 __A, __m512 __B)
+{
+  return (__m512) __builtin_ia32_unpcklps512_mask ((__v16sf) __A,
+               (__v16sf) __B,
+               (__v16sf)
+               _mm512_setzero_ps (),
+               (__mmask16) __U);
+}
+
+#define _mm_cvt_roundsd_si32( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvtsd2si32 ((__v2df)( __A),( __R));\
+})
+
+#define _mm_cvt_roundsd_i32( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvtsd2si32 ((__v2df)( __A),( __R));\
+})
+
+#define _mm_cvt_roundsd_si64( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvtsd2si64 ((__v2df)( __A),( __R));\
+})
+
+#define _mm_cvt_roundsd_i64( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvtsd2si64 ((__v2df)( __A),( __R));\
+})
+
+#define _mm_cvt_roundsd_u32( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvtsd2usi32 ((__v2df)( __A),( __R));\
+})
+
+static __inline__ unsigned __DEFAULT_FN_ATTRS
+_mm_cvtsd_u32 (__m128d __A)
+{
+  return (unsigned) __builtin_ia32_vcvtsd2usi32 ((__v2df) __A,
+             _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_cvt_roundsd_u64( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvtsd2usi64 ((__v2df)( __A),( __R));\
+})
+
+static __inline__ unsigned long long __DEFAULT_FN_ATTRS
+_mm_cvtsd_u64 (__m128d __A)
+{
+  return (unsigned long long) __builtin_ia32_vcvtsd2usi64 ((__v2df)
+                 __A,
+                 _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_cvt_roundss_si32( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvtss2si32 ((__v4sf)( __A),( __R));\
+})
+
+#define _mm_cvt_roundss_i32( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvtss2si32 ((__v4sf)( __A),( __R));\
+})
+
+#define _mm_cvt_roundss_si64( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvtss2si64 ((__v4sf)( __A),( __R));\
+})
+
+#define _mm_cvt_roundss_i64( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvtss2si64 ((__v4sf)( __A),( __R));\
+})
+
+#define _mm_cvt_roundss_u32( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvtss2usi32 ((__v4sf)( __A),( __R));\
+})
+
+static __inline__ unsigned __DEFAULT_FN_ATTRS
+_mm_cvtss_u32 (__m128 __A)
+{
+  return (unsigned) __builtin_ia32_vcvtss2usi32 ((__v4sf) __A,
+             _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_cvt_roundss_u64( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvtss2usi64 ((__v4sf)( __A),( __R));\
+})
+
+static __inline__ unsigned long long __DEFAULT_FN_ATTRS
+_mm_cvtss_u64 (__m128 __A)
+{
+  return (unsigned long long) __builtin_ia32_vcvtss2usi64 ((__v4sf)
+                 __A,
+                 _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_cvtt_roundsd_i32( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvttsd2si32 ((__v2df)( __A),( __R));\
+})
+
+#define _mm_cvtt_roundsd_si32( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvttsd2si32 ((__v2df)( __A),( __R));\
+})
+
+static __inline__ int __DEFAULT_FN_ATTRS
+_mm_cvttsd_i32 (__m128d __A)
+{
+  return (int) __builtin_ia32_vcvttsd2si32 ((__v2df) __A,
+              _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_cvtt_roundsd_si64( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvttsd2si64 ((__v2df)( __A),( __R));\
+})
+
+#define _mm_cvtt_roundsd_i64( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvttsd2si64 ((__v2df)( __A),( __R));\
+})
+
+static __inline__ long long __DEFAULT_FN_ATTRS
+_mm_cvttsd_i64 (__m128d __A)
+{
+  return (long long) __builtin_ia32_vcvttsd2si64 ((__v2df) __A,
+              _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_cvtt_roundsd_u32( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvttsd2usi32 ((__v2df)( __A),( __R));\
+})
+
+static __inline__ unsigned __DEFAULT_FN_ATTRS
+_mm_cvttsd_u32 (__m128d __A)
+{
+  return (unsigned) __builtin_ia32_vcvttsd2usi32 ((__v2df) __A,
+              _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_cvtt_roundsd_u64( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvttsd2usi64 ((__v2df)( __A),( __R));\
+})
+
+static __inline__ unsigned long long __DEFAULT_FN_ATTRS
+_mm_cvttsd_u64 (__m128d __A)
+{
+  return (unsigned long long) __builtin_ia32_vcvttsd2usi64 ((__v2df)
+                  __A,
+                  _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_cvtt_roundss_i32( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvttss2si32 ((__v4sf)( __A),( __R));\
+})
+
+#define _mm_cvtt_roundss_si32( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvttss2si32 ((__v4sf)( __A),( __R));\
+})
+
+static __inline__ int __DEFAULT_FN_ATTRS
+_mm_cvttss_i32 (__m128 __A)
+{
+  return (int) __builtin_ia32_vcvttss2si32 ((__v4sf) __A,
+              _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_cvtt_roundss_i64( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvttss2si64 ((__v4sf)( __A),( __R));\
+})
+
+#define _mm_cvtt_roundss_si64( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvttss2si64 ((__v4sf)( __A),( __R));\
+})
+
+static __inline__ long long __DEFAULT_FN_ATTRS
+_mm_cvttss_i64 (__m128 __A)
+{
+  return (long long) __builtin_ia32_vcvttss2si64 ((__v4sf) __A,
+              _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_cvtt_roundss_u32( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvttss2usi32 ((__v4sf)( __A),( __R));\
+})
+
+static __inline__ unsigned __DEFAULT_FN_ATTRS
+_mm_cvttss_u32 (__m128 __A)
+{
+  return (unsigned) __builtin_ia32_vcvttss2usi32 ((__v4sf) __A,
+              _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_cvtt_roundss_u64( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvttss2usi64 ((__v4sf)( __A),( __R));\
+})
+
+static __inline__ unsigned long long __DEFAULT_FN_ATTRS
+_mm_cvttss_u64 (__m128 __A)
+{
+  return (unsigned long long) __builtin_ia32_vcvttss2usi64 ((__v4sf)
+                  __A,
+                  _MM_FROUND_CUR_DIRECTION);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_mask2_permutex2var_epi32 (__m512i __A, __m512i __I,
+         __mmask16 __U, __m512i __B)
+{
+  return (__m512i) __builtin_ia32_vpermi2vard512_mask ((__v16si) __A,
+                   (__v16si) __I
+                   /* idx */ ,
+                   (__v16si) __B,
+                   (__mmask16) __U);
+}
+
+static __inline__ __m512d __DEFAULT_FN_ATTRS
+_mm512_mask2_permutex2var_pd (__m512d __A, __m512i __I, __mmask8 __U,
+            __m512d __B)
+{
+  return (__m512d) __builtin_ia32_vpermi2varpd512_mask ((__v8df) __A,
+              (__v8di) __I
+              /* idx */ ,
+              (__v8df) __B,
+              (__mmask8) __U);
+}
+
+static __inline__ __m512 __DEFAULT_FN_ATTRS
+_mm512_mask2_permutex2var_ps (__m512 __A, __m512i __I, __mmask16 __U,
+            __m512 __B)
+{
+  return (__m512) __builtin_ia32_vpermi2varps512_mask ((__v16sf) __A,
+                   (__v16si) __I
+                   /* idx */ ,
+                   (__v16sf) __B,
+                   (__mmask16) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_mask2_permutex2var_epi64 (__m512i __A, __m512i __I,
+         __mmask8 __U, __m512i __B)
+{
+  return (__m512i) __builtin_ia32_vpermi2varq512_mask ((__v8di) __A,
+                   (__v8di) __I
+                   /* idx */ ,
+                   (__v8di) __B,
+                   (__mmask8) __U);
+}
+
+#define _mm512_permute_pd( __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilpd512_mask ((__v8df)( __X),( __C),\
+                 (__v8df)\
+                 _mm512_undefined_pd (),\
+                 (__mmask8) -1);\
+})
+
+#define _mm512_mask_permute_pd( __W, __U, __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilpd512_mask ((__v8df)( __X),( __C),\
+                 (__v8df)( __W),\
+                 (__mmask8)( __U));\
+})
+
+#define _mm512_maskz_permute_pd( __U, __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilpd512_mask ((__v8df)( __X),( __C),\
+                 (__v8df)\
+                 _mm512_setzero_pd (),\
+                 (__mmask8)( __U));\
+})
+
+#define _mm512_permute_ps( __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilps512_mask ((__v16sf)( __X),( __C),\
+                (__v16sf)\
+                _mm512_undefined_ps (),\
+                (__mmask16) -1);\
+})
+
+#define _mm512_mask_permute_ps( __W, __U, __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilps512_mask ((__v16sf)( __X),( __C),\
+                (__v16sf)( __W),\
+                (__mmask16)( __U));\
+})
+
+#define _mm512_maskz_permute_ps( __U, __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilps512_mask ((__v16sf)( __X),( __C),\
+                (__v16sf)\
+                _mm512_setzero_ps (),\
+                (__mmask16)( __U));\
+})
+
+static __inline__ __m512d __DEFAULT_FN_ATTRS
+_mm512_permutevar_pd (__m512d __A, __m512i __C)
+{
+  return (__m512d) __builtin_ia32_vpermilvarpd512_mask ((__v8df) __A,
+              (__v8di) __C,
+              (__v8df)
+              _mm512_undefined_pd (),
+              (__mmask8) -1);
+}
+
+static __inline__ __m512d __DEFAULT_FN_ATTRS
+_mm512_mask_permutevar_pd (__m512d __W, __mmask8 __U, __m512d __A, __m512i __C)
+{
+  return (__m512d) __builtin_ia32_vpermilvarpd512_mask ((__v8df) __A,
+              (__v8di) __C,
+              (__v8df) __W,
+              (__mmask8) __U);
+}
+
+static __inline__ __m512d __DEFAULT_FN_ATTRS
+_mm512_maskz_permutevar_pd (__mmask8 __U, __m512d __A, __m512i __C)
+{
+  return (__m512d) __builtin_ia32_vpermilvarpd512_mask ((__v8df) __A,
+              (__v8di) __C,
+              (__v8df)
+              _mm512_setzero_pd (),
+              (__mmask8) __U);
+}
+
+static __inline__ __m512 __DEFAULT_FN_ATTRS
+_mm512_permutevar_ps (__m512 __A, __m512i __C)
+{
+  return (__m512) __builtin_ia32_vpermilvarps512_mask ((__v16sf) __A,
+                   (__v16si) __C,
+                   (__v16sf)
+                   _mm512_undefined_ps (),
+                   (__mmask16) -1);
+}
+
+static __inline__ __m512 __DEFAULT_FN_ATTRS
+_mm512_mask_permutevar_ps (__m512 __W, __mmask16 __U, __m512 __A, __m512i __C)
+{
+  return (__m512) __builtin_ia32_vpermilvarps512_mask ((__v16sf) __A,
+                   (__v16si) __C,
+                   (__v16sf) __W,
+                   (__mmask16) __U);
+}
+
+static __inline__ __m512 __DEFAULT_FN_ATTRS
+_mm512_maskz_permutevar_ps (__mmask16 __U, __m512 __A, __m512i __C)
+{
+  return (__m512) __builtin_ia32_vpermilvarps512_mask ((__v16sf) __A,
+                   (__v16si) __C,
+                   (__v16sf)
+                   _mm512_setzero_ps (),
+                   (__mmask16) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_maskz_permutex2var_epi32 (__mmask16 __U, __m512i __A,
+         __m512i __I, __m512i __B)
+{
+  return (__m512i) __builtin_ia32_vpermt2vard512_maskz ((__v16si) __I
+              /* idx */ ,
+              (__v16si) __A,
+              (__v16si) __B,
+              (__mmask16) __U);
+}
+
+static __inline__ __m512d __DEFAULT_FN_ATTRS
+_mm512_maskz_permutex2var_pd (__mmask8 __U, __m512d __A, __m512i __I,
+            __m512d __B)
+{
+  return (__m512d) __builtin_ia32_vpermt2varpd512_maskz ((__v8di) __I
+               /* idx */ ,
+               (__v8df) __A,
+               (__v8df) __B,
+               (__mmask8) __U);
+}
+
+static __inline__ __m512 __DEFAULT_FN_ATTRS
+_mm512_maskz_permutex2var_ps (__mmask16 __U, __m512 __A, __m512i __I,
+            __m512 __B)
+{
+  return (__m512) __builtin_ia32_vpermt2varps512_maskz ((__v16si) __I
+              /* idx */ ,
+              (__v16sf) __A,
+              (__v16sf) __B,
+              (__mmask16) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_maskz_permutex2var_epi64 (__mmask8 __U, __m512i __A,
+         __m512i __I, __m512i __B)
+{
+  return (__m512i) __builtin_ia32_vpermt2varq512_maskz ((__v8di) __I
+              /* idx */ ,
+              (__v8di) __A,
+              (__v8di) __B,
+              (__mmask8) __U);
+}
 
 #undef __DEFAULT_FN_ATTRS
 
 #endif // __AVX512FINTRIN_H

lib/Headers/avx512vlbwintrin.h

 #endif
 
 #ifndef __AVX512VLBWINTRIN_H
 #define __AVX512VLBWINTRIN_H
 
 /* Define the default attributes for the functions in this file. */
 #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512vl,avx512bw")))
 
-static  __inline __m128i __DEFAULT_FN_ATTRS
-_mm_setzero_hi(void){
-    return (__m128i){ 0LL, 0LL };
-}
-
 /* Integer compare */
 
 static __inline__ __mmask16 __DEFAULT_FN_ATTRS
 _mm_cmpeq_epi8_mask(__m128i __a, __m128i __b) {
   return (__mmask16)__builtin_ia32_pcmpeqb128_mask((__v16qi)__a, (__v16qi)__b,
                                                    (__mmask16)-1);
 }
 

 static __inline__ __m256i __DEFAULT_FN_ATTRS
 _mm256_maskz_set1_epi8 (__mmask32 __M, char __A)
 {
   return (__m256i) __builtin_ia32_pbroadcastb256_gpr_mask (__A,
                  (__v32qi)
                  _mm256_setzero_si256 (),
                  __M);
 }
+
+static __inline__ __mmask16 __DEFAULT_FN_ATTRS
+_mm_test_epi8_mask (__m128i __A, __m128i __B)
+{
+  return (__mmask16) __builtin_ia32_ptestmb128 ((__v16qi) __A,
+            (__v16qi) __B,
+            (__mmask16) -1);
+}
+
+static __inline__ __mmask16 __DEFAULT_FN_ATTRS
+_mm_mask_test_epi8_mask (__mmask16 __U, __m128i __A, __m128i __B)
+{
+  return (__mmask16) __builtin_ia32_ptestmb128 ((__v16qi) __A,
+            (__v16qi) __B, __U);
+}
+
+static __inline__ __mmask32 __DEFAULT_FN_ATTRS
+_mm256_test_epi8_mask (__m256i __A, __m256i __B)
+{
+  return (__mmask32) __builtin_ia32_ptestmb256 ((__v32qi) __A,
+            (__v32qi) __B,
+            (__mmask32) -1);
+}
+
+static __inline__ __mmask32 __DEFAULT_FN_ATTRS
+_mm256_mask_test_epi8_mask (__mmask32 __U, __m256i __A, __m256i __B)
+{
+  return (__mmask32) __builtin_ia32_ptestmb256 ((__v32qi) __A,
+            (__v32qi) __B, __U);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm_test_epi16_mask (__m128i __A, __m128i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestmw128 ((__v8hi) __A,
+                 (__v8hi) __B,
+                 (__mmask8) -1);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm_mask_test_epi16_mask (__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestmw128 ((__v8hi) __A,
+                 (__v8hi) __B, __U);
+}
+
+static __inline__ __mmask16 __DEFAULT_FN_ATTRS
+_mm256_test_epi16_mask (__m256i __A, __m256i __B)
+{
+  return (__mmask16) __builtin_ia32_ptestmw256 ((__v16hi) __A,
+            (__v16hi) __B,
+            (__mmask16) -1);
+}
+
+static __inline__ __mmask16 __DEFAULT_FN_ATTRS
+_mm256_mask_test_epi16_mask (__mmask16 __U, __m256i __A, __m256i __B)
+{
+  return (__mmask16) __builtin_ia32_ptestmw256 ((__v16hi) __A,
+            (__v16hi) __B, __U);
+}
+
+static __inline__ __mmask16 __DEFAULT_FN_ATTRS
+_mm_testn_epi8_mask (__m128i __A, __m128i __B)
+{
+  return (__mmask16) __builtin_ia32_ptestnmb128 ((__v16qi) __A,
+             (__v16qi) __B,
+             (__mmask16) -1);
+}
+
+static __inline__ __mmask16 __DEFAULT_FN_ATTRS
+_mm_mask_testn_epi8_mask (__mmask16 __U, __m128i __A, __m128i __B)
+{
+  return (__mmask16) __builtin_ia32_ptestnmb128 ((__v16qi) __A,
+             (__v16qi) __B, __U);
+}
+
+static __inline__ __mmask32 __DEFAULT_FN_ATTRS
+_mm256_testn_epi8_mask (__m256i __A, __m256i __B)
+{
+  return (__mmask32) __builtin_ia32_ptestnmb256 ((__v32qi) __A,
+             (__v32qi) __B,
+             (__mmask32) -1);
+}
+
+static __inline__ __mmask32 __DEFAULT_FN_ATTRS
+_mm256_mask_testn_epi8_mask (__mmask32 __U, __m256i __A, __m256i __B)
+{
+  return (__mmask32) __builtin_ia32_ptestnmb256 ((__v32qi) __A,
+             (__v32qi) __B, __U);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm_testn_epi16_mask (__m128i __A, __m128i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestnmw128 ((__v8hi) __A,
+            (__v8hi) __B,
+            (__mmask8) -1);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm_mask_testn_epi16_mask (__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestnmw128 ((__v8hi) __A,
+            (__v8hi) __B, __U);
+}
+
+static __inline__ __mmask16 __DEFAULT_FN_ATTRS
+_mm256_testn_epi16_mask (__m256i __A, __m256i __B)
+{
+  return (__mmask16) __builtin_ia32_ptestnmw256 ((__v16hi) __A,
+             (__v16hi) __B,
+             (__mmask16) -1);
+}
+
+static __inline__ __mmask16 __DEFAULT_FN_ATTRS
+_mm256_mask_testn_epi16_mask (__mmask16 __U, __m256i __A, __m256i __B)
+{
+  return (__mmask16) __builtin_ia32_ptestnmw256 ((__v16hi) __A,
+             (__v16hi) __B, __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_storeu_epi16 (void *__P, __mmask8 __U, __m128i __A)
+{
+  __builtin_ia32_storedquhi128_mask ((__v8hi *) __P,
+             (__v8hi) __A,
+             (__mmask8) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_storeu_epi16 (void *__P, __mmask16 __U, __m256i __A)
+{
+  __builtin_ia32_storedquhi256_mask ((__v16hi *) __P,
+             (__v16hi) __A,
+             (__mmask16) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_storeu_epi8 (void *__P, __mmask16 __U, __m128i __A)
+{
+  __builtin_ia32_storedquqi128_mask ((__v16qi *) __P,
+             (__v16qi) __A,
+             (__mmask16) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_storeu_epi8 (void *__P, __mmask32 __U, __m256i __A)
+{
+  __builtin_ia32_storedquqi256_mask ((__v32qi *) __P,
+             (__v32qi) __A,
+             (__mmask32) __U);
+}
+
+
 #undef __DEFAULT_FN_ATTRS
 
 #endif /* __AVX512VLBWINTRIN_H */

lib/Headers/avx512vlintrin.h

 #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512vl")))
 #define __DEFAULT_FN_ATTRS_BOTH __attribute__((__always_inline__, __nodebug__, __target__("avx512vl, avx512bw")))
 
 static  __inline __v2di __DEFAULT_FN_ATTRS
 _mm_setzero_di(void) {
   return (__v2di){ 0, 0};
 }
 
+static  __inline __m128i __DEFAULT_FN_ATTRS
+_mm_setzero_hi(void){
+    return (__m128i){ 0LL, 0LL };
+}
+
 /* Integer compare */
 
 static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
 _mm_cmpeq_epi32_mask(__m128i __a, __m128i __b) {
   return (__mmask8)__builtin_ia32_pcmpeqd128_mask((__v4si)__a, (__v4si)__b,
                                                   (__mmask8)-1);
 }
 

 static __inline__ __m256i __DEFAULT_FN_ATTRS
 _mm256_maskz_set1_epi64 (__mmask8 __M, long long __A)
 {
   return (__m256i) __builtin_ia32_pbroadcastq256_gpr_mask (__A,
                  (__v4di)
                  _mm256_setzero_si256 (),
                  __M);
 }
+
+#define _mm_ternarylogic_epi32( __A, __B, __C, imm) __extension__ ({ \
+__builtin_ia32_pternlogd128_mask ((__v4si)( __A),\
+                 (__v4si)( __B),\
+                 (__v4si)( __C),( imm),\
+                 (__mmask8) -1);\
+})
+
+#define _mm_mask_ternarylogic_epi32( __A, __U, __B, __C, imm) __extension__ ({ \
+__builtin_ia32_pternlogd128_mask ((__v4si)( __A),\
+                 (__v4si)( __B),\
+                 (__v4si)( __C),( imm),\
+                 (__mmask8)( __U));\
+})
+
+#define _mm_maskz_ternarylogic_epi32( __U, __A, __B, __C, imm) __extension__ ({ \
+__builtin_ia32_pternlogd128_maskz ((__v4si)( __A),\
+                  (__v4si)( __B),\
+                  (__v4si)( __C),\
+                 ( imm),\
+                  (__mmask8)( __U));\
+})
+
+#define _mm256_ternarylogic_epi32( __A, __B, __C, imm) __extension__ ({ \
+__builtin_ia32_pternlogd256_mask ((__v8si)( __A),\
+                 (__v8si)( __B),\
+                 (__v8si)( __C),( imm),\
+                 (__mmask8) -1);\
+})
+
+#define _mm256_mask_ternarylogic_epi32( __A, __U, __B, __C, imm) __extension__ ({ \
+__builtin_ia32_pternlogd256_mask ((__v8si)( __A),\
+                 (__v8si)( __B),\
+                 (__v8si)( __C),( imm),\
+                 (__mmask8)( __U));\
+})
+
+#define _mm256_maskz_ternarylogic_epi32( __U, __A, __B, __C, imm) __extension__ ({ \
+__builtin_ia32_pternlogd256_maskz ((__v8si)( __A),\
+                  (__v8si)( __B),\
+                  (__v8si)( __C),\
+                 ( imm),\
+                  (__mmask8)( __U));\
+})
+
+#define _mm_ternarylogic_epi64( __A, __B, __C, imm) __extension__ ({ \
+__builtin_ia32_pternlogq128_mask ((__v2di)( __A),\
+                 (__v2di)( __B),\
+                 (__v2di)( __C),( imm),\
+                 (__mmask8) -1);\
+})
+
+#define _mm_mask_ternarylogic_epi64( __A, __U, __B, __C, imm) __extension__ ({ \
+__builtin_ia32_pternlogq128_mask ((__v2di)( __A),\
+                 (__v2di)( __B),\
+                 (__v2di)( __C),( imm),\
+                 (__mmask8)( __U));\
+})
+
+#define _mm_maskz_ternarylogic_epi64( __U, __A, __B, __C, imm) __extension__ ({ \
+__builtin_ia32_pternlogq128_maskz ((__v2di)( __A),\
+                  (__v2di)( __B),\
+                  (__v2di)( __C),\
+                 ( imm),\
+                  (__mmask8)( __U));\
+})
+
+#define _mm256_ternarylogic_epi64( __A, __B, __C, imm) __extension__ ({ \
+__builtin_ia32_pternlogq256_mask ((__v4di)( __A),\
+                 (__v4di)( __B),\
+                 (__v4di)( __C),( imm),\
+                 (__mmask8) -1);\
+})
+
+
+#define _mm256_mask_ternarylogic_epi64( __A, __U, __B, __C, imm) __extension__ ({ \
+__builtin_ia32_pternlogq256_mask ((__v4di)( __A),\
+                 (__v4di)( __B),\
+                 (__v4di)( __C),( imm),\
+                 (__mmask8)( __U));\
+})
+
+
+#define _mm256_maskz_ternarylogic_epi64( __U, __A, __B, __C, imm) __extension__ ({ \
+__builtin_ia32_pternlogq256_maskz ((__v4di)( __A),\
+                  (__v4di)( __B),\
+                  (__v4di)( __C),\
+                 ( imm),\
+                  (__mmask8)( __U));\
+})
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm_test_epi32_mask (__m128i __A, __m128i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestmd128 ((__v4si) __A,
+                 (__v4si) __B,
+                 (__mmask8) -1);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm_mask_test_epi32_mask (__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestmd128 ((__v4si) __A,
+                 (__v4si) __B, __U);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm256_test_epi32_mask (__m256i __A, __m256i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestmd256 ((__v8si) __A,
+                 (__v8si) __B,
+                 (__mmask8) -1);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm256_mask_test_epi32_mask (__mmask8 __U, __m256i __A, __m256i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestmd256 ((__v8si) __A,
+                 (__v8si) __B, __U);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm_test_epi64_mask (__m128i __A, __m128i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestmq128 ((__v2di) __A,
+                 (__v2di) __B,
+                 (__mmask8) -1);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm_mask_test_epi64_mask (__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestmq128 ((__v2di) __A,
+                 (__v2di) __B, __U);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm256_test_epi64_mask (__m256i __A, __m256i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestmq256 ((__v4di) __A,
+                 (__v4di) __B,
+                 (__mmask8) -1);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm256_mask_test_epi64_mask (__mmask8 __U, __m256i __A, __m256i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestmq256 ((__v4di) __A,
+                 (__v4di) __B, __U);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm_testn_epi32_mask (__m128i __A, __m128i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestnmd128 ((__v4si) __A,
+            (__v4si) __B,
+            (__mmask8) -1);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm_mask_testn_epi32_mask (__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestnmd128 ((__v4si) __A,
+            (__v4si) __B, __U);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm256_testn_epi32_mask (__m256i __A, __m256i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestnmd256 ((__v8si) __A,
+            (__v8si) __B,
+            (__mmask8) -1);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm256_mask_testn_epi32_mask (__mmask8 __U, __m256i __A, __m256i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestnmd256 ((__v8si) __A,
+            (__v8si) __B, __U);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm_testn_epi64_mask (__m128i __A, __m128i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestnmq128 ((__v2di) __A,
+            (__v2di) __B,
+            (__mmask8) -1);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm_mask_testn_epi64_mask (__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestnmq128 ((__v2di) __A,
+            (__v2di) __B, __U);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm256_testn_epi64_mask (__m256i __A, __m256i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestnmq256 ((__v4di) __A,
+            (__v4di) __B,
+            (__mmask8) -1);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm256_mask_testn_epi64_mask (__mmask8 __U, __m256i __A, __m256i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestnmq256 ((__v4di) __A,
+            (__v4di) __B, __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_unpackhi_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
+       __m128i __B)
+{
+  return (__m128i) __builtin_ia32_punpckhdq128_mask ((__v4si) __A,
+                 (__v4si) __B,
+                 (__v4si) __W,
+                 (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_unpackhi_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__m128i) __builtin_ia32_punpckhdq128_mask ((__v4si) __A,
+                 (__v4si) __B,
+                 (__v4si)
+                 _mm_setzero_si128 (),
+                 (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_unpackhi_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
+          __m256i __B)
+{
+  return (__m256i) __builtin_ia32_punpckhdq256_mask ((__v8si) __A,
+                 (__v8si) __B,
+                 (__v8si) __W,
+                 (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_unpackhi_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
+{
+  return (__m256i) __builtin_ia32_punpckhdq256_mask ((__v8si) __A,
+                 (__v8si) __B,
+                 (__v8si)
+                 _mm256_setzero_si256 (),
+                 (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_unpackhi_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
+       __m128i __B)
+{
+  return (__m128i) __builtin_ia32_punpckhqdq128_mask ((__v2di) __A,
+                  (__v2di) __B,
+                  (__v2di) __W,
+                  (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_unpackhi_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__m128i) __builtin_ia32_punpckhqdq128_mask ((__v2di) __A,
+                  (__v2di) __B,
+                  (__v2di)
+                  _mm_setzero_di (),
+                  (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_unpackhi_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
+          __m256i __B)
+{
+  return (__m256i) __builtin_ia32_punpckhqdq256_mask ((__v4di) __A,
+                  (__v4di) __B,
+                  (__v4di) __W,
+                  (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_unpackhi_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
+{
+  return (__m256i) __builtin_ia32_punpckhqdq256_mask ((__v4di) __A,
+                  (__v4di) __B,
+                  (__v4di)
+                  _mm256_setzero_si256 (),
+                  (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_unpacklo_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
+       __m128i __B)
+{
+  return (__m128i) __builtin_ia32_punpckldq128_mask ((__v4si) __A,
+                 (__v4si) __B,
+                 (__v4si) __W,
+                 (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_unpacklo_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__m128i) __builtin_ia32_punpckldq128_mask ((__v4si) __A,
+                 (__v4si) __B,
+                 (__v4si)
+                 _mm_setzero_si128 (),
+                 (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_unpacklo_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
+          __m256i __B)
+{
+  return (__m256i) __builtin_ia32_punpckldq256_mask ((__v8si) __A,
+                 (__v8si) __B,
+                 (__v8si) __W,
+                 (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_unpacklo_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
+{
+  return (__m256i) __builtin_ia32_punpckldq256_mask ((__v8si) __A,
+                 (__v8si) __B,
+                 (__v8si)
+                 _mm256_setzero_si256 (),
+                 (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_unpacklo_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
+       __m128i __B)
+{
+  return (__m128i) __builtin_ia32_punpcklqdq128_mask ((__v2di) __A,
+                  (__v2di) __B,
+                  (__v2di) __W,
+                  (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_unpacklo_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__m128i) __builtin_ia32_punpcklqdq128_mask ((__v2di) __A,
+                  (__v2di) __B,
+                  (__v2di)
+                  _mm_setzero_di (),
+                  (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_unpacklo_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
+          __m256i __B)
+{
+  return (__m256i) __builtin_ia32_punpcklqdq256_mask ((__v4di) __A,
+                  (__v4di) __B,
+                  (__v4di) __W,
+                  (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_unpacklo_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
+{
+  return (__m256i) __builtin_ia32_punpcklqdq256_mask ((__v4di) __A,
+                  (__v4di) __B,
+                  (__v4di)
+                  _mm256_setzero_si256 (),
+                  (__mmask8) __U);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_rcp14_pd (__m128d __A)
+{
+  return (__m128d) __builtin_ia32_rcp14pd128_mask ((__v2df) __A,
+                (__v2df)
+                _mm_setzero_pd (),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_mask_rcp14_pd (__m128d __W, __mmask8 __U, __m128d __A)
+{
+  return (__m128d) __builtin_ia32_rcp14pd128_mask ((__v2df) __A,
+                (__v2df) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_rcp14_pd (__mmask8 __U, __m128d __A)
+{
+  return (__m128d) __builtin_ia32_rcp14pd128_mask ((__v2df) __A,
+                (__v2df)
+                _mm_setzero_pd (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_rcp14_pd (__m256d __A)
+{
+  return (__m256d) __builtin_ia32_rcp14pd256_mask ((__v4df) __A,
+                (__v4df)
+                _mm256_setzero_pd (),
+                (__mmask8) -1);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_mask_rcp14_pd (__m256d __W, __mmask8 __U, __m256d __A)
+{
+  return (__m256d) __builtin_ia32_rcp14pd256_mask ((__v4df) __A,
+                (__v4df) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_maskz_rcp14_pd (__mmask8 __U, __m256d __A)
+{
+  return (__m256d) __builtin_ia32_rcp14pd256_mask ((__v4df) __A,
+                (__v4df)
+                _mm256_setzero_pd (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_rcp14_ps (__m128 __A)
+{
+  return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A,
+               (__v4sf)
+               _mm_setzero_ps (),
+               (__mmask8) -1);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_rcp14_ps (__m128 __W, __mmask8 __U, __m128 __A)
+{
+  return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A,
+               (__v4sf) __W,
+               (__mmask8) __U);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_rcp14_ps (__mmask8 __U, __m128 __A)
+{
+  return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A,
+               (__v4sf)
+               _mm_setzero_ps (),
+               (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_rcp14_ps (__m256 __A)
+{
+  return (__m256) __builtin_ia32_rcp14ps256_mask ((__v8sf) __A,
+               (__v8sf)
+               _mm256_setzero_ps (),
+               (__mmask8) -1);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_mask_rcp14_ps (__m256 __W, __mmask8 __U, __m256 __A)
+{
+  return (__m256) __builtin_ia32_rcp14ps256_mask ((__v8sf) __A,
+               (__v8sf) __W,
+               (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_maskz_rcp14_ps (__mmask8 __U, __m256 __A)
+{
+  return (__m256) __builtin_ia32_rcp14ps256_mask ((__v8sf) __A,
+               (__v8sf)
+               _mm256_setzero_ps (),
+               (__mmask8) __U);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_rsqrt14_pd (__m128d __A)
+{
+  return (__m128d) __builtin_ia32_rsqrt14pd128_mask ((__v2df) __A,
+                 (__v2df)
+                 _mm_setzero_pd (),
+                 (__mmask8) -1);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_mask_rsqrt14_pd (__m128d __W, __mmask8 __U, __m128d __A)
+{
+  return (__m128d) __builtin_ia32_rsqrt14pd128_mask ((__v2df) __A,
+                 (__v2df) __W,
+                 (__mmask8) __U);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_rsqrt14_pd (__mmask8 __U, __m128d __A)
+{
+  return (__m128d) __builtin_ia32_rsqrt14pd128_mask ((__v2df) __A,
+                 (__v2df)
+                 _mm_setzero_pd (),
+                 (__mmask8) __U);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_rsqrt14_pd (__m256d __A)
+{
+  return (__m256d) __builtin_ia32_rsqrt14pd256_mask ((__v4df) __A,
+                 (__v4df)
+                 _mm256_setzero_pd (),
+                 (__mmask8) -1);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_mask_rsqrt14_pd (__m256d __W, __mmask8 __U, __m256d __A)
+{
+  return (__m256d) __builtin_ia32_rsqrt14pd256_mask ((__v4df) __A,
+                 (__v4df) __W,
+                 (__mmask8) __U);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_maskz_rsqrt14_pd (__mmask8 __U, __m256d __A)
+{
+  return (__m256d) __builtin_ia32_rsqrt14pd256_mask ((__v4df) __A,
+                 (__v4df)
+                 _mm256_setzero_pd (),
+                 (__mmask8) __U);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_rsqrt14_ps (__m128 __A)
+{
+  return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A,
+                (__v4sf)
+                _mm_setzero_ps (),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_rsqrt14_ps (__m128 __W, __mmask8 __U, __m128 __A)
+{
+  return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A,
+                (__v4sf) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_rsqrt14_ps (__mmask8 __U, __m128 __A)
+{
+  return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A,
+                (__v4sf)
+                _mm_setzero_ps (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_rsqrt14_ps (__m256 __A)
+{
+  return (__m256) __builtin_ia32_rsqrt14ps256_mask ((__v8sf) __A,
+                (__v8sf)
+                _mm256_setzero_ps (),
+                (__mmask8) -1);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_mask_rsqrt14_ps (__m256 __W, __mmask8 __U, __m256 __A)
+{
+  return (__m256) __builtin_ia32_rsqrt14ps256_mask ((__v8sf) __A,
+                (__v8sf) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_maskz_rsqrt14_ps (__mmask8 __U, __m256 __A)
+{
+  return (__m256) __builtin_ia32_rsqrt14ps256_mask ((__v8sf) __A,
+                (__v8sf)
+                _mm256_setzero_ps (),
+                (__mmask8) __U);
+}
+
+#define _mm256_shuffle_f32x4( __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_f32x4_256_mask ((__v8sf)( __A),\
+                  (__v8sf)( __B),\
+                 ( __imm),\
+                  (__v8sf)\
+                  _mm256_setzero_ps (),\
+                  (__mmask8) -1);\
+})
+
+#define _mm256_mask_shuffle_f32x4( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_f32x4_256_mask ((__v8sf)( __A),\
+                  (__v8sf)( __B),\
+                 ( __imm),\
+                  (__v8sf)( __W),\
+                  (__mmask8)( __U));\
+})
+
+#define _mm256_maskz_shuffle_f32x4( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_f32x4_256_mask ((__v8sf)( __A),\
+                  (__v8sf)( __B),\
+                 ( __imm),\
+                  (__v8sf)\
+                  _mm256_setzero_ps (),\
+                  (__mmask8)( __U));\
+})
+
+#define _mm256_shuffle_f64x2( __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_f64x2_256_mask ((__v4df)( __A),\
+                   (__v4df)( __B),\
+                  ( __imm),\
+                   (__v4df)\
+                   _mm256_setzero_pd (),\
+                   (__mmask8) -1);\
+})
+
+#define _mm256_mask_shuffle_f64x2( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_f64x2_256_mask ((__v4df)( __A),\
+                   (__v4df)( __B),\
+                  ( __imm),\
+                   (__v4df)( __W),\
+                   (__mmask8)( __U));\
+})
+
+#define _mm256_maskz_shuffle_f64x2( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_f64x2_256_mask ((__v4df)( __A),\
+                   (__v4df)( __B),\
+                  ( __imm),\
+                   (__v4df)\
+                   _mm256_setzero_pd (),\
+                   (__mmask8)( __U));\
+})
+
+#define _mm256_shuffle_i32x4( __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_i32x4_256_mask ((__v8si)( __A),\
+                   (__v8si)( __B),\
+                  ( __imm),\
+                   (__v8si)\
+                   _mm256_setzero_si256 (),\
+                   (__mmask8) -1);\
+})
+
+#define _mm256_mask_shuffle_i32x4( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_i32x4_256_mask ((__v8si)( __A),\
+                   (__v8si)( __B),\
+                  ( __imm),\
+                   (__v8si)( __W),\
+                   (__mmask8)( __U));\
+})
+
+#define _mm256_maskz_shuffle_i32x4( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_i32x4_256_mask ((__v8si)( __A),\
+                   (__v8si)( __B),\
+                  ( __imm),\
+                   (__v8si)\
+                   _mm256_setzero_si256 (),\
+                   (__mmask8)( __U));\
+})
+
+#define _mm256_shuffle_i64x2( __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_i64x2_256_mask ((__v4di)( __A),\
+                   (__v4di)( __B),\
+                  ( __imm),\
+                   (__v4di)\
+                   _mm256_setzero_si256 (),\
+                   (__mmask8) -1);\
+})
+
+#define _mm256_mask_shuffle_i64x2( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_i64x2_256_mask ((__v4di)( __A),\
+                   (__v4di)( __B),\
+                  ( __imm),\
+                   (__v4di)( __W),\
+                   (__mmask8)( __U));\
+})
+
+#define _mm256_maskz_shuffle_i64x2( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_i64x2_256_mask ((__v4di)( __A),\
+                   (__v4di)( __B),\
+                  ( __imm),\
+                   (__v4di)\
+                   _mm256_setzero_si256 (),\
+                   (__mmask8)( __U));\
+})
+
+#define _mm_mask_shuffle_pd( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shufpd128_mask ((__v2df)( __A),\
+              (__v2df)( __B),( __imm),\
+              (__v2df)( __W),\
+              (__mmask8)( __U));\
+})
+
+#define _mm_maskz_shuffle_pd( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shufpd128_mask ((__v2df)( __A),\
+              (__v2df)( __B),( __imm),\
+              (__v2df)\
+              _mm_setzero_pd (),\
+              (__mmask8)( __U));\
+})
+
+#define _mm256_mask_shuffle_pd( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shufpd256_mask ((__v4df)( __A),\
+              (__v4df)( __B),( __imm),\
+              (__v4df)( __W),\
+              (__mmask8)( __U));\
+})
+
+#define _mm256_maskz_shuffle_pd( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shufpd256_mask ((__v4df)( __A),\
+              (__v4df)( __B),( __imm),\
+              (__v4df)\
+              _mm256_setzero_pd (),\
+              (__mmask8)( __U));\
+})
+
+#define _mm_mask_shuffle_ps( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shufps128_mask ((__v4sf)( __A),\
+             (__v4sf)( __B),( __imm),\
+             (__v4sf)( __W),\
+             (__mmask8)( __U));\
+})
+
+#define _mm_maskz_shuffle_ps( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shufps128_mask ((__v4sf)( __A),\
+             (__v4sf)( __B),( __imm),\
+             (__v4sf)\
+             _mm_setzero_ps (),\
+             (__mmask8)( __U));\
+})
+
+#define _mm256_mask_shuffle_ps( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shufps256_mask ((__v8sf)( __A),\
+             (__v8sf)( __B),( __imm),\
+             (__v8sf)( __W),\
+             (__mmask8)( __U));\
+})
+
+#define _mm256_maskz_shuffle_ps( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shufps256_mask ((__v8sf)( __A),\
+             (__v8sf)( __B),( __imm),\
+             (__v8sf)\
+             _mm256_setzero_ps (),\
+             (__mmask8)( __U));\
+})
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_store_pd (void *__P, __mmask8 __U, __m128d __A)
+{
+  __builtin_ia32_storeapd128_mask ((__v2df *) __P,
+           (__v2df) __A,
+           (__mmask8) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_store_pd (void *__P, __mmask8 __U, __m256d __A)
+{
+  __builtin_ia32_storeapd256_mask ((__v4df *) __P,
+           (__v4df) __A,
+           (__mmask8) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_store_ps (void *__P, __mmask8 __U, __m128 __A)
+{
+  __builtin_ia32_storeaps128_mask ((__v4sf *) __P,
+           (__v4sf) __A,
+           (__mmask8) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_store_ps (void *__P, __mmask8 __U, __m256 __A)
+{
+  __builtin_ia32_storeaps256_mask ((__v8sf *) __P,
+           (__v8sf) __A,
+           (__mmask8) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_storeu_epi64 (void *__P, __mmask8 __U, __m128i __A)
+{
+  __builtin_ia32_storedqudi128_mask ((__v2di *) __P,
+             (__v2di) __A,
+             (__mmask8) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_storeu_epi64 (void *__P, __mmask8 __U, __m256i __A)
+{
+  __builtin_ia32_storedqudi256_mask ((__v4di *) __P,
+             (__v4di) __A,
+             (__mmask8) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_storeu_epi32 (void *__P, __mmask8 __U, __m128i __A)
+{
+  __builtin_ia32_storedqusi128_mask ((__v4si *) __P,
+             (__v4si) __A,
+             (__mmask8) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_storeu_epi32 (void *__P, __mmask8 __U, __m256i __A)
+{
+  __builtin_ia32_storedqusi256_mask ((__v8si *) __P,
+             (__v8si) __A,
+             (__mmask8) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_storeu_pd (void *__P, __mmask8 __U, __m128d __A)
+{
+  __builtin_ia32_storeupd128_mask ((__v2df *) __P,
+           (__v2df) __A,
+           (__mmask8) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_storeu_pd (void *__P, __mmask8 __U, __m256d __A)
+{
+  __builtin_ia32_storeupd256_mask ((__v4df *) __P,
+           (__v4df) __A,
+           (__mmask8) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_storeu_ps (void *__P, __mmask8 __U, __m128 __A)
+{
+  __builtin_ia32_storeups128_mask ((__v4sf *) __P,
+           (__v4sf) __A,
+           (__mmask8) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_storeu_ps (void *__P, __mmask8 __U, __m256 __A)
+{
+  __builtin_ia32_storeups256_mask ((__v8sf *) __P,
+           (__v8sf) __A,
+           (__mmask8) __U);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_mask_unpackhi_pd (__m128d __W, __mmask8 __U, __m128d __A,
+          __m128d __B)
+{
+  return (__m128d) __builtin_ia32_unpckhpd128_mask ((__v2df) __A,
+                (__v2df) __B,
+                (__v2df) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_unpackhi_pd (__mmask8 __U, __m128d __A, __m128d __B)
+{
+  return (__m128d) __builtin_ia32_unpckhpd128_mask ((__v2df) __A,
+                (__v2df) __B,
+                (__v2df)
+                _mm_setzero_pd (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_mask_unpackhi_pd (__m256d __W, __mmask8 __U, __m256d __A,
+       __m256d __B)
+{
+  return (__m256d) __builtin_ia32_unpckhpd256_mask ((__v4df) __A,
+                (__v4df) __B,
+                (__v4df) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_maskz_unpackhi_pd (__mmask8 __U, __m256d __A, __m256d __B)
+{
+  return (__m256d) __builtin_ia32_unpckhpd256_mask ((__v4df) __A,
+                (__v4df) __B,
+                (__v4df)
+                _mm256_setzero_pd (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_unpackhi_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B)
+{
+  return (__m128) __builtin_ia32_unpckhps128_mask ((__v4sf) __A,
+               (__v4sf) __B,
+               (__v4sf) __W,
+               (__mmask8) __U);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_unpackhi_ps (__mmask8 __U, __m128 __A, __m128 __B)
+{
+  return (__m128) __builtin_ia32_unpckhps128_mask ((__v4sf) __A,
+               (__v4sf) __B,
+               (__v4sf)
+               _mm_setzero_ps (),
+               (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_mask_unpackhi_ps (__m256 __W, __mmask8 __U, __m256 __A,
+       __m256 __B)
+{
+  return (__m256) __builtin_ia32_unpckhps256_mask ((__v8sf) __A,
+               (__v8sf) __B,
+               (__v8sf) __W,
+               (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_maskz_unpackhi_ps (__mmask8 __U, __m256 __A, __m256 __B)
+{
+  return (__m256) __builtin_ia32_unpckhps256_mask ((__v8sf) __A,
+               (__v8sf) __B,
+               (__v8sf)
+               _mm256_setzero_ps (),
+               (__mmask8) __U);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_mask_unpacklo_pd (__m128d __W, __mmask8 __U, __m128d __A,
+          __m128d __B)
+{
+  return (__m128d) __builtin_ia32_unpcklpd128_mask ((__v2df) __A,
+                (__v2df) __B,
+                (__v2df) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_unpacklo_pd (__mmask8 __U, __m128d __A, __m128d __B)
+{
+  return (__m128d) __builtin_ia32_unpcklpd128_mask ((__v2df) __A,
+                (__v2df) __B,
+                (__v2df)
+                _mm_setzero_pd (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_mask_unpacklo_pd (__m256d __W, __mmask8 __U, __m256d __A,
+       __m256d __B)
+{
+  return (__m256d) __builtin_ia32_unpcklpd256_mask ((__v4df) __A,
+                (__v4df) __B,
+                (__v4df) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_maskz_unpacklo_pd (__mmask8 __U, __m256d __A, __m256d __B)
+{
+  return (__m256d) __builtin_ia32_unpcklpd256_mask ((__v4df) __A,
+                (__v4df) __B,
+                (__v4df)
+                _mm256_setzero_pd (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_unpacklo_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B)
+{
+  return (__m128) __builtin_ia32_unpcklps128_mask ((__v4sf) __A,
+               (__v4sf) __B,
+               (__v4sf) __W,
+               (__mmask8) __U);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_unpacklo_ps (__mmask8 __U, __m128 __A, __m128 __B)
+{
+  return (__m128) __builtin_ia32_unpcklps128_mask ((__v4sf) __A,
+               (__v4sf) __B,
+               (__v4sf)
+               _mm_setzero_ps (),
+               (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_mask_unpacklo_ps (__m256 __W, __mmask8 __U, __m256 __A,
+       __m256 __B)
+{
+  return (__m256) __builtin_ia32_unpcklps256_mask ((__v8sf) __A,
+               (__v8sf) __B,
+               (__v8sf) __W,
+               (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_maskz_unpacklo_ps (__mmask8 __U, __m256 __A, __m256 __B)
+{
+  return (__m256) __builtin_ia32_unpcklps256_mask ((__v8sf) __A,
+               (__v8sf) __B,
+               (__v8sf)
+               _mm256_setzero_ps (),
+               (__mmask8) __U);
+}
+
+#define _mm_mask_permute_pd( __W, __U, __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilpd_mask ((__v2df)( __X),( __C),\
+              (__v2df)( __W),\
+              (__mmask8)( __U));\
+})
+
+#define _mm_maskz_permute_pd( __U, __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilpd_mask ((__v2df)( __X),( __C),\
+              (__v2df)\
+              _mm_setzero_pd (),\
+              (__mmask8)( __U));\
+})
+
+#define _mm256_mask_permute_pd( __W, __U, __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilpd256_mask ((__v4df)( __X),( __C),\
+                 (__v4df)( __W),\
+                 (__mmask8)( __U));\
+})
+
+#define _mm256_maskz_permute_pd( __U, __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilpd256_mask ((__v4df)( __X),( __C),\
+                 (__v4df)\
+                 _mm256_setzero_pd (),\
+                 (__mmask8)( __U));\
+})
+
+#define _mm_mask_permute_ps( __W, __U, __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilps_mask ((__v4sf)( __X),( __C),\
+             (__v4sf)( __W),\
+             (__mmask8)( __U));\
+})
+
+#define _mm_maskz_permute_ps( __U, __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilps_mask ((__v4sf)( __X),( __C),\
+             (__v4sf)\
+             _mm_setzero_ps (),\
+             (__mmask8)( __U));\
+})
+
+#define _mm256_mask_permute_ps( __W, __U, __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilps256_mask ((__v8sf)( __X),( __C),\
+                (__v8sf)( __W),\
+                (__mmask8)( __U));\
+})
+
+#define _mm256_maskz_permute_ps( __U, __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilps256_mask ((__v8sf)( __X),( __C),\
+                (__v8sf)\
+                _mm256_setzero_ps (),\
+                (__mmask8)( __U));\
+})
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_mask_permutevar_pd (__m128d __W, __mmask8 __U, __m128d __A,
+      __m128i __C)
+{
+  return (__m128d) __builtin_ia32_vpermilvarpd_mask ((__v2df) __A,
+                 (__v2di) __C,
+                 (__v2df) __W,
+                 (__mmask8) __U);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_permutevar_pd (__mmask8 __U, __m128d __A, __m128i __C)
+{
+  return (__m128d) __builtin_ia32_vpermilvarpd_mask ((__v2df) __A,
+                 (__v2di) __C,
+                 (__v2df)
+                 _mm_setzero_pd (),
+                 (__mmask8) __U);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_mask_permutevar_pd (__m256d __W, __mmask8 __U, __m256d __A,
+         __m256i __C)
+{
+  return (__m256d) __builtin_ia32_vpermilvarpd256_mask ((__v4df) __A,
+              (__v4di) __C,
+              (__v4df) __W,
+              (__mmask8)
+              __U);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_maskz_permutevar_pd (__mmask8 __U, __m256d __A, __m256i __C)
+{
+  return (__m256d) __builtin_ia32_vpermilvarpd256_mask ((__v4df) __A,
+              (__v4di) __C,
+              (__v4df)
+              _mm256_setzero_pd (),
+              (__mmask8)
+              __U);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_permutevar_ps (__m128 __W, __mmask8 __U, __m128 __A,
+      __m128i __C)
+{
+  return (__m128) __builtin_ia32_vpermilvarps_mask ((__v4sf) __A,
+                (__v4si) __C,
+                (__v4sf) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_permutevar_ps (__mmask8 __U, __m128 __A, __m128i __C)
+{
+  return (__m128) __builtin_ia32_vpermilvarps_mask ((__v4sf) __A,
+                (__v4si) __C,
+                (__v4sf)
+                _mm_setzero_ps (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_mask_permutevar_ps (__m256 __W, __mmask8 __U, __m256 __A,
+         __m256i __C)
+{
+  return (__m256) __builtin_ia32_vpermilvarps256_mask ((__v8sf) __A,
+                   (__v8si) __C,
+                   (__v8sf) __W,
+                   (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_maskz_permutevar_ps (__mmask8 __U, __m256 __A, __m256i __C)
+{
+  return (__m256) __builtin_ia32_vpermilvarps256_mask ((__v8sf) __A,
+                   (__v8si) __C,
+                   (__v8sf)
+                   _mm256_setzero_ps (),
+                   (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_lzcnt_epi32 (__m128i __A)
+{
+  return (__m128i) __builtin_ia32_vplzcntd_128_mask ((__v4si) __A,
+                 (__v4si)
+                 _mm_setzero_si128 (),
+                 (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_lzcnt_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_vplzcntd_128_mask ((__v4si) __A,
+                 (__v4si) __W,
+                 (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_lzcnt_epi32 (__mmask8 __U, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_vplzcntd_128_mask ((__v4si) __A,
+                 (__v4si)
+                 _mm_setzero_si128 (),
+                 (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_lzcnt_epi32 (__m256i __A)
+{
+  return (__m256i) __builtin_ia32_vplzcntd_256_mask ((__v8si) __A,
+                 (__v8si)
+                 _mm256_setzero_si256 (),
+                 (__mmask8) -1);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_lzcnt_epi32 (__m256i __W, __mmask8 __U, __m256i __A)
+{
+  return (__m256i) __builtin_ia32_vplzcntd_256_mask ((__v8si) __A,
+                 (__v8si) __W,
+                 (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_lzcnt_epi32 (__mmask8 __U, __m256i __A)
+{
+  return (__m256i) __builtin_ia32_vplzcntd_256_mask ((__v8si) __A,
+                 (__v8si)
+                 _mm256_setzero_si256 (),
+                 (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_lzcnt_epi64 (__m128i __A)
+{
+  return (__m128i) __builtin_ia32_vplzcntq_128_mask ((__v2di) __A,
+                 (__v2di)
+                 _mm_setzero_di (),
+                 (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_lzcnt_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_vplzcntq_128_mask ((__v2di) __A,
+                 (__v2di) __W,
+                 (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_lzcnt_epi64 (__mmask8 __U, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_vplzcntq_128_mask ((__v2di) __A,
+                 (__v2di)
+                 _mm_setzero_di (),
+                 (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_lzcnt_epi64 (__m256i __A)
+{
+  return (__m256i) __builtin_ia32_vplzcntq_256_mask ((__v4di) __A,
+                 (__v4di)
+                 _mm256_setzero_si256 (),
+                 (__mmask8) -1);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_lzcnt_epi64 (__m256i __W, __mmask8 __U, __m256i __A)
+{
+  return (__m256i) __builtin_ia32_vplzcntq_256_mask ((__v4di) __A,
+                 (__v4di) __W,
+                 (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_lzcnt_epi64 (__mmask8 __U, __m256i __A)
+{
+  return (__m256i) __builtin_ia32_vplzcntq_256_mask ((__v4di) __A,
+                 (__v4di)
+                 _mm256_setzero_si256 (),
+                 (__mmask8) __U);
+}
 #undef __DEFAULT_FN_ATTRS
 #undef __DEFAULT_FN_ATTRS_BOTH
 
 #endif /* __AVX512VLINTRIN_H */

test/CodeGen/avx512bw-builtins.c

__m512i test_mm512_maskz_min_epu16(__mmask32 __M, __m512i __A, __m512i __B) {
   // CHECK: @llvm.x86.avx512.mask.pminu.w.512
   return _mm512_maskz_min_epu16(__M,__A,__B); 
 }
 __m512i test_mm512_mask_min_epu16(__m512i __W, __mmask32 __M, __m512i __A, __m512i __B) {
   // CHECK-LABEL: @test_mm512_mask_min_epu16
   // CHECK: @llvm.x86.avx512.mask.pminu.w.512
   return _mm512_mask_min_epu16(__W,__M,__A,__B); 
 }
 __m512i test_mm512_shuffle_epi8(__m512i __A, __m512i __B) {
   // CHECK-LABEL: @test_mm512_shuffle_epi8
   // CHECK: @llvm.x86.avx512.mask.pshuf.b.512
   return _mm512_shuffle_epi8(__A,__B); 
 }
 __m512i test_mm512_mask_shuffle_epi8(__m512i __W, __mmask64 __U, __m512i __A, __m512i __B) {
   // CHECK-LABEL: @test_mm512_mask_shuffle_epi8
   // CHECK: @llvm.x86.avx512.mask.pshuf.b.512
   return _mm512_mask_shuffle_epi8(__W,__U,__A,__B); 
 }
 __m512i test_mm512_maskz_shuffle_epi8(__mmask64 __U, __m512i __A, __m512i __B) {
   // CHECK-LABEL: @test_mm512_maskz_shuffle_epi8
   // CHECK: @llvm.x86.avx512.mask.pshuf.b.512
   return _mm512_maskz_shuffle_epi8(__U,__A,__B); 

 }
 
 __m512i test_mm512_maskz_set1_epi8(__mmask64 __M, char __A) {
   // CHECK-LABEL: @test_mm512_maskz_set1_epi8
   // CHECK: @llvm.x86.avx512.mask.pbroadcast.b.gpr.512
   return _mm512_maskz_set1_epi8(__M, __A); 
 }
 
+__mmask64 test_mm512_test_epi8_mask(__m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_test_epi8_mask
+  // CHECK: @llvm.x86.avx512.ptestm.b.512
+  return _mm512_test_epi8_mask(__A, __B); 
+}
+
+__mmask64 test_mm512_mask_test_epi8_mask(__mmask64 __U, __m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_mask_test_epi8_mask
+  // CHECK: @llvm.x86.avx512.ptestm.b.512
+  return _mm512_mask_test_epi8_mask(__U, __A, __B); 
+}
+
+__mmask32 test_mm512_test_epi16_mask(__m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_test_epi16_mask
+  // CHECK: @llvm.x86.avx512.ptestm.w.
+  return _mm512_test_epi16_mask(__A, __B); 
+}
+
+__mmask32 test_mm512_mask_test_epi16_mask(__mmask32 __U, __m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_mask_test_epi16_mask
+  // CHECK: @llvm.x86.avx512.ptestm.w.
+  return _mm512_mask_test_epi16_mask(__U, __A, __B); 
+}
+
+__mmask64 test_mm512_testn_epi8_mask(__m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_testn_epi8_mask
+  // CHECK: @llvm.x86.avx512.ptestnm.b.
+  return _mm512_testn_epi8_mask(__A, __B); 
+}
+
+__mmask64 test_mm512_mask_testn_epi8_mask(__mmask64 __U, __m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_mask_testn_epi8_mask
+  // CHECK: @llvm.x86.avx512.ptestnm.b.
+  return _mm512_mask_testn_epi8_mask(__U, __A, __B); 
+}
+
+__mmask32 test_mm512_testn_epi16_mask(__m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_testn_epi16_mask
+  // CHECK: @llvm.x86.avx512.ptestnm.w.
+  return _mm512_testn_epi16_mask(__A, __B); 
+}
+
+__mmask32 test_mm512_mask_testn_epi16_mask(__mmask32 __U, __m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_mask_testn_epi16_mask
+  // CHECK: @llvm.x86.avx512.ptestnm.w.
+  return _mm512_mask_testn_epi16_mask(__U, __A, __B); 
+}
+
+void test_mm512_mask_storeu_epi16(void *__P, __mmask32 __U, __m512i __A) {
+  // CHECK-LABEL: @test_mm512_mask_storeu_epi16
+  // CHECK: @llvm.x86.avx512.mask.storeu.w.
+  return _mm512_mask_storeu_epi16(__P, __U, __A); 
+}
+
+void test_mm512_mask_storeu_epi8(void *__P, __mmask64 __U, __m512i __A) {
+  // CHECK-LABEL: @test_mm512_mask_storeu_epi8
+  // CHECK: @llvm.x86.avx512.mask.storeu.b.
+  return _mm512_mask_storeu_epi8(__P, __U, __A); 
+}
+
 

test/CodeGen/avx512f-builtins.c

__m512 test_mm512_unpacklo_ps(__m512 a, __m512 b)
   // CHECK-LABEL: @test_mm512_unpacklo_ps
   // CHECK: shufflevector <16 x float> {{.*}} <i32 0, i32 16, i32 1, i32 17, i32 4, i32 20, i32 5, i32 21, i32 8, i32 24, i32 9, i32 25, i32 12, i32 28, i32 13, i32 29>
   return _mm512_unpacklo_ps(a, b);
 }
 
 __mmask16 test_mm512_cmp_round_ps_mask(__m512 a, __m512 b) {
   // CHECK-LABEL: @test_mm512_cmp_round_ps_mask
   // CHECK: @llvm.x86.avx512.mask.cmp.ps.512
-  return _mm512_cmp_round_ps_mask(a, b, 0, _MM_FROUND_TO_NEAREST_INT);
+  return _mm512_cmp_round_ps_mask(a, b, 0, _MM_FROUND_CUR_DIRECTION);
 }
 
 __mmask16 test_mm512_mask_cmp_round_ps_mask(__mmask16 m, __m512 a, __m512 b) {
   // CHECK-LABEL: @test_mm512_mask_cmp_round_ps_mask
   // CHECK: @llvm.x86.avx512.mask.cmp.ps.512
-  return _mm512_mask_cmp_round_ps_mask(m, a, b, 0, _MM_FROUND_TO_NEAREST_INT);
+  return _mm512_mask_cmp_round_ps_mask(m, a, b, 0, _MM_FROUND_CUR_DIRECTION);
 }
 
 __mmask16 test_mm512_cmp_ps_mask(__m512 a, __m512 b) {
   // CHECK-LABEL: @test_mm512_cmp_ps_mask
   // CHECKn: @llvm.x86.avx512.mask.cmp.ps.512
   return _mm512_cmp_ps_mask(a, b, 0);
 }
 
 __mmask16 test_mm512_mask_cmp_ps_mask(__mmask16 m, __m512 a, __m512 b) {
   // CHECK-LABEL: @test_mm512_mask_cmp_ps_mask
   // CHECK: @llvm.x86.avx512.mask.cmp.ps.512
   return _mm512_mask_cmp_ps_mask(m, a, b, 0);
 }
 
 __mmask8 test_mm512_cmp_round_pd_mask(__m512d a, __m512d b) {
   // CHECK-LABEL: @test_mm512_cmp_round_pd_mask
   // CHECK: @llvm.x86.avx512.mask.cmp.pd.512
-  return _mm512_cmp_round_pd_mask(a, b, 0, _MM_FROUND_TO_NEAREST_INT);
+  return _mm512_cmp_round_pd_mask(a, b, 0, _MM_FROUND_CUR_DIRECTION);
 }
 
 __mmask8 test_mm512_mask_cmp_round_pd_mask(__mmask8 m, __m512d a, __m512d b) {
   // CHECK-LABEL: @test_mm512_mask_cmp_round_pd_mask
   // CHECK: @llvm.x86.avx512.mask.cmp.pd.512
-  return _mm512_mask_cmp_round_pd_mask(m, a, b, 0, _MM_FROUND_TO_NEAREST_INT);
+  return _mm512_mask_cmp_round_pd_mask(m, a, b, 0, _MM_FROUND_CUR_DIRECTION);
 }
 
 __mmask8 test_mm512_cmp_pd_mask(__m512d a, __m512d b) {
   // CHECK-LABEL: @test_mm512_cmp_pd_mask
   // CHECK: @llvm.x86.avx512.mask.cmp.pd.512
   return _mm512_cmp_pd_mask(a, b, 0);
 }
 
__m128 test_mm_mask_sub_round_ss(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
   // CHECK: @llvm.x86.avx512.mask.sub.ss.round
   return _mm_mask_sub_round_ss(__W,__U,__A,__B,_MM_FROUND_TO_NEAREST_INT); 
 }
 __m128 test_mm_maskz_sub_round_ss(__mmask8 __U, __m128 __A, __m128 __B) {
   // CHECK-LABEL: @test_mm_maskz_sub_round_ss
   // CHECK: @llvm.x86.avx512.mask.sub.ss.round
   return _mm_maskz_sub_round_ss(__U,__A,__B,_MM_FROUND_TO_NEAREST_INT); 
 }
 __m128 test_mm_mask_sub_ss(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
   // CHECK-LABEL: @test_mm_mask_sub_ss
   // CHECK: @llvm.x86.avx512.mask.sub.ss.round
   return _mm_mask_sub_ss(__W,__U,__A,__B); 
 }
 __m128 test_mm_maskz_sub_ss(__mmask8 __U, __m128 __A, __m128 __B) {
   // CHECK-LABEL: @test_mm_maskz_sub_ss
   // CHECK: @llvm.x86.avx512.mask.sub.ss.round
   return _mm_maskz_sub_ss(__U,__A,__B); 
int test_mm_comi_round_sd(__m128d __A, __m128d __B) {
   return _mm_comi_round_sd(__A, __B, 5, 3); 
 }
 
 int test_mm_comi_round_ss(__m128 __A, __m128 __B) {
   // CHECK-LABEL: @test_mm_comi_round_ss
   // CHECK: @llvm.x86.avx512.vcomi.ss
   return _mm_comi_round_ss(__A, __B, 5, 3); 
 }
+__m512i test_mm512_sll_epi32(__m512i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm512_sll_epi32
+  // CHECK: @llvm.x86.avx512.mask.psll.d
+  return _mm512_sll_epi32(__A, __B); 
+}
+
+__m512i test_mm512_mask_sll_epi32(__m512i __W, __mmask16 __U, __m512i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm512_mask_sll_epi32
+  // CHECK: @llvm.x86.avx512.mask.psll.d
+  return _mm512_mask_sll_epi32(__W, __U, __A, __B); 
+}
+
+__m512i test_mm512_maskz_sll_epi32(__mmask16 __U, __m512i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm512_maskz_sll_epi32
+  // CHECK: @llvm.x86.avx512.mask.psll.d
+  return _mm512_maskz_sll_epi32(__U, __A, __B); 
+}
+
+__m512i test_mm512_sll_epi64(__m512i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm512_sll_epi64
+  // CHECK: @llvm.x86.avx512.mask.psll.q
+  return _mm512_sll_epi64(__A, __B); 
+}
+
+__m512i test_mm512_mask_sll_epi64(__m512i __W, __mmask8 __U, __m512i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm512_mask_sll_epi64
+  // CHECK: @llvm.x86.avx512.mask.psll.q
+  return _mm512_mask_sll_epi64(__W, __U, __A, __B); 
+}
+
+__m512i test_mm512_maskz_sll_epi64(__mmask8 __U, __m512i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm512_maskz_sll_epi64
+  // CHECK: @llvm.x86.avx512.mask.psll.q
+  return _mm512_maskz_sll_epi64(__U, __A, __B); 
+}
+
+__m512i test_mm512_sllv_epi32(__m512i __X, __m512i __Y) {
+  // CHECK-LABEL: @test_mm512_sllv_epi32
+  // CHECK: @llvm.x86.avx512.mask.psllv.d
+  return _mm512_sllv_epi32(__X, __Y); 
+}
+
+__m512i test_mm512_mask_sllv_epi32(__m512i __W, __mmask16 __U, __m512i __X, __m512i __Y) {
+  // CHECK-LABEL: @test_mm512_mask_sllv_epi32
+  // CHECK: @llvm.x86.avx512.mask.psllv.d
+  return _mm512_mask_sllv_epi32(__W, __U, __X, __Y); 
+}
+
+__m512i test_mm512_maskz_sllv_epi32(__mmask16 __U, __m512i __X, __m512i __Y) {
+  // CHECK-LABEL: @test_mm512_maskz_sllv_epi32
+  // CHECK: @llvm.x86.avx512.mask.psllv.d
+  return _mm512_maskz_sllv_epi32(__U, __X, __Y); 
+}
+
+__m512i test_mm512_sllv_epi64(__m512i __X, __m512i __Y) {
+  // CHECK-LABEL: @test_mm512_sllv_epi64
+  // CHECK: @llvm.x86.avx512.mask.psllv.q
+  return _mm512_sllv_epi64(__X, __Y); 
+}
+
+__m512i test_mm512_mask_sllv_epi64(__m512i __W, __mmask8 __U, __m512i __X, __m512i __Y) {
+  // CHECK-LABEL: @test_mm512_mask_sllv_epi64
+  // CHECK: @llvm.x86.avx512.mask.psllv.q
+  return _mm512_mask_sllv_epi64(__W, __U, __X, __Y); 
+}
+
+__m512i test_mm512_maskz_sllv_epi64(__mmask8 __U, __m512i __X, __m512i __Y) {
+  // CHECK-LABEL: @test_mm512_maskz_sllv_epi64
+  // CHECK: @llvm.x86.avx512.mask.psllv.q
+  return _mm512_maskz_sllv_epi64(__U, __X, __Y); 
+}
+
+__m512i test_mm512_sra_epi32(__m512i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm512_sra_epi32
+  // CHECK: @llvm.x86.avx512.mask.psra.d
+  return _mm512_sra_epi32(__A, __B); 
+}
+
+__m512i test_mm512_mask_sra_epi32(__m512i __W, __mmask16 __U, __m512i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm512_mask_sra_epi32
+  // CHECK: @llvm.x86.avx512.mask.psra.d
+  return _mm512_mask_sra_epi32(__W, __U, __A, __B); 
+}
+
+__m512i test_mm512_maskz_sra_epi32(__mmask16 __U, __m512i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm512_maskz_sra_epi32
+  // CHECK: @llvm.x86.avx512.mask.psra.d
+  return _mm512_maskz_sra_epi32(__U, __A, __B); 
+}
+
+__m512i test_mm512_sra_epi64(__m512i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm512_sra_epi64
+  // CHECK: @llvm.x86.avx512.mask.psra.q
+  return _mm512_sra_epi64(__A, __B); 
+}
+
+__m512i test_mm512_mask_sra_epi64(__m512i __W, __mmask8 __U, __m512i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm512_mask_sra_epi64
+  // CHECK: @llvm.x86.avx512.mask.psra.q
+  return _mm512_mask_sra_epi64(__W, __U, __A, __B); 
+}
+
+__m512i test_mm512_maskz_sra_epi64(__mmask8 __U, __m512i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm512_maskz_sra_epi64
+  // CHECK: @llvm.x86.avx512.mask.psra.q
+  return _mm512_maskz_sra_epi64(__U, __A, __B); 
+}
+
+__m512i test_mm512_srav_epi32(__m512i __X, __m512i __Y) {
+  // CHECK-LABEL: @test_mm512_srav_epi32
+  // CHECK: @llvm.x86.avx512.mask.psrav.d
+  return _mm512_srav_epi32(__X, __Y); 
+}
+
+__m512i test_mm512_mask_srav_epi32(__m512i __W, __mmask16 __U, __m512i __X, __m512i __Y) {
+  // CHECK-LABEL: @test_mm512_mask_srav_epi32
+  // CHECK: @llvm.x86.avx512.mask.psrav.d
+  return _mm512_mask_srav_epi32(__W, __U, __X, __Y); 
+}
+
+__m512i test_mm512_maskz_srav_epi32(__mmask16 __U, __m512i __X, __m512i __Y) {
+  // CHECK-LABEL: @test_mm512_maskz_srav_epi32
+  // CHECK: @llvm.x86.avx512.mask.psrav.d
+  return _mm512_maskz_srav_epi32(__U, __X, __Y); 
+}
+
+__m512i test_mm512_srav_epi64(__m512i __X, __m512i __Y) {
+  // CHECK-LABEL: @test_mm512_srav_epi64
+  // CHECK: @llvm.x86.avx512.mask.psrav.q
+  return _mm512_srav_epi64(__X, __Y); 
+}
+
+__m512i test_mm512_mask_srav_epi64(__m512i __W, __mmask8 __U, __m512i __X, __m512i __Y) {
+  // CHECK-LABEL: @test_mm512_mask_srav_epi64
+  // CHECK: @llvm.x86.avx512.mask.psrav.q
+  return _mm512_mask_srav_epi64(__W, __U, __X, __Y); 
+}
+
+__m512i test_mm512_maskz_srav_epi64(__mmask8 __U, __m512i __X, __m512i __Y) {
+  // CHECK-LABEL: @test_mm512_maskz_srav_epi64
+  // CHECK: @llvm.x86.avx512.mask.psrav.q
+  return _mm512_maskz_srav_epi64(__U, __X, __Y); 
+}
+
+__m512i test_mm512_srl_epi32(__m512i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm512_srl_epi32
+  // CHECK: @llvm.x86.avx512.mask.psrl.d
+  return _mm512_srl_epi32(__A, __B); 
+}
+
+__m512i test_mm512_mask_srl_epi32(__m512i __W, __mmask16 __U, __m512i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm512_mask_srl_epi32
+  // CHECK: @llvm.x86.avx512.mask.psrl.d
+  return _mm512_mask_srl_epi32(__W, __U, __A, __B); 
+}
+
+__m512i test_mm512_maskz_srl_epi32(__mmask16 __U, __m512i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm512_maskz_srl_epi32
+  // CHECK: @llvm.x86.avx512.mask.psrl.d
+  return _mm512_maskz_srl_epi32(__U, __A, __B); 
+}
+
+__m512i test_mm512_srl_epi64(__m512i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm512_srl_epi64
+  // CHECK: @llvm.x86.avx512.mask.psrl.q
+  return _mm512_srl_epi64(__A, __B); 
+}
+
+__m512i test_mm512_mask_srl_epi64(__m512i __W, __mmask8 __U, __m512i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm512_mask_srl_epi64
+  // CHECK: @llvm.x86.avx512.mask.psrl.q
+  return _mm512_mask_srl_epi64(__W, __U, __A, __B); 
+}
+
+__m512i test_mm512_maskz_srl_epi64(__mmask8 __U, __m512i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm512_maskz_srl_epi64
+  // CHECK: @llvm.x86.avx512.mask.psrl.q
+  return _mm512_maskz_srl_epi64(__U, __A, __B); 
+}
+
+__m512i test_mm512_srlv_epi32(__m512i __X, __m512i __Y) {
+  // CHECK-LABEL: @test_mm512_srlv_epi32
+  // CHECK: @llvm.x86.avx512.mask.psrlv.d
+  return _mm512_srlv_epi32(__X, __Y); 
+}
+
+__m512i test_mm512_mask_srlv_epi32(__m512i __W, __mmask16 __U, __m512i __X, __m512i __Y) {
+  // CHECK-LABEL: @test_mm512_mask_srlv_epi32
+  // CHECK: @llvm.x86.avx512.mask.psrlv.d
+  return _mm512_mask_srlv_epi32(__W, __U, __X, __Y); 
+}
+
+__m512i test_mm512_maskz_srlv_epi32(__mmask16 __U, __m512i __X, __m512i __Y) {
+  // CHECK-LABEL: @test_mm512_maskz_srlv_epi32
+  // CHECK: @llvm.x86.avx512.mask.psrlv.d
+  return _mm512_maskz_srlv_epi32(__U, __X, __Y); 
+}
+
+__m512i test_mm512_srlv_epi64(__m512i __X, __m512i __Y) {
+  // CHECK-LABEL: @test_mm512_srlv_epi64
+  // CHECK: @llvm.x86.avx512.mask.psrlv.q
+  return _mm512_srlv_epi64(__X, __Y); 
+}
+
+__m512i test_mm512_mask_srlv_epi64(__m512i __W, __mmask8 __U, __m512i __X, __m512i __Y) {
+  // CHECK-LABEL: @test_mm512_mask_srlv_epi64
+  // CHECK: @llvm.x86.avx512.mask.psrlv.q
+  return _mm512_mask_srlv_epi64(__W, __U, __X, __Y); 
+}
+
+__m512i test_mm512_maskz_srlv_epi64(__mmask8 __U, __m512i __X, __m512i __Y) {
+  // CHECK-LABEL: @test_mm512_maskz_srlv_epi64
+  // CHECK: @llvm.x86.avx512.mask.psrlv.q
+  return _mm512_maskz_srlv_epi64(__U, __X, __Y); 
+}
+
+__m512i test_mm512_ternarylogic_epi32(__m512i __A, __m512i __B, __m512i __C) {
+  // CHECK-LABEL: @test_mm512_ternarylogic_epi32
+  // CHECK: @llvm.x86.avx512.mask.pternlog.d.512
+  return _mm512_ternarylogic_epi32(__A, __B, __C, 4); 
+}
+
+__m512i test_mm512_mask_ternarylogic_epi32(__m512i __A, __mmask16 __U, __m512i __B, __m512i __C) {
+  // CHECK-LABEL: @test_mm512_mask_ternarylogic_epi32
+  // CHECK: @llvm.x86.avx512.mask.pternlog.d.512
+  return _mm512_mask_ternarylogic_epi32(__A, __U, __B, __C, 4); 
+}
+
+__m512i test_mm512_maskz_ternarylogic_epi32(__mmask16 __U, __m512i __A, __m512i __B, __m512i __C) {
+  // CHECK-LABEL: @test_mm512_maskz_ternarylogic_epi32
+  // CHECK: @llvm.x86.avx512.maskz.pternlog.d.512
+  return _mm512_maskz_ternarylogic_epi32(__U, __A, __B, __C, 4); 
+}
+
+__m512i test_mm512_ternarylogic_epi64(__m512i __A, __m512i __B, __m512i __C) {
+  // CHECK-LABEL: @test_mm512_ternarylogic_epi64
+  // CHECK: @llvm.x86.avx512.mask.pternlog.q.512
+  return _mm512_ternarylogic_epi64(__A, __B, __C, 4); 
+}
+
+__m512i test_mm512_mask_ternarylogic_epi64(__m512i __A, __mmask8 __U, __m512i __B, __m512i __C) {
+  // CHECK-LABEL: @test_mm512_mask_ternarylogic_epi64
+  // CHECK: @llvm.x86.avx512.mask.pternlog.q.512
+  return _mm512_mask_ternarylogic_epi64(__A, __U, __B, __C, 4); 
+}
+
+__m512i test_mm512_maskz_ternarylogic_epi64(__mmask8 __U, __m512i __A, __m512i __B, __m512i __C) {
+  // CHECK-LABEL: @test_mm512_maskz_ternarylogic_epi64
+  // CHECK: @llvm.x86.avx512.maskz.pternlog.q.512
+  return _mm512_maskz_ternarylogic_epi64(__U, __A, __B, __C, 4); 
+}
+
+__mmask16 test_mm512_testn_epi32_mask(__m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_testn_epi32_mask
+  // CHECK: @llvm.x86.avx512.ptestnm.d.512
+  return _mm512_testn_epi32_mask(__A, __B); 
+}
+
+__mmask16 test_mm512_mask_testn_epi32_mask(__mmask16 __U, __m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_mask_testn_epi32_mask
+  // CHECK: @llvm.x86.avx512.ptestnm.d.512
+  return _mm512_mask_testn_epi32_mask(__U, __A, __B); 
+}
+
+__mmask8 test_mm512_testn_epi64_mask(__m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_testn_epi64_mask
+  // CHECK: @llvm.x86.avx512.ptestnm.q.512
+  return _mm512_testn_epi64_mask(__A, __B); 
+}
+
+__mmask8 test_mm512_mask_testn_epi64_mask(__mmask8 __U, __m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_mask_testn_epi64_mask
+  // CHECK: @llvm.x86.avx512.ptestnm.q.512
+  return _mm512_mask_testn_epi64_mask(__U, __A, __B); 
+}
+
+__m512i test_mm512_unpackhi_epi32(__m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_unpackhi_epi32
+  // CHECK: @llvm.x86.avx512.mask.punpckhd.q.512
+  return _mm512_unpackhi_epi32(__A, __B); 
+}
+
+__m512i test_mm512_mask_unpackhi_epi32(__m512i __W, __mmask16 __U, __m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_mask_unpackhi_epi32
+  // CHECK: @llvm.x86.avx512.mask.punpckhd.q.512
+  return _mm512_mask_unpackhi_epi32(__W, __U, __A, __B); 
+}
+
+__m512i test_mm512_maskz_unpackhi_epi32(__mmask16 __U, __m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_maskz_unpackhi_epi32
+  // CHECK: @llvm.x86.avx512.mask.punpckhd.q.512
+  return _mm512_maskz_unpackhi_epi32(__U, __A, __B); 
+}
+
+__m512i test_mm512_unpackhi_epi64(__m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_unpackhi_epi64
+  // CHECK: @llvm.x86.avx512.mask.punpckhqd.q.512
+  return _mm512_unpackhi_epi64(__A, __B); 
+}
+
+__m512i test_mm512_mask_unpackhi_epi64(__m512i __W, __mmask8 __U, __m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_mask_unpackhi_epi64
+  // CHECK: @llvm.x86.avx512.mask.punpckhqd.q.512
+  return _mm512_mask_unpackhi_epi64(__W, __U, __A, __B); 
+}
+
+__m512i test_mm512_maskz_unpackhi_epi64(__mmask8 __U, __m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_maskz_unpackhi_epi64
+  // CHECK: @llvm.x86.avx512.mask.punpckhqd.q.512
+  return _mm512_maskz_unpackhi_epi64(__U, __A, __B); 
+}
+
+__m512i test_mm512_unpacklo_epi32(__m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_unpacklo_epi32
+  // CHECK: @llvm.x86.avx512.mask.punpckld.q.512
+  return _mm512_unpacklo_epi32(__A, __B); 
+}
+
+__m512i test_mm512_mask_unpacklo_epi32(__m512i __W, __mmask16 __U, __m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_mask_unpacklo_epi32
+  // CHECK: @llvm.x86.avx512.mask.punpckld.q.512
+  return _mm512_mask_unpacklo_epi32(__W, __U, __A, __B); 
+}
+
+__m512i test_mm512_maskz_unpacklo_epi32(__mmask16 __U, __m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_maskz_unpacklo_epi32
+  // CHECK: @llvm.x86.avx512.mask.punpckld.q.512
+  return _mm512_maskz_unpacklo_epi32(__U, __A, __B); 
+}
+
+__m512i test_mm512_unpacklo_epi64(__m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_unpacklo_epi64
+  // CHECK: @llvm.x86.avx512.mask.punpcklqd.q.512
+  return _mm512_unpacklo_epi64(__A, __B); 
+}
+
+__m512i test_mm512_mask_unpacklo_epi64(__m512i __W, __mmask8 __U, __m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_mask_unpacklo_epi64
+  // CHECK: @llvm.x86.avx512.mask.punpcklqd.q.512
+  return _mm512_mask_unpacklo_epi64(__W, __U, __A, __B); 
+}
+
+__m512i test_mm512_maskz_unpacklo_epi64(__mmask8 __U, __m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_maskz_unpacklo_epi64
+  // CHECK: @llvm.x86.avx512.mask.punpcklqd.q.512
+  return _mm512_maskz_unpacklo_epi64(__U, __A, __B); 
+}
+
+__m128d test_mm_rcp14_sd(__m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_rcp14_sd
+  // CHECK: @llvm.x86.avx512.rcp14.sd
+  return _mm_rcp14_sd(__A, __B); 
+}
+
+__m128 test_mm_rcp14_ss(__m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_rcp14_ss
+  // CHECK: @llvm.x86.avx512.rcp14.ss
+  return _mm_rcp14_ss(__A, __B);
+}
+
+__m128d test_mm_roundscale_round_sd(__m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_roundscale_round_sd
+  // CHECK: @llvm.x86.avx512.mask.rndscale.sd
+  return _mm_roundscale_round_sd(__A, __B, 4, 4); 
+}
+
+__m128d test_mm_roundscale_sd(__m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_roundscale_sd
+  // CHECK: @llvm.x86.avx512.mask.rndscale.sd
+  return _mm_roundscale_sd(__A, __B, 4); 
+}
+
+__m128d test_mm_mask_roundscale_sd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B){
+  // CHECK: @llvm.x86.avx512.mask.rndscale.sd
+    return _mm_mask_roundscale_sd(__W,__U,__A,__B,3);
+}
+
+__m128d test_mm_mask_roundscale_round_sd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B){
+  // CHECK: @llvm.x86.avx512.mask.rndscale.sd
+    return _mm_mask_roundscale_round_sd(__W,__U,__A,__B,3,_MM_FROUND_CUR_DIRECTION);
+}
+
+__m128d test_mm_maskz_roundscale_sd(__mmask8 __U, __m128d __A, __m128d __B){
+  // CHECK: @llvm.x86.avx512.mask.rndscale.sd
+    return _mm_maskz_roundscale_sd(__U,__A,__B,3);
+}
+
+__m128d test_mm_maskz_roundscale_round_sd(__mmask8 __U, __m128d __A, __m128d __B){
+  // CHECK: @llvm.x86.avx512.mask.rndscale.sd
+    return _mm_maskz_roundscale_round_sd(__U,__A,__B,3,_MM_FROUND_CUR_DIRECTION );
+}
+
+__m128 test_mm_roundscale_round_ss(__m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_roundscale_round_ss
+  // CHECK: @llvm.x86.avx512.mask.rndscale.ss
+  return _mm_roundscale_round_ss(__A, __B, 4, 4); 
+}
+
+__m128 test_mm_roundscale_ss(__m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_roundscale_ss
+  // CHECK: @llvm.x86.avx512.mask.rndscale.ss
+  return _mm_roundscale_ss(__A, __B, 4); 
+}
+
+__m128d test_mm_rsqrt14_sd(__m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_rsqrt14_sd
+  // CHECK: @llvm.x86.avx512.rsqrt14.sd
+  return _mm_rsqrt14_sd(__A, __B); 
+}
+
+__m128 test_mm_rsqrt14_ss(__m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_rsqrt14_ss
+  // CHECK: @llvm.x86.avx512.rsqrt14.ss
+  return _mm_rsqrt14_ss(__A, __B); 
+}
+
+__m512d test_mm512_scalef_round_pd(__m512d __A, __m512d __B) {
+  // CHECK-LABEL: @test_mm512_scalef_round_pd
+  // CHECK: @llvm.x86.avx512.mask.scalef.pd.512
+  return _mm512_scalef_round_pd(__A, __B, 4); 
+}
+
+__m512d test_mm512_mask_scalef_round_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) {
+  // CHECK-LABEL: @test_mm512_mask_scalef_round_pd
+  // CHECK: @llvm.x86.avx512.mask.scalef.pd.512
+  return _mm512_mask_scalef_round_pd(__W, __U, __A, __B, 4); 
+}
+
+__m512d test_mm512_maskz_scalef_round_pd(__mmask8 __U, __m512d __A, __m512d __B) {
+  // CHECK-LABEL: @test_mm512_maskz_scalef_round_pd
+  // CHECK: @llvm.x86.avx512.mask.scalef.pd.512
+  return _mm512_maskz_scalef_round_pd(__U, __A, __B, 4); 
+}
+
+__m512d test_mm512_scalef_pd(__m512d __A, __m512d __B) {
+  // CHECK-LABEL: @test_mm512_scalef_pd
+  // CHECK: @llvm.x86.avx512.mask.scalef.pd.512
+  return _mm512_scalef_pd(__A, __B); 
+}
+
+__m512d test_mm512_mask_scalef_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) {
+  // CHECK-LABEL: @test_mm512_mask_scalef_pd
+  // CHECK: @llvm.x86.avx512.mask.scalef.pd.512
+  return _mm512_mask_scalef_pd(__W, __U, __A, __B); 
+}
+
+__m512d test_mm512_maskz_scalef_pd(__mmask8 __U, __m512d __A, __m512d __B) {
+  // CHECK-LABEL: @test_mm512_maskz_scalef_pd
+  // CHECK: @llvm.x86.avx512.mask.scalef.pd.512
+  return _mm512_maskz_scalef_pd(__U, __A, __B); 
+}
+
+__m512 test_mm512_scalef_round_ps(__m512 __A, __m512 __B) {
+  // CHECK-LABEL: @test_mm512_scalef_round_ps
+  // CHECK: @llvm.x86.avx512.mask.scalef.ps.512
+  return _mm512_scalef_round_ps(__A, __B, 4); 
+}
+
+__m512 test_mm512_mask_scalef_round_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) {
+  // CHECK-LABEL: @test_mm512_mask_scalef_round_ps
+  // CHECK: @llvm.x86.avx512.mask.scalef.ps.512
+  return _mm512_mask_scalef_round_ps(__W, __U, __A, __B, 4); 
+}
+
+__m512 test_mm512_maskz_scalef_round_ps(__mmask16 __U, __m512 __A, __m512 __B) {
+  // CHECK-LABEL: @test_mm512_maskz_scalef_round_ps
+  // CHECK: @llvm.x86.avx512.mask.scalef.ps.512
+  return _mm512_maskz_scalef_round_ps(__U, __A, __B, 4); 
+}
+
+__m512 test_mm512_scalef_ps(__m512 __A, __m512 __B) {
+  // CHECK-LABEL: @test_mm512_scalef_ps
+  // CHECK: @llvm.x86.avx512.mask.scalef.ps.512
+  return _mm512_scalef_ps(__A, __B); 
+}
+
+__m512 test_mm512_mask_scalef_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) {
+  // CHECK-LABEL: @test_mm512_mask_scalef_ps
+  // CHECK: @llvm.x86.avx512.mask.scalef.ps.512
+  return _mm512_mask_scalef_ps(__W, __U, __A, __B); 
+}
+
+__m512 test_mm512_maskz_scalef_ps(__mmask16 __U, __m512 __A, __m512 __B) {
+  // CHECK-LABEL: @test_mm512_maskz_scalef_ps
+  // CHECK: @llvm.x86.avx512.mask.scalef.ps.512
+  return _mm512_maskz_scalef_ps(__U, __A, __B); 
+}
+
+__m128d test_mm_scalef_round_sd(__m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_scalef_round_sd
+  // CHECK: @llvm.x86.avx512.mask.scalef
+  return _mm_scalef_round_sd(__A, __B, 4); 
+}
+
+__m128d test_mm_scalef_sd(__m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_scalef_sd
+  // CHECK: @llvm.x86.avx512.mask.scalef
+  return _mm_scalef_sd(__A, __B); 
+}
+
+__m128d test_mm_mask_scalef_sd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B){
+    return _mm_mask_scalef_sd(__W,__U,__A,__B);
+}
+
+__m128d test_mm_mask_scalef_round_sd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B){
+    return _mm_mask_scalef_round_sd(__W,__U,__A,__B,_MM_FROUND_CUR_DIRECTION);
+}
+
+__m128d test_mm_maskz_scalef_sd(__mmask8 __U, __m128d __A, __m128d __B){
+    return _mm_maskz_scalef_sd(__U,__A,__B);
+}
+
+__m128d test_mm_maskz_scalef_round_sd(__mmask8 __U, __m128d __A, __m128d __B){
+    return _mm_maskz_scalef_round_sd(__U,__A,__B,_MM_FROUND_CUR_DIRECTION);
+}
+
+__m128 test_mm_scalef_round_ss(__m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_scalef_round_ss
+  // CHECK: @llvm.x86.avx512.mask.scalef
+  return _mm_scalef_round_ss(__A, __B, 4); 
+}
+
+__m128 test_mm_scalef_ss(__m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_scalef_ss
+  // CHECK: @llvm.x86.avx512.mask.scalef
+  return _mm_scalef_ss(__A, __B); 
+}
+
+__m128 test_mm_mask_scalef_ss(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B){
+// CHECK: @llvm.x86.avx512.mask.scalef.ss
+    return _mm_mask_scalef_ss(__W,__U,__A,__B);
+}
+
+__m128 test_mm_mask_scalef_round_ss(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B){
+// CHECK: @llvm.x86.avx512.mask.scalef.ss
+    return _mm_mask_scalef_round_ss(__W,__U,__A,__B,_MM_FROUND_CUR_DIRECTION);
+}
+
+__m128 test_mm_maskz_scalef_ss(__mmask8 __U, __m128 __A, __m128 __B){
+// CHECK: @llvm.x86.avx512.mask.scalef.ss
+    return _mm_maskz_scalef_ss(__U,__A,__B);
+}
+
+__m128 test_mm_maskz_scalef_round_ss(__mmask8 __U, __m128 __A, __m128 __B){
+// CHECK: @llvm.x86.avx512.mask.scalef.ss
+    return _mm_maskz_scalef_round_ss(__U,__A,__B,_MM_FROUND_CUR_DIRECTION);
+}
+
+void test_mm512_i64scatter_ps(float *__addr, __m512i __index, __m256 __v1) {
+  // CHECK-LABEL: @test_mm512_i64scatter_ps
+  // CHECK: @llvm.x86.avx512.scatter.qps.512
+  return _mm512_i64scatter_ps(__addr, __index, __v1, 4); 
+}
+
+void test_mm512_mask_i64scatter_ps(float *__addr, __mmask8 __mask, __m512i __index, __m256 __v1) {
+  // CHECK-LABEL: @test_mm512_mask_i64scatter_ps
+  // CHECK: @llvm.x86.avx512.scatter.qps.512
+  return _mm512_mask_i64scatter_ps(__addr, __mask, __index, __v1, 4); 
+}
+
+void test_mm512_i64scatter_epi32(int *__addr, __m512i __index, __m256i __v1) {
+  // CHECK-LABEL: @test_mm512_i64scatter_epi32
+  // CHECK: @llvm.x86.avx512.scatter.qpi.512
+  return _mm512_i64scatter_epi32(__addr, __index, __v1, 4); 
+}
+
+void test_mm512_mask_i64scatter_epi32(int *__addr, __mmask8 __mask, __m512i __index, __m256i __v1) {
+  // CHECK-LABEL: @test_mm512_mask_i64scatter_epi32
+  // CHECK: @llvm.x86.avx512.scatter.qpi.512
+  return _mm512_mask_i64scatter_epi32(__addr, __mask, __index, __v1, 4); 
+}
+
+void test_mm512_i64scatter_pd(double *__addr, __m512i __index, __m512d __v1) {
+  // CHECK-LABEL: @test_mm512_i64scatter_pd
+  // CHECK: @llvm.x86.avx512.scatter.qpd.512
+  return _mm512_i64scatter_pd(__addr, __index, __v1, 4); 
+}
+
+void test_mm512_mask_i64scatter_pd(double *__addr, __mmask8 __mask, __m512i __index, __m512d __v1) {
+  // CHECK-LABEL: @test_mm512_mask_i64scatter_pd
+  // CHECK: @llvm.x86.avx512.scatter.qpd.512
+  return _mm512_mask_i64scatter_pd(__addr, __mask, __index, __v1, 4); 
+}
+
+void test_mm512_i64scatter_epi64(long long *__addr, __m512i __index, __m512i __v1) {
+  // CHECK-LABEL: @test_mm512_i64scatter_epi64
+  // CHECK: @llvm.x86.avx512.scatter.qpq.512
+  return _mm512_i64scatter_epi64(__addr, __index, __v1, 4); 
+}
+
+void test_mm512_mask_i64scatter_epi64(long long *__addr, __mmask8 __mask, __m512i __index, __m512i __v1) {
+  // CHECK-LABEL: @test_mm512_mask_i64scatter_epi64
+  // CHECK: @llvm.x86.avx512.scatter.qpq.512
+  return _mm512_mask_i64scatter_epi64(__addr, __mask, __index, __v1, 4); 
+}
+
+void test_mm512_i32scatter_ps(float *__addr, __m512i __index, __m512 __v1) {
+  // CHECK-LABEL: @test_mm512_i32scatter_ps
+  // CHECK: @llvm.x86.avx512.scatter.dps.512
+  return _mm512_i32scatter_ps(__addr, __index, __v1, 4); 
+}
+
+void test_mm512_mask_i32scatter_ps(float *__addr, __mmask16 __mask, __m512i __index, __m512 __v1) {
+  // CHECK-LABEL: @test_mm512_mask_i32scatter_ps
+  // CHECK: @llvm.x86.avx512.scatter.dps.512
+  return _mm512_mask_i32scatter_ps(__addr, __mask, __index, __v1, 4); 
+}
+
+void test_mm512_i32scatter_epi32(int *__addr, __m512i __index, __m512i __v1) {
+  // CHECK-LABEL: @test_mm512_i32scatter_epi32
+  // CHECK: @llvm.x86.avx512.scatter.dpi.512
+  return _mm512_i32scatter_epi32(__addr, __index, __v1, 4); 
+}
+
+void test_mm512_mask_i32scatter_epi32(int *__addr, __mmask16 __mask, __m512i __index, __m512i __v1) {
+  // CHECK-LABEL: @test_mm512_mask_i32scatter_epi32
+  // CHECK: @llvm.x86.avx512.scatter.dpi.512
+  return _mm512_mask_i32scatter_epi32(__addr, __mask, __index, __v1, 4); 
+}
+
+void test_mm512_i32scatter_pd(double *__addr, __m256i __index, __m512d __v1) {
+  // CHECK-LABEL: @test_mm512_i32scatter_pd
+  // CHECK: @llvm.x86.avx512.scatter.dpd.512
+  return _mm512_i32scatter_pd(__addr, __index, __v1, 4); 
+}
+
+void test_mm512_mask_i32scatter_pd(double *__addr, __mmask8 __mask, __m256i __index, __m512d __v1) {
+  // CHECK-LABEL: @test_mm512_mask_i32scatter_pd
+  // CHECK: @llvm.x86.avx512.scatter.dpd.512
+  return _mm512_mask_i32scatter_pd(__addr, __mask, __index, __v1, 4); 
+}
+
+void test_mm512_i32scatter_epi64(long long *__addr, __m256i __index, __m512i __v1) {
+  // CHECK-LABEL: @test_mm512_i32scatter_epi64
+  // CHECK: @llvm.x86.avx512.scatter.dpq.512
+  return _mm512_i32scatter_epi64(__addr, __index, __v1, 4); 
+}
+
+void test_mm512_mask_i32scatter_epi64(long long *__addr, __mmask8 __mask, __m256i __index, __m512i __v1) {
+  // CHECK-LABEL: @test_mm512_mask_i32scatter_epi64
+  // CHECK: @llvm.x86.avx512.scatter.dpq.512
+  return _mm512_mask_i32scatter_epi64(__addr, __mask, __index, __v1, 4); 
+}
+
+__m512 test_mm512_shuffle_f32x4(__m512 __A, __m512 __B) {
+  // CHECK-LABEL: @test_mm512_shuffle_f32x4
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm512_shuffle_f32x4(__A, __B, 4); 
+}
+
+__m512 test_mm512_mask_shuffle_f32x4(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) {
+  // CHECK-LABEL: @test_mm512_mask_shuffle_f32x4
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm512_mask_shuffle_f32x4(__W, __U, __A, __B, 4); 
+}
+
+__m512 test_mm512_maskz_shuffle_f32x4(__mmask16 __U, __m512 __A, __m512 __B) {
+  // CHECK-LABEL: @test_mm512_maskz_shuffle_f32x4
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm512_maskz_shuffle_f32x4(__U, __A, __B, 4); 
+}
+
+__m512d test_mm512_shuffle_f64x2(__m512d __A, __m512d __B) {
+  // CHECK-LABEL: @test_mm512_shuffle_f64x2
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm512_shuffle_f64x2(__A, __B, 4); 
+}
+
+__m512d test_mm512_mask_shuffle_f64x2(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) {
+  // CHECK-LABEL: @test_mm512_mask_shuffle_f64x2
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm512_mask_shuffle_f64x2(__W, __U, __A, __B, 4); 
+}
+
+__m512d test_mm512_maskz_shuffle_f64x2(__mmask8 __U, __m512d __A, __m512d __B) {
+  // CHECK-LABEL: @test_mm512_maskz_shuffle_f64x2
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm512_maskz_shuffle_f64x2(__U, __A, __B, 4); 
+}
+
+__m512i test_mm512_shuffle_i32x4(__m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_shuffle_i32x4
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm512_shuffle_i32x4(__A, __B, 4); 
+}
+
+__m512i test_mm512_mask_shuffle_i32x4(__m512i __W, __mmask16 __U, __m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_mask_shuffle_i32x4
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm512_mask_shuffle_i32x4(__W, __U, __A, __B, 4); 
+}
+
+__m512i test_mm512_maskz_shuffle_i32x4(__mmask16 __U, __m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_maskz_shuffle_i32x4
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm512_maskz_shuffle_i32x4(__U, __A, __B, 4); 
+}
+
+__m512i test_mm512_shuffle_i64x2(__m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_shuffle_i64x2
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm512_shuffle_i64x2(__A, __B, 4); 
+}
+
+__m512i test_mm512_mask_shuffle_i64x2(__m512i __W, __mmask8 __U, __m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_mask_shuffle_i64x2
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm512_mask_shuffle_i64x2(__W, __U, __A, __B, 4); 
+}
+
+__m512i test_mm512_maskz_shuffle_i64x2(__mmask8 __U, __m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_maskz_shuffle_i64x2
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm512_maskz_shuffle_i64x2(__U, __A, __B, 4); 
+}
+
+__m512d test_mm512_shuffle_pd(__m512d __M, __m512d __V) {
+  // CHECK-LABEL: @test_mm512_shuffle_pd
+  // CHECK: @llvm.x86.avx512.mask.shuf.pd.512
+  return _mm512_shuffle_pd(__M, __V, 4); 
+}
+
+__m512d test_mm512_mask_shuffle_pd(__m512d __W, __mmask8 __U, __m512d __M, __m512d __V) {
+  // CHECK-LABEL: @test_mm512_mask_shuffle_pd
+  // CHECK: @llvm.x86.avx512.mask.shuf.pd.512
+  return _mm512_mask_shuffle_pd(__W, __U, __M, __V, 4); 
+}
+
+__m512d test_mm512_maskz_shuffle_pd(__mmask8 __U, __m512d __M, __m512d __V) {
+  // CHECK-LABEL: @test_mm512_maskz_shuffle_pd
+  // CHECK: @llvm.x86.avx512.mask.shuf.pd.512
+  return _mm512_maskz_shuffle_pd(__U, __M, __V, 4); 
+}
+
+__m512 test_mm512_shuffle_ps(__m512 __M, __m512 __V) {
+  // CHECK-LABEL: @test_mm512_shuffle_ps
+  // CHECK: @llvm.x86.avx512.mask.shuf.ps.512
+  return _mm512_shuffle_ps(__M, __V, 4); 
+}
+
+__m512 test_mm512_mask_shuffle_ps(__m512 __W, __mmask16 __U, __m512 __M, __m512 __V) {
+  // CHECK-LABEL: @test_mm512_mask_shuffle_ps
+  // CHECK: @llvm.x86.avx512.mask.shuf.ps.512
+  return _mm512_mask_shuffle_ps(__W, __U, __M, __V, 4); 
+}
+
+__m512 test_mm512_maskz_shuffle_ps(__mmask16 __U, __m512 __M, __m512 __V) {
+  // CHECK-LABEL: @test_mm512_maskz_shuffle_ps
+  // CHECK: @llvm.x86.avx512.mask.shuf.ps.512
+  return _mm512_maskz_shuffle_ps(__U, __M, __V, 4); 
+}
+
+__m128d test_mm_sqrt_round_sd(__m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_sqrt_round_sd
+  // CHECK: @llvm.x86.avx512.mask.sqrt.sd
+  return _mm_sqrt_round_sd(__A, __B, 4); 
+}
+
+__m128d test_mm_mask_sqrt_sd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B){
+  // CHECK: @llvm.x86.avx512.mask.sqrt.sd
+    return _mm_mask_sqrt_sd(__W,__U,__A,__B);
+}
+
+__m128d test_mm_mask_sqrt_round_sd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B){
+  // CHECK: @llvm.x86.avx512.mask.sqrt.sd
+    return _mm_mask_sqrt_round_sd(__W,__U,__A,__B,_MM_FROUND_CUR_DIRECTION);
+}
+
+__m128d test_mm_maskz_sqrt_sd(__mmask8 __U, __m128d __A, __m128d __B){
+  // CHECK: @llvm.x86.avx512.mask.sqrt.sd
+    return _mm_maskz_sqrt_sd(__U,__A,__B);
+}
+
+__m128d test_mm_maskz_sqrt_round_sd(__mmask8 __U, __m128d __A, __m128d __B){
+  // CHECK: @llvm.x86.avx512.mask.sqrt.sd
+    return _mm_maskz_sqrt_round_sd(__U,__A,__B,_MM_FROUND_CUR_DIRECTION);
+}
+
+__m128 test_mm_sqrt_round_ss(__m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_sqrt_round_ss
+  // CHECK: @llvm.x86.avx512.mask.sqrt.ss
+  return _mm_sqrt_round_ss(__A, __B, 4); 
+}
+
+__m128 test_mm_mask_sqrt_ss(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B){
+  // CHECK: @llvm.x86.avx512.mask.sqrt.ss
+    return _mm_mask_sqrt_ss(__W,__U,__A,__B);
+}
+
+__m128 test_mm_mask_sqrt_round_ss(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B){
+  // CHECK: @llvm.x86.avx512.mask.sqrt.ss
+    return _mm_mask_sqrt_round_ss(__W,__U,__A,__B,_MM_FROUND_CUR_DIRECTION);
+}
+
+__m128 test_mm_maskz_sqrt_ss(__mmask8 __U, __m128 __A, __m128 __B){
+  // CHECK: @llvm.x86.avx512.mask.sqrt.ss
+    return _mm_maskz_sqrt_ss(__U,__A,__B);
+}
+
+__m128 test_mm_maskz_sqrt_round_ss(__mmask8 __U, __m128 __A, __m128 __B){
+  // CHECK: @llvm.x86.avx512.mask.sqrt.ss
+    return _mm_maskz_sqrt_round_ss(__U,__A,__B,_MM_FROUND_CUR_DIRECTION);
+}
+
+__m512d test_mm512_mask_unpackhi_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) {
+  // CHECK-LABEL: @test_mm512_mask_unpackhi_pd
+  // CHECK: @llvm.x86.avx512.mask.unpckh.pd.512
+  return _mm512_mask_unpackhi_pd(__W, __U, __A, __B); 
+}
+
+__m512d test_mm512_maskz_unpackhi_pd(__mmask8 __U, __m512d __A, __m512d __B) {
+  // CHECK-LABEL: @test_mm512_maskz_unpackhi_pd
+  // CHECK: @llvm.x86.avx512.mask.unpckh.pd.512
+  return _mm512_maskz_unpackhi_pd(__U, __A, __B); 
+}
+
+__m512 test_mm512_mask_unpackhi_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) {
+  // CHECK-LABEL: @test_mm512_mask_unpackhi_ps
+  // CHECK: @llvm.x86.avx512.mask.unpckh.ps.512
+  return _mm512_mask_unpackhi_ps(__W, __U, __A, __B); 
+}
+
+__m512 test_mm512_maskz_unpackhi_ps(__mmask16 __U, __m512 __A, __m512 __B) {
+  // CHECK-LABEL: @test_mm512_maskz_unpackhi_ps
+  // CHECK: @llvm.x86.avx512.mask.unpckh.ps.512
+  return _mm512_maskz_unpackhi_ps(__U, __A, __B); 
+}
+
+__m512d test_mm512_mask_unpacklo_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) {
+  // CHECK-LABEL: @test_mm512_mask_unpacklo_pd
+  // CHECK: @llvm.x86.avx512.mask.unpckl.pd.512
+  return _mm512_mask_unpacklo_pd(__W, __U, __A, __B); 
+}
+
+__m512d test_mm512_maskz_unpacklo_pd(__mmask8 __U, __m512d __A, __m512d __B) {
+  // CHECK-LABEL: @test_mm512_maskz_unpacklo_pd
+  // CHECK: @llvm.x86.avx512.mask.unpckl.pd.512
+  return _mm512_maskz_unpacklo_pd(__U, __A, __B); 
+}
+
+__m512 test_mm512_mask_unpacklo_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) {
+  // CHECK-LABEL: @test_mm512_mask_unpacklo_ps
+  // CHECK: @llvm.x86.avx512.mask.unpckl.ps.512
+  return _mm512_mask_unpacklo_ps(__W, __U, __A, __B); 
+}
+
+__m512 test_mm512_maskz_unpacklo_ps(__mmask16 __U, __m512 __A, __m512 __B) {
+  // CHECK-LABEL: @test_mm512_maskz_unpacklo_ps
+  // CHECK: @llvm.x86.avx512.mask.unpckl.ps.512
+  return _mm512_maskz_unpacklo_ps(__U, __A, __B); 
+}
+
+int test_mm_cvt_roundsd_si32(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvt_roundsd_si32
+  // CHECK: @llvm.x86.avx512.vcvtsd2si32
+  return _mm_cvt_roundsd_si32(__A, 4); 
+}
+
+int test_mm_cvt_roundsd_i32(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvt_roundsd_i32
+  // CHECK: @llvm.x86.avx512.vcvtsd2si32
+  return _mm_cvt_roundsd_i32(__A, 4); 
+}
+
+unsigned long long test_mm_cvt_roundsd_si64(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvt_roundsd_si64
+  // CHECK: @llvm.x86.avx512.vcvtsd2si64
+  return _mm_cvt_roundsd_si64(__A, 4); 
+}
+
+long long test_mm_cvt_roundsd_i64(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvt_roundsd_i64
+  // CHECK: @llvm.x86.avx512.vcvtsd2si64
+  return _mm_cvt_roundsd_i64(__A, 4); 
+}
+
+unsigned test_mm_cvt_roundsd_u32(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvt_roundsd_u32
+  // CHECK: @llvm.x86.avx512.vcvtsd2usi32
+  return _mm_cvt_roundsd_u32(__A, 4); 
+}
+
+unsigned test_mm_cvtsd_u32(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvtsd_u32
+  // CHECK: @llvm.x86.avx512.vcvtsd2usi32
+  return _mm_cvtsd_u32(__A); 
+}
+
+unsigned long long test_mm_cvt_roundsd_u64(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvt_roundsd_u64
+  // CHECK: @llvm.x86.avx512.vcvtsd2usi64
+  return _mm_cvt_roundsd_u64(__A, 4); 
+}
+
+unsigned long long test_mm_cvtsd_u64(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvtsd_u64
+  // CHECK: @llvm.x86.avx512.vcvtsd2usi64
+  return _mm_cvtsd_u64(__A); 
+}
+
+int test_mm_cvt_roundss_si32(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvt_roundss_si32
+  // CHECK: @llvm.x86.avx512.vcvtss2si32
+  return _mm_cvt_roundss_si32(__A, 4); 
+}
+
+int test_mm_cvt_roundss_i32(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvt_roundss_i32
+  // CHECK: @llvm.x86.avx512.vcvtss2si32
+  return _mm_cvt_roundss_i32(__A, 4); 
+}
+
+int test_mm_cvt_roundss_si64(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvt_roundss_si64
+  // CHECK: @llvm.x86.avx512.vcvtss2si64
+  return _mm_cvt_roundss_si64(__A, 4); 
+}
+
+long long test_mm_cvt_roundss_i64(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvt_roundss_i64
+  // CHECK: @llvm.x86.avx512.vcvtss2si64
+  return _mm_cvt_roundss_i64(__A, 4); 
+}
+
+unsigned test_mm_cvt_roundss_u32(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvt_roundss_u32
+  // CHECK: @llvm.x86.avx512.vcvtss2usi32
+  return _mm_cvt_roundss_u32(__A, 4); 
+}
+
+unsigned test_mm_cvtss_u32(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvtss_u32
+  // CHECK: @llvm.x86.avx512.vcvtss2usi32
+  return _mm_cvtss_u32(__A); 
+}
+
+unsigned long long test_mm_cvt_roundss_u64(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvt_roundss_u64
+  // CHECK: @llvm.x86.avx512.vcvtss2usi64
+  return _mm_cvt_roundss_u64(__A, 4); 
+}
+
+unsigned long long test_mm_cvtss_u64(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvtss_u64
+  // CHECK: @llvm.x86.avx512.vcvtss2usi64
+  return _mm_cvtss_u64(__A); 
+}
+
+int test_mm_cvtt_roundsd_i32(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvtt_roundsd_i32
+  // CHECK: @llvm.x86.avx512.cvttsd2si
+  return _mm_cvtt_roundsd_i32(__A, 4); 
+}
+
+int test_mm_cvtt_roundsd_si32(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvtt_roundsd_si32
+  // CHECK: @llvm.x86.avx512.cvttsd2si
+  return _mm_cvtt_roundsd_si32(__A, 4); 
+}
+
+int test_mm_cvttsd_i32(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvttsd_i32
+  // CHECK: @llvm.x86.avx512.cvttsd2si
+  return _mm_cvttsd_i32(__A); 
+}
+
+unsigned long long test_mm_cvtt_roundsd_si64(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvtt_roundsd_si64
+  // CHECK: @llvm.x86.avx512.cvttsd2si64
+  return _mm_cvtt_roundsd_si64(__A, 4); 
+}
+
+long long test_mm_cvtt_roundsd_i64(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvtt_roundsd_i64
+  // CHECK: @llvm.x86.avx512.cvttsd2si64
+  return _mm_cvtt_roundsd_i64(__A, 4); 
+}
+
+long long test_mm_cvttsd_i64(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvttsd_i64
+  // CHECK: @llvm.x86.avx512.cvttsd2si64
+  return _mm_cvttsd_i64(__A); 
+}
+
+unsigned test_mm_cvtt_roundsd_u32(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvtt_roundsd_u32
+  // CHECK: @llvm.x86.avx512.cvttsd2usi
+  return _mm_cvtt_roundsd_u32(__A, 4); 
+}
+
+unsigned test_mm_cvttsd_u32(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvttsd_u32
+  // CHECK: @llvm.x86.avx512.cvttsd2usi
+  return _mm_cvttsd_u32(__A); 
+}
+
+unsigned long long test_mm_cvtt_roundsd_u64(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvtt_roundsd_u64
+  // CHECK: @llvm.x86.avx512.cvttsd2usi64
+  return _mm_cvtt_roundsd_u64(__A, 4); 
+}
+
+unsigned long long test_mm_cvttsd_u64(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvttsd_u64
+  // CHECK: @llvm.x86.avx512.cvttsd2usi64
+  return _mm_cvttsd_u64(__A); 
+}
+
+int test_mm_cvtt_roundss_i32(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvtt_roundss_i32
+  // CHECK: @llvm.x86.avx512.cvttss2si
+  return _mm_cvtt_roundss_i32(__A, 4); 
+}
+
+int test_mm_cvtt_roundss_si32(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvtt_roundss_si32
+  // CHECK: @llvm.x86.avx512.cvttss2si
+  return _mm_cvtt_roundss_si32(__A, 4); 
+}
+
+int test_mm_cvttss_i32(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvttss_i32
+  // CHECK: @llvm.x86.avx512.cvttss2si
+  return _mm_cvttss_i32(__A); 
+}
+
+float test_mm_cvtt_roundss_i64(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvtt_roundss_i64
+  // CHECK: @llvm.x86.avx512.cvttss2si64
+  return _mm_cvtt_roundss_i64(__A, 4); 
+}
+
+long long test_mm_cvtt_roundss_si64(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvtt_roundss_si64
+  // CHECK: @llvm.x86.avx512.cvttss2si64
+  return _mm_cvtt_roundss_si64(__A, 4); 
+}
+
+long long test_mm_cvttss_i64(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvttss_i64
+  // CHECK: @llvm.x86.avx512.cvttss2si64
+  return _mm_cvttss_i64(__A); 
+}
+
+unsigned test_mm_cvtt_roundss_u32(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvtt_roundss_u32
+  // CHECK: @llvm.x86.avx512.cvttss2usi
+  return _mm_cvtt_roundss_u32(__A, 4); 
+}
+
+unsigned test_mm_cvttss_u32(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvttss_u32
+  // CHECK: @llvm.x86.avx512.cvttss2usi
+  return _mm_cvttss_u32(__A); 
+}
+
+unsigned long long test_mm_cvtt_roundss_u64(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvtt_roundss_u64
+  // CHECK: @llvm.x86.avx512.cvttss2usi64
+  return _mm_cvtt_roundss_u64(__A, 4); 
+}
+
+unsigned long long test_mm_cvttss_u64(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvttss_u64
+  // CHECK: @llvm.x86.avx512.cvttss2usi64
+  return _mm_cvttss_u64(__A); 
+}
+
+__m512i test_mm512_mask2_permutex2var_epi32(__m512i __A, __m512i __I, __mmask16 __U, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_mask2_permutex2var_epi32
+  // CHECK: @llvm.x86.avx512.mask.vpermi2var.d.512
+  return _mm512_mask2_permutex2var_epi32(__A, __I, __U, __B); 
+}
+
+__m512d test_mm512_mask2_permutex2var_pd(__m512d __A, __m512i __I, __mmask8 __U, __m512d __B) {
+  // CHECK-LABEL: @test_mm512_mask2_permutex2var_pd
+  // CHECK: @llvm.x86.avx512.mask.vpermi2var.pd.512
+  return _mm512_mask2_permutex2var_pd(__A, __I, __U, __B); 
+}
+
+__m512 test_mm512_mask2_permutex2var_ps(__m512 __A, __m512i __I, __mmask16 __U, __m512 __B) {
+  // CHECK-LABEL: @test_mm512_mask2_permutex2var_ps
+  // CHECK: @llvm.x86.avx512.mask.vpermi2var.ps.512
+  return _mm512_mask2_permutex2var_ps(__A, __I, __U, __B); 
+}
+
+__m512i test_mm512_mask2_permutex2var_epi64(__m512i __A, __m512i __I, __mmask8 __U, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_mask2_permutex2var_epi64
+  // CHECK: @llvm.x86.avx512.mask.vpermi2var.q.512
+  return _mm512_mask2_permutex2var_epi64(__A, __I, __U, __B); 
+}
+
+__m512d test_mm512_permute_pd(__m512d __X) {
+  // CHECK-LABEL: @test_mm512_permute_pd
+  // CHECK: @llvm.x86.avx512.mask.vpermil.pd.512
+  return _mm512_permute_pd(__X, 4); 
+}
+
+__m512d test_mm512_mask_permute_pd(__m512d __W, __mmask8 __U, __m512d __X) {
+  // CHECK-LABEL: @test_mm512_mask_permute_pd
+  // CHECK: @llvm.x86.avx512.mask.vpermil.pd.512
+  return _mm512_mask_permute_pd(__W, __U, __X, 4); 
+}
+
+__m512d test_mm512_maskz_permute_pd(__mmask8 __U, __m512d __X) {
+  // CHECK-LABEL: @test_mm512_maskz_permute_pd
+  // CHECK: @llvm.x86.avx512.mask.vpermil.pd.512
+  return _mm512_maskz_permute_pd(__U, __X, 4); 
+}
+
+__m512 test_mm512_permute_ps(__m512 __X) {
+  // CHECK-LABEL: @test_mm512_permute_ps
+  // CHECK: @llvm.x86.avx512.mask.vpermil.ps.512
+  return _mm512_permute_ps(__X, 4); 
+}
+
+__m512 test_mm512_mask_permute_ps(__m512 __W, __mmask16 __U, __m512 __X) {
+  // CHECK-LABEL: @test_mm512_mask_permute_ps
+  // CHECK: @llvm.x86.avx512.mask.vpermil.ps.512
+  return _mm512_mask_permute_ps(__W, __U, __X, 4); 
+}
+
+__m512 test_mm512_maskz_permute_ps(__mmask16 __U, __m512 __X) {
+  // CHECK-LABEL: @test_mm512_maskz_permute_ps
+  // CHECK: @llvm.x86.avx512.mask.vpermil.ps.512
+  return _mm512_maskz_permute_ps(__U, __X, 4); 
+}
+
+__m512d test_mm512_permutevar_pd(__m512d __A, __m512i __C) {
+  // CHECK-LABEL: @test_mm512_permutevar_pd
+  // CHECK: @llvm.x86.avx512.mask.vpermilvar.pd.512
+  return _mm512_permutevar_pd(__A, __C); 
+}
+
+__m512d test_mm512_mask_permutevar_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512i __C) {
+  // CHECK-LABEL: @test_mm512_mask_permutevar_pd
+  // CHECK: @llvm.x86.avx512.mask.vpermilvar.pd.512
+  return _mm512_mask_permutevar_pd(__W, __U, __A, __C); 
+}
+
+__m512d test_mm512_maskz_permutevar_pd(__mmask8 __U, __m512d __A, __m512i __C) {
+  // CHECK-LABEL: @test_mm512_maskz_permutevar_pd
+  // CHECK: @llvm.x86.avx512.mask.vpermilvar.pd.512
+  return _mm512_maskz_permutevar_pd(__U, __A, __C); 
+}
+
+__m512 test_mm512_permutevar_ps(__m512 __A, __m512i __C) {
+  // CHECK-LABEL: @test_mm512_permutevar_ps
+  // CHECK: @llvm.x86.avx512.mask.vpermilvar.ps.512
+  return _mm512_permutevar_ps(__A, __C); 
+}
+
+__m512 test_mm512_mask_permutevar_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512i __C) {
+  // CHECK-LABEL: @test_mm512_mask_permutevar_ps
+  // CHECK: @llvm.x86.avx512.mask.vpermilvar.ps.512
+  return _mm512_mask_permutevar_ps(__W, __U, __A, __C); 
+}
+
+__m512 test_mm512_maskz_permutevar_ps(__mmask16 __U, __m512 __A, __m512i __C) {
+  // CHECK-LABEL: @test_mm512_maskz_permutevar_ps
+  // CHECK: @llvm.x86.avx512.mask.vpermilvar.ps.512
+  return _mm512_maskz_permutevar_ps(__U, __A, __C); 
+}
+
+__m512i test_mm512_maskz_permutex2var_epi32(__mmask16 __U, __m512i __A, __m512i __I, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_maskz_permutex2var_epi32
+  // CHECK: @llvm.x86.avx512.maskz.vpermt2var.d.512
+  return _mm512_maskz_permutex2var_epi32(__U, __A, __I, __B); 
+}
+
+__m512d test_mm512_maskz_permutex2var_pd(__mmask8 __U, __m512d __A, __m512i __I, __m512d __B) {
+  // CHECK-LABEL: @test_mm512_maskz_permutex2var_pd
+  // CHECK: @llvm.x86.avx512.maskz.vpermt2var.pd.512
+  return _mm512_maskz_permutex2var_pd(__U, __A, __I, __B); 
+}
+
+__m512 test_mm512_maskz_permutex2var_ps(__mmask16 __U, __m512 __A, __m512i __I, __m512 __B) {
+  // CHECK-LABEL: @test_mm512_maskz_permutex2var_ps
+  // CHECK: @llvm.x86.avx512.maskz.vpermt2var.ps.512
+  return _mm512_maskz_permutex2var_ps(__U, __A, __I, __B); 
+}
+
+__m512i test_mm512_maskz_permutex2var_epi64(__mmask8 __U, __m512i __A, __m512i __I, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_maskz_permutex2var_epi64
+  // CHECK: @llvm.x86.avx512.maskz.vpermt2var.q.512
+  return _mm512_maskz_permutex2var_epi64(__U, __A, __I, __B); 
+}

test/CodeGen/avx512vl-builtins.c

 __m256i test_mm256_maskz_min_epu64(__mmask8 __M, __m256i __A, __m256i __B) {
   // CHECK-LABEL: @test_mm256_maskz_min_epu64
   // CHECK: @llvm.x86.avx512.mask.pminu.q.256
   return _mm256_maskz_min_epu64(__M,__A,__B); 
 }
 __m128d test_mm_roundscale_pd(__m128d __A) {
   // CHECK-LABEL: @test_mm_roundscale_pd
   // CHECK: @llvm.x86.avx512.mask.rndscale.pd.128
-  return _mm_roundscale_pd(__A,4); 
+  return _mm_roundscale_pd(__A, _MM_FROUND_CUR_DIRECTION); 
 }
 __m128d test_mm_mask_roundscale_pd(__m128d __W, __mmask8 __U, __m128d __A) {
   // CHECK-LABEL: @test_mm_mask_roundscale_pd
   // CHECK: @llvm.x86.avx512.mask.rndscale.pd.128
-  return _mm_mask_roundscale_pd(__W,__U,__A,4); 
+  return _mm_mask_roundscale_pd(__W,__U,__A, _MM_FROUND_CUR_DIRECTION); 
 }
 __m128d test_mm_maskz_roundscale_pd(__mmask8 __U, __m128d __A) {
   // CHECK-LABEL: @test_mm_maskz_roundscale_pd
   // CHECK: @llvm.x86.avx512.mask.rndscale.pd.128
-  return _mm_maskz_roundscale_pd(__U,__A,4); 
+  return _mm_maskz_roundscale_pd(__U,__A, _MM_FROUND_CUR_DIRECTION); 
 }
 __m256d test_mm256_roundscale_pd(__m256d __A) {
   // CHECK-LABEL: @test_mm256_roundscale_pd
   // CHECK: @llvm.x86.avx512.mask.rndscale.pd.256
-  return _mm256_roundscale_pd(__A,4); 
+  return _mm256_roundscale_pd(__A, _MM_FROUND_CUR_DIRECTION); 
 }
 __m256d test_mm256_mask_roundscale_pd(__m256d __W, __mmask8 __U, __m256d __A) {
   // CHECK-LABEL: @test_mm256_mask_roundscale_pd
   // CHECK: @llvm.x86.avx512.mask.rndscale.pd.256
-  return _mm256_mask_roundscale_pd(__W,__U,__A,4); 
+  return _mm256_mask_roundscale_pd(__W,__U,__A, _MM_FROUND_CUR_DIRECTION); 
 }
 __m256d test_mm256_maskz_roundscale_pd(__mmask8 __U, __m256d __A) {
   // CHECK-LABEL: @test_mm256_maskz_roundscale_pd
   // CHECK: @llvm.x86.avx512.mask.rndscale.pd.256
-  return _mm256_maskz_roundscale_pd(__U,__A,4); 
+  return _mm256_maskz_roundscale_pd(__U,__A, _MM_FROUND_CUR_DIRECTION); 
 }
 __m128 test_mm_roundscale_ps(__m128 __A) {
   // CHECK-LABEL: @test_mm_roundscale_ps
   // CHECK: @llvm.x86.avx512.mask.rndscale.ps.128
-  return _mm_roundscale_ps(__A,4); 
+  return _mm_roundscale_ps(__A, _MM_FROUND_CUR_DIRECTION); 
 }
 __m128 test_mm_mask_roundscale_ps(__m128 __W, __mmask8 __U, __m128 __A) {
   // CHECK-LABEL: @test_mm_mask_roundscale_ps
   // CHECK: @llvm.x86.avx512.mask.rndscale.ps.128
-  return _mm_mask_roundscale_ps(__W,__U,__A,4); 
+  return _mm_mask_roundscale_ps(__W,__U,__A, _MM_FROUND_CUR_DIRECTION); 
 }
 __m128 test_mm_maskz_roundscale_ps(__mmask8 __U, __m128 __A) {
   // CHECK-LABEL: @test_mm_maskz_roundscale_ps
   // CHECK: @llvm.x86.avx512.mask.rndscale.ps.128
-  return _mm_maskz_roundscale_ps(__U,__A, 4); 
+  return _mm_maskz_roundscale_ps(__U,__A, _MM_FROUND_CUR_DIRECTION); 
 }
 __m256 test_mm256_roundscale_ps(__m256 __A) {
   // CHECK-LABEL: @test_mm256_roundscale_ps
   // CHECK: @llvm.x86.avx512.mask.rndscale.ps.256
-  return _mm256_roundscale_ps(__A,4); 
+  return _mm256_roundscale_ps(__A, _MM_FROUND_CUR_DIRECTION); 
 }
 __m256 test_mm256_mask_roundscale_ps(__m256 __W, __mmask8 __U, __m256 __A) {
   // CHECK-LABEL: @test_mm256_mask_roundscale_ps
   // CHECK: @llvm.x86.avx512.mask.rndscale.ps.256
-  return _mm256_mask_roundscale_ps(__W,__U,__A,4); 
+  return _mm256_mask_roundscale_ps(__W,__U,__A, _MM_FROUND_CUR_DIRECTION); 
 }
 __m256 test_mm256_maskz_roundscale_ps(__mmask8 __U, __m256 __A) {
   // CHECK-LABEL: @test_mm256_maskz_roundscale_ps
   // CHECK: @llvm.x86.avx512.mask.rndscale.ps.256
-  return _mm256_maskz_roundscale_ps(__U,__A,4); 
+  return _mm256_maskz_roundscale_ps(__U,__A, _MM_FROUND_CUR_DIRECTION); 
 }
 __m128d test_mm_scalef_pd(__m128d __A, __m128d __B) {
   // CHECK-LABEL: @test_mm_scalef_pd
   // CHECK: @llvm.x86.avx512.mask.scalef.pd.128
   return _mm_scalef_pd(__A,__B); 
 }
 __m128d test_mm_mask_scalef_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
   // CHECK-LABEL: @test_mm_mask_scalef_pd

 }
 
 __m256i test_mm256_maskz_set1_epi64(__mmask8 __M, long long __A) {
   // CHECK-LABEL: @test_mm256_maskz_set1_epi64
   // CHECK: @llvm.x86.avx512.mask.pbroadcast.q.gpr.256
   return _mm256_maskz_set1_epi64(__M, __A); 
 }
 
+__m128i test_mm_ternarylogic_epi32(__m128i __A, __m128i __B, __m128i __C) {
+  // CHECK-LABEL: @test_mm_ternarylogic_epi32
+  // CHECK: @llvm.x86.avx512.mask.pternlog.d.128
+  return _mm_ternarylogic_epi32(__A, __B, __C, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m128i test_mm_mask_ternarylogic_epi32(__m128i __A, __mmask8 __U, __m128i __B, __m128i __C) {
+  // CHECK-LABEL: @test_mm_mask_ternarylogic_epi32
+  // CHECK: @llvm.x86.avx512.mask.pternlog.d.128
+  return _mm_mask_ternarylogic_epi32(__A, __U, __B, __C, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m128i test_mm_maskz_ternarylogic_epi32(__mmask8 __U, __m128i __A, __m128i __B, __m128i __C) {
+  // CHECK-LABEL: @test_mm_maskz_ternarylogic_epi32
+  // CHECK: @llvm.x86.avx512.maskz.pternlog.d.128
+  return _mm_maskz_ternarylogic_epi32(__U, __A, __B, __C, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256i test_mm256_ternarylogic_epi32(__m256i __A, __m256i __B, __m256i __C) {
+  // CHECK-LABEL: @test_mm256_ternarylogic_epi32
+  // CHECK: @llvm.x86.avx512.mask.pternlog.d.256
+  return _mm256_ternarylogic_epi32(__A, __B, __C, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256i test_mm256_mask_ternarylogic_epi32(__m256i __A, __mmask8 __U, __m256i __B, __m256i __C) {
+  // CHECK-LABEL: @test_mm256_mask_ternarylogic_epi32
+  // CHECK: @llvm.x86.avx512.mask.pternlog.d.256
+  return _mm256_mask_ternarylogic_epi32(__A, __U, __B, __C, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256i test_mm256_maskz_ternarylogic_epi32(__mmask8 __U, __m256i __A, __m256i __B, __m256i __C) {
+  // CHECK-LABEL: @test_mm256_maskz_ternarylogic_epi32
+  // CHECK: @llvm.x86.avx512.maskz.pternlog.d.256
+  return _mm256_maskz_ternarylogic_epi32(__U, __A, __B, __C, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m128i test_mm_ternarylogic_epi64(__m128i __A, __m128i __B, __m128i __C) {
+  // CHECK-LABEL: @test_mm_ternarylogic_epi64
+  // CHECK: @llvm.x86.avx512.mask.pternlog.q.128
+  return _mm_ternarylogic_epi64(__A, __B, __C, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m128i test_mm_mask_ternarylogic_epi64(__m128i __A, __mmask8 __U, __m128i __B, __m128i __C) {
+  // CHECK-LABEL: @test_mm_mask_ternarylogic_epi64
+  // CHECK: @llvm.x86.avx512.mask.pternlog.q.128
+  return _mm_mask_ternarylogic_epi64(__A, __U, __B, __C, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m128i test_mm_maskz_ternarylogic_epi64(__mmask8 __U, __m128i __A, __m128i __B, __m128i __C) {
+  // CHECK-LABEL: @test_mm_maskz_ternarylogic_epi64
+  // CHECK: @llvm.x86.avx512.maskz.pternlog.q.128
+  return _mm_maskz_ternarylogic_epi64(__U, __A, __B, __C, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256i test_mm256_ternarylogic_epi64(__m256i __A, __m256i __B, __m256i __C) {
+  // CHECK-LABEL: @test_mm256_ternarylogic_epi64
+  // CHECK: @llvm.x86.avx512.mask.pternlog.q.256
+  return _mm256_ternarylogic_epi64(__A, __B, __C, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256i test_mm256_mask_ternarylogic_epi64(__m256i __A, __mmask8 __U, __m256i __B, __m256i __C) {
+  // CHECK-LABEL: @test_mm256_mask_ternarylogic_epi64
+  // CHECK: @llvm.x86.avx512.mask.pternlog.q.256
+  return _mm256_mask_ternarylogic_epi64(__A, __U, __B, __C, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256i test_mm256_maskz_ternarylogic_epi64(__mmask8 __U, __m256i __A, __m256i __B, __m256i __C) {
+  // CHECK-LABEL: @test_mm256_maskz_ternarylogic_epi64
+  // CHECK: @llvm.x86.avx512.maskz.pternlog.q.256
+  return _mm256_maskz_ternarylogic_epi64(__U, __A, __B, __C, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__mmask8 test_mm_test_epi32_mask(__m128i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm_test_epi32_mask
+  // CHECK: @llvm.x86.avx512.ptestm.d.128
+  return _mm_test_epi32_mask(__A, __B); 
+}
+
+__mmask8 test_mm_mask_test_epi32_mask(__mmask8 __U, __m128i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm_mask_test_epi32_mask
+  // CHECK: @llvm.x86.avx512.ptestm.d.128
+  return _mm_mask_test_epi32_mask(__U, __A, __B); 
+}
+
+__mmask8 test_mm256_test_epi32_mask(__m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_test_epi32_mask
+  // CHECK: @llvm.x86.avx512.ptestm.d.256
+  return _mm256_test_epi32_mask(__A, __B); 
+}
+
+__mmask8 test_mm256_mask_test_epi32_mask(__mmask8 __U, __m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_mask_test_epi32_mask
+  // CHECK: @llvm.x86.avx512.ptestm.d.256
+  return _mm256_mask_test_epi32_mask(__U, __A, __B); 
+}
+
+__mmask8 test_mm_test_epi64_mask(__m128i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm_test_epi64_mask
+  // CHECK: @llvm.x86.avx512.ptestm.q.128
+  return _mm_test_epi64_mask(__A, __B); 
+}
+
+__mmask8 test_mm_mask_test_epi64_mask(__mmask8 __U, __m128i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm_mask_test_epi64_mask
+  // CHECK: @llvm.x86.avx512.ptestm.q.128
+  return _mm_mask_test_epi64_mask(__U, __A, __B); 
+}
+
+__mmask8 test_mm256_test_epi64_mask(__m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_test_epi64_mask
+  // CHECK: @llvm.x86.avx512.ptestm.q.256
+  return _mm256_test_epi64_mask(__A, __B); 
+}
+
+__mmask8 test_mm256_mask_test_epi64_mask(__mmask8 __U, __m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_mask_test_epi64_mask
+  // CHECK: @llvm.x86.avx512.ptestm.q.256
+  return _mm256_mask_test_epi64_mask(__U, __A, __B); 
+}
+
+__mmask8 test_mm_testn_epi32_mask(__m128i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm_testn_epi32_mask
+  // CHECK: @llvm.x86.avx512.ptestnm.d.128
+  return _mm_testn_epi32_mask(__A, __B); 
+}
+
+__mmask8 test_mm_mask_testn_epi32_mask(__mmask8 __U, __m128i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm_mask_testn_epi32_mask
+  // CHECK: @llvm.x86.avx512.ptestnm.d.128
+  return _mm_mask_testn_epi32_mask(__U, __A, __B); 
+}
+
+__mmask8 test_mm256_testn_epi32_mask(__m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_testn_epi32_mask
+  // CHECK: @llvm.x86.avx512.ptestnm.d.256
+  return _mm256_testn_epi32_mask(__A, __B); 
+}
+
+__mmask8 test_mm256_mask_testn_epi32_mask(__mmask8 __U, __m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_mask_testn_epi32_mask
+  // CHECK: @llvm.x86.avx512.ptestnm.d.256
+  return _mm256_mask_testn_epi32_mask(__U, __A, __B); 
+}
+
+__mmask8 test_mm_testn_epi64_mask(__m128i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm_testn_epi64_mask
+  // CHECK: @llvm.x86.avx512.ptestnm.q.128
+  return _mm_testn_epi64_mask(__A, __B); 
+}
+
+__mmask8 test_mm_mask_testn_epi64_mask(__mmask8 __U, __m128i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm_mask_testn_epi64_mask
+  // CHECK: @llvm.x86.avx512.ptestnm.q.128
+  return _mm_mask_testn_epi64_mask(__U, __A, __B); 
+}
+
+__mmask8 test_mm256_testn_epi64_mask(__m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_testn_epi64_mask
+  // CHECK: @llvm.x86.avx512.ptestnm.q.256
+  return _mm256_testn_epi64_mask(__A, __B); 
+}
+
+__mmask8 test_mm256_mask_testn_epi64_mask(__mmask8 __U, __m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_mask_testn_epi64_mask
+  // CHECK: @llvm.x86.avx512.ptestnm.q.256
+  return _mm256_mask_testn_epi64_mask(__U, __A, __B); 
+}
+
+__m128i test_mm_mask_unpackhi_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm_mask_unpackhi_epi32
+  // CHECK: @llvm.x86.avx512.mask.punpckhd.q.128
+  return _mm_mask_unpackhi_epi32(__W, __U, __A, __B); 
+}
+
+__m128i test_mm_maskz_unpackhi_epi32(__mmask8 __U, __m128i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm_maskz_unpackhi_epi32
+  // CHECK: @llvm.x86.avx512.mask.punpckhd.q.128
+  return _mm_maskz_unpackhi_epi32(__U, __A, __B); 
+}
+
+__m256i test_mm256_mask_unpackhi_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_mask_unpackhi_epi32
+  // CHECK: @llvm.x86.avx512.mask.punpckhd.q.256
+  return _mm256_mask_unpackhi_epi32(__W, __U, __A, __B); 
+}
+
+__m256i test_mm256_maskz_unpackhi_epi32(__mmask8 __U, __m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_maskz_unpackhi_epi32
+  // CHECK: @llvm.x86.avx512.mask.punpckhd.q.256
+  return _mm256_maskz_unpackhi_epi32(__U, __A, __B); 
+}
+
+__m128i test_mm_mask_unpackhi_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm_mask_unpackhi_epi64
+  // CHECK: @llvm.x86.avx512.mask.punpckhqd.q.128
+  return _mm_mask_unpackhi_epi64(__W, __U, __A, __B); 
+}
+
+__m128i test_mm_maskz_unpackhi_epi64(__mmask8 __U, __m128i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm_maskz_unpackhi_epi64
+  // CHECK: @llvm.x86.avx512.mask.punpckhqd.q.128
+  return _mm_maskz_unpackhi_epi64(__U, __A, __B); 
+}
+
+__m256i test_mm256_mask_unpackhi_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_mask_unpackhi_epi64
+  // CHECK: @llvm.x86.avx512.mask.punpckhqd.q.256
+  return _mm256_mask_unpackhi_epi64(__W, __U, __A, __B); 
+}
+
+__m256i test_mm256_maskz_unpackhi_epi64(__mmask8 __U, __m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_maskz_unpackhi_epi64
+  // CHECK: @llvm.x86.avx512.mask.punpckhqd.q.256
+  return _mm256_maskz_unpackhi_epi64(__U, __A, __B); 
+}
+
+__m128i test_mm_mask_unpacklo_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm_mask_unpacklo_epi32
+  // CHECK: @llvm.x86.avx512.mask.punpckld.q.128
+  return _mm_mask_unpacklo_epi32(__W, __U, __A, __B); 
+}
+
+__m128i test_mm_maskz_unpacklo_epi32(__mmask8 __U, __m128i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm_maskz_unpacklo_epi32
+  // CHECK: @llvm.x86.avx512.mask.punpckld.q.128
+  return _mm_maskz_unpacklo_epi32(__U, __A, __B); 
+}
+
+__m256i test_mm256_mask_unpacklo_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_mask_unpacklo_epi32
+  // CHECK: @llvm.x86.avx512.mask.punpckld.q.256
+  return _mm256_mask_unpacklo_epi32(__W, __U, __A, __B); 
+}
+
+__m256i test_mm256_maskz_unpacklo_epi32(__mmask8 __U, __m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_maskz_unpacklo_epi32
+  // CHECK: @llvm.x86.avx512.mask.punpckld.q.256
+  return _mm256_maskz_unpacklo_epi32(__U, __A, __B); 
+}
+
+__m128i test_mm_mask_unpacklo_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm_mask_unpacklo_epi64
+  // CHECK: @llvm.x86.avx512.mask.punpcklqd.q.128
+  return _mm_mask_unpacklo_epi64(__W, __U, __A, __B); 
+}
+
+__m128i test_mm_maskz_unpacklo_epi64(__mmask8 __U, __m128i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm_maskz_unpacklo_epi64
+  // CHECK: @llvm.x86.avx512.mask.punpcklqd.q.128
+  return _mm_maskz_unpacklo_epi64(__U, __A, __B); 
+}
+
+__m256i test_mm256_mask_unpacklo_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_mask_unpacklo_epi64
+  // CHECK: @llvm.x86.avx512.mask.punpcklqd.q.256
+  return _mm256_mask_unpacklo_epi64(__W, __U, __A, __B); 
+}
+
+__m256i test_mm256_maskz_unpacklo_epi64(__mmask8 __U, __m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_maskz_unpacklo_epi64
+  // CHECK: @llvm.x86.avx512.mask.punpcklqd.q.256
+  return _mm256_maskz_unpacklo_epi64(__U, __A, __B); 
+}
+
+__m128d test_mm_rcp14_pd(__m128d __A) {
+  // CHECK-LABEL: @test_mm_rcp14_pd
+  // CHECK: @llvm.x86.avx512.rcp14.pd.128
+  return _mm_rcp14_pd(__A); 
+}
+
+__m128d test_mm_mask_rcp14_pd(__m128d __W, __mmask8 __U, __m128d __A) {
+  // CHECK-LABEL: @test_mm_mask_rcp14_pd
+  // CHECK: @llvm.x86.avx512.rcp14.pd.128
+  return _mm_mask_rcp14_pd(__W, __U, __A); 
+}
+
+__m128d test_mm_maskz_rcp14_pd(__mmask8 __U, __m128d __A) {
+  // CHECK-LABEL: @test_mm_maskz_rcp14_pd
+  // CHECK: @llvm.x86.avx512.rcp14.pd.128
+  return _mm_maskz_rcp14_pd(__U, __A); 
+}
+
+__m256d test_mm256_rcp14_pd(__m256d __A) {
+  // CHECK-LABEL: @test_mm256_rcp14_pd
+  // CHECK: @llvm.x86.avx512.rcp14.pd.256
+  return _mm256_rcp14_pd(__A); 
+}
+
+__m256d test_mm256_mask_rcp14_pd(__m256d __W, __mmask8 __U, __m256d __A) {
+  // CHECK-LABEL: @test_mm256_mask_rcp14_pd
+  // CHECK: @llvm.x86.avx512.rcp14.pd.256
+  return _mm256_mask_rcp14_pd(__W, __U, __A); 
+}
+
+__m256d test_mm256_maskz_rcp14_pd(__mmask8 __U, __m256d __A) {
+  // CHECK-LABEL: @test_mm256_maskz_rcp14_pd
+  // CHECK: @llvm.x86.avx512.rcp14.pd.256
+  return _mm256_maskz_rcp14_pd(__U, __A); 
+}
+
+__m128 test_mm_rcp14_ps(__m128 __A) {
+  // CHECK-LABEL: @test_mm_rcp14_ps
+  // CHECK: @llvm.x86.avx512.rcp14.ps.128
+  return _mm_rcp14_ps(__A); 
+}
+
+__m128 test_mm_mask_rcp14_ps(__m128 __W, __mmask8 __U, __m128 __A) {
+  // CHECK-LABEL: @test_mm_mask_rcp14_ps
+  // CHECK: @llvm.x86.avx512.rcp14.ps.128
+  return _mm_mask_rcp14_ps(__W, __U, __A); 
+}
+
+__m128 test_mm_maskz_rcp14_ps(__mmask8 __U, __m128 __A) {
+  // CHECK-LABEL: @test_mm_maskz_rcp14_ps
+  // CHECK: @llvm.x86.avx512.rcp14.ps.128
+  return _mm_maskz_rcp14_ps(__U, __A); 
+}
+
+__m256 test_mm256_rcp14_ps(__m256 __A) {
+  // CHECK-LABEL: @test_mm256_rcp14_ps
+  // CHECK: @llvm.x86.avx512.rcp14.ps.256
+  return _mm256_rcp14_ps(__A); 
+}
+
+__m256 test_mm256_mask_rcp14_ps(__m256 __W, __mmask8 __U, __m256 __A) {
+  // CHECK-LABEL: @test_mm256_mask_rcp14_ps
+  // CHECK: @llvm.x86.avx512.rcp14.ps.256
+  return _mm256_mask_rcp14_ps(__W, __U, __A); 
+}
+
+__m256 test_mm256_maskz_rcp14_ps(__mmask8 __U, __m256 __A) {
+  // CHECK-LABEL: @test_mm256_maskz_rcp14_ps
+  // CHECK: @llvm.x86.avx512.rcp14.ps.256
+  return _mm256_maskz_rcp14_ps(__U, __A); 
+}
+
+__m128d test_mm_rsqrt14_pd(__m128d __A) {
+  // CHECK-LABEL: @test_mm_rsqrt14_pd
+  // CHECK: @llvm.x86.avx512.rsqrt14.pd.128
+  return _mm_rsqrt14_pd(__A); 
+}
+
+__m128d test_mm_mask_rsqrt14_pd(__m128d __W, __mmask8 __U, __m128d __A) {
+  // CHECK-LABEL: @test_mm_mask_rsqrt14_pd
+  // CHECK: @llvm.x86.avx512.rsqrt14.pd.128
+  return _mm_mask_rsqrt14_pd(__W, __U, __A); 
+}
+
+__m128d test_mm_maskz_rsqrt14_pd(__mmask8 __U, __m128d __A) {
+  // CHECK-LABEL: @test_mm_maskz_rsqrt14_pd
+  // CHECK: @llvm.x86.avx512.rsqrt14.pd.128
+  return _mm_maskz_rsqrt14_pd(__U, __A); 
+}
+
+__m256d test_mm256_rsqrt14_pd(__m256d __A) {
+  // CHECK-LABEL: @test_mm256_rsqrt14_pd
+  // CHECK: @llvm.x86.avx512.rsqrt14.pd.256
+  return _mm256_rsqrt14_pd(__A); 
+}
+
+__m256d test_mm256_mask_rsqrt14_pd(__m256d __W, __mmask8 __U, __m256d __A) {
+  // CHECK-LABEL: @test_mm256_mask_rsqrt14_pd
+  // CHECK: @llvm.x86.avx512.rsqrt14.pd.256
+  return _mm256_mask_rsqrt14_pd(__W, __U, __A); 
+}
+
+__m256d test_mm256_maskz_rsqrt14_pd(__mmask8 __U, __m256d __A) {
+  // CHECK-LABEL: @test_mm256_maskz_rsqrt14_pd
+  // CHECK: @llvm.x86.avx512.rsqrt14.pd.256
+  return _mm256_maskz_rsqrt14_pd(__U, __A); 
+}
+
+__m128 test_mm_rsqrt14_ps(__m128 __A) {
+  // CHECK-LABEL: @test_mm_rsqrt14_ps
+  // CHECK: @llvm.x86.avx512.rsqrt14.ps.128
+  return _mm_rsqrt14_ps(__A); 
+}
+
+__m128 test_mm_mask_rsqrt14_ps(__m128 __W, __mmask8 __U, __m128 __A) {
+  // CHECK-LABEL: @test_mm_mask_rsqrt14_ps
+  // CHECK: @llvm.x86.avx512.rsqrt14.ps.128
+  return _mm_mask_rsqrt14_ps(__W, __U, __A); 
+}
+
+__m128 test_mm_maskz_rsqrt14_ps(__mmask8 __U, __m128 __A) {
+  // CHECK-LABEL: @test_mm_maskz_rsqrt14_ps
+  // CHECK: @llvm.x86.avx512.rsqrt14.ps.128
+  return _mm_maskz_rsqrt14_ps(__U, __A); 
+}
+
+__m256 test_mm256_rsqrt14_ps(__m256 __A) {
+  // CHECK-LABEL: @test_mm256_rsqrt14_ps
+  // CHECK: @llvm.x86.avx512.rsqrt14.ps.256
+  return _mm256_rsqrt14_ps(__A); 
+}
+
+__m256 test_mm256_mask_rsqrt14_ps(__m256 __W, __mmask8 __U, __m256 __A) {
+  // CHECK-LABEL: @test_mm256_mask_rsqrt14_ps
+  // CHECK: @llvm.x86.avx512.rsqrt14.ps.256
+  return _mm256_mask_rsqrt14_ps(__W, __U, __A); 
+}
+
+__m256 test_mm256_maskz_rsqrt14_ps(__mmask8 __U, __m256 __A) {
+  // CHECK-LABEL: @test_mm256_maskz_rsqrt14_ps
+  // CHECK: @llvm.x86.avx512.rsqrt14.ps.256
+  return _mm256_maskz_rsqrt14_ps(__U, __A); 
+}
+
+__m256 test_mm256_shuffle_f32x4(__m256 __A, __m256 __B) {
+  // CHECK-LABEL: @test_mm256_shuffle_f32x4
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm256_shuffle_f32x4(__A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256 test_mm256_mask_shuffle_f32x4(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) {
+  // CHECK-LABEL: @test_mm256_mask_shuffle_f32x4
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm256_mask_shuffle_f32x4(__W, __U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256 test_mm256_maskz_shuffle_f32x4(__mmask8 __U, __m256 __A, __m256 __B) {
+  // CHECK-LABEL: @test_mm256_maskz_shuffle_f32x4
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm256_maskz_shuffle_f32x4(__U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256d test_mm256_shuffle_f64x2(__m256d __A, __m256d __B) {
+  // CHECK-LABEL: @test_mm256_shuffle_f64x2
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm256_shuffle_f64x2(__A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256d test_mm256_mask_shuffle_f64x2(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) {
+  // CHECK-LABEL: @test_mm256_mask_shuffle_f64x2
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm256_mask_shuffle_f64x2(__W, __U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256d test_mm256_maskz_shuffle_f64x2(__mmask8 __U, __m256d __A, __m256d __B) {
+  // CHECK-LABEL: @test_mm256_maskz_shuffle_f64x2
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm256_maskz_shuffle_f64x2(__U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256i test_mm256_shuffle_i32x4(__m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_shuffle_i32x4
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm256_shuffle_i32x4(__A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256i test_mm256_mask_shuffle_i32x4(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_mask_shuffle_i32x4
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm256_mask_shuffle_i32x4(__W, __U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256i test_mm256_maskz_shuffle_i32x4(__mmask8 __U, __m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_maskz_shuffle_i32x4
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm256_maskz_shuffle_i32x4(__U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256i test_mm256_shuffle_i64x2(__m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_shuffle_i64x2
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm256_shuffle_i64x2(__A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256i test_mm256_mask_shuffle_i64x2(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_mask_shuffle_i64x2
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm256_mask_shuffle_i64x2(__W, __U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256i test_mm256_maskz_shuffle_i64x2(__mmask8 __U, __m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_maskz_shuffle_i64x2
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm256_maskz_shuffle_i64x2(__U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m128d test_mm_mask_shuffle_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_mask_shuffle_pd
+  // CHECK: @llvm.x86.avx512.mask.shuf.pd.128
+  return _mm_mask_shuffle_pd(__W, __U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m128d test_mm_maskz_shuffle_pd(__mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_maskz_shuffle_pd
+  // CHECK: @llvm.x86.avx512.mask.shuf.pd.128
+  return _mm_maskz_shuffle_pd(__U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256d test_mm256_mask_shuffle_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) {
+  // CHECK-LABEL: @test_mm256_mask_shuffle_pd
+  // CHECK: @llvm.x86.avx512.mask.shuf.pd.256
+  return _mm256_mask_shuffle_pd(__W, __U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256d test_mm256_maskz_shuffle_pd(__mmask8 __U, __m256d __A, __m256d __B) {
+  // CHECK-LABEL: @test_mm256_maskz_shuffle_pd
+  // CHECK: @llvm.x86.avx512.mask.shuf.pd.256
+  return _mm256_maskz_shuffle_pd(__U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m128 test_mm_mask_shuffle_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_mask_shuffle_ps
+  // CHECK: @llvm.x86.avx512.mask.shuf.ps.128
+  return _mm_mask_shuffle_ps(__W, __U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m128 test_mm_maskz_shuffle_ps(__mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_maskz_shuffle_ps
+  // CHECK: @llvm.x86.avx512.mask.shuf.ps.128
+  return _mm_maskz_shuffle_ps(__U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256 test_mm256_mask_shuffle_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) {
+  // CHECK-LABEL: @test_mm256_mask_shuffle_ps
+  // CHECK: @llvm.x86.avx512.mask.shuf.ps.256
+  return _mm256_mask_shuffle_ps(__W, __U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256 test_mm256_maskz_shuffle_ps(__mmask8 __U, __m256 __A, __m256 __B) {
+  // CHECK-LABEL: @test_mm256_maskz_shuffle_ps
+  // CHECK: @llvm.x86.avx512.mask.shuf.ps.256
+  return _mm256_maskz_shuffle_ps(__U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+void test_mm_mask_store_pd(void *__P, __mmask8 __U, __m128d __A) {
+  // CHECK-LABEL: @test_mm_mask_store_pd
+  // CHECK: @llvm.x86.avx512.mask.store.pd.128
+  return _mm_mask_store_pd(__P, __U, __A); 
+}
+
+void test_mm256_mask_store_pd(void *__P, __mmask8 __U, __m256d __A) {
+  // CHECK-LABEL: @test_mm256_mask_store_pd
+  // CHECK: @llvm.x86.avx512.mask.store.pd.256
+  return _mm256_mask_store_pd(__P, __U, __A); 
+}
+
+void test_mm_mask_store_ps(void *__P, __mmask8 __U, __m128 __A) {
+  // CHECK-LABEL: @test_mm_mask_store_ps
+  // CHECK: @llvm.x86.avx512.mask.store.ps.128
+  return _mm_mask_store_ps(__P, __U, __A); 
+}
+
+void test_mm256_mask_store_ps(void *__P, __mmask8 __U, __m256 __A) {
+  // CHECK-LABEL: @test_mm256_mask_store_ps
+  // CHECK: @llvm.x86.avx512.mask.store.ps.256
+  return _mm256_mask_store_ps(__P, __U, __A); 
+}
+
+void test_mm_mask_storeu_epi64(void *__P, __mmask8 __U, __m128i __A) {
+  // CHECK-LABEL: @test_mm_mask_storeu_epi64
+  // CHECK: @llvm.x86.avx512.mask.storeu
+  return _mm_mask_storeu_epi64(__P, __U, __A); 
+}
+
+void test_mm256_mask_storeu_epi64(void *__P, __mmask8 __U, __m256i __A) {
+  // CHECK-LABEL: @test_mm256_mask_storeu_epi64
+  // CHECK: @llvm.x86.avx512.mask.storeu
+  return _mm256_mask_storeu_epi64(__P, __U, __A); 
+}
+
+void test_mm_mask_storeu_epi32(void *__P, __mmask8 __U, __m128i __A) {
+  // CHECK-LABEL: @test_mm_mask_storeu_epi32
+  // CHECK: @llvm.x86.avx512.mask.storeu
+  return _mm_mask_storeu_epi32(__P, __U, __A); 
+}
+
+void test_mm256_mask_storeu_epi32(void *__P, __mmask8 __U, __m256i __A) {
+  // CHECK-LABEL: @test_mm256_mask_storeu_epi32
+  // CHECK: @llvm.x86.avx512.mask.storeu
+  return _mm256_mask_storeu_epi32(__P, __U, __A); 
+}
+
+void test_mm_mask_storeu_pd(void *__P, __mmask8 __U, __m128d __A) {
+  // CHECK-LABEL: @test_mm_mask_storeu_pd
+  // CHECK: @llvm.x86.avx512.mask.storeu.pd.128
+  return _mm_mask_storeu_pd(__P, __U, __A); 
+}
+
+void test_mm256_mask_storeu_pd(void *__P, __mmask8 __U, __m256d __A) {
+  // CHECK-LABEL: @test_mm256_mask_storeu_pd
+  // CHECK: @llvm.x86.avx512.mask.storeu.pd.256
+  return _mm256_mask_storeu_pd(__P, __U, __A); 
+}
+
+void test_mm_mask_storeu_ps(void *__P, __mmask8 __U, __m128 __A) {
+  // CHECK-LABEL: @test_mm_mask_storeu_ps
+  // CHECK: @llvm.x86.avx512.mask.storeu.ps.128
+  return _mm_mask_storeu_ps(__P, __U, __A); 
+}
+
+void test_mm256_mask_storeu_ps(void *__P, __mmask8 __U, __m256 __A) {
+  // CHECK-LABEL: @test_mm256_mask_storeu_ps
+  // CHECK: @llvm.x86.avx512.mask.storeu.ps.256
+  return _mm256_mask_storeu_ps(__P, __U, __A); 
+}
+
+__m128d test_mm_mask_unpackhi_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_mask_unpackhi_pd
+  // CHECK: @llvm.x86.avx512.mask.unpckh.pd.128
+  return _mm_mask_unpackhi_pd(__W, __U, __A, __B); 
+}
+
+__m128d test_mm_maskz_unpackhi_pd(__mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_maskz_unpackhi_pd
+  // CHECK: @llvm.x86.avx512.mask.unpckh.pd.128
+  return _mm_maskz_unpackhi_pd(__U, __A, __B); 
+}
+
+__m256d test_mm256_mask_unpackhi_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) {
+  // CHECK-LABEL: @test_mm256_mask_unpackhi_pd
+  // CHECK: @llvm.x86.avx512.mask.unpckh.pd.256
+  return _mm256_mask_unpackhi_pd(__W, __U, __A, __B); 
+}
+
+__m256d test_mm256_maskz_unpackhi_pd(__mmask8 __U, __m256d __A, __m256d __B) {
+  // CHECK-LABEL: @test_mm256_maskz_unpackhi_pd
+  // CHECK: @llvm.x86.avx512.mask.unpckh.pd.256
+  return _mm256_maskz_unpackhi_pd(__U, __A, __B); 
+}
+
+__m128 test_mm_mask_unpackhi_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_mask_unpackhi_ps
+  // CHECK: @llvm.x86.avx512.mask.unpckh.ps.128
+  return _mm_mask_unpackhi_ps(__W, __U, __A, __B); 
+}
+
+__m128 test_mm_maskz_unpackhi_ps(__mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_maskz_unpackhi_ps
+  // CHECK: @llvm.x86.avx512.mask.unpckh.ps.128
+  return _mm_maskz_unpackhi_ps(__U, __A, __B); 
+}
+
+__m256 test_mm256_mask_unpackhi_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) {
+  // CHECK-LABEL: @test_mm256_mask_unpackhi_ps
+  // CHECK: @llvm.x86.avx512.mask.unpckh.ps.256
+  return _mm256_mask_unpackhi_ps(__W, __U, __A, __B); 
+}
+
+__m256 test_mm256_maskz_unpackhi_ps(__mmask8 __U, __m256 __A, __m256 __B) {
+  // CHECK-LABEL: @test_mm256_maskz_unpackhi_ps
+  // CHECK: @llvm.x86.avx512.mask.unpckh.ps.256
+  return _mm256_maskz_unpackhi_ps(__U, __A, __B); 
+}
+
+__m128d test_mm_mask_unpacklo_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_mask_unpacklo_pd
+  // CHECK: @llvm.x86.avx512.mask.unpckl.pd.128
+  return _mm_mask_unpacklo_pd(__W, __U, __A, __B); 
+}
+
+__m128d test_mm_maskz_unpacklo_pd(__mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_maskz_unpacklo_pd
+  // CHECK: @llvm.x86.avx512.mask.unpckl.pd.128
+  return _mm_maskz_unpacklo_pd(__U, __A, __B); 
+}
+
+__m256d test_mm256_mask_unpacklo_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) {
+  // CHECK-LABEL: @test_mm256_mask_unpacklo_pd
+  // CHECK: @llvm.x86.avx512.mask.unpckl.pd.256
+  return _mm256_mask_unpacklo_pd(__W, __U, __A, __B); 
+}
+
+__m256d test_mm256_maskz_unpacklo_pd(__mmask8 __U, __m256d __A, __m256d __B) {
+  // CHECK-LABEL: @test_mm256_maskz_unpacklo_pd
+  // CHECK: @llvm.x86.avx512.mask.unpckl.pd.256
+  return _mm256_maskz_unpacklo_pd(__U, __A, __B); 
+}
+
+__m128 test_mm_mask_unpacklo_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_mask_unpacklo_ps
+  // CHECK: @llvm.x86.avx512.mask.unpckl.ps.128
+  return _mm_mask_unpacklo_ps(__W, __U, __A, __B); 
+}
+
+__m128 test_mm_maskz_unpacklo_ps(__mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_maskz_unpacklo_ps
+  // CHECK: @llvm.x86.avx512.mask.unpckl.ps.128
+  return _mm_maskz_unpacklo_ps(__U, __A, __B); 
+}
+
+__m256 test_mm256_mask_unpacklo_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) {
+  // CHECK-LABEL: @test_mm256_mask_unpacklo_ps
+  // CHECK: @llvm.x86.avx512.mask.unpckl.ps.256
+  return _mm256_mask_unpacklo_ps(__W, __U, __A, __B); 
+}
+
+__m256 test_mm256_maskz_unpacklo_ps(__mmask8 __U, __m256 __A, __m256 __B) {
+  // CHECK-LABEL: @test_mm256_maskz_unpacklo_ps
+  // CHECK: @llvm.x86.avx512.mask.unpckl.ps.256
+  return _mm256_maskz_unpacklo_ps(__U, __A, __B); 
+}
+
+__m128d test_mm_mask_permute_pd(__m128d __W, __mmask8 __U, __m128d __X) {
+  // CHECK-LABEL: @test_mm_mask_permute_pd
+  // CHECK: @llvm.x86.avx512.mask.vpermil.pd
+  return _mm_mask_permute_pd(__W, __U, __X, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m128d test_mm_maskz_permute_pd(__mmask8 __U, __m128d __X) {
+  // CHECK-LABEL: @test_mm_maskz_permute_pd
+  // CHECK: @llvm.x86.avx512.mask.vpermil.pd
+  return _mm_maskz_permute_pd(__U, __X, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256d test_mm256_mask_permute_pd(__m256d __W, __mmask8 __U, __m256d __X) {
+  // CHECK-LABEL: @test_mm256_mask_permute_pd
+  // CHECK: @llvm.x86.avx512.mask.vpermil.pd.256
+  return _mm256_mask_permute_pd(__W, __U, __X, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256d test_mm256_maskz_permute_pd(__mmask8 __U, __m256d __X) {
+  // CHECK-LABEL: @test_mm256_maskz_permute_pd
+  // CHECK: @llvm.x86.avx512.mask.vpermil.pd.256
+  return _mm256_maskz_permute_pd(__U, __X, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m128 test_mm_mask_permute_ps(__m128 __W, __mmask8 __U, __m128 __X) {
+  // CHECK-LABEL: @test_mm_mask_permute_ps
+  // CHECK: @llvm.x86.avx512.mask.vpermil.ps
+  return _mm_mask_permute_ps(__W, __U, __X, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m128 test_mm_maskz_permute_ps(__mmask8 __U, __m128 __X) {
+  // CHECK-LABEL: @test_mm_maskz_permute_ps
+  // CHECK: @llvm.x86.avx512.mask.vpermil.ps
+  return _mm_maskz_permute_ps(__U, __X, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256 test_mm256_mask_permute_ps(__m256 __W, __mmask8 __U, __m256 __X) {
+  // CHECK-LABEL: @test_mm256_mask_permute_ps
+  // CHECK: @llvm.x86.avx512.mask.vpermil.ps.256
+  return _mm256_mask_permute_ps(__W, __U, __X, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256 test_mm256_maskz_permute_ps(__mmask8 __U, __m256 __X) {
+  // CHECK-LABEL: @test_mm256_maskz_permute_ps
+  // CHECK: @llvm.x86.avx512.mask.vpermil.ps.256
+  return _mm256_maskz_permute_ps(__U, __X, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m128d test_mm_mask_permutevar_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128i __C) {
+  // CHECK-LABEL: @test_mm_mask_permutevar_pd
+  // CHECK: @llvm.x86.avx512.mask.vpermilvar.pd
+  return _mm_mask_permutevar_pd(__W, __U, __A, __C); 
+}
+
+__m128d test_mm_maskz_permutevar_pd(__mmask8 __U, __m128d __A, __m128i __C) {
+  // CHECK-LABEL: @test_mm_maskz_permutevar_pd
+  // CHECK: @llvm.x86.avx512.mask.vpermilvar.pd
+  return _mm_maskz_permutevar_pd(__U, __A, __C); 
+}
+
+__m256d test_mm256_mask_permutevar_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256i __C) {
+  // CHECK-LABEL: @test_mm256_mask_permutevar_pd
+  // CHECK: @llvm.x86.avx512.mask.vpermilvar.pd.256
+  return _mm256_mask_permutevar_pd(__W, __U, __A, __C); 
+}
+
+__m256d test_mm256_maskz_permutevar_pd(__mmask8 __U, __m256d __A, __m256i __C) {
+  // CHECK-LABEL: @test_mm256_maskz_permutevar_pd
+  // CHECK: @llvm.x86.avx512.mask.vpermilvar.pd.256
+  return _mm256_maskz_permutevar_pd(__U, __A, __C); 
+}
+
+__m128 test_mm_mask_permutevar_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128i __C) {
+  // CHECK-LABEL: @test_mm_mask_permutevar_ps
+  // CHECK: @llvm.x86.avx512.mask.vpermilvar.ps
+  return _mm_mask_permutevar_ps(__W, __U, __A, __C); 
+}
+
+__m128 test_mm_maskz_permutevar_ps(__mmask8 __U, __m128 __A, __m128i __C) {
+  // CHECK-LABEL: @test_mm_maskz_permutevar_ps
+  // CHECK: @llvm.x86.avx512.mask.vpermilvar.ps
+  return _mm_maskz_permutevar_ps(__U, __A, __C); 
+}
+
+__m256 test_mm256_mask_permutevar_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256i __C) {
+  // CHECK-LABEL: @test_mm256_mask_permutevar_ps
+  // CHECK: @llvm.x86.avx512.mask.vpermilvar.ps.256
+  return _mm256_mask_permutevar_ps(__W, __U, __A, __C); 
+}
+
+__m256 test_mm256_maskz_permutevar_ps(__mmask8 __U, __m256 __A, __m256i __C) {
+  // CHECK-LABEL: @test_mm256_maskz_permutevar_ps
+  // CHECK: @llvm.x86.avx512.mask.vpermilvar.ps.256
+  return _mm256_maskz_permutevar_ps(__U, __A, __C); 
+}
+
+__m128i test_mm_lzcnt_epi32(__m128i __A) {
+  // CHECK-LABEL: @test_mm_lzcnt_epi32
+  // CHECK: @llvm.x86.avx512.mask.lzcnt.d
+  return _mm_lzcnt_epi32(__A); 
+}
+
+__m128i test_mm_mask_lzcnt_epi32(__m128i __W, __mmask8 __U, __m128i __A) {
+  // CHECK-LABEL: @test_mm_mask_lzcnt_epi32
+  // CHECK: @llvm.x86.avx512.mask.lzcnt.d
+  return _mm_mask_lzcnt_epi32(__W, __U, __A); 
+}
+
+__m128i test_mm_maskz_lzcnt_epi32(__mmask8 __U, __m128i __A) {
+  // CHECK-LABEL: @test_mm_maskz_lzcnt_epi32
+  // CHECK: @llvm.x86.avx512.mask.lzcnt.d
+  return _mm_maskz_lzcnt_epi32(__U, __A); 
+}
+
+__m256i test_mm256_lzcnt_epi32(__m256i __A) {
+  // CHECK-LABEL: @test_mm256_lzcnt_epi32
+  // CHECK: @llvm.x86.avx512.mask.lzcnt.d
+  return _mm256_lzcnt_epi32(__A); 
+}
+
+__m256i test_mm256_mask_lzcnt_epi32(__m256i __W, __mmask8 __U, __m256i __A) {
+  // CHECK-LABEL: @test_mm256_mask_lzcnt_epi32
+  // CHECK: @llvm.x86.avx512.mask.lzcnt.d
+  return _mm256_mask_lzcnt_epi32(__W, __U, __A); 
+}
+
+__m256i test_mm256_maskz_lzcnt_epi32(__mmask8 __U, __m256i __A) {
+  // CHECK-LABEL: @test_mm256_maskz_lzcnt_epi32
+  // CHECK: @llvm.x86.avx512.mask.lzcnt.d
+  return _mm256_maskz_lzcnt_epi32(__U, __A); 
+}
+
+__m128i test_mm_lzcnt_epi64(__m128i __A) {
+  // CHECK-LABEL: @test_mm_lzcnt_epi64
+  // CHECK: @llvm.x86.avx512.mask.lzcnt.q
+  return _mm_lzcnt_epi64(__A); 
+}
+
+__m128i test_mm_mask_lzcnt_epi64(__m128i __W, __mmask8 __U, __m128i __A) {
+  // CHECK-LABEL: @test_mm_mask_lzcnt_epi64
+  // CHECK: @llvm.x86.avx512.mask.lzcnt.q
+  return _mm_mask_lzcnt_epi64(__W, __U, __A); 
+}
+
+__m128i test_mm_maskz_lzcnt_epi64(__mmask8 __U, __m128i __A) {
+  // CHECK-LABEL: @test_mm_maskz_lzcnt_epi64
+  // CHECK: @llvm.x86.avx512.mask.lzcnt.q
+  return _mm_maskz_lzcnt_epi64(__U, __A); 
+}
+
+__m256i test_mm256_lzcnt_epi64(__m256i __A) {
+  // CHECK-LABEL: @test_mm256_lzcnt_epi64
+  // CHECK: @llvm.x86.avx512.mask.lzcnt.q
+  return _mm256_lzcnt_epi64(__A); 
+}
+
+__m256i test_mm256_mask_lzcnt_epi64(__m256i __W, __mmask8 __U, __m256i __A) {
+  // CHECK-LABEL: @test_mm256_mask_lzcnt_epi64
+  // CHECK: @llvm.x86.avx512.mask.lzcnt.q
+  return _mm256_mask_lzcnt_epi64(__W, __U, __A); 
+}
+
+__m256i test_mm256_maskz_lzcnt_epi64(__mmask8 __U, __m256i __A) {
+  // CHECK-LABEL: @test_mm256_maskz_lzcnt_epi64
+  // CHECK: @llvm.x86.avx512.mask.lzcnt.q
+  return _mm256_maskz_lzcnt_epi64(__U, __A); 
+}
+

test/CodeGen/avx512vlbw-builtins.c

__m256i test_mm256_maskz_min_epu16(__mmask16 __M, __m256i __A, __m256i __B) {
   // CHECK: @llvm.x86.avx512.mask.pminu.w.256
   return _mm256_maskz_min_epu16(__M,__A,__B); 
 }
 __m256i test_mm256_mask_min_epu16(__m256i __W, __mmask16 __M, __m256i __A,           __m256i __B) {
   // CHECK-LABEL: @test_mm256_mask_min_epu16
   // CHECK: @llvm.x86.avx512.mask.pminu.w.256
   return _mm256_mask_min_epu16(__W,__M,__A,__B); 
 }
+
 __m128i test_mm_mask_shuffle_epi8(__m128i __W, __mmask16 __U, __m128i __A,           __m128i __B) {
   // CHECK-LABEL: @test_mm_mask_shuffle_epi8
   // CHECK: @llvm.x86.avx512.mask.pshuf.b.128
   return _mm_mask_shuffle_epi8(__W,__U,__A,__B); 
 }
 __m128i test_mm_maskz_shuffle_epi8(__mmask16 __U, __m128i __A, __m128i __B) {
   // CHECK-LABEL: @test_mm_maskz_shuffle_epi8
   // CHECK: @llvm.x86.avx512.mask.pshuf.b.128

 }
 
 __m256i test_mm256_maskz_mov_epi8(__mmask32 __U, __m256i __A) {
   // CHECK-LABEL: @test_mm256_maskz_mov_epi8
   // CHECK: @llvm.x86.avx512.mask.movu.b.256
   return _mm256_maskz_mov_epi8(__U, __A); 
 }
 
+__mmask16 test_mm_test_epi8_mask(__m128i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm_test_epi8_mask
+  // CHECK: @llvm.x86.avx512.ptestm.b.128
+  return _mm_test_epi8_mask(__A, __B); 
+}
+
+__mmask16 test_mm_mask_test_epi8_mask(__mmask16 __U, __m128i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm_mask_test_epi8_mask
+  // CHECK: @llvm.x86.avx512.ptestm.b.128
+  return _mm_mask_test_epi8_mask(__U, __A, __B); 
+}
+
+__mmask32 test_mm256_test_epi8_mask(__m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_test_epi8_mask
+  // CHECK: @llvm.x86.avx512.ptestm.b.256
+  return _mm256_test_epi8_mask(__A, __B); 
+}
+
+__mmask32 test_mm256_mask_test_epi8_mask(__mmask32 __U, __m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_mask_test_epi8_mask
+  // CHECK: @llvm.x86.avx512.ptestm.b.256
+  return _mm256_mask_test_epi8_mask(__U, __A, __B); 
+}
+
+__mmask8 test_mm_test_epi16_mask(__m128i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm_test_epi16_mask
+  // CHECK: @llvm.x86.avx512.ptestm.w.128
+  return _mm_test_epi16_mask(__A, __B); 
+}
+
+__mmask8 test_mm_mask_test_epi16_mask(__mmask8 __U, __m128i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm_mask_test_epi16_mask
+  // CHECK: @llvm.x86.avx512.ptestm.w.128
+  return _mm_mask_test_epi16_mask(__U, __A, __B); 
+}
+
+__mmask16 test_mm256_test_epi16_mask(__m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_test_epi16_mask
+  // CHECK: @llvm.x86.avx512.ptestm.w.256
+  return _mm256_test_epi16_mask(__A, __B); 
+}
+
+__mmask16 test_mm256_mask_test_epi16_mask(__mmask16 __U, __m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_mask_test_epi16_mask
+  // CHECK: @llvm.x86.avx512.ptestm.w.256
+  return _mm256_mask_test_epi16_mask(__U, __A, __B); 
+}
+
+__mmask16 test_mm_testn_epi8_mask(__m128i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm_testn_epi8_mask
+  // CHECK: @llvm.x86.avx512.ptestnm.b.128
+  return _mm_testn_epi8_mask(__A, __B); 
+}
+
+__mmask16 test_mm_mask_testn_epi8_mask(__mmask16 __U, __m128i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm_mask_testn_epi8_mask
+  // CHECK: @llvm.x86.avx512.ptestnm.b.128
+  return _mm_mask_testn_epi8_mask(__U, __A, __B); 
+}
+
+__mmask32 test_mm256_testn_epi8_mask(__m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_testn_epi8_mask
+  // CHECK: @llvm.x86.avx512.ptestnm.b.256
+  return _mm256_testn_epi8_mask(__A, __B); 
+}
+
+__mmask32 test_mm256_mask_testn_epi8_mask(__mmask32 __U, __m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_mask_testn_epi8_mask
+  // CHECK: @llvm.x86.avx512.ptestnm.b.256
+  return _mm256_mask_testn_epi8_mask(__U, __A, __B); 
+}
+
+__mmask8 test_mm_testn_epi16_mask(__m128i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm_testn_epi16_mask
+  // CHECK: @llvm.x86.avx512.ptestnm.w.128
+  return _mm_testn_epi16_mask(__A, __B); 
+}
+
+__mmask8 test_mm_mask_testn_epi16_mask(__mmask8 __U, __m128i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm_mask_testn_epi16_mask
+  // CHECK: @llvm.x86.avx512.ptestnm.w.128
+  return _mm_mask_testn_epi16_mask(__U, __A, __B); 
+}
+
+__mmask16 test_mm256_testn_epi16_mask(__m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_testn_epi16_mask
+  // CHECK: @llvm.x86.avx512.ptestnm.w.256
+  return _mm256_testn_epi16_mask(__A, __B); 
+}
+
+__mmask16 test_mm256_mask_testn_epi16_mask(__mmask16 __U, __m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_mask_testn_epi16_mask
+  // CHECK: @llvm.x86.avx512.ptestnm.w.256
+  return _mm256_mask_testn_epi16_mask(__U, __A, __B); 
+}
+
+void test_mm_mask_storeu_epi16(void *__P, __mmask8 __U, __m128i __A) {
+  // CHECK-LABEL: @test_mm_mask_storeu_epi16
+  // CHECK: @llvm.x86.avx512.mask.storeu.w.128
+  return _mm_mask_storeu_epi16(__P, __U, __A); 
+}
+
+void test_mm256_mask_storeu_epi16(void *__P, __mmask16 __U, __m256i __A) {
+  // CHECK-LABEL: @test_mm256_mask_storeu_epi16
+  // CHECK: @llvm.x86.avx512.mask.storeu.w.256
+  return _mm256_mask_storeu_epi16(__P, __U, __A); 
+}
+
+void test_mm_mask_storeu_epi8(void *__P, __mmask16 __U, __m128i __A) {
+  // CHECK-LABEL: @test_mm_mask_storeu_epi8
+  // CHECK: @llvm.x86.avx512.mask.storeu.b.128
+  return _mm_mask_storeu_epi8(__P, __U, __A); 
+}
+
+void test_mm256_mask_storeu_epi8(void *__P, __mmask32 __U, __m256i __A) {
+  // CHECK-LABEL: @test_mm256_mask_storeu_epi8
+  // CHECK: @llvm.x86.avx512.mask.storeu.b.256
+  return _mm256_mask_storeu_epi8(__P, __U, __A); 
+}