libbps-suf.cpp

//Module name: libbps-suf
//Author: Alcaro
//Date: See Git history
//Licence: GPL v3.0 or higher


#include "libbps.h"
#include "crc32.h"
#include <stdint.h>
#include <stdlib.h>
#include <string.h>
#include <stdio.h>

//These two give minor performance penalties and will print some random stuff to stdout.
//The former will verify the correctness of the output patch, the latter will print some performance data.
//Can be useful for debugging, but should be disabled for release builds.
#ifdef BPS_STANDALONE
#endif
//#define TEST_CORRECT
//#define TEST_PERF

//If the suffix array of [0, 0, 0, 0] is [3, 2, 1, 0], set to true. If it's [0, 1, 2, 3], this is false.
//If it's [4, 3, 2, 1, 0] or [0, 1, 2, 3, 4], remove the 4 (easily done with some pointer math), and follow the above.
//If it's something else, get a non-broken array calculator.
#define EOF_IS_LAST false

#if defined(TEST_CORRECT) || defined(TEST_PERF)
#include <stdio.h>
#endif

//Algorithm description:
//
//This is heavily built upon suffix sorting; the implementation I use, libdivsufsort, claims
// O(n log n) complexity, so I'll believe that. There is also SA-IS, which claims O(n), but if that
// is true, its constant factors are ridiculously high.
//
//The program starts by taking an equal amount of the source file and target file, concatenates that
// with target first, and suffix sorts it.
//It also calculates a reverse index, such that reverse[sorted[i]]==i.
//
//To find a match, it goes to reverse[outpos], and scans sorted[] up and down for the closest entry
// that either starts before the current output position, or is somewhere in the source file.
//As the source file comes last, the end-of-file marker (whose value is outside the range of a byte)
// is guaranteed to not be in the way for a better match.
//This is called O(n) times, and averages O(1) as at least 50% of sorted[] is in range. However, it
// is worst-case O(n) for sorted inputs, giving a total of O(n^2).
//
//It then checks which of the two candidates are superior, by checking how far they match each
// other, and then checking if the upper one has another correct byte.
//This is potentially O(n), but for each matched byte, another iteration is removed from the outer
// loop, so the sum of all calls is O(n).
//
//When the program approaches the end of the sorted area, it re-sorts twice as much as last time.
// This gives O(log n) calls to the suffix sorter.
//Given O(n log n) for one sorting step, the time taken is O(n/1 log n/1 + n/2 log n/2 +
// n/4 log n/4 + ...), which is strictly less than O(n/1 log n + n/2 log n + n/4 log n + ...), which
// equals O(2n log n), which is O(n log n). (The exact value of that infinite sum is 2n*log(n/2).)
//
//Many details were omitted from the above, but that's the basic setup.
//
//Thus, the program is O(max(n log n, n, n) = n log n) average and O(max(n log n, n^2, n) = n^2)
// worst case.
//
//I conclude that the task of finding, understanding and implementing a sub-O(n^2) algorithm for
// delta patching is resolved.


//Known cases where this function does not emit the optimal encoding:
//If a match in the target file would extend further than target_search_size, it is often skipped.
// Penalty: O(log n), with extremely low constants (it'd require a >256B match to be exactly there).
// Even for big files, the penalty is very likely to remain zero; even hitting double-digit bytes
// would require a file designed exactly for that.
//If multiple matches are equally good, it picks one at random, not the one that's cheaper to encode.
// Penalty: Likely O(n) or O(n log log n), with low constants. I'd guess ~1.4% for my 48MB test file.
//However, due to better heuristics and others' performance optimizations, this one still beats its
// competitors.


//Possible optimizations:
//divsufsort() takes approximately 2/3 of the total time. create_reverse_index() takes roughly a third of the remainder.
//Each iteration takes four times as long as the previous one.
//If each iteration takes 4 times as long as the previous one, then the last one takes 3/4 of the total time.
//Since divsufsort+create_reverse_index doesn't depend on anything else, the last iteration can be split off to its own thread.
//This would split it to
//Search, non-final:   2/9 * 1/4 = 2/36
//Search, final:       2/9 * 3/4 = 6/36
//Sort+rev, non-final: 7/9 * 1/4 = 7/36
//Sort+rev, final:     7/9 * 3/4 = 21/36
//All non-final must be done sequentially. Both Sort Final and non-final must be done before Search Final can start.
//This means the final time, if Sort Final is split off, is
//max(7/36+2/36, 21/36) + 6/36 = 27/36 = 3/4
//of the original time.
//Due to
//- the considerable complexity costs (OpenMP doesn't seem able to represent the "insert a wait in
//   the middle of this while loop" I would need)
//- the added memory use, approximately 25% higher - it's already high enough
//- libdivsufsort already using threads, which would make the gains lower
//   and would increase complexity, as I have to ensure the big one remains threaded -
//    and that the small ones are not, as that'd starve the big one
//I deem a possible 25% boost not worthwhile.


//Both sorting algorithms claim O(1) memory use (in addition to the bytes and the output). In
// addition to that, this algorithm uses (source.len*target.len)*(sizeof(uint8_t)+2*sizeof(off_t))
// bytes of memory, plus the patch (the input/output files are read from disk).
//For most hardware, this is 9*(source.len+target.len), or 5*(source+target) for the slim one.


//I don't need 64bit support, it'd take 20GB RAM and way too long.
//#include "sais.cpp"
//template<typename sais_index_type>
//static void sufsort(sais_index_type* SA, const uint8_t* T, sais_index_type n) {
//  if(n <= 1) { if(n == 1) SA[0] = 0; return; }
//  sais_main<sais_index_type>(T, SA, 0, n, 256);
//}

//According to <https://code.google.com/p/libdivsufsort/wiki/SACA_Benchmarks>, divsufsort achieves
// approximately half the time of SAIS for nearly all files, despite SAIS' promises of linear
// performance (divsufsort claims O(n log n)).

//divsufsort only allocates O(1) for some radix/bucket sorting. SAIS seems constant too.
//I'd prefer to let them allocate from an array I give it, but divsuf doesn't allow that, and there
// are only half a dozen allocations per call anyways.

#include "divsufsort.h"
static void sufsort(int32_t* SA, uint8_t* T, int32_t n)
{
	divsufsort(T, SA, n);
}

#ifdef USE_DIVSUFSORT64
#include "divsufsort64.h"

static void sufsort(int64_t* SA, uint8_t* T, int64_t n)
{
	divsufsort(T, SA, n);
}
#endif



template<typename T> static T min(T a, T b) { return a<b ? a : b; }
template<typename T> static T max(T a, T b) { return a<b ? b : a; }



namespace {
struct bps_creator {
	uint8_t* out;
	size_t outlen;
	size_t outbuflen;
	
	void reserve(size_t len)
	{
		if (outlen+len > outbuflen)
		{
			if (!outbuflen) outbuflen = 128;
			while (outlen+len > outbuflen) outbuflen *= 2;
			out = (uint8_t*)realloc(out, outbuflen);
		}
	}
	
	void append(const uint8_t * data, size_t len)
	{
		reserve(len);
		memcpy(out+outlen, data, len);
		outlen+=len;
	}
	
	void appendnum(size_t num)
	{
#ifdef TEST_CORRECT
		if (num > 1000000000)
			printf("ERROR: Attempt to write %.8lX\n",(unsigned long)num),abort();
#endif
		reserve(sizeof(size_t)*8/7+1);
		
		while (num >= 128)
		{
			out[outlen++]=(num&0x7F);
			num>>=7;
			num--;
		}
		out[outlen++]=num|0x80;
	}
	
	void appendnum32(uint32_t num)
	{
		reserve(4);
		out[outlen++] = num>>0;
		out[outlen++] = num>>8;
		out[outlen++] = num>>16;
		out[outlen++] = num>>24;
	}
	
	static size_t maxsize()
	{
		return SIZE_MAX>>2; // can be reduced to SIZE_MAX>>1 by amending append_cmd, but the mallocs overflow at that point anyways.
	}
	
	size_t sourcelen;
	size_t targetlen;
	const uint8_t* targetmem;
	
	enum bpscmd { SourceRead, TargetRead, SourceCopy, TargetCopy };
	
	size_t outpos;
	
	size_t sourcecopypos;
	size_t targetcopypos;
	
	size_t numtargetread;
	
	bps_creator(file* source, file* target, struct mem metadata)
	{
		outlen = 0;
		outbuflen = 128;
		out = (uint8_t*)malloc(outbuflen);
		
		outpos = 0;
		
		sourcelen = source->len();
		targetlen = target->len();
		
		sourcecopypos = 0;
		targetcopypos = 0;
		
		numtargetread = 0;
		
		append((const uint8_t*)"BPS1", 4);
		appendnum(sourcelen);
		appendnum(targetlen);
		appendnum(metadata.len);
		append(metadata.ptr, metadata.len);
		
		setProgress(NULL, NULL);
	}
	
	
	void move_target(const uint8_t* ptr)
	{
		targetmem = ptr;
	}
	
	size_t encode_delta(size_t prev, size_t next)
	{
		bool negative = (next<prev);
		size_t offset = negative ? prev-next : next-prev;
		return (negative?1:0) | (offset<<1);
	}
	
	void append_delta(size_t prev, size_t next)
	{
		appendnum(encode_delta(prev, next));
	}
	
	void append_cmd(bpscmd command, size_t count)
	{
		appendnum((count-1)<<2 | command);
	}
	
	void flush_target_read()
	{
		if (!numtargetread) return;
		append_cmd(TargetRead, numtargetread);
		append(targetmem+outpos-numtargetread, numtargetread);
		numtargetread = 0;
	}
	
	size_t emit_source_copy(size_t location, size_t count)
	{
		if (location == outpos) return emit_source_read(location, count);
		flush_target_read();
		append_cmd(SourceCopy, count);
		append_delta(sourcecopypos, location);
		sourcecopypos = location+count;
		outpos += count;
		return count;
	}
	
	size_t emit_source_read(size_t location, size_t count)
	{
		flush_target_read();
#ifdef TEST_CORRECT
		if (location != outpos)
			puts("ERROR: SourceRead not from source pointer"),abort();
#endif
		append_cmd(SourceRead, count);
		outpos+=count;
		return count;
	}
	
	size_t emit_target_copy(size_t location, size_t count)
	{
		flush_target_read();
		append_cmd(TargetCopy, count);
		append_delta(targetcopypos, location);
		targetcopypos = location+count;
		outpos += count;
		return count;
	}
	
	size_t emit_target_read()
	{
		numtargetread++;
		outpos++;
		return 1;
	}
	
	
	size_t abs_diff(size_t a, size_t b)
	{
		return (b<a) ? (a-b) : (b-a);
	}
	size_t num_cost(size_t num)
	{
		if (num<128) return 1;
		if (num<128*128) return 2; // 32KB
		if (num<128*128*128) return 3; // 2MB
		if (num<128*128*128*128) return 4; // 256MB
		return 5; // 128^5 is 32GB, let's just assume the sizes don't go any higher...
	}
	
	bool use_match(bool hastargetread, size_t cost, size_t len)
	{
		//numbers calculated via trial and error; checking for each cost, optimizing 'len' for each, and checking what happens
		//then a pattern was identified and used
		//yes, it looks weird
		return len >= 1+cost+hastargetread+(len==1);
	}
	
	
	//Return value is how many bytes were used. If you believe the given one sucks, use TargetRead and return 1.
	size_t match(bool is_target, size_t pos, size_t len)
	{
		if (!use_match(
		     numtargetread,
		     (!is_target && pos==outpos) ? 1 : // SourceRead
		         (num_cost(abs_diff(pos, (is_target ? targetcopypos : sourcecopypos)))+1),
		     len
		   ))
		{
			return emit_target_read();
		}
		
		if (is_target) return emit_target_copy(pos, len);
		else return emit_source_copy(pos, len);
	}
	
	
	bool (*prog_func)(void* userdata, size_t done, size_t total);
	void* prog_dat;
	
	static bool prog_func_null(void* userdata, size_t done, size_t total) { return true; }
	
	void setProgress(bool (*progress)(void* userdata, size_t done, size_t total), void* userdata)
	{
		if (!progress) progress = prog_func_null;
		
		prog_func=progress;
		prog_dat=userdata;
	}
	
	bool progress(size_t done, size_t total)
	{
		return prog_func(prog_dat, done, total);
	}
	
	
	void finish(const uint8_t* source, const uint8_t* target)
	{
		flush_target_read();
#ifdef TEST_CORRECT
		if (outpos != targetlen)
			puts("ERROR: patch creates wrong ROM size"),abort();
#endif
		
		appendnum32(crc32(source, sourcelen));
		appendnum32(crc32(target, targetlen));
		appendnum32(crc32(out, outlen));
	}
	
	struct mem getpatch()
	{
		struct mem ret = { out, outlen };
		out = NULL;
		return ret;
	}
	
	~bps_creator() { free(out); }
};
}



#ifdef TEST_PERF
static int match_len_n=0;
static int match_len_tot=0;
#endif

template<typename off_t>
static off_t match_len(const uint8_t* a, const uint8_t* b, off_t len)
{
	off_t i;
	for (i=0;i<len && a[i]==b[i];i++) {}
#ifdef TEST_PERF
	match_len_n++;
	match_len_tot+=i;
#endif
	return i;
}

//This one assumes that the longest common prefix of 'a' and 'b' is shared also by 'search'.
//In practice, lexographically, a < search < b, which is a stronger guarantee.
template<typename off_t>
static off_t pick_best_of_two(const uint8_t* search, off_t searchlen,
                              const uint8_t* data, off_t datalen,
                              off_t a, off_t b,
                              off_t* bestlen)
{
	off_t commonlen = match_len(data+a, data+b, min(datalen-a, datalen-b));
	if (commonlen>=searchlen)
	{
		*bestlen=searchlen;
		return a;
	}
	
	if (a+commonlen<datalen && search[commonlen]==data[a+commonlen])
	{
		// a is better
		*bestlen = commonlen + match_len(search+commonlen, data+a+commonlen, min(searchlen, datalen-a)-commonlen);
		return a;
	}
	else
	{
		// b is better, or they're equal
		*bestlen = commonlen + match_len(search+commonlen, data+b+commonlen, min(searchlen, datalen-b)-commonlen);
		return b;
	}
}

//This one takes a match, which is assumed optimal, and looks for the lexographically closest one
// that either starts before 'maxstart', or starts at or after 'minstart'.
template<typename off_t>
static off_t adjust_match(off_t match, const uint8_t* search, off_t searchlen,
                          const uint8_t* data,off_t datalen, off_t maxstart,off_t minstart,
                          const off_t* sorted, off_t sortedlen,
                          off_t* bestlen)
{
	off_t match_up = match;
	off_t match_dn = match;
	while (match_up>=0        && sorted[match_up]>=maxstart && sorted[match_up]<minstart) match_up--;
	while (match_dn<sortedlen && sorted[match_dn]>=maxstart && sorted[match_dn]<minstart) match_dn++;
	if (match_up<0 || match_dn>=sortedlen)
	{
		if (match_up<0 && match_dn>=sortedlen)
		{
			*bestlen=0;
			return 0;
		}
		off_t pos = sorted[match_up<0 ? match_dn : match_up];
		*bestlen = match_len(search, data+pos, min(searchlen, datalen-pos));
		return pos;
	}
	
	return pick_best_of_two(search,searchlen, data,datalen, sorted[match_up],sorted[match_dn], bestlen);
}



static uint16_t read2_uc(const uint8_t* data)
{
	return data[0]<<8 | data[1];
}

template<typename off_t>
static uint16_t read2(const uint8_t* data, off_t len)
{
	if (len>=2) return read2_uc(data);
	else
	{
		uint16_t out = (EOF_IS_LAST ? 0xFFFF : 0x0000);
		if (len==1) out = (data[0]<<8) | (out&0x00FF);
		return out;
	}
}

template<typename off_t>
static void create_buckets(const uint8_t* data, off_t* index, off_t len, off_t* buckets)
{
	off_t low = 0;
	off_t high;
	
	for (int n=0;n<65536;n++)
	{
		//'low' remains from the previous iteration and is a known minimum
		high = low+(len/131072)+1; // optimal value: slightly above a third of the distance to the next one
		while (true)
		{
			if (high > len-1) break;
			
			off_t pos = index[high];
			uint16_t here = read2(data+pos, len-pos);
			
			if (here >= n) break;
			else
			{
				off_t diff = high-low;
				low = high;
				high = high+diff*2;
			}
		}
		if (high > len-1) high = len-1;
		
		
		while (low < high)
		{
			off_t mid = low + (high-low)/2;
			off_t midpos = index[mid];
			
			uint16_t here = read2(data+midpos, len-midpos);
			if (here < n) low = mid+1;
			else high = mid;
		}
		buckets[n] = low;
	}
	
	buckets[65536] = len;
	
#ifdef TEST_CORRECT
	if (buckets[0]!=0)
	{
		printf("e: buckets suck, [0]=%i\n", buckets[0]);
		abort();
	}
	for (int n=0;n<65536;n++)
	{
		off_t low = buckets[n];
		off_t high = buckets[n+1];
		for (off_t i=low;i<high;i++)
		{
			if (read2(data+index[i], len-index[i])!=n)
			{
				printf("e: buckets suck, %i != (%i)[%i]%i [%i-%i]", n, i,index[i],read2(data+index[i],len-index[i]),low,high);
				abort();
			}
		}
		//printf("%i:[%i]%i\n",n,low,read2(data+index[low],len-low));
	}
#endif
}

template<typename off_t>
static off_t find_index(off_t pos, const uint8_t* data, off_t datalen, const off_t* index, const off_t* reverse, off_t* buckets)
{
	if (reverse) return reverse[pos];
	
	uint16_t bucket = read2(data+pos, datalen-pos);
//printf("p=%i b=%i\n",pos,bucket);
	
	off_t low = buckets[bucket];
	off_t high = buckets[bucket+1]-1;
	
	off_t lowmatch = 2;
	off_t highmatch = 2;
	
//printf("b=%i r=%i(%i)-%i(%i)\n",bucket,low,read2(data+index[low],datalen-index[low]),high,read2(data+index[high],datalen-index[high]));
//fflush(stdout);
	while (true)
	{
		off_t mid = low + (high-low)/2;
		off_t midpos = index[mid];
		if (midpos == pos) return mid;
//printf("r=[%i]%i-%i \n",high-low,low,high,);
//fflush(stdout);
#ifdef TEST_CORRECT
		if (low >= high)
		{
			printf("E: [%i](%i): stuck at %i(%i)-%i(%i)\n", pos, read2_uc(data+pos),
				low, read2_uc(data+index[low]), high, read2_uc(data+index[high]));
			int n=0;
			while (index[n]!=pos) n++;
			printf("correct one is %i(%i)\n",n, read2_uc(data+index[n]));
			abort();
		}
#endif
		
		off_t matchlenstart = min(lowmatch, highmatch);
		
		off_t len = datalen - max(pos, midpos) - matchlenstart;
		
		const uint8_t* search = data+pos+matchlenstart;
		const uint8_t* here = data+midpos+matchlenstart;
		
		while (len>0 && *search==*here)
		{
			search++;
			here++;
			len--;
		}
		
		off_t matchlen = search-data-pos;
		
		bool less;
		if (len > 0) less = (*here<*search);
		else less = (here > search) ^ EOF_IS_LAST;
		
		if (less)
		{
			low = mid+1;
			lowmatch = matchlen;
		}
		else
		{
			high = mid-1;
			highmatch = matchlen;
		}
		
		if (low+256 > high)
		{
			off_t i=low;
			while (true)
			{
				if (index[i]==pos) return i;
				i++;
			}
		}
	}
}


template<typename off_t>
static void create_reverse_index(off_t* index, off_t* reverse, off_t len)
{
//testcase: linux 3.18.14 -> 4.0.4 .xz
//without: real23.544 user32.930
//with:    real22.636 user40.168
//'user' jumps up quite a lot, while 'real' only moves a short bit
//I'm not sure why the tradeoff is so bad (do the cachelines bounce THAT badly?), but I deem it not worth it.
//#pragma omp parallel for
	for (off_t i=0;i<len;i++) reverse[index[i]]=i;
}

template<typename off_t>
static off_t nextsize(off_t outpos, off_t sortedsize, off_t targetlen)
{
	while (outpos >= sortedsize-256 && sortedsize < targetlen)
		sortedsize = min(sortedsize*4+3, targetlen);
	return sortedsize;
}

template<typename off_t>
off_t lerp(off_t x, off_t y, float frac)
{
	return x + (y-x)*frac;
}

template<typename off_t>
static bpserror bps_create_suf_core(file* source, file* target, bool moremem, struct bps_creator * out)
{
#define error(which) do { err = which; goto error; } while(0)
	bpserror err;
	
	size_t realsourcelen = source->len();
	size_t realtargetlen = target->len();
	
	size_t overflowtest = realsourcelen + realtargetlen;
	
	//source+target length is bigger than size_t (how did that manage to get allocated?)
	if (overflowtest < realsourcelen) return bps_too_big;
	
	//source+target doesn't fit in unsigned off_t
	if ((size_t)(off_t)overflowtest != overflowtest) return bps_too_big;
	
	//source+target doesn't fit in signed off_t
	if ((off_t)overflowtest < 0) return bps_too_big;
	
	//the mallocs would overflow
	if (realsourcelen+realtargetlen >= SIZE_MAX/sizeof(off_t)) return bps_too_big;
	
	if (realsourcelen+realtargetlen >= out->maxsize()) return bps_too_big;
	
	
	off_t sourcelen = realsourcelen;
	off_t targetlen = realtargetlen;
	
	uint8_t* mem_joined = (uint8_t*)malloc(sizeof(uint8_t)*(realsourcelen+realtargetlen));
	
	off_t* sorted = (off_t*)malloc(sizeof(off_t)*(realsourcelen+realtargetlen));
	
	off_t* sorted_inverse = NULL;
	if (moremem) sorted_inverse = (off_t*)malloc(sizeof(off_t)*(realsourcelen+realtargetlen));
	
	off_t* buckets = NULL;
	if (!sorted_inverse) buckets = (off_t*)malloc(sizeof(off_t)*65537);
	
	if (!sorted || !mem_joined || (!sorted_inverse && !buckets))
	{
		free(mem_joined);
		free(sorted);
		free(sorted_inverse);
		free(buckets);
		return bps_out_of_mem;
	}
	
	//sortedsize is how much of the target file is sorted
	off_t sortedsize = targetlen;
	//divide by 4 for each iteration, to avoid sorting 50% of the file (the sorter is slow)
	while (sortedsize/4 > sourcelen && sortedsize > 1024) sortedsize >>= 2;
	
	off_t prevsortedsize = 0;
	off_t outpos = 0;
	
	goto reindex; // jump into the middle so I won't need a special case to enter it
	
	while (outpos < targetlen)
	{
		if (outpos >= sortedsize-256 && sortedsize < targetlen)
		{
			sortedsize = nextsize(outpos, sortedsize, targetlen);
			
		reindex:
			
			//this isn't an exact science
			const float percSort = sorted_inverse ? 0.67 : 0.50;
			const float percInv = sorted_inverse ? 0.11 : 0.10;
			//const float percFind = sorted_inverse ? 0.22 : 0.40; // unused
			
			const size_t progPreSort = lerp(prevsortedsize, sortedsize, 0);
			const size_t progPreInv = lerp(prevsortedsize, sortedsize, percSort);
			const size_t progPreFind = lerp(prevsortedsize, sortedsize, percSort+percInv);
			
			prevsortedsize = sortedsize;
			
			if (!out->progress(progPreSort, targetlen)) error(bps_canceled);
			
			if (!target->read(mem_joined, 0, sortedsize)) error(bps_io);
			if (!source->read(mem_joined+sortedsize, 0, sourcelen)) error(bps_io);
			out->move_target(mem_joined);
			sufsort(sorted, mem_joined, sortedsize+sourcelen);
			
			if (!out->progress(progPreInv, targetlen)) error(bps_canceled);
			
			if (sorted_inverse)
				create_reverse_index(sorted, sorted_inverse, sortedsize+sourcelen);
			else
				create_buckets(mem_joined, sorted, sortedsize+sourcelen, buckets);
			
			if (!out->progress(progPreFind, targetlen)) error(bps_canceled);
		}
		
		off_t matchlen = 0;
		off_t matchpos = adjust_match(find_index(outpos, mem_joined, sortedsize+sourcelen, sorted, sorted_inverse, buckets),
		                              mem_joined+outpos, sortedsize-outpos,
		                              mem_joined,sortedsize+sourcelen, outpos,sortedsize,
		                              sorted, sortedsize+sourcelen,
		                              &matchlen);
		
#ifdef TEST_CORRECT
		if (matchlen && matchpos >= outpos && matchpos < sortedsize) puts("ERROR: found match in invalid location"),abort();
		if (memcmp(mem_joined+matchpos, mem_joined+outpos, matchlen)) puts("ERROR: found match doesn't match"),abort();
#endif
		
		off_t taken;
		if (matchpos >= sortedsize) taken = out->match(false, matchpos-sortedsize, matchlen);
		else taken = out->match(true, matchpos, matchlen);
#ifdef TEST_CORRECT
		if (taken < 0) puts("ERROR: match() returned negative"),abort();
		if (matchlen >= 7 && taken < matchlen) printf("ERROR: match() took %i bytes, offered %i\n", taken, matchlen),abort();
#endif
		outpos += taken;
	}
	
	out->finish(mem_joined+sortedsize, mem_joined);
	
	err = bps_ok;
	
error:
	free(buckets);
	free(sorted_inverse);
	free(sorted);
	free(mem_joined);
	
	return err;
}


//template<typename T> static bpserror bps_create_suf_pick(file* source, file* target, bool moremem, struct bps_creator * bps);
//template<> bpserror bps_create_suf_pick<uint32_t>(file* source, file* target, bool moremem, struct bps_creator * bps)
//{
//	return bps_create_suf_core<int32_t>(source, target, moremem, bps);
//}
//template<> bpserror bps_create_suf_pick<uint64_t>(file* source, file* target, bool moremem, struct bps_creator * bps)
//{
//	bpserror err = bps_create_suf_core<int32_t>(source, target, moremem, bps);
//	if (err==bps_too_big) err = bps_create_suf_core<int64_t>(source, target, moremem, bps);
//	return err;
//}

//This one picks a function based on 32-bit integers if that fits. This halves memory use for common inputs.
//It also handles some stuff related to the BPS headers and footers.
bpserror bps_create_delta(file* source, file* target, struct mem metadata, struct mem * patchmem,
                          bool (*progress)(void* userdata, size_t done, size_t total), void* userdata, bool moremem)
{
	bps_creator bps(source, target, metadata);
	bps.setProgress(progress, userdata);
	
	size_t maindata = bps.outlen;
	
	//off_t must be signed
	bpserror err = bps_create_suf_core<int32_t>(source, target, moremem, &bps);
	if (err!=bps_ok) return err;
	
	*patchmem = bps.getpatch();
	
	while ((patchmem->ptr[maindata]&0x80) == 0x00) maindata++;
	if (maindata==patchmem->len-12-1) return bps_identical;
	return bps_ok;
}

enum bpserror bps_create_delta_inmem(struct mem source, struct mem target, struct mem metadata, struct mem * patch,
                               bool (*progress)(void* userdata, size_t done, size_t total), void* userdata,
                               bool moremem)
{
	class memfile : public file {
	public:
		const uint8_t * m_ptr;
		size_t m_len;
		
		size_t len() { return m_len; }
		bool read(uint8_t* target, size_t start, size_t len) { memcpy(target, m_ptr+start, len); return true; }
		
		memfile(const uint8_t * ptr, size_t len) : m_ptr(ptr), m_len(len) {}
	};
	
	memfile sourcef(source.ptr, source.len);
	memfile targetf(target.ptr, target.len);
	
	return bps_create_delta(&sourcef, &targetf, metadata, patch, progress, userdata, moremem);
}




#ifdef BPS_STANDALONE
#include <stdio.h>
static struct mem ReadWholeFile(const char * filename)
{
	struct mem null = {NULL, 0};
	
	FILE * file=fopen(filename, "rb");
	if (!file) return null;
	fseek(file, 0, SEEK_END);
	size_t len=ftell(file);
	fseek(file, 0, SEEK_SET);
	unsigned char * data=(unsigned char*)malloc(len);
	size_t truelen=fread(data, 1,len, file);
	fclose(file);
	if (len!=truelen)
	{
		free(data);
		return null;
	}
	
	struct mem ret = { (unsigned char*)data, len };
	return ret;
}
static bool WriteWholeFile(const char * filename, struct mem data)
{
	FILE * file=fopen(filename, "wb");
	if (!file) return false;
	unsigned int truelen=fwrite(data.ptr, 1,data.len, file);
	fclose(file);
	return (truelen==data.len);
}
int main(int argc, char * argv[])
{
//struct mem out = ReadWholeFile(argv[2]);
//printf("check=%.8X\n",crc32(out.ptr, out.len));

struct mem in = ReadWholeFile(argv[1]);
struct mem out = ReadWholeFile(argv[2]);
struct mem null = {NULL, 0};
struct mem p={NULL,0};
//int n=50;
//for(int i=0;i<n;i++)
//printf("%i/%i\n",i,n),
bps_create_delta(in,out,null,&p, NULL,NULL);
printf("len=%lu        \n",p.len);
printf("check=%.8X\n",*(uint32_t*)(p.ptr+p.len-4));
WriteWholeFile(argv[3], p);
free(in.ptr);
free(out.ptr);
free(p.ptr);

#ifdef TEST_PERF
printf("%i/%i=%f\n",match_len_tot,match_len_n,(float)match_len_tot/match_len_n);
#endif
}
#endif