package fr32

import (
	"math/bits"
	"runtime"
	"sync"

	"github.com/filecoin-project/specs-actors/actors/abi"
)

var mtTresh = 32 << 20

func mt(in, out []byte, padLen int, op func(in, out []byte)) {
	threads := padLen / mtTresh
	if threads > runtime.NumCPU() {
		threads = 1 << (32 - bits.LeadingZeros32(uint32(runtime.NumCPU())))
	}
	threadBytes := abi.PaddedPieceSize(padLen / threads)

	var wg sync.WaitGroup
	wg.Add(threads)

	for i := 0; i < threads; i++ {
		go func(thread int) {
			defer wg.Done()

			start := threadBytes * abi.PaddedPieceSize(thread)
			end := start + threadBytes

			op(in[start.Unpadded():end.Unpadded()], out[start:end])
		}(i)
	}
	wg.Wait()
}

// Assumes len(in)%127==0 and len(out)%128==0
func Pad(in, out []byte) {
	if len(out) > mtTresh {
		mt(in, out, len(out), Pad)
		return
	}

	chunks := len(out) / 128
	for chunk := 0; chunk < chunks; chunk++ {
		inOff := chunk * 127
		outOff := chunk * 128

		copy(out[outOff:outOff+31], in[inOff:inOff+31])

		t := in[inOff+31] >> 6
		out[outOff+31] = in[inOff+31] & 0x3f
		var v byte

		for i := 32; i < 64; i++ {
			v = in[inOff+i]
			out[outOff+i] = (v << 2) | t
			t = v >> 6
		}

		t = v >> 4
		out[outOff+63] &= 0x3f

		for i := 64; i < 96; i++ {
			v = in[inOff+i]
			out[outOff+i] = (v << 4) | t
			t = v >> 4
		}

		t = v >> 2
		out[outOff+95] &= 0x3f

		for i := 96; i < 127; i++ {
			v = in[inOff+i]
			out[outOff+i] = (v << 6) | t
			t = v >> 2
		}

		out[outOff+127] = t & 0x3f
	}
}

// Assumes len(in)%128==0 and len(out)%127==0
func Unpad(in []byte, out []byte) {
	if len(out) > mtTresh {
		mt(in, out, len(in), Unpad)
		return
	}

	chunks := len(in) / 128
	for chunk := 0; chunk < chunks; chunk++ {
		inOffNext := chunk*128 + 1
		outOff := chunk * 127

		at := in[chunk*128]

		for i := 0; i < 32; i++ {
			next := in[i+inOffNext]

			out[outOff+i] = at
			//out[i] |= next << 8

			at = next
		}

		out[outOff+31] |= at << 6

		for i := 32; i < 64; i++ {
			next := in[i+inOffNext]

			out[outOff+i] = at >> 2
			out[outOff+i] |= next << 6

			at = next
		}

		out[outOff+63] ^= (at << 6) ^ (at << 4)

		for i := 64; i < 96; i++ {
			next := in[i+inOffNext]

			out[outOff+i] = at >> 4
			out[outOff+i] |= next << 4

			at = next
		}

		out[outOff+95] ^= (at << 4) ^ (at << 2)

		for i := 96; i < 127; i++ {
			next := in[i+inOffNext]

			out[outOff+i] = at >> 6
			out[outOff+i] |= next << 2

			at = next
		}
	}
}