add arm vector types in union, avoid -flax-conversions, fix castings

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 05e6a5c767b944c1b65901ec063bcb16d3f3e508..92abf6dc7306d4bbed0b970f1e616f823d372e78 100644 (file)
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -277,8 +277,6 @@ elseif (ARCH_ARM32 OR ARCH_AARCH64)
       message(FATAL_ERROR "arm_sve.h is required to build for SVE.")
     endif()
   endif()
-  set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -flax-vector-conversions")
-  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -flax-vector-conversions")
 endif()
 
 CHECK_FUNCTION_EXISTS(posix_memalign HAVE_POSIX_MEMALIGN)

diff --git a/src/util/arch/arm/match.hpp b/src/util/arch/arm/match.hpp
index 46d84d060d9a9b2f64c81235cb96984c291945b8..e7f757bd1a372a872e6e61b0fcf64f1b0de8acaa 100644 (file)
--- a/src/util/arch/arm/match.hpp
+++ b/src/util/arch/arm/match.hpp
@@ -30,8 +30,8 @@
 template <>
 really_really_inline
 const u8 *firstMatch<16>(const u8 *buf, SuperVector<16> mask) {
-    uint32x4_t res_t = vreinterpretq_u32_u8(mask.u.v128[0]);
-    uint64_t vmax = vgetq_lane_u64 (vreinterpretq_u64_u32 (vpmaxq_u32(res_t, res_t)), 0);
+    uint32x4_t m = mask.u.u32x4[0];
+    uint64_t vmax = vgetq_lane_u64 (vreinterpretq_u64_u32 (vpmaxq_u32(m, m)), 0);
     if (vmax != 0) {
        typename SuperVector<16>::movemask_type z = mask.movemask();
         DEBUG_PRINTF("z %08x\n", z);
@@ -49,8 +49,8 @@ const u8 *firstMatch<16>(const u8 *buf, SuperVector<16> mask) {
 template <>
 really_really_inline
 const u8 *lastMatch<16>(const u8 *buf, SuperVector<16> mask) {
-    uint32x4_t res_t = vreinterpretq_u32_u8(mask.u.v128[0]);
-    uint64_t vmax = vgetq_lane_u64 (vreinterpretq_u64_u32 (vpmaxq_u32(res_t, res_t)), 0);
+    uint32x4_t m = mask.u.u32x4[0];
+    uint64_t vmax = vgetq_lane_u64 (vreinterpretq_u64_u32 (vpmaxq_u32(m, m)), 0);
     if (vmax != 0) {
        typename SuperVector<16>::movemask_type z = mask.movemask();
         DEBUG_PRINTF("buf %p z %08x \n", buf, z);

diff --git a/src/util/arch/arm/simd_utils.h b/src/util/arch/arm/simd_utils.h
index 2485177340e39f9c13c2221d6b6bddd3784f255c..630cac932ada086d1bb6625a561f8f3d1278a22b 100644 (file)
--- a/src/util/arch/arm/simd_utils.h
+++ b/src/util/arch/arm/simd_utils.h
@@ -100,7 +100,7 @@ static really_inline int isnonzero128(m128 a) {
  */
 static really_inline u32 diffrich128(m128 a, m128 b) {
     static const uint32x4_t movemask = { 1, 2, 4, 8 };
-    return vaddvq_u32(vandq_u32(vmvnq_s32(vceqq_s32((int32x4_t)a, (int32x4_t)b)), movemask));
+    return vaddvq_u32(vandq_u32(vmvnq_u32(vceqq_u32((uint32x4_t)a, (uint32x4_t)b)), movemask));
 }
 
 /**
@@ -109,53 +109,53 @@ static really_inline u32 diffrich128(m128 a, m128 b) {
  */
 static really_inline u32 diffrich64_128(m128 a, m128 b) {
     static const uint64x2_t movemask = { 1, 4 };
-    return vaddvq_u64(vandq_u64(vmvnq_s32(vceqq_s64((int64x2_t)a, (int64x2_t)b)), movemask));
+    return (u32) vaddvq_u64(vandq_u64((uint64x2_t)vmvnq_u32((uint32x4_t)vceqq_u64((uint64x2_t)a, (uint64x2_t)b)), movemask));
 }
 
 static really_really_inline
 m128 add_2x64(m128 a, m128 b) {
-    return (m128) vaddq_u64((int64x2_t)a, (int64x2_t)b);
+    return (m128) vaddq_u64((uint64x2_t)a, (uint64x2_t)b);
 }
 
 static really_really_inline
 m128 sub_2x64(m128 a, m128 b) {
-    return (m128) vsubq_u64((int64x2_t)a, (int64x2_t)b);
+    return (m128) vsubq_u64((uint64x2_t)a, (uint64x2_t)b);
 }
 
 static really_really_inline
 m128 lshift_m128(m128 a, unsigned b) {
-    return (m128) vshlq_n_s32((int64x2_t)a, b);
+    return (m128) vshlq_n_u32((uint32x4_t)a, b);
 }
 
 static really_really_inline
 m128 rshift_m128(m128 a, unsigned b) {
-    return (m128) vshrq_n_s32((int64x2_t)a, b);
+    return (m128) vshrq_n_u32((uint32x4_t)a, b);
 }
 
 static really_really_inline
 m128 lshift64_m128(m128 a, unsigned b) {
-    return (m128) vshlq_n_s64((int64x2_t)a, b);
+    return (m128) vshlq_n_u64((uint64x2_t)a, b);
 }
 
 static really_really_inline
 m128 rshift64_m128(m128 a, unsigned b) {
-    return (m128) vshrq_n_s64((int64x2_t)a, b);
+    return (m128) vshrq_n_u64((uint64x2_t)a, b);
 }
 
 static really_inline m128 eq128(m128 a, m128 b) {
-    return (m128) vceqq_s8((int8x16_t)a, (int8x16_t)b);
+    return (m128) vceqq_u8((uint8x16_t)a, (uint8x16_t)b);
 }
 
 static really_inline m128 eq64_m128(m128 a, m128 b) {
-    return (m128) vceqq_u64((int64x2_t)a, (int64x2_t)b);
+    return (m128) vceqq_u64((uint64x2_t)a, (uint64x2_t)b);
 }
 
 static really_inline u32 movemask128(m128 a) {
     static const uint8x16_t powers = { 1, 2, 4, 8, 16, 32, 64, 128, 1, 2, 4, 8, 16, 32, 64, 128 };
 
     // Compute the mask from the input
-    uint64x2_t mask  = vpaddlq_u32(vpaddlq_u16(vpaddlq_u8(vandq_u8((uint8x16_t)a, powers))));
-    uint64x2_t mask1 = (m128)vextq_s8(mask, zeroes128(), 7);
+    uint8x16_t mask  = (uint8x16_t) vpaddlq_u32(vpaddlq_u16(vpaddlq_u8(vandq_u8((uint8x16_t)a, powers))));
+    uint8x16_t mask1 = vextq_u8(mask, (uint8x16_t)zeroes128(), 7);
     mask = vorrq_u8(mask, mask1);
 
     // Get the resulting bytes
@@ -187,7 +187,7 @@ static really_inline u64a movq(const m128 in) {
 /* another form of movq */
 static really_inline
 m128 load_m128_from_u64a(const u64a *p) {
-    return (m128) vsetq_lane_u64(*p, zeroes128(), 0);
+    return (m128) vsetq_lane_u64(*p, (uint64x2_t) zeroes128(), 0);
 }
 
 static really_inline u32 extract32from128(const m128 in, unsigned imm) {
@@ -220,10 +220,10 @@ static really_inline u64a extract64from128(const m128 in, unsigned imm) {
 #else
     switch (imm) {
     case 0:
-        return vgetq_lane_u64((uint32x4_t) in, 0);
+        return vgetq_lane_u64((uint64x2_t) in, 0);
        break;
     case 1:
-        return vgetq_lane_u64((uint32x4_t) in, 1);
+        return vgetq_lane_u64((uint64x2_t) in, 1);
        break;
     default:
        return 0;
@@ -233,11 +233,11 @@ static really_inline u64a extract64from128(const m128 in, unsigned imm) {
 }
 
 static really_inline m128 low64from128(const m128 in) {
-    return vcombine_u64(vget_low_u64(in), vdup_n_u64(0));
+    return (m128) vcombine_u64(vget_low_u64((uint64x2_t)in), vdup_n_u64(0));
 }
 
 static really_inline m128 high64from128(const m128 in) {
-    return vcombine_u64(vget_high_u64(in), vdup_n_u64(0));
+    return (m128) vcombine_u64(vget_high_u64((uint64x2_t)in), vdup_n_u64(0));
 }
 
 static really_inline m128 add128(m128 a, m128 b) {
@@ -257,7 +257,7 @@ static really_inline m128 or128(m128 a, m128 b) {
 }
 
 static really_inline m128 andnot128(m128 a, m128 b) {
-    return (m128) (m128) vandq_s8( vmvnq_s8(a), b);
+    return (m128) vandq_s8( vmvnq_s8((int8x16_t) a), (int8x16_t) b);
 }
 
 // aligned load
@@ -401,12 +401,12 @@ m128 pshufb_m128(m128 a, m128 b) {
 
 static really_inline
 m128 max_u8_m128(m128 a, m128 b) {
-    return (m128) vmaxq_u8((int8x16_t)a, (int8x16_t)b);
+    return (m128) vmaxq_u8((uint8x16_t)a, (uint8x16_t)b);
 }
 
 static really_inline
 m128 min_u8_m128(m128 a, m128 b) {
-    return (m128) vminq_u8((int8x16_t)a, (int8x16_t)b);
+    return (m128) vminq_u8((uint8x16_t)a, (uint8x16_t)b);
 }
 
 static really_inline

diff --git a/src/util/supervector/arch/arm/impl.cpp b/src/util/supervector/arch/arm/impl.cpp
index 34e5486d94144e9c94ac5bf4accee5dfac562087..f804abeb6abba229e10b3b45134a421999abec81 100644 (file)
--- a/src/util/supervector/arch/arm/impl.cpp
+++ b/src/util/supervector/arch/arm/impl.cpp
@@ -45,72 +45,114 @@ really_inline SuperVector<16>::SuperVector(typename base_type::type const v)
 
 template<>
 template<>
-really_inline SuperVector<16>::SuperVector<int8x16_t>(int8x16_t const other)
+really_inline SuperVector<16>::SuperVector<int8x16_t>(int8x16_t other)
 {
-    u.v128[0] = static_cast<m128>(other);
+    u.s8x16[0] = other;
 }
 
 template<>
 template<>
-really_inline SuperVector<16>::SuperVector<uint8x16_t>(uint8x16_t const other)
+really_inline SuperVector<16>::SuperVector<uint8x16_t>(uint8x16_t other)
 {
-    u.v128[0] = static_cast<m128>(other);
+    u.u8x16[0] = other;
+}
+
+template<>
+template<>
+really_inline SuperVector<16>::SuperVector<int16x8_t>(int16x8_t other)
+{
+    u.s16x8[0] = other;
+}
+
+template<>
+template<>
+really_inline SuperVector<16>::SuperVector<uint16x8_t>(uint16x8_t other)
+{
+    u.u16x8[0] = other;
+}
+
+template<>
+template<>
+really_inline SuperVector<16>::SuperVector<int32x4_t>(int32x4_t other)
+{
+    u.s32x4[0] = other;
+}
+
+template<>
+template<>
+really_inline SuperVector<16>::SuperVector<uint32x4_t>(uint32x4_t other)
+{
+    u.u32x4[0] = other;
+}
+
+template<>
+template<>
+really_inline SuperVector<16>::SuperVector<int64x2_t>(int64x2_t other)
+{
+    u.s64x2[0] = other;
+}
+
+template<>
+template<>
+really_inline SuperVector<16>::SuperVector<uint64x2_t>(uint64x2_t other)
+{
+    u.u64x2[0] = other;
 }
 
 template<>
 template<>
 really_inline SuperVector<16>::SuperVector<int8_t>(int8_t const other)
 {
-    u.v128[0] = vdupq_n_s8(other);
+    u.s8x16[0] = vdupq_n_s8(other);
 }
 
 template<>
 template<>
 really_inline SuperVector<16>::SuperVector<uint8_t>(uint8_t const other)
 {
-    u.v128[0] = vdupq_n_u8(other);
+    u.u8x16[0] = vdupq_n_u8(other);
 }
 
 template<>
 template<>
 really_inline SuperVector<16>::SuperVector<int16_t>(int16_t const other)
 {
-    u.v128[0] = vdupq_n_s16(other);
+    u.s16x8[0] = vdupq_n_s16(other);
 }
 
 template<>
 template<>
 really_inline SuperVector<16>::SuperVector<uint16_t>(uint16_t const other)
 {
-    u.v128[0] = vdupq_n_u16(other);
+    u.u16x8[0] = vdupq_n_u16(other);
 }
 
 template<>
 template<>
 really_inline SuperVector<16>::SuperVector<int32_t>(int32_t const other)
 {
-    u.v128[0] = vdupq_n_s32(other);
+    u.s32x4[0] = vdupq_n_s32(other);
 }
 
 template<>
 template<>
 really_inline SuperVector<16>::SuperVector<uint32_t>(uint32_t const other)
 {
-    u.v128[0] = vdupq_n_u32(other);
+    u.u32x4[0] = vdupq_n_u32(other);
 }
 
 template<>
 template<>
 really_inline SuperVector<16>::SuperVector<int64_t>(int64_t const other)
 {
-    u.v128[0] = vdupq_n_s64(other);
+    u.s64x2[0] = vdupq_n_s64(other);
 }
 
 template<>
 template<>
 really_inline SuperVector<16>::SuperVector<uint64_t>(uint64_t const other)
 {
-    u.v128[0] = vdupq_n_u64(other);
+    u.u64x2[0] = vdupq_n_u64(other);
 }
 
 // Constants
@@ -137,37 +179,37 @@ really_inline void SuperVector<16>::operator=(SuperVector<16> const &other)
 template <>
 really_inline SuperVector<16> SuperVector<16>::operator&(SuperVector<16> const &b) const
 {
-    return {vandq_s8(u.v128[0], b.u.v128[0])};
+    return {vandq_u8(u.u8x16[0], b.u.u8x16[0])};
 }
 
 template <>
 really_inline SuperVector<16> SuperVector<16>::operator|(SuperVector<16> const &b) const
 {
-    return {vorrq_s8(u.v128[0], b.u.v128[0])};
+    return {vorrq_u8(u.u8x16[0], b.u.u8x16[0])};
 }
 
 template <>
 really_inline SuperVector<16> SuperVector<16>::operator^(SuperVector<16> const &b) const
 {
-    return {veorq_s8(u.v128[0], b.u.v128[0])};
+    return {veorq_u8(u.u8x16[0], b.u.u8x16[0])};
 }
 
 template <>
 really_inline SuperVector<16> SuperVector<16>::operator!() const
 {
-    return {vmvnq_s8(u.v128[0])};
+    return {vmvnq_u8(u.u8x16[0])};
 }
 
 template <>
 really_inline SuperVector<16> SuperVector<16>::opandnot(SuperVector<16> const &b) const
 {
-    return {vandq_s8(vmvnq_s8(u.v128[0]), b.u.v128[0])};
+    return {vandq_u8(vmvnq_u8(u.u8x16[0]), b.u.u8x16[0])};
 }
 
 template <>
 really_inline SuperVector<16> SuperVector<16>::operator==(SuperVector<16> const &b) const
 {
-    return {vceqq_s8((int16x8_t)u.v128[0], (int16x8_t)b.u.v128[0])};
+    return {vceqq_u8(u.u8x16[0], b.u.u8x16[0])};
 }
 
 template <>
@@ -179,25 +221,25 @@ really_inline SuperVector<16> SuperVector<16>::operator!=(SuperVector<16> const
 template <>
 really_inline SuperVector<16> SuperVector<16>::operator>(SuperVector<16> const &b) const
 {
-    return {vcgtq_s8((int16x8_t)u.v128[0], (int16x8_t)b.u.v128[0])};
+    return {vcgtq_s8(u.s8x16[0], b.u.s8x16[0])};
 }
 
 template <>
 really_inline SuperVector<16> SuperVector<16>::operator>=(SuperVector<16> const &b) const
 {
-    return {vcgeq_s8((int16x8_t)u.v128[0], (int16x8_t)b.u.v128[0])};
+    return {vcgeq_u8(u.u8x16[0], b.u.u8x16[0])};
 }
 
 template <>
 really_inline SuperVector<16> SuperVector<16>::operator<(SuperVector<16> const &b) const
 {
-    return {vcltq_s8((int16x8_t)u.v128[0], (int16x8_t)b.u.v128[0])};
+    return {vcltq_s8(u.s8x16[0], b.u.s8x16[0])};
 }
 
 template <>
 really_inline SuperVector<16> SuperVector<16>::operator<=(SuperVector<16> const &b) const
 {
-    return {vcgeq_s8((int16x8_t)u.v128[0], (int16x8_t)b.u.v128[0])};
+    return {vcgeq_s8(u.s8x16[0], b.u.s8x16[0])};
 }
 
 template <>
@@ -212,9 +254,9 @@ really_inline typename SuperVector<16>::movemask_type SuperVector<16>::movemask(
     SuperVector powers{0x8040201008040201UL};
 
     // Compute the mask from the input
-    uint64x2_t mask  = vpaddlq_u32(vpaddlq_u16(vpaddlq_u8(vandq_u8((uint16x8_t)u.v128[0], powers.u.v128[0]))));
-    uint64x2_t mask1 = (m128)vextq_s8(mask, vdupq_n_u8(0), 7);
-    mask = vorrq_u8(mask, mask1);
+    uint8x16_t mask  = (uint8x16_t) vpaddlq_u32(vpaddlq_u16(vpaddlq_u8(vandq_u8(u.u8x16[0], powers.u.u8x16[0]))));
+    uint64x2_t mask1 = (uint64x2_t) vextq_u8(mask, vdupq_n_u8(0), 7);
+    mask = vorrq_u8(mask, (uint8x16_t) mask1);
 
     // Get the resulting bytes
     uint16_t output;
@@ -232,35 +274,35 @@ template <>
 template<uint8_t N>
 really_inline SuperVector<16> SuperVector<16>::vshl_8_imm() const
 {
-    return {(m128)vshlq_n_s8(u.v128[0], N)};
+    return {vshlq_n_u8(u.u8x16[0], N)};
 }
 
 template <>
 template<uint8_t N>
 really_inline SuperVector<16> SuperVector<16>::vshl_16_imm() const
 {
-    return {(m128)vshlq_n_s16(u.v128[0], N)};
+    return {vshlq_n_u16(u.u16x8[0], N)};
 }
 
 template <>
 template<uint8_t N>
 really_inline SuperVector<16> SuperVector<16>::vshl_32_imm() const
 {
-    return {(m128)vshlq_n_s32(u.v128[0], N)};
+    return {vshlq_n_u32(u.u32x4[0], N)};
 }
 
 template <>
 template<uint8_t N>
 really_inline SuperVector<16> SuperVector<16>::vshl_64_imm() const
 {
-    return {(m128)vshlq_n_s64(u.v128[0], N)};
+    return {vshlq_n_u64(u.u64x2[0], N)};
 }
 
 template <>
 template<uint8_t N>
 really_inline SuperVector<16> SuperVector<16>::vshl_128_imm() const
 {
-    return {vextq_s8(vdupq_n_u8(0), (int16x8_t)u.v128[0], 16 - N)};
+    return {vextq_u8(vdupq_n_u8(0), u.u8x16[0], 16 - N)};
 }
 
 template <>
@@ -274,35 +316,35 @@ template <>
 template<uint8_t N>
 really_inline SuperVector<16> SuperVector<16>::vshr_8_imm() const
 {
-    return {(m128)vshrq_n_s8(u.v128[0], N)};
+    return {vshrq_n_u8(u.u8x16[0], N)};
 }
 
 template <>
 template<uint8_t N>
 really_inline SuperVector<16> SuperVector<16>::vshr_16_imm() const
 {
-    return {(m128)vshrq_n_s16(u.v128[0], N)};
+    return {vshrq_n_u16(u.u16x8[0], N)};
 }
 
 template <>
 template<uint8_t N>
 really_inline SuperVector<16> SuperVector<16>::vshr_32_imm() const
 {
-    return {(m128)vshrq_n_s32(u.v128[0], N)};
+    return {vshrq_n_u32(u.u32x4[0], N)};
 }
 
 template <>
 template<uint8_t N>
 really_inline SuperVector<16> SuperVector<16>::vshr_64_imm() const
 {
-    return {(m128)vshrq_n_s64(u.v128[0], N)};
+    return {vshrq_n_u64(u.u64x2[0], N)};
 }
 
 template <>
 template<uint8_t N>
 really_inline SuperVector<16> SuperVector<16>::vshr_128_imm() const
 {
-    return {vextq_s8((int16x8_t)u.v128[0], vdupq_n_u8(0), N)};
+    return {vextq_u8(u.u8x16[0], vdupq_n_u8(0), N)};
 }
 
 template <>
@@ -334,7 +376,7 @@ really_inline SuperVector<16> SuperVector<16>::vshl_8  (uint8_t const N) const
     if (N == 0) return *this;
     if (N == 16) return Zeroes();
     SuperVector result;
-    Unroller<1, 16>::iterator([&,v=this](auto const i) { constexpr uint8_t n = i.value; if (N == n) result = {(m128)vshlq_n_s8(u.v128[0], n)}; });
+    Unroller<1, 16>::iterator([&,v=this](auto const i) { constexpr uint8_t n = i.value; if (N == n) result = {vshlq_n_u8(u.u8x16[0], n)}; });
     return result;
 }
 
@@ -344,7 +386,7 @@ really_inline SuperVector<16> SuperVector<16>::vshl_16 (uint8_t const N) const
     if (N == 0) return *this;
     if (N == 16) return Zeroes();
     SuperVector result;
-    Unroller<1, 16>::iterator([&,v=this](auto const i) { constexpr uint8_t n = i.value; if (N == n) result = {(m128)vshlq_n_s16(u.v128[0], n)}; });
+    Unroller<1, 16>::iterator([&,v=this](auto const i) { constexpr uint8_t n = i.value; if (N == n) result = {vshlq_n_u16(u.u16x8[0], n)}; });
     return result;
 }
 
@@ -354,7 +396,7 @@ really_inline SuperVector<16> SuperVector<16>::vshl_32 (uint8_t const N) const
     if (N == 0) return *this;
     if (N == 16) return Zeroes();
     SuperVector result;
-    Unroller<1, 16>::iterator([&,v=this](auto const i) { constexpr uint8_t n = i.value; if (N == n) result = {(m128)vshlq_n_s32(u.v128[0], n)}; });
+    Unroller<1, 16>::iterator([&,v=this](auto const i) { constexpr uint8_t n = i.value; if (N == n) result = {vshlq_n_u32(u.u32x4[0], n)}; });
     return result;
 }
 
@@ -364,7 +406,7 @@ really_inline SuperVector<16> SuperVector<16>::vshl_64 (uint8_t const N) const
     if (N == 0) return *this;
     if (N == 16) return Zeroes();
     SuperVector result;
-    Unroller<1, 16>::iterator([&,v=this](auto const i) { constexpr uint8_t n = i.value; if (N == n) result = {(m128)vshlq_n_s64(u.v128[0], n)}; });
+    Unroller<1, 16>::iterator([&,v=this](auto const i) { constexpr uint8_t n = i.value; if (N == n) result = {vshlq_n_u64(u.u64x2[0], n)}; });
     return result;
 }
 
@@ -374,7 +416,7 @@ really_inline SuperVector<16> SuperVector<16>::vshl_128(uint8_t const N) const
     if (N == 0) return *this;
     if (N == 16) return Zeroes();
     SuperVector result;
-    Unroller<1, 16>::iterator([&,v=this](auto const i) { constexpr uint8_t n = i.value; if (N == n) result = {vextq_s8(vdupq_n_u8(0), (int16x8_t)u.v128[0], 16 - n)}; });
+    Unroller<1, 16>::iterator([&,v=this](auto const i) { constexpr uint8_t n = i.value; if (N == n) result = {vextq_u8(vdupq_n_u8(0), u.u8x16[0], 16 - n)}; });
     return result;
 }
 
@@ -390,7 +432,7 @@ really_inline SuperVector<16> SuperVector<16>::vshr_8  (uint8_t const N) const
     if (N == 0) return *this;
     if (N == 16) return Zeroes();
     SuperVector result;
-    Unroller<1, 16>::iterator([&,v=this](auto const i) { constexpr uint8_t n = i.value; if (N == n) result = {(m128)vshrq_n_s8(u.v128[0], n)}; });
+    Unroller<1, 16>::iterator([&,v=this](auto const i) { constexpr uint8_t n = i.value; if (N == n) result = {vshrq_n_u8(u.u8x16[0], n)}; });
     return result;
 }
 
@@ -400,7 +442,7 @@ really_inline SuperVector<16> SuperVector<16>::vshr_16 (uint8_t const N) const
     if (N == 0) return *this;
     if (N == 16) return Zeroes();
     SuperVector result;
-    Unroller<1, 16>::iterator([&,v=this](auto const i) { constexpr uint8_t n = i.value; if (N == n) result = {(m128)vshrq_n_s16(u.v128[0], n)}; });
+    Unroller<1, 16>::iterator([&,v=this](auto const i) { constexpr uint8_t n = i.value; if (N == n) result = {vshrq_n_u16(u.u16x8[0], n)}; });
     return result;
 }
 
@@ -410,7 +452,7 @@ really_inline SuperVector<16> SuperVector<16>::vshr_32 (uint8_t const N) const
     if (N == 0) return *this;
     if (N == 16) return Zeroes();
     SuperVector result;
-    Unroller<1, 16>::iterator([&,v=this](auto const i) { constexpr uint8_t n = i.value; if (N == n) result = {(m128)vshrq_n_s32(u.v128[0], n)}; });
+    Unroller<1, 16>::iterator([&,v=this](auto const i) { constexpr uint8_t n = i.value; if (N == n) result = {vshrq_n_u32(u.u32x4[0], n)}; });
     return result;
 }
 
@@ -420,7 +462,7 @@ really_inline SuperVector<16> SuperVector<16>::vshr_64 (uint8_t const N) const
     if (N == 0) return *this;
     if (N == 16) return Zeroes();
     SuperVector result;
-    Unroller<1, 16>::iterator([&,v=this](auto const i) { constexpr uint8_t n = i.value; if (N == n) result = {(m128)vshrq_n_s64(u.v128[0], n)}; });
+    Unroller<1, 16>::iterator([&,v=this](auto const i) { constexpr uint8_t n = i.value; if (N == n) result = {vshrq_n_u64(u.u64x2[0], n)}; });
     return result;
 }
 
@@ -430,7 +472,7 @@ really_inline SuperVector<16> SuperVector<16>::vshr_128(uint8_t const N) const
     if (N == 0) return *this;
     if (N == 16) return Zeroes();
     SuperVector result;
-    Unroller<1, 16>::iterator([&,v=this](auto const i) { constexpr uint8_t n = i.value; if (N == n) result = {vextq_s8((int16x8_t)u.v128[0], vdupq_n_u8(0), n)}; });
+    Unroller<1, 16>::iterator([&,v=this](auto const i) { constexpr uint8_t n = i.value; if (N == n) result = {vextq_u8(u.u8x16[0], vdupq_n_u8(0), n)}; });
     return result;
 }
 
@@ -444,7 +486,7 @@ really_inline SuperVector<16> SuperVector<16>::vshr(uint8_t const N) const
 template <>
 really_inline SuperVector<16> SuperVector<16>::operator>>(uint8_t const N) const
 {
-    return {vextq_s8((int16x8_t)u.v128[0], vdupq_n_u8(0), N)};
+    return {vextq_u8(u.u8x16[0], vdupq_n_u8(0), N)};
 }
 #else
 template <>
@@ -458,7 +500,7 @@ really_inline SuperVector<16> SuperVector<16>::operator>>(uint8_t const N) const
 template <>
 really_inline SuperVector<16> SuperVector<16>::operator<<(uint8_t const N) const
 {
-    return {vextq_s8(vdupq_n_u8(0), (int16x8_t)u.v128[0], 16 - N)};
+    return {vextq_u8(vdupq_n_u8(0), u.u8x16[0], 16 - N)};
 }
 #else
 template <>
@@ -512,7 +554,7 @@ really_inline SuperVector<16> SuperVector<16>::alignr(SuperVector<16> &other, in
     if (offset == 16) {
         return *this;
     } else {
-        return {vextq_s8((int16x8_t)other.u.v128[0], (int16x8_t)u.v128[0], offset)};
+        return {vextq_u8(other.u.u8x16[0], u.u8x16[0], offset)};
     }
 }
 #else
@@ -521,21 +563,21 @@ really_inline SuperVector<16> SuperVector<16>::alignr(SuperVector<16> &other, in
 {
     switch(offset) {
     case 0: return other; break;
-    case 1: return {vextq_s8((int16x8_t) other.u.v128[0], (int16x8_t) u.v128[0], 1)}; break;
-    case 2: return {vextq_s8((int16x8_t) other.u.v128[0], (int16x8_t) u.v128[0], 2)}; break;
-    case 3: return {vextq_s8((int16x8_t) other.u.v128[0], (int16x8_t) u.v128[0], 3)}; break;
-    case 4: return {vextq_s8((int16x8_t) other.u.v128[0], (int16x8_t) u.v128[0], 4)}; break;
-    case 5: return {vextq_s8((int16x8_t) other.u.v128[0], (int16x8_t) u.v128[0], 5)}; break;
-    case 6: return {vextq_s8((int16x8_t) other.u.v128[0], (int16x8_t) u.v128[0], 6)}; break;
-    case 7: return {vextq_s8((int16x8_t) other.u.v128[0], (int16x8_t) u.v128[0], 7)}; break;
-    case 8: return {vextq_s8((int16x8_t) other.u.v128[0], (int16x8_t) u.v128[0], 8)}; break;
-    case 9: return {vextq_s8((int16x8_t) other.u.v128[0], (int16x8_t) u.v128[0], 9)}; break;
-    case 10: return {vextq_s8((int16x8_t) other.u.v128[0], (int16x8_t) u.v128[0], 10)}; break;
-    case 11: return {vextq_s8((int16x8_t) other.u.v128[0], (int16x8_t) u.v128[0], 11)}; break;
-    case 12: return {vextq_s8((int16x8_t) other.u.v128[0], (int16x8_t) u.v128[0], 12)}; break;
-    case 13: return {vextq_s8((int16x8_t) other.u.v128[0], (int16x8_t) u.v128[0], 13)}; break;
-    case 14: return {vextq_s8((int16x8_t) other.u.v128[0], (int16x8_t) u.v128[0], 14)}; break;
-    case 15: return {vextq_s8((int16x8_t) other.u.v128[0], (int16x8_t) u.v128[0], 15)}; break;
+    case 1: return {vextq_u8( other.u.u8x16[0], u.u8x16[0], 1)}; break;
+    case 2: return {vextq_u8( other.u.u8x16[0], u.u8x16[0], 2)}; break;
+    case 3: return {vextq_u8( other.u.u8x16[0], u.u8x16[0], 3)}; break;
+    case 4: return {vextq_u8( other.u.u8x16[0], u.u8x16[0], 4)}; break;
+    case 5: return {vextq_u8( other.u.u8x16[0], u.u8x16[0], 5)}; break;
+    case 6: return {vextq_u8( other.u.u8x16[0], u.u8x16[0], 6)}; break;
+    case 7: return {vextq_u8( other.u.u8x16[0], u.u8x16[0], 7)}; break;
+    case 8: return {vextq_u8( other.u.u8x16[0], u.u8x16[0], 8)}; break;
+    case 9: return {vextq_u8( other.u.u8x16[0], u.u8x16[0], 9)}; break;
+    case 10: return {vextq_u8( other.u.u8x16[0], u.u8x16[0], 10)}; break;
+    case 11: return {vextq_u8( other.u.u8x16[0], u.u8x16[0], 11)}; break;
+    case 12: return {vextq_u8( other.u.u8x16[0], u.u8x16[0], 12)}; break;
+    case 13: return {vextq_u8( other.u.u8x16[0], u.u8x16[0], 13)}; break;
+    case 14: return {vextq_u8( other.u.u8x16[0], u.u8x16[0], 14)}; break;
+    case 15: return {vextq_u8( other.u.u8x16[0], u.u8x16[0], 15)}; break;
     case 16: return *this; break;
     default: break;
     }
@@ -547,7 +589,7 @@ template<>
 template<>
 really_inline SuperVector<16> SuperVector<16>::pshufb<false>(SuperVector<16> b)
 {
-    return {vqtbl1q_s8((int8x16_t)u.v128[0], (uint8x16_t)b.u.v128[0])};
+    return {vqtbl1q_u8(u.u8x16[0], b.u.u8x16[0])};
 }
 
 template<>
@@ -565,7 +607,7 @@ template<>
 really_inline SuperVector<16> SuperVector<16>::pshufb_maskz(SuperVector<16> b, uint8_t const len)
 {
     SuperVector mask = Ones_vshr(16 -len);
-    return mask & pshufb<true>(b);
+    return mask & pshufb(b);
 }
 
 #endif // SIMD_IMPL_HPP

diff --git a/src/util/supervector/supervector.hpp b/src/util/supervector/supervector.hpp
index 76e167ce399ae1ecb3f29b733212533d8d612420..e69e4b42a255c523d69b7331827b46b6a5f82b9c 100644 (file)
--- a/src/util/supervector/supervector.hpp
+++ b/src/util/supervector/supervector.hpp
@@ -162,6 +162,18 @@ public:
     typename BaseVector<16>::type ALIGN_ATTR(BaseVector<16>::size) v128[SIZE / BaseVector<16>::size];
     typename BaseVector<32>::type ALIGN_ATTR(BaseVector<32>::size) v256[SIZE / BaseVector<32>::size];
     typename BaseVector<64>::type ALIGN_ATTR(BaseVector<64>::size) v512[SIZE / BaseVector<64>::size];
+
+#if defined(ARCH_ARM32) || defined(ARCH_AARCH64)
+    uint64x2_t ALIGN_ATTR(BaseVector<16>::size) u64x2[SIZE / BaseVector<16>::size];
+    int64x2_t  ALIGN_ATTR(BaseVector<16>::size) s64x2[SIZE / BaseVector<16>::size];
+    uint32x4_t ALIGN_ATTR(BaseVector<16>::size) u32x4[SIZE / BaseVector<16>::size];
+    int32x4_t  ALIGN_ATTR(BaseVector<16>::size) s32x4[SIZE / BaseVector<16>::size];
+    uint16x8_t ALIGN_ATTR(BaseVector<16>::size) u16x8[SIZE / BaseVector<16>::size];
+    int16x8_t  ALIGN_ATTR(BaseVector<16>::size) s16x8[SIZE / BaseVector<16>::size];
+    uint8x16_t ALIGN_ATTR(BaseVector<16>::size) u8x16[SIZE / BaseVector<16>::size];
+    int8x16_t  ALIGN_ATTR(BaseVector<16>::size) s8x16[SIZE / BaseVector<16>::size];
+#endif
+
     uint64_t u64[SIZE / sizeof(uint64_t)];
     int64_t  s64[SIZE / sizeof(int64_t)];
     uint32_t u32[SIZE / sizeof(uint32_t)];
@@ -180,7 +192,7 @@ public:
   SuperVector(typename base_type::type const v);
 
   template<typename T>
-  SuperVector(T const other);
+  SuperVector(T other);
 
   SuperVector(SuperVector<SIZE/2> const lo, SuperVector<SIZE/2> const hi);
   SuperVector(previous_type const lo, previous_type const hi);

diff --git a/unit/internal/simd_utils.cpp b/unit/internal/simd_utils.cpp
index 2a9accae3042b62033e645aff5ca7062339c615e..9b206e1b80fc13818e4043ae997918a676cbbbf0 100644 (file)
--- a/unit/internal/simd_utils.cpp
+++ b/unit/internal/simd_utils.cpp
@@ -667,7 +667,7 @@ TEST(SimdUtilsTest, movq) {
     simd = _mm_set_epi64x(~0LL, 0x123456789abcdef);
 #elif defined(ARCH_ARM32) || defined(ARCH_AARCH64)
     int64x2_t a = { 0x123456789abcdefLL, ~0LL };
-    simd = vreinterpretq_s64_s8(a);
+    simd = vreinterpretq_s32_s64(a);
 #endif
 #endif
     r = movq(simd);