instrset.h

Go to the documentation of this file.

/****************************  instrset.h   **********************************
 * Author:        Agner Fog
 * Date created:  2012-05-30
 * Last modified: 2020-11-11
 * Version:       2.01.02
 * Project:       vector class library
 * Description:
 * Header file for various compiler-specific tasks as well as common
 * macros and templates. This file contains:
 *
 * > Selection of the supported instruction set
 * > Defines compiler version macros
 * > Undefines certain macros that prevent function overloading
 * > Helper functions that depend on instruction set, compiler, or platform
 * > Common templates for permute, blend, etc.
 *
 * For instructions, see vcl_manual.pdf
 *
 * (c) Copyright 2012-2020 Agner Fog.
 * Apache License version 2.0 or later.
 ******************************************************************************/
 
#pragma once
 
// Allow the use of floating point permute instructions on integer vectors.
// Some CPU's have an extra latency of 1 or 2 clock cycles for this, but
// it may still be faster than alternative implementations:
#define ALLOW_FP_PERMUTE true
 
// Macro to indicate 64 bit mode
#if ( defined( _M_AMD64 ) || defined( _M_X64 ) || defined( __amd64 ) ) && !defined( __x86_64__ )
#  define __x86_64__ 1 // There are many different macros for this, decide on only one
#endif
 
// The following values of INSTRSET are currently defined:
// 2:  SSE2
// 3:  SSE3
// 4:  SSSE3
// 5:  SSE4.1
// 6:  SSE4.2
// 7:  AVX
// 8:  AVX2
// 9:  AVX512F
// 10: AVX512BW/DQ/VL
// In the future, INSTRSET = 11 may include AVX512VBMI and AVX512VBMI2, but this
// decision cannot be made before the market situation for CPUs with these
// instruction sets is known (these future instruction set extensions are already
// used in some VCL functions and tested with an emulator)
 
// Find instruction set from compiler macros if INSTRSET is not defined.
// Note: Most of these macros are not defined in Microsoft compilers
#ifndef INSTRSET
#  if defined( __AVX512VL__ ) && defined( __AVX512BW__ ) && defined( __AVX512DQ__ )
#    define INSTRSET 10
#  elif defined( __AVX512F__ ) || defined( __AVX512__ )
#    define INSTRSET 9
#  elif defined( __AVX2__ )
#    define INSTRSET 8
#  elif defined( __AVX__ )
#    define INSTRSET 7
#  elif defined( __SSE4_2__ )
#    define INSTRSET 6
#  elif defined( __SSE4_1__ )
#    define INSTRSET 5
#  elif defined( __SSSE3__ )
#    define INSTRSET 4
#  elif defined( __SSE3__ )
#    define INSTRSET 3
#  elif defined( __SSE2__ ) || defined( __x86_64__ )
#    define INSTRSET 2
#  elif defined( __SSE__ )
#    define INSTRSET 1
#  elif defined( _M_IX86_FP ) // Defined in MS compiler. 1: SSE, 2: SSE2
#    define INSTRSET _M_IX86_FP
#  else
#    define INSTRSET 0
#  endif // instruction set defines
#endif   // INSTRSET
 
// Include the appropriate header file for intrinsic functions
#if INSTRSET > 7 // AVX2 and later
#  if defined( __GNUC__ ) && !defined( __INTEL_COMPILER )
#    include <x86intrin.h> // x86intrin.h includes header files for whatever instruction
                           // sets are specified on the compiler command line, such as:
                           // xopintrin.h, fma4intrin.h
#  else
#    include <immintrin.h> // MS/Intel version of immintrin.h covers AVX and later
#  endif                   // __GNUC__
#elif INSTRSET == 7
#  include <immintrin.h> // AVX
#elif INSTRSET == 6
#  include <nmmintrin.h> // SSE4.2
#elif INSTRSET == 5
#  include <smmintrin.h> // SSE4.1
#elif INSTRSET == 4
#  include <tmmintrin.h> // SSSE3
#elif INSTRSET == 3
#  include <pmmintrin.h> // SSE3
#elif INSTRSET == 2
#  include <emmintrin.h> // SSE2
#elif INSTRSET == 1
#  include <xmmintrin.h> // SSE
#endif                   // INSTRSET
 
#if INSTRSET >= 8 && !defined( __FMA__ )
// Assume that all processors that have AVX2 also have FMA3
#  if defined( __GNUC__ ) && !defined( __INTEL_COMPILER )
// Prevent error message in g++ and Clang when using FMA intrinsics with avx2:
#    if !defined( DISABLE_WARNING_AVX2_WITHOUT_FMA )
#      pragma message "It is recommended to specify also option -mfma when using -mavx2 or higher"
#    endif
#  elif !defined( __clang__ )
#    define __FMA__ 1
#  endif
#endif
 
// AMD  instruction sets
#if defined( __XOP__ ) || defined( __FMA4__ )
#  ifdef __GNUC__
#    include <x86intrin.h> // AMD XOP (Gnu)
#  else
#    include <ammintrin.h> // AMD XOP (Microsoft)
#  endif                   //  __GNUC__
#elif defined( __SSE4A__ ) // AMD SSE4A
#  include <ammintrin.h>
#endif // __XOP__
 
// FMA3 instruction set
#if defined( __FMA__ ) && ( defined( __GNUC__ ) || defined( __clang__ ) ) && !defined( __INTEL_COMPILER )
#  include <fmaintrin.h>
#endif // __FMA__
 
// FMA4 instruction set
#if defined( __FMA4__ ) && ( defined( __GNUC__ ) || defined( __clang__ ) )
#  include <fma4intrin.h> // must have both x86intrin.h and fma4intrin.h, don't know why
#endif                    // __FMA4__
 
#include <stdint.h> // Define integer types with known size
#include <stdlib.h> // define abs(int)
 
#ifdef _MSC_VER       // Microsoft compiler or compatible Intel compiler
#  include <intrin.h> // define _BitScanReverse(int), __cpuid(int[4],int), _xgetbv(int)
#endif                // _MSC_VER
 
// functions in instrset_detect.cpp:
#ifdef VCL_NAMESPACE
namespace VCL_NAMESPACE {
#endif
  int  instrset_detect( void ); // tells which instruction sets are supported
  bool hasFMA3( void );         // true if FMA3 instructions supported
  bool hasFMA4( void );         // true if FMA4 instructions supported
  bool hasXOP( void );          // true if XOP  instructions supported
  bool hasAVX512ER( void );     // true if AVX512ER instructions supported
  bool hasAVX512VBMI( void );   // true if AVX512VBMI instructions supported
  bool hasAVX512VBMI2( void );  // true if AVX512VBMI2 instructions supported
#ifdef VCL_NAMESPACE
}
#endif
 
// functions in physical_processors.cpp:
int physicalProcessors( int* logical_processors = 0 );
 
// GCC version
#if defined( __GNUC__ ) && !defined( GCC_VERSION ) && !defined( __clang__ )
#  define GCC_VERSION ( (__GNUC__)*10000 + (__GNUC_MINOR__)*100 + ( __GNUC_PATCHLEVEL__ ) )
#endif
 
// Clang version
#if defined( __clang__ )
#  define CLANG_VERSION ( (__clang_major__)*10000 + (__clang_minor__)*100 + ( __clang_patchlevel__ ) )
// Problem: The version number is not consistent across platforms
// http://llvm.org/bugs/show_bug.cgi?id=12643
// Apple bug 18746972
#endif
 
// Fix problem with non-overloadable macros named min and max in WinDef.h
#ifdef _MSC_VER
#  if defined( _WINDEF_ ) && defined( min ) && defined( max )
#    undef min
#    undef max
#  endif
#  ifndef NOMINMAX
#    define NOMINMAX
#  endif
#endif
 
/* Intel compiler problem:
The Intel compiler currently cannot compile version 2.00 of VCL. It seems to have
a problem with constexpr function returns not being constant enough.
*/
#if defined( __INTEL_COMPILER ) && __INTEL_COMPILER < 9999
#  error The Intel compiler version 19.00 cannot compile VCL version 2. Use Version 1.xx of VCL instead
#endif
 
/* Clang problem:
The Clang compiler treats the intrinsic vector types __m128, __m128i, and __m128d as identical.
See the bug report at https://bugs.llvm.org/show_bug.cgi?id=17164
Additional problem: The version number is not consistent across platforms. The Apple build has
different version numbers. We have to rely on __apple_build_version__ on the Mac platform:
http://llvm.org/bugs/show_bug.cgi?id=12643
We have to make switches here when - hopefully - the error some day has been fixed.
We need different version checks with and whithout __apple_build_version__
*/
#if ( defined( __clang__ ) || defined( __apple_build_version__ ) ) && !defined( __INTEL_COMPILER )
#  define FIX_CLANG_VECTOR_ALIAS_AMBIGUITY
#endif
 
#if defined( GCC_VERSION ) && GCC_VERSION < 99999 && !defined( __clang__ )
#  define ZEXT_MISSING // Gcc 7.4.0 does not have _mm256_zextsi128_si256 and similar functions
#endif
 
#ifdef VCL_NAMESPACE
namespace VCL_NAMESPACE {
#endif
 
  // Constant for indicating don't care in permute and blend functions.
  // V_DC is -256 in Vector class library version 1.xx
  // V_DC can be any value less than -1 in Vector class library version 2.00
  constexpr int V_DC = -256;
 
  /*****************************************************************************
   *
   *    Helper functions that depend on instruction set, compiler, or platform
   *
   *****************************************************************************/
 
  // Define interface to cpuid instruction.
  // input:  functionnumber = leaf (eax), ecxleaf = subleaf(ecx)
  // output: output[0] = eax, output[1] = ebx, output[2] = ecx, output[3] = edx
  static inline void cpuid( int output[4], int functionnumber, int ecxleaf = 0 ) {
#ifdef __x86_64__
#  if defined( __GNUC__ ) || defined( __clang__ ) // use inline assembly, Gnu/AT&T syntax
    int a, b, c, d;
    __asm( "cpuid" : "=a"( a ), "=b"( b ), "=c"( c ), "=d"( d ) : "a"( functionnumber ), "c"( ecxleaf ) : );
    output[0] = a;
    output[1] = b;
    output[2] = c;
    output[3] = d;
 
#  elif defined( _MSC_VER ) // Microsoft compiler, intrin.h included
    __cpuidex( output, functionnumber, ecxleaf ); // intrinsic function for CPUID
 
#  else  // unknown platform. try inline assembly with masm/intel syntax
    __asm {
        mov eax, functionnumber
        mov ecx, ecxleaf
        cpuid;
        mov esi, output
        mov[esi], eax
        mov[esi + 4], ebx
        mov[esi + 8], ecx
        mov[esi + 12], edx
    }
#  endif // compiler/platform
#endif   // __x86_64__
  }
 
// Define popcount function. Gives sum of bits
#if INSTRSET >= 6 // SSE4.2
  // popcnt instruction is not officially part of the SSE4.2 instruction set,
  // but available in all known processors with SSE4.2
  static inline uint32_t vml_popcnt( uint32_t a ) {
    return (uint32_t)_mm_popcnt_u32( a ); // Intel intrinsic. Supported by gcc and clang
  }
#  ifdef __x86_64__
  static inline int64_t vml_popcnt( uint64_t a ) {
    return _mm_popcnt_u64( a ); // Intel intrinsic.
  }
#  else // 32 bit mode
  static inline int64_t vml_popcnt( uint64_t a ) {
    return _mm_popcnt_u32( uint32_t( a >> 32 ) ) + _mm_popcnt_u32( uint32_t( a ) );
  }
#  endif
#else // no SSE4.2
static inline uint32_t vml_popcnt( uint32_t a ) {
  // popcnt instruction not available
  uint32_t b = a - ( ( a >> 1 ) & 0x55555555 );
  uint32_t c = ( b & 0x33333333 ) + ( ( b >> 2 ) & 0x33333333 );
  uint32_t d = ( c + ( c >> 4 ) ) & 0x0F0F0F0F;
  uint32_t e = d * 0x01010101;
  return e >> 24;
}
 
static inline int32_t vml_popcnt( uint64_t a ) {
  return vml_popcnt( uint32_t( a >> 32 ) ) + vml_popcnt( uint32_t( a ) );
}
 
#endif
 
// Define bit-scan-forward function. Gives index to lowest set bit
#if defined( __GNUC__ ) || defined( __clang__ )
  // gcc and Clang have no bit_scan_forward intrinsic
#  if defined( __clang__ ) // fix clang bug
                           // Clang uses a k register as parameter a when inlined from horizontal_find_first
  __attribute__( ( noinline ) )
#  endif
  static uint32_t
  bit_scan_forward( uint32_t a ) {
    uint32_t r;
    __asm( "bsfl %1, %0" : "=r"( r ) : "r"( a ) : );
    return r;
  }
  static inline uint32_t bit_scan_forward( uint64_t a ) {
    uint32_t lo = uint32_t( a );
    if ( lo ) return bit_scan_forward( lo );
    uint32_t hi = uint32_t( a >> 32 );
    return bit_scan_forward( hi ) + 32;
  }
 
#else // other compilers
static inline uint32_t bit_scan_forward( uint32_t a ) {
  unsigned long r;
  _BitScanForward( &r, a ); // defined in intrin.h for MS and Intel compilers
  return r;
}
#  ifdef __x86_64__
static inline uint32_t bit_scan_forward( uint64_t a ) {
  unsigned long r;
  _BitScanForward64( &r, a ); // defined in intrin.h for MS and Intel compilers
  return (uint32_t)r;
}
#  else
static inline uint32_t bit_scan_forward( uint64_t a ) {
  uint32_t lo = uint32_t( a );
  if ( lo ) return bit_scan_forward( lo );
  uint32_t hi = uint32_t( a >> 32 );
  return bit_scan_forward( hi ) + 32;
}
#  endif
#endif
 
// Define bit-scan-reverse function. Gives index to highest set bit = floor(log2(a))
#if defined( __GNUC__ ) || defined( __clang__ )
  static inline uint32_t bit_scan_reverse( uint32_t a ) __attribute__( ( pure ) );
  static inline uint32_t bit_scan_reverse( uint32_t a ) {
    uint32_t r;
    __asm( "bsrl %1, %0" : "=r"( r ) : "r"( a ) : );
    return r;
  }
#  ifdef __x86_64__
  static inline uint32_t bit_scan_reverse( uint64_t a ) {
    uint64_t r;
    __asm( "bsrq %1, %0" : "=r"( r ) : "r"( a ) : );
    return r;
  }
#  else // 32 bit mode
  static inline uint32_t bit_scan_reverse( uint64_t a ) {
    uint64_t ahi = a >> 32;
    if ( ahi == 0 )
      return bit_scan_reverse( uint32_t( a ) );
    else
      return bit_scan_reverse( uint32_t( ahi ) ) + 32;
  }
#  endif
#else
static inline uint32_t bit_scan_reverse( uint32_t a ) {
  unsigned long r;
  _BitScanReverse( &r, a ); // defined in intrin.h for MS and Intel compilers
  return r;
}
#  ifdef __x86_64__
static inline uint32_t bit_scan_reverse( uint64_t a ) {
  unsigned long r;
  _BitScanReverse64( &r, a ); // defined in intrin.h for MS and Intel compilers
  return r;
}
#  else // 32 bit mode
static inline uint32_t bit_scan_reverse( uint64_t a ) {
  uint64_t ahi = a >> 32;
  if ( ahi == 0 )
    return bit_scan_reverse( uint32_t( a ) );
  else
    return bit_scan_reverse( uint32_t( ahi ) ) + 32;
}
#  endif
#endif
 
  // Same function, for compile-time constants
  constexpr int bit_scan_reverse_const( uint64_t const n ) {
    if ( n == 0 ) return -1;
    uint64_t a = n, b = 0, j = 64, k = 0;
    do {
      j >>= 1;
      k = (uint64_t)1 << j;
      if ( a >= k ) {
        a >>= j;
        b += j;
      }
    } while ( j > 0 );
    return int( b );
  }
 
  /*****************************************************************************
   *
   *    Common templates
   *
   *****************************************************************************/
 
  // Template class to represent compile-time integer constant
  template <int32_t n>
  class Const_int_t {}; // represent compile-time signed integer constant
  template <uint32_t n>
  class Const_uint_t {};                      // represent compile-time unsigned integer constant
#define const_int( n ) ( Const_int_t<n>() )   // n must be compile-time integer constant
#define const_uint( n ) ( Const_uint_t<n>() ) // n must be compile-time unsigned integer constant
 
  // template for producing quiet NAN
  template <class VTYPE>
  static inline VTYPE nan_vec( uint32_t payload = 0x100 ) {
    if constexpr ( ( VTYPE::elementtype() & 1 ) != 0 ) { // double
      union {
        uint64_t q;
        double   f;
      } ud;
      // n is left justified to avoid loss of NAN payload when converting to float
      ud.q = 0x7FF8000000000000 | uint64_t( payload ) << 29;
      return VTYPE( ud.f );
    }
    // float will be converted to double if necessary
    union {
      uint32_t i;
      float    f;
    } uf;
    uf.i = 0x7FC00000 | ( payload & 0x003FFFFF );
    return VTYPE( uf.f );
  }
 
  // Test if a parameter is a compile-time constant
  /* Unfortunately, this works only for macro parameters, not for inline function parameters.
     I hope that some solution will appear in the future, but for now it appears to be
     impossible to check if a function parameter is a compile-time constant.
     This would be useful in operator / and in function pow:
     #if defined(__GNUC__) || defined (__clang__)
     #define is_constant(a) __builtin_constant_p(a)
     #else
     #define is_constant(a) false
     #endif
  */
 
  /*****************************************************************************
  *
  *    Helper functions for permute and blend functions
  *
  ******************************************************************************
  Rules for constexpr functions:
 
  > All variable declarations must include initialization
 
  > Do not put variable declarations inside a for-clause, e.g. avoid: for (int i=0; ..
    Instead, you have to declare the loop counter before the for-loop.
 
  > Do not make constexpr functions that return vector types. This requires type
    punning with a union, which is not allowed in constexpr functions under C++17.
    It may be possible under C++20
 
  *****************************************************************************/
 
  // Define type for Encapsulated array to use as return type:
  template <typename T, int N>
  struct EList {
    T a[N];
  };
 
  // get_inttype: get an integer of a size that matches the element size
  // of vector class V with the value -1
  template <typename V>
  constexpr auto get_inttype() {
    constexpr int elementsize = sizeof( V ) / V::size(); // size of vector elements
 
    if constexpr ( elementsize >= 8 ) {
      return -int64_t( 1 );
    } else if constexpr ( elementsize >= 4 ) {
      return int32_t( -1 );
    } else if constexpr ( elementsize >= 2 ) {
      return int16_t( -1 );
    } else {
      return int8_t( -1 );
    }
  }
 
  // zero_mask: return a compact bit mask mask for zeroing using AVX512 mask.
  // Parameter a is a reference to a constexpr int array of permutation indexes
  template <int N>
  constexpr auto zero_mask( int const ( &a )[N] ) {
    uint64_t mask = 0;
    int      i    = 0;
 
    for ( i = 0; i < N; i++ ) {
      if ( a[i] >= 0 ) mask |= uint64_t( 1 ) << i;
    }
    if constexpr ( N <= 8 )
      return uint8_t( mask );
    else if constexpr ( N <= 16 )
      return uint16_t( mask );
    else if constexpr ( N <= 32 )
      return uint32_t( mask );
    else
      return mask;
  }
 
  // zero_mask_broad: return a broad byte mask for zeroing.
  // Parameter a is a reference to a constexpr int array of permutation indexes
  template <typename V>
  constexpr auto zero_mask_broad( int const ( &A )[V::size()] ) {
    constexpr int                        N = V::size(); // number of vector elements
    typedef decltype( get_inttype<V>() ) Etype;         // element type
    EList<Etype, N>                      u = { { 0 } }; // list for return
    int                                  i = 0;
    for ( i = 0; i < N; i++ ) { u.a[i] = A[i] >= 0 ? get_inttype<V>() : 0; }
    return u; // return encapsulated array
  }
 
  // make_bit_mask: return a compact mask of bits from a list of N indexes:
  // B contains options indicating how to gather the mask
  // bit 0-7 in B indicates which bit in each index to collect
  // bit 8 = 0x100:  set 1 in the lower half of the bit mask if the indicated bit is 1.
  // bit 8 = 0    :  set 1 in the lower half of the bit mask if the indicated bit is 0.
  // bit 9 = 0x200:  set 1 in the upper half of the bit mask if the indicated bit is 1.
  // bit 9 = 0    :  set 1 in the upper half of the bit mask if the indicated bit is 0.
  // bit 10 = 0x400: set 1 in the bit mask if the corresponding index is -1 or V_DC
  // Parameter a is a reference to a constexpr int array of permutation indexes
  template <int N, int B>
  constexpr uint64_t make_bit_mask( int const ( &a )[N] ) {
    uint64_t r = 0;                   // return value
    uint8_t  j = uint8_t( B & 0xFF ); // index to selected bit
    uint64_t s = 0;                   // bit number i in r
    uint64_t f = 0;                   // 1 if bit not flipped
    int      i = 0;
    for ( i = 0; i < N; i++ ) {
      int ix = a[i];
      if ( ix < 0 ) { // -1 or V_DC
        s = ( B >> 10 ) & 1;
      } else {
        s = ( (uint32_t)ix >> j ) & 1; // extract selected bit
        if ( i < N / 2 ) {
          f = ( B >> 8 ) & 1; // lower half
        } else {
          f = ( B >> 9 ) & 1; // upper half
        }
        s ^= f ^ 1; // flip bit if needed
      }
      r |= uint64_t( s ) << i; // set bit in return value
    }
    return r;
  }
 
  // make_broad_mask: Convert a bit mask m to a broad mask
  // The return value will be a broad boolean mask with elementsize matching vector class V
  template <typename V>
  constexpr auto make_broad_mask( uint64_t const m ) {
    constexpr int                        N = V::size(); // number of vector elements
    typedef decltype( get_inttype<V>() ) Etype;         // element type
    EList<Etype, N>                      u = { { 0 } }; // list for returning
    int                                  i = 0;
    for ( i = 0; i < N; i++ ) { u.a[i] = ( ( m >> i ) & 1 ) != 0 ? get_inttype<V>() : 0; }
    return u; // return encapsulated array
  }
 
  // perm_mask_broad: return a mask for permutation by a vector register index.
  // Parameter A is a reference to a constexpr int array of permutation indexes
  template <typename V>
  constexpr auto perm_mask_broad( int const ( &A )[V::size()] ) {
    constexpr int                        N = V::size(); // number of vector elements
    typedef decltype( get_inttype<V>() ) Etype;         // vector element type
    EList<Etype, N>                      u = { { 0 } }; // list for returning
    int                                  i = 0;
    for ( i = 0; i < N; i++ ) { u.a[i] = Etype( A[i] ); }
    return u; // return encapsulated array
  }
 
  // perm_flags: returns information about how a permute can be implemented.
  // The return value is composed of these flag bits:
  const int perm_zeroing      = 1;     // needs zeroing
  const int perm_perm         = 2;     // permutation needed
  const int perm_allzero      = 4;     // all is zero or don't care
  const int perm_largeblock   = 8;     // fits permute with a larger block size (e.g permute Vec2q instead of Vec4i)
  const int perm_addz         = 0x10;  // additional zeroing needed after permute with larger block size or shift
  const int perm_addz2        = 0x20;  // additional zeroing needed after perm_zext, perm_compress, or perm_expand
  const int perm_cross_lane   = 0x40;  // permutation crossing 128-bit lanes
  const int perm_same_pattern = 0x80;  // same permute pattern in all 128-bit lanes
  const int perm_punpckh      = 0x100; // permutation pattern fits punpckh instruction
  const int perm_punpckl      = 0x200; // permutation pattern fits punpckl instruction
  const int perm_rotate =
      0x400; // permutation pattern fits rotation within lanes. 4 bit count returned in bit perm_rot_count
  const int perm_shright =
      0x1000; // permutation pattern fits shift right within lanes. 4 bit count returned in bit perm_rot_count
  const int perm_shleft =
      0x2000; // permutation pattern fits shift left within lanes. negative count returned in bit perm_rot_count
  const int perm_rotate_big =
      0x4000; // permutation pattern fits rotation across lanes. 6 bit count returned in bit perm_rot_count
  const int perm_broadcast  = 0x8000;     // permutation pattern fits broadcast of a single element.
  const int perm_zext       = 0x10000;    // permutation pattern fits zero extension
  const int perm_compress   = 0x20000;    // permutation pattern fits vpcompress instruction
  const int perm_expand     = 0x40000;    // permutation pattern fits vpexpand instruction
  const int perm_outofrange = 0x10000000; // index out of range
  const int perm_rot_count  = 32;         // rotate or shift count is in bits perm_rot_count to perm_rot_count+3
  const int perm_ipattern =
      40; // pattern for pshufd is in bit perm_ipattern to perm_ipattern + 7 if perm_same_pattern and elementsize >= 4
 
  template <typename V>
  constexpr uint64_t perm_flags( int const ( &a )[V::size()] ) {
    // a is a reference to a constexpr array of permutation indexes
    // V is a vector class
    constexpr int  N             = V::size();                                          // number of elements
    uint64_t       r             = perm_largeblock | perm_same_pattern | perm_allzero; // return value
    uint32_t       i             = 0;                                                  // loop counter
    int            j             = 0;                                                  // loop counter
    int            ix            = 0;                                                  // index number i
    const uint32_t nlanes        = sizeof( V ) / 16;                                   // number of 128-bit lanes
    const uint32_t lanesize      = N / nlanes;                                         // elements per lane
    const uint32_t elementsize   = sizeof( V ) / N;                                    // size of each vector element
    uint32_t       lane          = 0;                                                  // current lane
    uint32_t       rot           = 999;                                                // rotate left count
    int32_t        broadc        = 999;                                                // index to broadcasted element
    uint32_t       patfail       = 0;  // remember certain patterns that do not fit
    uint32_t       addz2         = 0;  // remember certain patterns need extra zeroing
    int32_t        compresslasti = -1; // last index in perm_compress fit
    int32_t        compresslastp = -1; // last position in perm_compress fit
    int32_t        expandlasti   = -1; // last index in perm_expand fit
    int32_t        expandlastp   = -1; // last position in perm_expand fit
 
    int lanepattern[lanesize] = { 0 }; // pattern in each lane
 
    for ( i = 0; i < N; i++ ) { // loop through indexes
      ix = a[i];                // current index
      // meaning of ix: -1 = set to zero, V_DC = don't care, non-negative value = permute.
      if ( ix == -1 ) {
        r |= perm_zeroing; // zeroing requested
      } else if ( ix != V_DC && uint32_t( ix ) >= N ) {
        r |= perm_outofrange; // index out of range
      }
      if ( ix >= 0 ) {
        r &= ~perm_allzero;                 // not all zero
        if ( ix != (int)i ) r |= perm_perm; // needs permutation
        if ( broadc == 999 )
          broadc = ix; // remember broadcast index
        else if ( broadc != ix )
          broadc = 1000; // does not fit broadcast
      }
      // check if pattern fits a larger block size:
      // even indexes must be even, odd indexes must fit the preceding even index + 1
      if ( ( i & 1 ) == 0 ) {                                            // even index
        if ( ix >= 0 && ( ix & 1 ) ) r &= ~perm_largeblock;              // not even. does not fit larger block size
        int iy = a[i + 1];                                               // next odd index
        if ( iy >= 0 && ( iy & 1 ) == 0 ) r &= ~perm_largeblock;         // not odd. does not fit larger block size
        if ( ix >= 0 && iy >= 0 && iy != ix + 1 ) r &= ~perm_largeblock; // does not fit preceding index + 1
        if ( ix == -1 && iy >= 0 ) r |= perm_addz; // needs additional zeroing at current block size
        if ( iy == -1 && ix >= 0 ) r |= perm_addz; // needs additional zeroing at current block size
      }
      lane = i / lanesize;   // current lane
      if ( lane == 0 ) {     // first lane, or no pattern yet
        lanepattern[i] = ix; // save pattern
      }
      // check if crossing lanes
      if ( ix >= 0 ) {
        uint32_t lanei = (uint32_t)ix / lanesize;  // source lane
        if ( lanei != lane ) r |= perm_cross_lane; // crossing lane
      }
      // check if same pattern in all lanes
      if ( lane != 0 && ix >= 0 ) {                                   // not first lane
        int j1 = i - int( lane * lanesize );                          // index into lanepattern
        int jx = ix - int( lane * lanesize );                         // pattern within lane
        if ( jx < 0 || jx >= (int)lanesize ) r &= ~perm_same_pattern; // source is in another lane
        if ( lanepattern[j1] < 0 ) {
          lanepattern[j1] = jx; // pattern not known from previous lane
        } else {
          if ( lanepattern[j1] != jx ) r &= ~perm_same_pattern; // not same pattern
        }
      }
      if ( ix >= 0 ) {
        // check if pattern fits zero extension (perm_zext)
        if ( uint32_t( ix * 2 ) != i ) {
          patfail |= 1; // does not fit zero extension
        }
        // check if pattern fits compress (perm_compress)
        if ( ix > compresslasti && ix - compresslasti >= (int)i - compresslastp ) {
          if ( (int)i - compresslastp > 1 ) addz2 |= 2; // perm_compress may need additional zeroing
          compresslasti = ix;
          compresslastp = i;
        } else {
          patfail |= 2; // does not fit perm_compress
        }
        // check if pattern fits expand (perm_expand)
        if ( ix > expandlasti && ix - expandlasti <= (int)i - expandlastp ) {
          if ( ix - expandlasti > 1 ) addz2 |= 4; // perm_expand may need additional zeroing
          expandlasti = ix;
          expandlastp = i;
        } else {
          patfail |= 4; // does not fit perm_compress
        }
      } else if ( ix == -1 ) {
        if ( ( i & 1 ) == 0 ) addz2 |= 1; // zero extension needs additional zeroing
      }
    }
    if ( !( r & perm_perm ) ) return r; // more checks are superfluous
 
    if ( !( r & perm_largeblock ) ) r &= ~perm_addz;    // remove irrelevant flag
    if ( r & perm_cross_lane ) r &= ~perm_same_pattern; // remove irrelevant flag
    if ( ( patfail & 1 ) == 0 ) {
      r |= perm_zext; // fits zero extension
      if ( ( addz2 & 1 ) != 0 ) r |= perm_addz2;
    } else if ( ( patfail & 2 ) == 0 ) {
      r |= perm_compress;         // fits compression
      if ( ( addz2 & 2 ) != 0 ) { // check if additional zeroing needed
        for ( j = 0; j < compresslastp; j++ ) {
          if ( a[j] == -1 ) r |= perm_addz2;
        }
      }
    } else if ( ( patfail & 4 ) == 0 ) {
      r |= perm_expand;           // fits expansion
      if ( ( addz2 & 4 ) != 0 ) { // check if additional zeroing needed
        for ( j = 0; j < expandlastp; j++ ) {
          if ( a[j] == -1 ) r |= perm_addz2;
        }
      }
    }
 
    if ( r & perm_same_pattern ) {
      // same pattern in all lanes. check if it fits specific patterns
      bool fit = true;
      // fit shift or rotate
      for ( i = 0; i < lanesize; i++ ) {
        if ( lanepattern[i] >= 0 ) {
          uint32_t rot1 = uint32_t( lanepattern[i] + lanesize - i ) % lanesize;
          if ( rot == 999 ) {
            rot = rot1;
          } else { // check if fit
            if ( rot != rot1 ) fit = false;
          }
        }
      }
      rot &= lanesize - 1;                           // prevent out of range values
      if ( fit ) {                                   // fits rotate, and possibly shift
        uint64_t rot2 = ( rot * elementsize ) & 0xF; // rotate right count in bytes
        r |= rot2 << perm_rot_count;                 // put shift/rotate count in output bit 16-19
#if INSTRSET >= 4                                    // SSSE3
        r |= perm_rotate;                            // allow palignr
#endif
        // fit shift left
        fit = true;
        for ( i = 0; i < lanesize - rot; i++ ) { // check if first rot elements are zero or don't care
          if ( lanepattern[i] >= 0 ) fit = false;
        }
        if ( fit ) {
          r |= perm_shleft;
          for ( ; i < lanesize; i++ )
            if ( lanepattern[i] == -1 ) r |= perm_addz; // additional zeroing needed
        }
        // fit shift right
        fit = true;
        for ( i = lanesize - (uint32_t)rot; i < lanesize;
              i++ ) { // check if last (lanesize-rot) elements are zero or don't care
          if ( lanepattern[i] >= 0 ) fit = false;
        }
        if ( fit ) {
          r |= perm_shright;
          for ( i = 0; i < lanesize - rot; i++ ) {
            if ( lanepattern[i] == -1 ) r |= perm_addz; // additional zeroing needed
          }
        }
      }
      // fit punpckhi
      fit         = true;
      uint32_t j2 = lanesize / 2;
      for ( i = 0; i < lanesize; i++ ) {
        if ( lanepattern[i] >= 0 && lanepattern[i] != (int)j2 ) fit = false;
        if ( ( i & 1 ) != 0 ) j2++;
      }
      if ( fit ) r |= perm_punpckh;
      // fit punpcklo
      fit = true;
      j2  = 0;
      for ( i = 0; i < lanesize; i++ ) {
        if ( lanepattern[i] >= 0 && lanepattern[i] != (int)j2 ) fit = false;
        if ( ( i & 1 ) != 0 ) j2++;
      }
      if ( fit ) r |= perm_punpckl;
      // fit pshufd
      if ( elementsize >= 4 ) {
        uint64_t p = 0;
        for ( i = 0; i < lanesize; i++ ) {
          if ( lanesize == 4 ) {
            p |= ( lanepattern[i] & 3 ) << 2 * i;
          } else { // lanesize = 2
            p |= ( ( lanepattern[i] & 1 ) * 10 + 4 ) << 4 * i;
          }
        }
        r |= p << perm_ipattern;
      }
    }
#if INSTRSET >= 7
    else {                          // not same pattern in all lanes
      if constexpr ( nlanes > 1 ) { // Try if it fits big rotate
        for ( i = 0; i < N; i++ ) {
          ix = a[i];
          if ( ix >= 0 ) {
            uint32_t rot2 = ( ix + N - i ) % N; // rotate count
            if ( rot == 999 ) {
              rot = rot2; // save rotate count
            } else if ( rot != rot2 ) {
              rot = 1000;
              break; // does not fit big rotate
            }
          }
        }
        if ( rot < N ) { // fits big rotate
          r |= perm_rotate_big | (uint64_t)rot << perm_rot_count;
        }
      }
    }
#endif
    if ( broadc < 999 && ( r & ( perm_rotate | perm_shright | perm_shleft | perm_rotate_big ) ) == 0 ) {
      r |= perm_broadcast | (uint64_t)broadc << perm_rot_count; // fits broadcast
    }
    return r;
  }
 
  // compress_mask: returns a bit mask to use for compression instruction.
  // It is presupposed that perm_flags indicates perm_compress.
  // Additional zeroing is needed if perm_flags indicates perm_addz2
  template <int N>
  constexpr uint64_t compress_mask( int const ( &a )[N] ) {
    // a is a reference to a constexpr array of permutation indexes
    int      ix = 0, lasti = -1, lastp = -1;
    uint64_t m = 0;
    int      i = 0;
    int      j = 1; // loop counters
    for ( i = 0; i < N; i++ ) {
      ix = a[i]; // permutation index
      if ( ix >= 0 ) {
        m |= (uint64_t)1 << ix; // mask for compression source
        for ( j = 1; j < i - lastp; j++ ) {
          m |= (uint64_t)1 << ( lasti + j ); // dummy filling source
        }
        lastp = i;
        lasti = ix;
      }
    }
    return m;
  }
 
  // expand_mask: returns a bit mask to use for expansion instruction.
  // It is presupposed that perm_flags indicates perm_expand.
  // Additional zeroing is needed if perm_flags indicates perm_addz2
  template <int N>
  constexpr uint64_t expand_mask( int const ( &a )[N] ) {
    // a is a reference to a constexpr array of permutation indexes
    int      ix = 0, lasti = -1, lastp = -1;
    uint64_t m = 0;
    int      i = 0;
    int      j = 1;
    for ( i = 0; i < N; i++ ) {
      ix = a[i]; // permutation index
      if ( ix >= 0 ) {
        m |= (uint64_t)1 << i; // mask for expansion destination
        for ( j = 1; j < ix - lasti; j++ ) {
          m |= (uint64_t)1 << ( lastp + j ); // dummy filling destination
        }
        lastp = i;
        lasti = ix;
      }
    }
    return m;
  }
 
  // perm16_flags: returns information about how to permute a vector of 16-bit integers
  // Note: It is presupposed that perm_flags reports perm_same_pattern
  // The return value is composed of these bits:
  // 1:  data from low  64 bits to low  64 bits. pattern in bit 32-39
  // 2:  data from high 64 bits to high 64 bits. pattern in bit 40-47
  // 4:  data from high 64 bits to low  64 bits. pattern in bit 48-55
  // 8:  data from low  64 bits to high 64 bits. pattern in bit 56-63
  template <typename V>
  constexpr uint64_t perm16_flags( int const ( &a )[V::size()] ) {
    // a is a reference to a constexpr array of permutation indexes
    // V is a vector class
    constexpr int N = V::size(); // number of elements
 
    uint64_t       retval                = 0;              // return value
    uint32_t       pat[4]                = { 0, 0, 0, 0 }; // permute patterns
    uint32_t       i                     = 0;              // loop counter
    int            ix                    = 0;              // index number i
    const uint32_t lanesize              = 8;              // elements per lane
    uint32_t       lane                  = 0;              // current lane
    int            lanepattern[lanesize] = { 0 };          // pattern in each lane
 
    for ( i = 0; i < N; i++ ) {
      ix   = a[i];
      lane = i / lanesize; // current lane
      if ( lane == 0 ) {
        lanepattern[i] = ix;                // save pattern
      } else if ( ix >= 0 ) {               // not first lane
        uint32_t j  = i - lane * lanesize;  // index into lanepattern
        int      jx = ix - lane * lanesize; // pattern within lane
        if ( lanepattern[j] < 0 ) {
          lanepattern[j] = jx; // pattern not known from previous lane
        }
      }
    }
    // four patterns: low2low, high2high, high2low, low2high
    for ( i = 0; i < 4; i++ ) {
      // loop through low pattern
      if ( lanepattern[i] >= 0 ) {
        if ( lanepattern[i] < 4 ) { // low2low
          retval |= 1;
          pat[0] |= uint32_t( lanepattern[i] & 3 ) << ( 2 * i );
        } else { // high2low
          retval |= 4;
          pat[2] |= uint32_t( lanepattern[i] & 3 ) << ( 2 * i );
        }
      }
      // loop through high pattern
      if ( lanepattern[i + 4] >= 0 ) {
        if ( lanepattern[i + 4] < 4 ) { // low2high
          retval |= 8;
          pat[3] |= uint32_t( lanepattern[i + 4] & 3 ) << ( 2 * i );
        } else { // high2high
          retval |= 2;
          pat[1] |= uint32_t( lanepattern[i + 4] & 3 ) << ( 2 * i );
        }
      }
    }
    // join return data
    for ( i = 0; i < 4; i++ ) { retval |= (uint64_t)pat[i] << ( 32 + i * 8 ); }
    return retval;
  }
 
  // pshufb_mask: return a broad byte mask for permutation within lanes
  // for use with the pshufb instruction (_mm..._shuffle_epi8).
  // The pshufb instruction provides fast permutation and zeroing,
  // allowing different patterns in each lane but no crossing of lane boundaries
  template <typename V, int oppos = 0>
  constexpr auto pshufb_mask( int const ( &A )[V::size()] ) {
    // Parameter a is a reference to a constexpr array of permutation indexes
    // V is a vector class
    // oppos = 1 for data from the opposite 128-bit lane in 256-bit vectors
    constexpr uint32_t N                 = V::size();        // number of vector elements
    constexpr uint32_t elementsize       = sizeof( V ) / N;  // size of each vector element
    constexpr uint32_t nlanes            = sizeof( V ) / 16; // number of 128 bit lanes in vector
    constexpr uint32_t elements_per_lane = N / nlanes;       // number of vector elements per lane
 
    EList<int8_t, sizeof( V )> u = { { 0 } }; // list for returning
 
    uint32_t i    = 0; // loop counters
    uint32_t j    = 0;
    int      m    = 0;
    int      k    = 0;
    uint32_t lane = 0;
 
    for ( lane = 0; lane < nlanes; lane++ ) {     // loop through lanes
      for ( i = 0; i < elements_per_lane; i++ ) { // loop through elements in lane
        // permutation index for element within lane
        int8_t p  = -1;
        int    ix = A[m];
        if ( ix >= 0 ) {
          ix ^= oppos * elements_per_lane; // flip bit if opposite lane
        }
        ix -= int( lane * elements_per_lane );          // index relative to lane
        if ( ix >= 0 && ix < (int)elements_per_lane ) { // index points to desired lane
          p = ix * elementsize;
        }
        for ( j = 0; j < elementsize; j++ ) { // loop through bytes in element
          u.a[k++] = p < 0 ? -1 : p + j;      // store byte permutation index
        }
        m++;
      }
    }
    return u; // return encapsulated array
  }
 
  // largeblock_perm: return indexes for replacing a permute or blend with
  // a certain block size by a permute or blend with the double block size.
  // Note: it is presupposed that perm_flags() indicates perm_largeblock
  // It is required that additional zeroing is added if perm_flags() indicates perm_addz
  template <int N>
  constexpr EList<int, N / 2> largeblock_perm( int const ( &a )[N] ) {
    // Parameter a is a reference to a constexpr array of permutation indexes
    EList<int, N / 2> list     = { { 0 } }; // result indexes
    int               ix       = 0;         // even index
    int               iy       = 0;         // odd index
    int               iz       = 0;         // combined index
    bool              fit_addz = false;     // additional zeroing needed at the lower block level
    int               i        = 0;         // loop counter
 
    // check if additional zeroing is needed at current block size
    for ( i = 0; i < N; i += 2 ) {
      ix = a[i];     // even index
      iy = a[i + 1]; // odd index
      if ( ( ix == -1 && iy >= 0 ) || ( iy == -1 && ix >= 0 ) ) { fit_addz = true; }
    }
 
    // loop through indexes
    for ( i = 0; i < N; i += 2 ) {
      ix = a[i];     // even index
      iy = a[i + 1]; // odd index
      if ( ix >= 0 ) {
        iz = ix / 2; // half index
      } else if ( iy >= 0 ) {
        iz = iy / 2;
      } else {
        iz = ix | iy;              // -1 or V_DC. -1 takes precedence
        if ( fit_addz ) iz = V_DC; // V_DC, because result will be zeroed later
      }
      list.a[i / 2] = iz; // save to list
    }
    return list;
  }
 
  // blend_flags: returns information about how a blend function can be implemented
  // The return value is composed of these flag bits:
  const int blend_zeroing      = 1;       // needs zeroing
  const int blend_allzero      = 2;       // all is zero or don't care
  const int blend_largeblock   = 4;       // fits blend with a larger block size (e.g permute Vec2q instead of Vec4i)
  const int blend_addz         = 8;       // additional zeroing needed after blend with larger block size or shift
  const int blend_a            = 0x10;    // has data from a
  const int blend_b            = 0x20;    // has data from b
  const int blend_perma        = 0x40;    // permutation of a needed
  const int blend_permb        = 0x80;    // permutation of b needed
  const int blend_cross_lane   = 0x100;   // permutation crossing 128-bit lanes
  const int blend_same_pattern = 0x200;   // same permute/blend pattern in all 128-bit lanes
  const int blend_punpckhab    = 0x1000;  // pattern fits punpckh(a,b)
  const int blend_punpckhba    = 0x2000;  // pattern fits punpckh(b,a)
  const int blend_punpcklab    = 0x4000;  // pattern fits punpckl(a,b)
  const int blend_punpcklba    = 0x8000;  // pattern fits punpckl(b,a)
  const int blend_rotateab     = 0x10000; // pattern fits palignr(a,b)
  const int blend_rotateba     = 0x20000; // pattern fits palignr(b,a)
  const int blend_shufab       = 0x40000; // pattern fits shufps/shufpd(a,b)
  const int blend_shufba       = 0x80000; // pattern fits shufps/shufpd(b,a)
  const int blend_rotate_big  = 0x100000; // pattern fits rotation across lanes. count returned in bits blend_rotpattern
  const int blend_outofrange  = 0x10000000; // index out of range
  const int blend_shufpattern = 32; // pattern for shufps/shufpd is in bit blend_shufpattern to blend_shufpattern + 7
  const int blend_rotpattern  = 40; // pattern for palignr is in bit blend_rotpattern to blend_rotpattern + 7
 
  template <typename V>
  constexpr uint64_t blend_flags( int const ( &a )[V::size()] ) {
    // a is a reference to a constexpr array of permutation indexes
    // V is a vector class
    constexpr int  N                     = V::size();                                             // number of elements
    uint64_t       r                     = blend_largeblock | blend_same_pattern | blend_allzero; // return value
    uint32_t       iu                    = 0;                                                     // loop counter
    int32_t        ii                    = 0;                                                     // loop counter
    int            ix                    = 0;                                                     // index number i
    const uint32_t nlanes                = sizeof( V ) / 16; // number of 128-bit lanes
    const uint32_t lanesize              = N / nlanes;       // elements per lane
    uint32_t       lane                  = 0;                // current lane
    uint32_t       rot                   = 999;              // rotate left count
    int            lanepattern[lanesize] = { 0 };            // pattern in each lane
    if ( lanesize == 2 && N <= 8 ) {
      r |= blend_shufab | blend_shufba; // check if it fits shufpd
    }
 
    for ( ii = 0; ii < N; ii++ ) { // loop through indexes
      ix = a[ii];                  // index
      if ( ix < 0 ) {
        if ( ix == -1 )
          r |= blend_zeroing; // set to zero
        else if ( ix != V_DC ) {
          r = blend_outofrange;
          break; // illegal index
        }
      } else { // ix >= 0
        r &= ~blend_allzero;
        if ( ix < N ) {
          r |= blend_a;                     // data from a
          if ( ix != ii ) r |= blend_perma; // permutation of a
        } else if ( ix < 2 * N ) {
          r |= blend_b;                         // data from b
          if ( ix != ii + N ) r |= blend_permb; // permutation of b
        } else {
          r = blend_outofrange;
          break; // illegal index
        }
      }
      // check if pattern fits a larger block size:
      // even indexes must be even, odd indexes must fit the preceding even index + 1
      if ( ( ii & 1 ) == 0 ) {                                            // even index
        if ( ix >= 0 && ( ix & 1 ) ) r &= ~blend_largeblock;              // not even. does not fit larger block size
        int iy = a[ii + 1];                                               // next odd index
        if ( iy >= 0 && ( iy & 1 ) == 0 ) r &= ~blend_largeblock;         // not odd. does not fit larger block size
        if ( ix >= 0 && iy >= 0 && iy != ix + 1 ) r &= ~blend_largeblock; // does not fit preceding index + 1
        if ( ix == -1 && iy >= 0 ) r |= blend_addz; // needs additional zeroing at current block size
        if ( iy == -1 && ix >= 0 ) r |= blend_addz; // needs additional zeroing at current block size
      }
      lane = (uint32_t)ii / lanesize; // current lane
      if ( lane == 0 ) {              // first lane, or no pattern yet
        lanepattern[ii] = ix;         // save pattern
      }
      // check if crossing lanes
      if ( ix >= 0 ) {
        uint32_t lanei = uint32_t( ix & ~N ) / lanesize; // source lane
        if ( lanei != lane ) {
          r |= blend_cross_lane; // crossing lane
        }
        if ( lanesize == 2 ) { // check if it fits pshufd
          if ( lanei != lane ) r &= ~( blend_shufab | blend_shufba );
          if ( ( ( ( ix & N ) != 0 ) ^ ii ) & 1 )
            r &= ~blend_shufab;
          else
            r &= ~blend_shufba;
        }
      }
      // check if same pattern in all lanes
      if ( lane != 0 && ix >= 0 ) {                                             // not first lane
        int j  = ii - int( lane * lanesize );                                   // index into lanepattern
        int jx = ix - int( lane * lanesize );                                   // pattern within lane
        if ( jx < 0 || ( jx & ~N ) >= (int)lanesize ) r &= ~blend_same_pattern; // source is in another lane
        if ( lanepattern[j] < 0 ) {
          lanepattern[j] = jx; // pattern not known from previous lane
        } else {
          if ( lanepattern[j] != jx ) r &= ~blend_same_pattern; // not same pattern
        }
      }
    }
    if ( !( r & blend_largeblock ) ) r &= ~blend_addz;    // remove irrelevant flag
    if ( r & blend_cross_lane ) r &= ~blend_same_pattern; // remove irrelevant flag
    if ( !( r & ( blend_perma | blend_permb ) ) ) {
      return r; // no permutation. more checks are superfluous
    }
    if ( r & blend_same_pattern ) {
      // same pattern in all lanes. check if it fits unpack patterns
      r |= blend_punpckhab | blend_punpckhba | blend_punpcklab | blend_punpcklba;
      for ( iu = 0; iu < lanesize; iu++ ) { // loop through lanepattern
        ix = lanepattern[iu];
        if ( ix >= 0 ) {
          if ( (uint32_t)ix != iu / 2 + ( iu & 1 ) * N ) r &= ~blend_punpcklab;
          if ( (uint32_t)ix != iu / 2 + ( ( iu & 1 ) ^ 1 ) * N ) r &= ~blend_punpcklba;
          if ( (uint32_t)ix != ( iu + lanesize ) / 2 + ( iu & 1 ) * N ) r &= ~blend_punpckhab;
          if ( (uint32_t)ix != ( iu + lanesize ) / 2 + ( ( iu & 1 ) ^ 1 ) * N ) r &= ~blend_punpckhba;
        }
      }
#if INSTRSET >= 4 // SSSE3. check if it fits palignr
      for ( iu = 0; iu < lanesize; iu++ ) {
        ix = lanepattern[iu];
        if ( ix >= 0 ) {
          uint32_t t = ix & ~N;
          if ( ix & N ) t += lanesize;
          uint32_t tb = ( t + 2 * lanesize - iu ) % ( lanesize * 2 );
          if ( rot == 999 ) {
            rot = tb;
          } else { // check if fit
            if ( rot != tb ) rot = 1000;
          }
        }
      }
      if ( rot < 999 ) { // firs palignr
        if ( rot < lanesize ) {
          r |= blend_rotateba;
        } else {
          r |= blend_rotateab;
        }
        const uint32_t elementsize = sizeof( V ) / N;
        r |= uint64_t( ( rot & ( lanesize - 1 ) ) * elementsize ) << blend_rotpattern;
      }
#endif
      if ( lanesize == 4 ) {
        // check if it fits shufps
        r |= blend_shufab | blend_shufba;
        for ( ii = 0; ii < 2; ii++ ) {
          ix = lanepattern[ii];
          if ( ix >= 0 ) {
            if ( ix & N )
              r &= ~blend_shufab;
            else
              r &= ~blend_shufba;
          }
        }
        for ( ; ii < 4; ii++ ) {
          ix = lanepattern[ii];
          if ( ix >= 0 ) {
            if ( ix & N )
              r &= ~blend_shufba;
            else
              r &= ~blend_shufab;
          }
        }
        if ( r & ( blend_shufab | blend_shufba ) ) { // fits shufps/shufpd
          uint8_t shufpattern = 0;                   // get pattern
          for ( iu = 0; iu < lanesize; iu++ ) { shufpattern |= ( lanepattern[iu] & 3 ) << iu * 2; }
          r |= (uint64_t)shufpattern << blend_shufpattern; // return pattern
        }
      }
    } else if ( nlanes > 1 ) { // not same pattern in all lanes
      rot = 999;               // check if it fits big rotate
      for ( ii = 0; ii < N; ii++ ) {
        ix = a[ii];
        if ( ix >= 0 ) {
          uint32_t rot2 = ( ix + 2 * N - ii ) % ( 2 * N ); // rotate count
          if ( rot == 999 ) {
            rot = rot2; // save rotate count
          } else if ( rot != rot2 ) {
            rot = 1000;
            break; // does not fit big rotate
          }
        }
      }
      if ( rot < 2 * N ) { // fits big rotate
        r |= blend_rotate_big | (uint64_t)rot << blend_rotpattern;
      }
    }
    if ( lanesize == 2 && ( r & ( blend_shufab | blend_shufba ) ) ) { // fits shufpd. Get pattern
      for ( ii = 0; ii < N; ii++ ) { r |= uint64_t( a[ii] & 1 ) << ( blend_shufpattern + ii ); }
    }
    return r;
  }
 
  // blend_perm_indexes: return an Indexlist for implementing a blend function as
  // two permutations. N = vector size.
  // dozero = 0: let unused elements be don't care. The two permutation results must be blended
  // dozero = 1: zero unused elements in each permuation. The two permutation results can be OR'ed
  // dozero = 2: indexes that are -1 or V_DC are preserved
  template <int N, int dozero>
  constexpr EList<int, 2 * N> blend_perm_indexes( int const ( &a )[N] ) {
    // a is a reference to a constexpr array of permutation indexes
    EList<int, 2 * N> list = { { 0 } };          // list to return
    int               u    = dozero ? -1 : V_DC; // value to use for unused entries
    int               j    = 0;
 
    for ( j = 0; j < N; j++ ) { // loop through indexes
      int ix = a[j];            // current index
      if ( ix < 0 ) {           // zero or don't care
        if ( dozero == 2 ) {
          // list.a[j] = list.a[j + N] = ix;  // fails in gcc in complicated cases
          list.a[j]     = ix;
          list.a[j + N] = ix;
        } else {
          // list.a[j] = list.a[j + N] = u;
          list.a[j]     = u;
          list.a[j + N] = u;
        }
      } else if ( ix < N ) { // value from a
        list.a[j]     = ix;
        list.a[j + N] = u;
      } else {
        list.a[j]     = u; // value from b
        list.a[j + N] = ix - N;
      }
    }
    return list;
  }
 
  // largeblock_indexes: return indexes for replacing a permute or blend with a
  // certain block size by a permute or blend with the double block size.
  // Note: it is presupposed that perm_flags or blend_flags indicates _largeblock
  // It is required that additional zeroing is added if perm_flags or blend_flags
  // indicates _addz
  template <int N>
  constexpr EList<int, N / 2> largeblock_indexes( int const ( &a )[N] ) {
    // Parameter a is a reference to a constexpr array of N permutation indexes
    EList<int, N / 2> list = { { 0 } }; // list to return
 
    bool fit_addz = false; // additional zeroing needed at the lower block level
    int  ix       = 0;     // even index
    int  iy       = 0;     // odd index
    int  iz       = 0;     // combined index
    int  i        = 0;     // loop counter
 
    for ( i = 0; i < N; i += 2 ) {
      ix = a[i];     // even index
      iy = a[i + 1]; // odd index
      if ( ix >= 0 ) {
        iz = ix / 2; // half index
      } else if ( iy >= 0 ) {
        iz = iy / 2; // half index
      } else
        iz = ix | iy;     // -1 or V_DC. -1 takes precedence
      list.a[i / 2] = iz; // save to list
      // check if additional zeroing is needed at current block size
      if ( ( ix == -1 && iy >= 0 ) || ( iy == -1 && ix >= 0 ) ) { fit_addz = true; }
    }
    // replace -1 by V_DC if fit_addz
    if ( fit_addz ) {
      for ( i = 0; i < N / 2; i++ ) {
        if ( list.a[i] < 0 ) list.a[i] = V_DC;
      }
    }
    return list;
  }
 
  /****************************************************************************************
   *
   *          Vector blend helper function templates
   *
   * These templates are for emulating a blend with a vector size that is not supported by
   * the instruction set, using multiple blends or permutations of half the vector size
   *
   ****************************************************************************************/
 
  // Make dummy blend function templates to avoid error messages when the blend funtions are not yet defined
  template <typename dummy>
  void blend2() {}
  template <typename dummy>
  void blend4() {}
  template <typename dummy>
  void blend8() {}
  template <typename dummy>
  void blend16() {}
  template <typename dummy>
  void blend32() {}
 
  // blend_half_indexes: return an Indexlist for emulating a blend function as
  // blends or permutations from multiple sources
  // dozero = 0: let unused elements be don't care. Multiple permutation results must be blended
  // dozero = 1: zero unused elements in each permuation. Multiple permutation results can be OR'ed
  // dozero = 2: indexes that are -1 or V_DC are preserved
  // src1, src2: sources to blend in a partial implementation
  template <int N, int dozero, int src1, int src2>
  constexpr EList<int, N> blend_half_indexes( int const ( &a )[N] ) {
    // a is a reference to a constexpr array of permutation indexes
    EList<int, N> list = { { 0 } };          // list to return
    int           u    = dozero ? -1 : V_DC; // value to use for unused entries
    int           j    = 0;                  // loop counter
 
    for ( j = 0; j < N; j++ ) { // loop through indexes
      int ix = a[j];            // current index
      if ( ix < 0 ) {           // zero or don't care
        list.a[j] = ( dozero == 2 ) ? ix : u;
      } else {
        int src = ix / N; // source
        if ( src == src1 ) {
          list.a[j] = ix & ( N - 1 );
        } else if ( src == src2 ) {
          list.a[j] = ( ix & ( N - 1 ) ) + N;
        } else
          list.a[j] = u;
      }
    }
    return list;
  }
 
  // selectblend: select one of four sources for blending
  template <typename W, int s>
  static inline auto selectblend( W const a, W const b ) {
    if constexpr ( s == 0 )
      return a.get_low();
    else if constexpr ( s == 1 )
      return a.get_high();
    else if constexpr ( s == 2 )
      return b.get_low();
    else
      return b.get_high();
  }
 
  // blend_half: Emulate a blend with a vector size that is not supported
  // by multiple blends with half the vector size.
  // blend_half is called twice, to give the low and high half of the result
  // Parameters: W: type of full-size vector
  // i0...: indexes for low or high half
  // a, b: full size input vectors
  // return value: half-size vector for lower or upper part
  template <typename W, int... i0>
  auto blend_half( W const& a, W const& b ) {
    typedef decltype( a.get_low() ) V;             // type for half-size vector
    constexpr int                   N = V::size(); // size of half-size vector
    static_assert( sizeof...( i0 ) == N, "wrong number of indexes in blend_half" );
    constexpr int ind[N] = { i0... }; // array of indexes
 
    // lambda to find which of the four possible sources are used
    // return: EList<int, 5> containing a list of up to 4 sources. The last element is the number of sources used
    auto listsources = []( int const n, int const( &ind )[N] ) constexpr {
      bool source_used[4] = { false, false, false, false }; // list of sources used
      int  i              = 0;
      for ( i = 0; i < n; i++ ) {
        int ix = ind[i]; // index
        if ( ix >= 0 ) {
          int src              = ix / n; // source used
          source_used[src & 3] = true;
        }
      }
      // return a list of sources used. The last element is the number of sources used
      EList<int, 5> sources = { { 0 } };
      int           nsrc    = 0; // number of sources
      for ( i = 0; i < 4; i++ ) {
        if ( source_used[i] ) { sources.a[nsrc++] = i; }
      }
      sources.a[4] = nsrc;
      return sources;
    };
    // list of sources used
    constexpr EList<int, 5> sources = listsources( N, ind );
    constexpr int           nsrc    = sources.a[4]; // number of sources used
 
    if constexpr ( nsrc == 0 ) { // no sources
      return V( 0 );
    }
    // get indexes for the first one or two sources
    constexpr int           uindex = ( nsrc > 2 ) ? 1 : 2; // unused elements set to zero if two blends are combined
    constexpr EList<int, N> L      = blend_half_indexes<N, uindex, sources.a[0], sources.a[1]>( ind );
    V                       x0;
    V                       src0 = selectblend<W, sources.a[0]>( a, b ); // first source
    V                       src1 = selectblend<W, sources.a[1]>( a, b ); // second source
    if constexpr ( N == 2 ) {
      x0 = blend2<L.a[0], L.a[1]>( src0, src1 );
    } else if constexpr ( N == 4 ) {
      x0 = blend4<L.a[0], L.a[1], L.a[2], L.a[3]>( src0, src1 );
    } else if constexpr ( N == 8 ) {
      x0 = blend8<L.a[0], L.a[1], L.a[2], L.a[3], L.a[4], L.a[5], L.a[6], L.a[7]>( src0, src1 );
    } else if constexpr ( N == 16 ) {
      x0 = blend16<L.a[0], L.a[1], L.a[2], L.a[3], L.a[4], L.a[5], L.a[6], L.a[7], L.a[8], L.a[9], L.a[10], L.a[11],
                   L.a[12], L.a[13], L.a[14], L.a[15]>( src0, src1 );
    } else if constexpr ( N == 32 ) {
      x0 = blend32<L.a[0], L.a[1], L.a[2], L.a[3], L.a[4], L.a[5], L.a[6], L.a[7], L.a[8], L.a[9], L.a[10], L.a[11],
                   L.a[12], L.a[13], L.a[14], L.a[15], L.a[16], L.a[17], L.a[18], L.a[19], L.a[20], L.a[21], L.a[22],
                   L.a[23], L.a[24], L.a[25], L.a[26], L.a[27], L.a[28], L.a[29], L.a[30], L.a[31]>( src0, src1 );
    }
    if constexpr ( nsrc > 2 ) { // get last one or two sources
      constexpr EList<int, N> M = blend_half_indexes<N, 1, sources.a[2], sources.a[3]>( ind );
      V                       x1;
      V                       src2 = selectblend<W, sources.a[2]>( a, b ); // third source
      V                       src3 = selectblend<W, sources.a[3]>( a, b ); // fourth source
      if constexpr ( N == 2 ) {
        x1 = blend2<M.a[0], M.a[1]>( src0, src1 );
      } else if constexpr ( N == 4 ) {
        x1 = blend4<M.a[0], M.a[1], M.a[2], M.a[3]>( src2, src3 );
      } else if constexpr ( N == 8 ) {
        x1 = blend8<M.a[0], M.a[1], M.a[2], M.a[3], M.a[4], M.a[5], M.a[6], M.a[7]>( src2, src3 );
      } else if constexpr ( N == 16 ) {
        x1 = blend16<M.a[0], M.a[1], M.a[2], M.a[3], M.a[4], M.a[5], M.a[6], M.a[7], M.a[8], M.a[9], M.a[10], M.a[11],
                     M.a[12], M.a[13], M.a[14], M.a[15]>( src2, src3 );
      } else if constexpr ( N == 32 ) {
        x1 = blend32<M.a[0], M.a[1], M.a[2], M.a[3], M.a[4], M.a[5], M.a[6], M.a[7], M.a[8], M.a[9], M.a[10], M.a[11],
                     M.a[12], M.a[13], M.a[14], M.a[15], M.a[16], M.a[17], M.a[18], M.a[19], M.a[20], M.a[21], M.a[22],
                     M.a[23], M.a[24], M.a[25], M.a[26], M.a[27], M.a[28], M.a[29], M.a[30], M.a[31]>( src2, src3 );
      }
      x0 |= x1; // combine result of two blends. Unused elements are zero
    }
    return x0;
  }
 
#ifdef VCL_NAMESPACE
}
#endif