reverse calldata in chunks of 32 bytes

rev-calldata-chunks.huff

// Reverse bytes received in calldata (abcd -> dcba)
// https://twitter.com/huff_language/status/1583894073487654913
// code length 206

/// @author Philippe Dumonet <[email protected]> -- https://twitter.com/real_philogy/status/1584304102418223104
/// @author karma (@0xkarmacoma) -- https://twitter.com/0xkarmacoma/status/1584239664310779904
/// @author kaden.eth (@0xKaden) -- https://twitter.com/0xKaden/status/1584280521089376256/
#define macro reverse_word() = takes(1) returns(1) {
  //               [x0]
  0x00ff00ff00ff00ff00ff00ff00ff00ff00ff00ff00ff00ff00ff00ff00ff00ff
  dup2 dup2 and // [x1r, mask1, x0]
  0x08 shl      // [x1r', mask1, x0]
  swap2         // [x1, mask1, x1r']
  0x08 shr and  // [x1l', x1r']
  or            // [x1]

  0x0000ffff0000ffff0000ffff0000ffff0000ffff0000ffff0000ffff0000ffff
  dup2 dup2 and
  0x10 shl
  swap2
  0x10 shr and
  or            // [x2]

  0x00000000ffffffff00000000ffffffff00000000ffffffff00000000ffffffff
  dup2 dup2 and
  0x20 shl
  swap2
  0x20 shr and
  or            // [x4]

  0x0000000000000000ffffffffffffffff0000000000000000ffffffffffffffff
  dup2 dup2 and
  0x40 shl
  swap2
  0x40 shr and
  or            // [x8]

  dup1
  0x80 shl
  swap1
  0x80 shr
  or
}

/// @dev 0xffffffe0 only accomodates ~4 GB calldata max
#define macro start_chunk_offset() = takes(0) returns(1) {
  0x20                  // [0x20]
  calldatasize 0x1f add // [cds + 31, 0x20]
  0xffffffe0 and        // [(cds + 31) / 32 * 32, 0x20]
  sub                   // [(cds + 31) / 32 * 32 - 32]
}

#define macro MAIN() = takes(0) returns(0) {
  start_chunk_offset() // [start_chunk_offset = 32 * num_chunks - 32]
  calldatasize         // [cds, start_chunk_offset]
  calldatasize         // [cds, cds, start_chunk_offset]

  next_iter:
    msize dup4 sub     // [cd_offset, cds, cds, start_chunk_offset]
    calldataload       // [cd[i'], cds, cds, start_chunk_offset]
    reverse_word()     // [rev(cd[i']), cds, cds, start_chunk_offset]
    msize mstore       // [cds, cds, start_chunk_offset]
  calldatasize msize lt next_iter jumpi

  msize sub return
}