comparison contrib/python-zstandard/c-ext/decompressor.c @ 37495:b1fb341d8a61

zstandard: vendor python-zstandard 0.9.0

This was just released. It features a number of goodies. More info at
https://gregoryszorc.com/blog/2018/04/09/release-of-python-zstandard-0.9/.

The clang-format ignore list was updated to reflect the new source
of files.

The project contains a vendored copy of zstandard 1.3.4. The old
version was 1.1.3. One of the changes between those versions is that
zstandard is now dual licensed BSD + GPLv2 and the patent rights grant
has been removed. Good riddance.

The API should be backwards compatible. So no changes in core
should be needed. However, there were a number of changes in the
library that we'll want to adapt to. Those will be addressed in
subsequent commits.

Differential Revision: https://phab.mercurial-scm.org/D3198

37494:1ce7a55b09d1

#include "pool.h"

extern PyObject* ZstdError;

/**
  * Ensure the ZSTD_DStream on a ZstdDecompressor is initialized and reset.
  *
  * This should be called before starting a decompression operation with a
  * ZSTD_DStream on a ZstdDecompressor.
  */
int init_dstream(ZstdDecompressor* decompressor) {
	void* dictData = NULL;
	size_t dictSize = 0;
	size_t zresult;

	/* Simple case of dstream already exists. Just reset it. */
	if (decompressor->dstream) {
		zresult = ZSTD_resetDStream(decompressor->dstream);
		if (ZSTD_isError(zresult)) {
			PyErr_Format(ZstdError, "could not reset DStream: %s",
				ZSTD_getErrorName(zresult));
			return -1;
		}

		return 0;
	}

	decompressor->dstream = ZSTD_createDStream();
	if (!decompressor->dstream) {
		PyErr_SetString(ZstdError, "could not create DStream");
		return -1;
	}

	if (decompressor->dict) {
		dictData = decompressor->dict->dictData;
		dictSize = decompressor->dict->dictSize;
	}

	if (dictData) {
		zresult = ZSTD_initDStream_usingDict(decompressor->dstream, dictData, dictSize);
	}
	else {
		zresult = ZSTD_initDStream(decompressor->dstream);
	}

	if (ZSTD_isError(zresult)) {
		/* Don't leave a reference to an invalid object. */
		ZSTD_freeDStream(decompressor->dstream);
		decompressor->dstream = NULL;

		PyErr_Format(ZstdError, "could not initialize DStream: %s",
			ZSTD_getErrorName(zresult));
		return -1;
	}

	return 0;
}

);

static int Decompressor_init(ZstdDecompressor* self, PyObject* args, PyObject* kwargs) {
	static char* kwlist[] = {
		"dict_data",
		NULL
	};

	ZstdCompressionDict* dict = NULL;

	self->dctx = NULL;
	self->dict = NULL;
	self->ddict = NULL;

	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|O!:ZstdDecompressor", kwlist,
		&ZstdCompressionDictType, &dict)) {
		return -1;
	}

	/* TODO lazily initialize the reference ZSTD_DCtx on first use since
	   not instances of ZstdDecompressor will use a ZSTD_DCtx. */
	self->dctx = ZSTD_createDCtx();
	if (!self->dctx) {
		PyErr_NoMemory();
		goto except;
	}

	if (dict) {
		self->dict = dict;
		Py_INCREF(dict);
	}

	return 0;

except:
	if (self->dctx) {
		ZSTD_freeDCtx(self->dctx);
		self->dctx = NULL;
	}

}

static void Decompressor_dealloc(ZstdDecompressor* self) {
	Py_CLEAR(self->dict);

	if (self->ddict) {
		ZSTD_freeDDict(self->ddict);
		self->ddict = NULL;
	}

	if (self->dstream) {
		ZSTD_freeDStream(self->dstream);
		self->dstream = NULL;
	}

	if (self->dctx) {
		ZSTD_freeDCtx(self->dctx);
		self->dctx = NULL;
	}

	PyObject_Del(self);
}

PyDoc_STRVAR(Decompressor_copy_stream__doc__,
	"copy_stream(ifh, ofh[, read_size=default, write_size=default]) -- decompress data between streams\n"
	"\n"

	ZSTD_outBuffer output;
	Py_ssize_t totalRead = 0;
	Py_ssize_t totalWrite = 0;
	char* readBuffer;
	Py_ssize_t readSize;
	PyObject* readResult;
	PyObject* res = NULL;
	size_t zresult = 0;
	PyObject* writeResult;
	PyObject* totalReadPy;
	PyObject* totalWritePy;

	}

	/* Prevent free on uninitialized memory in finally. */
	output.dst = NULL;

	if (0 != init_dstream(self)) {
		res = NULL;
		goto finally;
	}

	output.dst = PyMem_Malloc(outSize);

		input.size = readSize;
		input.pos = 0;

		while (input.pos < input.size) {
			Py_BEGIN_ALLOW_THREADS
			zresult = ZSTD_decompressStream(self->dstream, &output, &input);
			Py_END_ALLOW_THREADS

			if (ZSTD_isError(zresult)) {
				PyErr_Format(ZstdError, "zstd decompressor error: %s",
					ZSTD_getErrorName(zresult));

				Py_XDECREF(writeResult);
				totalWrite += output.pos;
				output.pos = 0;
			}
		}
	}

	/* Source stream is exhausted. Finish up. */

	totalReadPy = PyLong_FromSsize_t(totalRead);

finally:
	if (output.dst) {
		PyMem_Free(output.dst);
	}

	return res;
}

PyDoc_STRVAR(Decompressor_decompress__doc__,

		"data",
		"max_output_size",
		NULL
	};

	const char* source;
	Py_ssize_t sourceSize;
	Py_ssize_t maxOutputSize = 0;
	unsigned long long decompressedSize;
	size_t destCapacity;
	PyObject* result = NULL;
	void* dictData = NULL;
	size_t dictSize = 0;
	size_t zresult;

#if PY_MAJOR_VERSION >= 3
	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y#|n:decompress",
#else
	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "s#|n:decompress",
#endif
		kwlist, &source, &sourceSize, &maxOutputSize)) {
		return NULL;
	}

	if (self->dict) {
		dictData = self->dict->dictData;
		dictSize = self->dict->dictSize;
	}

	if (dictData && !self->ddict) {
		Py_BEGIN_ALLOW_THREADS
		self->ddict = ZSTD_createDDict_byReference(dictData, dictSize);
		Py_END_ALLOW_THREADS

		if (!self->ddict) {
			PyErr_SetString(ZstdError, "could not create decompression dict");
			return NULL;
		}
	}

	decompressedSize = ZSTD_getDecompressedSize(source, sourceSize);
	/* 0 returned if content size not in the zstd frame header */
	if (0 == decompressedSize) {
		if (0 == maxOutputSize) {
			PyErr_SetString(ZstdError, "input data invalid or missing content size "
				"in frame header");
			return NULL;
		}
		else {
			result = PyBytes_FromStringAndSize(NULL, maxOutputSize);
			destCapacity = maxOutputSize;
		}
	}
	else {
		result = PyBytes_FromStringAndSize(NULL, decompressedSize);
		destCapacity = decompressedSize;
	}

	if (!result) {
		return NULL;
	}

	Py_BEGIN_ALLOW_THREADS
	if (self->ddict) {
		zresult = ZSTD_decompress_usingDDict(self->dctx,
			PyBytes_AsString(result), destCapacity,
			source, sourceSize, self->ddict);
	}
	else {
		zresult = ZSTD_decompressDCtx(self->dctx,
			PyBytes_AsString(result), destCapacity, source, sourceSize);
	}
	Py_END_ALLOW_THREADS

	if (ZSTD_isError(zresult)) {
		PyErr_Format(ZstdError, "decompression error: %s", ZSTD_getErrorName(zresult));
		Py_DECREF(result);
		return NULL;
	}
	else if (decompressedSize && zresult != decompressedSize) {
		PyErr_Format(ZstdError, "decompression error: decompressed %zu bytes; expected %llu",
			zresult, decompressedSize);
		Py_DECREF(result);
		return NULL;
	}
	else if (zresult < destCapacity) {
		if (_PyBytes_Resize(&result, zresult)) {
			Py_DECREF(result);
			return NULL;
		}
	}

	return result;
}

PyDoc_STRVAR(Decompressor_decompressobj__doc__,
"decompressobj()\n"
"\n"
"Incrementally feed data into a decompressor.\n"
"\n"
"The returned object exposes a ``decompress(data)`` method. This makes it\n"
"compatible with ``zlib.decompressobj`` and ``bz2.BZ2Decompressor`` so that\n"
"callers can swap in the zstd decompressor while using the same API.\n"
);

static ZstdDecompressionObj* Decompressor_decompressobj(ZstdDecompressor* self) {
	ZstdDecompressionObj* result = (ZstdDecompressionObj*)PyObject_CallObject((PyObject*)&ZstdDecompressionObjType, NULL);
	if (!result) {
		return NULL;
	}

	if (0 != init_dstream(self)) {
		Py_DECREF(result);
		return NULL;
	}

	result->decompressor = self;
	Py_INCREF(result->decompressor);

	return result;
}

PyDoc_STRVAR(Decompressor_read_from__doc__,
"read_from(reader[, read_size=default, write_size=default, skip_bytes=0])\n"
"Read compressed data and return an iterator\n"
"\n"
"Returns an iterator of decompressed data chunks produced from reading from\n"
"the ``reader``.\n"
"\n"

"\n"
"There is also support for skipping the first ``skip_bytes`` of data from\n"
"the source.\n"
);

static ZstdDecompressorIterator* Decompressor_read_from(ZstdDecompressor* self, PyObject* args, PyObject* kwargs) {
	static char* kwlist[] = {
		"reader",
		"read_size",
		"write_size",
		"skip_bytes",

	size_t inSize = ZSTD_DStreamInSize();
	size_t outSize = ZSTD_DStreamOutSize();
	ZstdDecompressorIterator* result;
	size_t skipBytes = 0;

	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O|kkk:read_from", kwlist,
		&reader, &inSize, &outSize, &skipBytes)) {
		return NULL;
	}

	if (skipBytes >= inSize) {

		result->reader = reader;
		Py_INCREF(result->reader);
	}
	else if (1 == PyObject_CheckBuffer(reader)) {
		/* Object claims it is a buffer. Try to get a handle to it. */
		result->buffer = PyMem_Malloc(sizeof(Py_buffer));
		if (!result->buffer) {
			goto except;
		}

		memset(result->buffer, 0, sizeof(Py_buffer));

		if (0 != PyObject_GetBuffer(reader, result->buffer, PyBUF_CONTIG_RO)) {
			goto except;
		}
	}
	else {
		PyErr_SetString(PyExc_ValueError,

	result->inSize = inSize;
	result->outSize = outSize;
	result->skipBytes = skipBytes;

	if (0 != init_dstream(self)) {
		goto except;
	}

	result->input.src = PyMem_Malloc(inSize);
	if (!result->input.src) {

	}

	goto finally;

except:
	Py_CLEAR(result->reader);

	if (result->buffer) {
		PyBuffer_Release(result->buffer);
		Py_CLEAR(result->buffer);
	}

	Py_CLEAR(result);

finally:

	return result;
}

PyDoc_STRVAR(Decompressor_write_to__doc__,
"Create a context manager to write decompressed data to an object.\n"
"\n"
"The passed object must have a ``write()`` method.\n"
"\n"
"The caller feeds intput data to the object by calling ``write(data)``.\n"

"An optional ``write_size`` argument defines the size of chunks to\n"
"``write()`` to the writer. It defaults to the default output size for a zstd\n"
"streaming decompressor.\n"
);

static ZstdDecompressionWriter* Decompressor_write_to(ZstdDecompressor* self, PyObject* args, PyObject* kwargs) {
	static char* kwlist[] = {
		"writer",
		"write_size",
		NULL
	};

	PyObject* writer;
	size_t outSize = ZSTD_DStreamOutSize();
	ZstdDecompressionWriter* result;

	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O|k:write_to", kwlist,
		&writer, &outSize)) {
		return NULL;
	}

	if (!PyObject_HasAttrString(writer, "write")) {

PyDoc_STRVAR(Decompressor_decompress_content_dict_chain__doc__,
"Decompress a series of chunks using the content dictionary chaining technique\n"
);

static PyObject* Decompressor_decompress_content_dict_chain(PyObject* self, PyObject* args, PyObject* kwargs) {
	static char* kwlist[] = {
		"frames",
		NULL
	};

	Py_ssize_t chunkIndex;
	char parity = 0;
	PyObject* chunk;
	char* chunkData;
	Py_ssize_t chunkSize;
	ZSTD_DCtx* dctx = NULL;
	size_t zresult;
	ZSTD_frameParams frameParams;
	void* buffer1 = NULL;
	size_t buffer1Size = 0;
	size_t buffer1ContentSize = 0;
	void* buffer2 = NULL;
	size_t buffer2Size = 0;
	size_t buffer2ContentSize = 0;
	void* destBuffer = NULL;
	PyObject* result = NULL;

	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O!:decompress_content_dict_chain",
		kwlist, &PyList_Type, &chunks)) {
		return NULL;
	}

		return NULL;
	}

	/* We require that all chunks be zstd frames and that they have content size set. */
	PyBytes_AsStringAndSize(chunk, &chunkData, &chunkSize);
	zresult = ZSTD_getFrameParams(&frameParams, (void*)chunkData, chunkSize);
	if (ZSTD_isError(zresult)) {
		PyErr_SetString(PyExc_ValueError, "chunk 0 is not a valid zstd frame");
		return NULL;
	}
	else if (zresult) {
		PyErr_SetString(PyExc_ValueError, "chunk 0 is too small to contain a zstd frame");
		return NULL;
	}

	if (0 == frameParams.frameContentSize) {
		PyErr_SetString(PyExc_ValueError, "chunk 0 missing content size in frame");
		return NULL;
	}

	dctx = ZSTD_createDCtx();
	if (!dctx) {
		PyErr_NoMemory();
		goto finally;
	}

	buffer1Size = frameParams.frameContentSize;
	buffer1 = PyMem_Malloc(buffer1Size);
	if (!buffer1) {
		goto finally;
	}

	Py_BEGIN_ALLOW_THREADS
	zresult = ZSTD_decompressDCtx(dctx, buffer1, buffer1Size, chunkData, chunkSize);
	Py_END_ALLOW_THREADS
	if (ZSTD_isError(zresult)) {
		PyErr_Format(ZstdError, "could not decompress chunk 0: %s", ZSTD_getErrorName(zresult));
		goto finally;
	}

	buffer1ContentSize = zresult;

	/* Special case of a simple chain. */
	if (1 == chunksLen) {
		result = PyBytes_FromStringAndSize(buffer1, buffer1Size);
		goto finally;
	}

	/* This should ideally look at next chunk. But this is slightly simpler. */
	buffer2Size = frameParams.frameContentSize;
	buffer2 = PyMem_Malloc(buffer2Size);
	if (!buffer2) {
		goto finally;
	}

			PyErr_Format(PyExc_ValueError, "chunk %zd must be bytes", chunkIndex);
			goto finally;
		}

		PyBytes_AsStringAndSize(chunk, &chunkData, &chunkSize);
		zresult = ZSTD_getFrameParams(&frameParams, (void*)chunkData, chunkSize);
		if (ZSTD_isError(zresult)) {
			PyErr_Format(PyExc_ValueError, "chunk %zd is not a valid zstd frame", chunkIndex);
			goto finally;
		}
		else if (zresult) {
			PyErr_Format(PyExc_ValueError, "chunk %zd is too small to contain a zstd frame", chunkIndex);
			goto finally;
		}

		if (0 == frameParams.frameContentSize) {
			PyErr_Format(PyExc_ValueError, "chunk %zd missing content size in frame", chunkIndex);
			goto finally;
		}

		parity = chunkIndex % 2;

		/* This could definitely be abstracted to reduce code duplication. */
		if (parity) {
			/* Resize destination buffer to hold larger content. */
			if (buffer2Size < frameParams.frameContentSize) {
				buffer2Size = frameParams.frameContentSize;
				destBuffer = PyMem_Realloc(buffer2, buffer2Size);
				if (!destBuffer) {
					goto finally;
				}
				buffer2 = destBuffer;
			}

			Py_BEGIN_ALLOW_THREADS
			zresult = ZSTD_decompress_usingDict(dctx, buffer2, buffer2Size,
				chunkData, chunkSize, buffer1, buffer1ContentSize);
			Py_END_ALLOW_THREADS
			if (ZSTD_isError(zresult)) {
				PyErr_Format(ZstdError, "could not decompress chunk %zd: %s",
					chunkIndex, ZSTD_getErrorName(zresult));
				goto finally;
			}
			buffer2ContentSize = zresult;
		}
		else {
			if (buffer1Size < frameParams.frameContentSize) {
				buffer1Size = frameParams.frameContentSize;
				destBuffer = PyMem_Realloc(buffer1, buffer1Size);
				if (!destBuffer) {
					goto finally;
				}
				buffer1 = destBuffer;
			}

			Py_BEGIN_ALLOW_THREADS
			zresult = ZSTD_decompress_usingDict(dctx, buffer1, buffer1Size,
				chunkData, chunkSize, buffer2, buffer2ContentSize);
			Py_END_ALLOW_THREADS
			if (ZSTD_isError(zresult)) {
				PyErr_Format(ZstdError, "could not decompress chunk %zd: %s",
					chunkIndex, ZSTD_getErrorName(zresult));
				goto finally;
			}
			buffer1ContentSize = zresult;
		}
	}

	result = PyBytes_FromStringAndSize(parity ? buffer2 : buffer1,
		parity ? buffer2ContentSize : buffer1ContentSize);

	}
	if (buffer1) {
		PyMem_Free(buffer1);
	}

	if (dctx) {
		ZSTD_freeDCtx(dctx);
	}

	return result;
}

typedef struct {
	void* sourceData;
	size_t sourceSize;
	unsigned long long destSize;
} FramePointer;

typedef struct {
	FramePointer* frames;
	Py_ssize_t framesSize;

	Py_ssize_t endOffset;
	unsigned long long totalSourceSize;

	/* Compression state and settings. */
	ZSTD_DCtx* dctx;
	ZSTD_DDict* ddict;
	int requireOutputSizes;

	/* Output storage. */
	DestBuffer* destBuffers;
	Py_ssize_t destCount;

	assert(NULL == state->destBuffers);
	assert(0 == state->destCount);
	assert(state->endOffset - state->startOffset >= 0);

	/*
	 * We need to allocate a buffer to hold decompressed data. How we do this
	 * depends on what we know about the output. The following scenarios are
	 * possible:
	 *

	 */

	/* Resolve ouput segments. */
	for (frameIndex = state->startOffset; frameIndex <= state->endOffset; frameIndex++) {
		FramePointer* fp = &framePointers[frameIndex];

		if (0 == fp->destSize) {
			fp->destSize = ZSTD_getDecompressedSize(fp->sourceData, fp->sourceSize);
			if (0 == fp->destSize && state->requireOutputSizes) {
				state->error = WorkerError_unknownSize;
				state->errorOffset = frameIndex;
				return;
			}
		}

		totalOutputSize += fp->destSize;
	}

	destBuffer = &state->destBuffers[state->destCount - 1];

	assert(framePointers[state->startOffset].destSize > 0); /* For now. */

	allocationSize = roundpow2(state->totalSourceSize);

	if (framePointers[state->startOffset].destSize > allocationSize) {
		allocationSize = roundpow2(framePointers[state->startOffset].destSize);
	}

	}

	destBuffer->segmentsSize = remainingItems;

	for (frameIndex = state->startOffset; frameIndex <= state->endOffset; frameIndex++) {
		const void* source = framePointers[frameIndex].sourceData;
		const size_t sourceSize = framePointers[frameIndex].sourceSize;
		void* dest;
		const size_t decompressedSize = framePointers[frameIndex].destSize;
		size_t destAvailable = destBuffer->destSize - destOffset;

			destBuffer = &state->destBuffers[state->destCount - 1];

			/* Don't take any chances will non-NULL pointers. */
			memset(destBuffer, 0, sizeof(DestBuffer));

			allocationSize = roundpow2(state->totalSourceSize);

			if (decompressedSize > allocationSize) {
				allocationSize = roundpow2(decompressedSize);
			}

			currentBufferStartIndex = frameIndex;
		}

		dest = (char*)destBuffer->dest + destOffset;

		if (state->ddict) {
			zresult = ZSTD_decompress_usingDDict(state->dctx, dest, decompressedSize,
				source, sourceSize, state->ddict);
		}
		else {
			zresult = ZSTD_decompressDCtx(state->dctx, dest, decompressedSize,
				source, sourceSize);
		}

		if (ZSTD_isError(zresult)) {
			state->error = WorkerError_zstd;
			state->zresult = zresult;
			state->errorOffset = frameIndex;
			return;
		}
		else if (zresult != decompressedSize) {
			state->error = WorkerError_sizeMismatch;
			state->zresult = zresult;
			state->errorOffset = frameIndex;
			return;
		}

		destBuffer->segments[localOffset].offset = destOffset;
		destBuffer->segments[localOffset].length = decompressedSize;
		destOffset += zresult;
		localOffset++;
		remainingItems--;
	}

	if (destBuffer->destSize > destOffset) {

		destBuffer->destSize = destOffset;
	}
}

ZstdBufferWithSegmentsCollection* decompress_from_framesources(ZstdDecompressor* decompressor, FrameSources* frames,
	unsigned int threadCount) {
	void* dictData = NULL;
	size_t dictSize = 0;
	Py_ssize_t i = 0;
	int errored = 0;
	Py_ssize_t segmentsCount;
	ZstdBufferWithSegments* bws = NULL;
	PyObject* resultArg = NULL;
	Py_ssize_t resultIndex;
	ZstdBufferWithSegmentsCollection* result = NULL;
	FramePointer* framePointers = frames->frames;
	unsigned long long workerBytes = 0;
	int currentThread = 0;
	Py_ssize_t workerStartOffset = 0;
	POOL_ctx* pool = NULL;
	WorkerState* workerStates = NULL;
	unsigned long long bytesPerWorker;

	/* Caller should normalize 0 and negative values to 1 or larger. */
	assert(threadCount >= 1);

	/* More threads than inputs makes no sense under any conditions. */
	threadCount = frames->framesSize < threadCount ? (unsigned int)frames->framesSize
												   : threadCount;

	/* TODO lower thread count if input size is too small and threads would just
	   add overhead. */

	if (decompressor->dict) {
		dictData = decompressor->dict->dictData;
		dictSize = decompressor->dict->dictSize;
	}

	if (dictData && !decompressor->ddict) {
		Py_BEGIN_ALLOW_THREADS
		decompressor->ddict = ZSTD_createDDict_byReference(dictData, dictSize);
		Py_END_ALLOW_THREADS

		if (!decompressor->ddict) {
			PyErr_SetString(ZstdError, "could not create decompression dict");
			return NULL;
		}
	}

	/* If threadCount==1, we don't start a thread pool. But we do leverage the

		}
	}

	bytesPerWorker = frames->compressedSize / threadCount;

	for (i = 0; i < threadCount; i++) {
		workerStates[i].dctx = ZSTD_createDCtx();
		if (NULL == workerStates[i].dctx) {
			PyErr_NoMemory();
			goto finally;
		}

		ZSTD_copyDCtx(workerStates[i].dctx, decompressor->dctx);

		workerStates[i].ddict = decompressor->ddict;
		workerStates[i].framePointers = framePointers;
		workerStates[i].requireOutputSizes = 1;
	}

	Py_BEGIN_ALLOW_THREADS

			PyErr_NoMemory();
			errored = 1;
			break;

		case WorkerError_sizeMismatch:
			PyErr_Format(ZstdError, "error decompressing item %zd: decompressed %zu bytes; expected %llu",
				workerStates[i].errorOffset, workerStates[i].zresult,
				framePointers[workerStates[i].errorOffset].destSize);
			errored = 1;
			break;

			if (frameSizesP) {
				decompressedSize = frameSizesP[i];
			}

			framePointers[i].sourceData = sourceData;
			framePointers[i].sourceSize = sourceSize;
			framePointers[i].destSize = decompressedSize;
		}
	}
	else if (PyObject_TypeCheck(frames, &ZstdBufferWithSegmentsCollectionType)) {
		Py_ssize_t offset = 0;
		ZstdBufferWithSegments* buffer;

		for (i = 0; i < collection->bufferCount; i++) {
			Py_ssize_t segmentIndex;
			buffer = collection->buffers[i];

			for (segmentIndex = 0; segmentIndex < buffer->segmentCount; segmentIndex++) {
				if (buffer->segments[segmentIndex].offset + buffer->segments[segmentIndex].length > buffer->dataSize) {
					PyErr_Format(PyExc_ValueError, "item %zd has offset outside memory area",
						offset);
					goto finally;
				}

				totalInputSize += buffer->segments[segmentIndex].length;

				framePointers[offset].sourceData = (char*)buffer->data + buffer->segments[segmentIndex].offset;
				framePointers[offset].sourceSize = buffer->segments[segmentIndex].length;
				framePointers[offset].destSize = frameSizesP ? frameSizesP[offset] : 0;

				offset++;
			}
		}
	}

		if (!framePointers) {
			PyErr_NoMemory();
			goto finally;
		}

		/*
		 * It is not clear whether Py_buffer.buf is still valid after
		 * PyBuffer_Release. So, we hold a reference to all Py_buffer instances
		 * for the duration of the operation.
		 */
		frameBuffers = PyMem_Malloc(frameCount * sizeof(Py_buffer));
		if (NULL == frameBuffers) {
			PyErr_NoMemory();
			goto finally;
		}

		memset(frameBuffers, 0, frameCount * sizeof(Py_buffer));

		/* Do a pass to assemble info about our input buffers and output sizes. */
		for (i = 0; i < frameCount; i++) {
			if (0 != PyObject_GetBuffer(PyList_GET_ITEM(frames, i),
				&frameBuffers[i], PyBUF_CONTIG_RO)) {
				PyErr_Clear();
				PyErr_Format(PyExc_TypeError, "item %zd not a bytes like object", i);
				goto finally;
			}

			totalInputSize += frameBuffers[i].len;

			framePointers[i].sourceData = frameBuffers[i].buf;
			framePointers[i].sourceSize = frameBuffers[i].len;
			framePointers[i].destSize = frameSizesP ? frameSizesP[i] : 0;
		}
	}
	else {
		PyErr_SetString(PyExc_TypeError, "argument must be list or BufferWithSegments");
		goto finally;

static PyMethodDef Decompressor_methods[] = {
	{ "copy_stream", (PyCFunction)Decompressor_copy_stream, METH_VARARGS | METH_KEYWORDS,
	Decompressor_copy_stream__doc__ },
	{ "decompress", (PyCFunction)Decompressor_decompress, METH_VARARGS | METH_KEYWORDS,
	Decompressor_decompress__doc__ },
	{ "decompressobj", (PyCFunction)Decompressor_decompressobj, METH_NOARGS,
	Decompressor_decompressobj__doc__ },
	{ "read_from", (PyCFunction)Decompressor_read_from, METH_VARARGS | METH_KEYWORDS,
	Decompressor_read_from__doc__ },
	{ "write_to", (PyCFunction)Decompressor_write_to, METH_VARARGS | METH_KEYWORDS,
	Decompressor_write_to__doc__ },
	{ "decompress_content_dict_chain", (PyCFunction)Decompressor_decompress_content_dict_chain,
	  METH_VARARGS | METH_KEYWORDS, Decompressor_decompress_content_dict_chain__doc__ },
	{ "multi_decompress_to_buffer", (PyCFunction)Decompressor_multi_decompress_to_buffer,
	  METH_VARARGS | METH_KEYWORDS, Decompressor_multi_decompress_to_buffer__doc__ },
	{ NULL, NULL }
};

PyTypeObject ZstdDecompressorType = {
	PyVarObject_HEAD_INIT(NULL, 0)

37495:b1fb341d8a61

#include "pool.h"

extern PyObject* ZstdError;

/**
 * Ensure the ZSTD_DCtx on a decompressor is initiated and ready for a new operation.
 */
int ensure_dctx(ZstdDecompressor* decompressor, int loadDict) {
	size_t zresult;

	ZSTD_DCtx_reset(decompressor->dctx);

	if (decompressor->maxWindowSize) {
		zresult = ZSTD_DCtx_setMaxWindowSize(decompressor->dctx, decompressor->maxWindowSize);
		if (ZSTD_isError(zresult)) {
			PyErr_Format(ZstdError, "unable to set max window size: %s",
				ZSTD_getErrorName(zresult));
			return 1;
		}
	}

	zresult = ZSTD_DCtx_setFormat(decompressor->dctx, decompressor->format);
	if (ZSTD_isError(zresult)) {
		PyErr_Format(ZstdError, "unable to set decoding format: %s",
			ZSTD_getErrorName(zresult));
		return 1;
	}

	if (loadDict && decompressor->dict) {
		if (ensure_ddict(decompressor->dict)) {
			return 1;
		}

		zresult = ZSTD_DCtx_refDDict(decompressor->dctx, decompressor->dict->ddict);
		if (ZSTD_isError(zresult)) {
			PyErr_Format(ZstdError, "unable to reference prepared dictionary: %s",
				ZSTD_getErrorName(zresult));
			return 1;
		}
	}

	return 0;
}

);

static int Decompressor_init(ZstdDecompressor* self, PyObject* args, PyObject* kwargs) {
	static char* kwlist[] = {
		"dict_data",
		"max_window_size",
		"format",
		NULL
	};

	ZstdCompressionDict* dict = NULL;
	size_t maxWindowSize = 0;
	ZSTD_format_e format = ZSTD_f_zstd1;

	self->dctx = NULL;
	self->dict = NULL;

	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|O!II:ZstdDecompressor", kwlist,
		&ZstdCompressionDictType, &dict, &maxWindowSize, &format)) {
		return -1;
	}

	self->dctx = ZSTD_createDCtx();
	if (!self->dctx) {
		PyErr_NoMemory();
		goto except;
	}

	self->maxWindowSize = maxWindowSize;
	self->format = format;

	if (dict) {
		self->dict = dict;
		Py_INCREF(dict);
	}

	if (ensure_dctx(self, 1)) {
		goto except;
	}

	return 0;

except:
	Py_CLEAR(self->dict);

	if (self->dctx) {
		ZSTD_freeDCtx(self->dctx);
		self->dctx = NULL;
	}

}

static void Decompressor_dealloc(ZstdDecompressor* self) {
	Py_CLEAR(self->dict);

	if (self->dctx) {
		ZSTD_freeDCtx(self->dctx);
		self->dctx = NULL;
	}

	PyObject_Del(self);
}

PyDoc_STRVAR(Decompressor_memory_size__doc__,
"memory_size() -- Size of decompression context, in bytes\n"
);

static PyObject* Decompressor_memory_size(ZstdDecompressor* self) {
	if (self->dctx) {
		return PyLong_FromSize_t(ZSTD_sizeof_DCtx(self->dctx));
	}
	else {
		PyErr_SetString(ZstdError, "no decompressor context found; this should never happen");
		return NULL;
	}
}

PyDoc_STRVAR(Decompressor_copy_stream__doc__,
	"copy_stream(ifh, ofh[, read_size=default, write_size=default]) -- decompress data between streams\n"
	"\n"

	ZSTD_outBuffer output;
	Py_ssize_t totalRead = 0;
	Py_ssize_t totalWrite = 0;
	char* readBuffer;
	Py_ssize_t readSize;
	PyObject* readResult = NULL;
	PyObject* res = NULL;
	size_t zresult = 0;
	PyObject* writeResult;
	PyObject* totalReadPy;
	PyObject* totalWritePy;

	}

	/* Prevent free on uninitialized memory in finally. */
	output.dst = NULL;

	if (ensure_dctx(self, 1)) {
		res = NULL;
		goto finally;
	}

	output.dst = PyMem_Malloc(outSize);

		input.size = readSize;
		input.pos = 0;

		while (input.pos < input.size) {
			Py_BEGIN_ALLOW_THREADS
			zresult = ZSTD_decompress_generic(self->dctx, &output, &input);
			Py_END_ALLOW_THREADS

			if (ZSTD_isError(zresult)) {
				PyErr_Format(ZstdError, "zstd decompressor error: %s",
					ZSTD_getErrorName(zresult));

				Py_XDECREF(writeResult);
				totalWrite += output.pos;
				output.pos = 0;
			}
		}

		Py_CLEAR(readResult);
	}

	/* Source stream is exhausted. Finish up. */

	totalReadPy = PyLong_FromSsize_t(totalRead);

finally:
	if (output.dst) {
		PyMem_Free(output.dst);
	}

	Py_XDECREF(readResult);

	return res;
}

PyDoc_STRVAR(Decompressor_decompress__doc__,

		"data",
		"max_output_size",
		NULL
	};

	Py_buffer source;
	Py_ssize_t maxOutputSize = 0;
	unsigned long long decompressedSize;
	size_t destCapacity;
	PyObject* result = NULL;
	size_t zresult;
	ZSTD_outBuffer outBuffer;
	ZSTD_inBuffer inBuffer;

#if PY_MAJOR_VERSION >= 3
	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "y*|n:decompress",
#else
	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "s*|n:decompress",
#endif
		kwlist, &source, &maxOutputSize)) {
		return NULL;
	}

	if (!PyBuffer_IsContiguous(&source, 'C') || source.ndim > 1) {
		PyErr_SetString(PyExc_ValueError,
			"data buffer should be contiguous and have at most one dimension");
		goto finally;
	}

	if (ensure_dctx(self, 1)) {
		goto finally;
	}

	decompressedSize = ZSTD_getFrameContentSize(source.buf, source.len);

	if (ZSTD_CONTENTSIZE_ERROR == decompressedSize) {
		PyErr_SetString(ZstdError, "error determining content size from frame header");
		goto finally;
	}
	/* Special case of empty frame. */
	else if (0 == decompressedSize) {
		result = PyBytes_FromStringAndSize("", 0);
		goto finally;
	}
	/* Missing content size in frame header. */
	if (ZSTD_CONTENTSIZE_UNKNOWN == decompressedSize) {
		if (0 == maxOutputSize) {
			PyErr_SetString(ZstdError, "could not determine content size in frame header");
			goto finally;
		}

		result = PyBytes_FromStringAndSize(NULL, maxOutputSize);
		destCapacity = maxOutputSize;
		decompressedSize = 0;
	}
	/* Size is recorded in frame header. */
	else {
		assert(SIZE_MAX >= PY_SSIZE_T_MAX);
		if (decompressedSize > PY_SSIZE_T_MAX) {
			PyErr_SetString(ZstdError, "frame is too large to decompress on this platform");
			goto finally;
		}

		result = PyBytes_FromStringAndSize(NULL, (Py_ssize_t)decompressedSize);
		destCapacity = (size_t)decompressedSize;
	}

	if (!result) {
		goto finally;
	}

	outBuffer.dst = PyBytes_AsString(result);
	outBuffer.size = destCapacity;
	outBuffer.pos = 0;

	inBuffer.src = source.buf;
	inBuffer.size = source.len;
	inBuffer.pos = 0;

	Py_BEGIN_ALLOW_THREADS
	zresult = ZSTD_decompress_generic(self->dctx, &outBuffer, &inBuffer);
	Py_END_ALLOW_THREADS

	if (ZSTD_isError(zresult)) {
		PyErr_Format(ZstdError, "decompression error: %s", ZSTD_getErrorName(zresult));
		Py_CLEAR(result);
		goto finally;
	}
	else if (zresult) {
		PyErr_Format(ZstdError, "decompression error: did not decompress full frame");
		Py_CLEAR(result);
		goto finally;
	}
	else if (decompressedSize && outBuffer.pos != decompressedSize) {
		PyErr_Format(ZstdError, "decompression error: decompressed %zu bytes; expected %llu",
			zresult, decompressedSize);
		Py_CLEAR(result);
		goto finally;
	}
	else if (outBuffer.pos < destCapacity) {
		if (safe_pybytes_resize(&result, outBuffer.pos)) {
			Py_CLEAR(result);
			goto finally;
		}
	}

finally:
	PyBuffer_Release(&source);
	return result;
}

PyDoc_STRVAR(Decompressor_decompressobj__doc__,
"decompressobj([write_size=default])\n"
"\n"
"Incrementally feed data into a decompressor.\n"
"\n"
"The returned object exposes a ``decompress(data)`` method. This makes it\n"
"compatible with ``zlib.decompressobj`` and ``bz2.BZ2Decompressor`` so that\n"
"callers can swap in the zstd decompressor while using the same API.\n"
);

static ZstdDecompressionObj* Decompressor_decompressobj(ZstdDecompressor* self, PyObject* args, PyObject* kwargs) {
	static char* kwlist[] = {
		"write_size",
		NULL
	};

	ZstdDecompressionObj* result = NULL;
	size_t outSize = ZSTD_DStreamOutSize();

	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|k:decompressobj", kwlist, &outSize)) {
		return NULL;
	}

	if (!outSize) {
		PyErr_SetString(PyExc_ValueError, "write_size must be positive");
		return NULL;
	}

	result = (ZstdDecompressionObj*)PyObject_CallObject((PyObject*)&ZstdDecompressionObjType, NULL);
	if (!result) {
		return NULL;
	}

	if (ensure_dctx(self, 1)) {
		Py_DECREF(result);
		return NULL;
	}

	result->decompressor = self;
	Py_INCREF(result->decompressor);
	result->outSize = outSize;

	return result;
}

PyDoc_STRVAR(Decompressor_read_to_iter__doc__,
"read_to_iter(reader[, read_size=default, write_size=default, skip_bytes=0])\n"
"Read compressed data and return an iterator\n"
"\n"
"Returns an iterator of decompressed data chunks produced from reading from\n"
"the ``reader``.\n"
"\n"

"\n"
"There is also support for skipping the first ``skip_bytes`` of data from\n"
"the source.\n"
);

static ZstdDecompressorIterator* Decompressor_read_to_iter(ZstdDecompressor* self, PyObject* args, PyObject* kwargs) {
	static char* kwlist[] = {
		"reader",
		"read_size",
		"write_size",
		"skip_bytes",

	size_t inSize = ZSTD_DStreamInSize();
	size_t outSize = ZSTD_DStreamOutSize();
	ZstdDecompressorIterator* result;
	size_t skipBytes = 0;

	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O|kkk:read_to_iter", kwlist,
		&reader, &inSize, &outSize, &skipBytes)) {
		return NULL;
	}

	if (skipBytes >= inSize) {

		result->reader = reader;
		Py_INCREF(result->reader);
	}
	else if (1 == PyObject_CheckBuffer(reader)) {
		/* Object claims it is a buffer. Try to get a handle to it. */
		if (0 != PyObject_GetBuffer(reader, &result->buffer, PyBUF_CONTIG_RO)) {
			goto except;
		}
	}
	else {
		PyErr_SetString(PyExc_ValueError,

	result->inSize = inSize;
	result->outSize = outSize;
	result->skipBytes = skipBytes;

	if (ensure_dctx(self, 1)) {
		goto except;
	}

	result->input.src = PyMem_Malloc(inSize);
	if (!result->input.src) {

	}

	goto finally;

except:
	Py_CLEAR(result);

finally:

	return result;
}

PyDoc_STRVAR(Decompressor_stream_reader__doc__,
"stream_reader(source, [read_size=default])\n"
"\n"
"Obtain an object that behaves like an I/O stream that can be used for\n"
"reading decompressed output from an object.\n"
"\n"
"The source object can be any object with a ``read(size)`` method or that\n"
"conforms to the buffer protocol.\n"
);

static ZstdDecompressionReader* Decompressor_stream_reader(ZstdDecompressor* self, PyObject* args, PyObject* kwargs) {
	static char* kwlist[] = {
		"source",
		"read_size",
		NULL
	};

	PyObject* source;
	size_t readSize = ZSTD_DStreamInSize();
	ZstdDecompressionReader* result;

	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O|k:stream_reader", kwlist,
		&source, &readSize)) {
		return NULL;
	}

	result = (ZstdDecompressionReader*)PyObject_CallObject((PyObject*)&ZstdDecompressionReaderType, NULL);
	if (NULL == result) {
		return NULL;
	}

	if (PyObject_HasAttrString(source, "read")) {
		result->reader = source;
		Py_INCREF(source);
		result->readSize = readSize;
	}
	else if (1 == PyObject_CheckBuffer(source)) {
		if (0 != PyObject_GetBuffer(source, &result->buffer, PyBUF_CONTIG_RO)) {
			Py_CLEAR(result);
			return NULL;
		}
	}
	else {
		PyErr_SetString(PyExc_TypeError,
			"must pass an object with a read() method or that conforms to the buffer protocol");
		Py_CLEAR(result);
		return NULL;
	}

	result->decompressor = self;
	Py_INCREF(self);

	return result;
}

PyDoc_STRVAR(Decompressor_stream_writer__doc__,
"Create a context manager to write decompressed data to an object.\n"
"\n"
"The passed object must have a ``write()`` method.\n"
"\n"
"The caller feeds intput data to the object by calling ``write(data)``.\n"

"An optional ``write_size`` argument defines the size of chunks to\n"
"``write()`` to the writer. It defaults to the default output size for a zstd\n"
"streaming decompressor.\n"
);

static ZstdDecompressionWriter* Decompressor_stream_writer(ZstdDecompressor* self, PyObject* args, PyObject* kwargs) {
	static char* kwlist[] = {
		"writer",
		"write_size",
		NULL
	};

	PyObject* writer;
	size_t outSize = ZSTD_DStreamOutSize();
	ZstdDecompressionWriter* result;

	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O|k:stream_writer", kwlist,
		&writer, &outSize)) {
		return NULL;
	}

	if (!PyObject_HasAttrString(writer, "write")) {

PyDoc_STRVAR(Decompressor_decompress_content_dict_chain__doc__,
"Decompress a series of chunks using the content dictionary chaining technique\n"
);

static PyObject* Decompressor_decompress_content_dict_chain(ZstdDecompressor* self, PyObject* args, PyObject* kwargs) {
	static char* kwlist[] = {
		"frames",
		NULL
	};

	Py_ssize_t chunkIndex;
	char parity = 0;
	PyObject* chunk;
	char* chunkData;
	Py_ssize_t chunkSize;
	size_t zresult;
	ZSTD_frameHeader frameHeader;
	void* buffer1 = NULL;
	size_t buffer1Size = 0;
	size_t buffer1ContentSize = 0;
	void* buffer2 = NULL;
	size_t buffer2Size = 0;
	size_t buffer2ContentSize = 0;
	void* destBuffer = NULL;
	PyObject* result = NULL;
	ZSTD_outBuffer outBuffer;
	ZSTD_inBuffer inBuffer;

	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "O!:decompress_content_dict_chain",
		kwlist, &PyList_Type, &chunks)) {
		return NULL;
	}

		return NULL;
	}

	/* We require that all chunks be zstd frames and that they have content size set. */
	PyBytes_AsStringAndSize(chunk, &chunkData, &chunkSize);
	zresult = ZSTD_getFrameHeader(&frameHeader, (void*)chunkData, chunkSize);
	if (ZSTD_isError(zresult)) {
		PyErr_SetString(PyExc_ValueError, "chunk 0 is not a valid zstd frame");
		return NULL;
	}
	else if (zresult) {
		PyErr_SetString(PyExc_ValueError, "chunk 0 is too small to contain a zstd frame");
		return NULL;
	}

	if (ZSTD_CONTENTSIZE_UNKNOWN == frameHeader.frameContentSize) {
		PyErr_SetString(PyExc_ValueError, "chunk 0 missing content size in frame");
		return NULL;
	}

	assert(ZSTD_CONTENTSIZE_ERROR != frameHeader.frameContentSize);

	/* We check against PY_SSIZE_T_MAX here because we ultimately cast the
	 * result to a Python object and it's length can be no greater than
	 * Py_ssize_t. In theory, we could have an intermediate frame that is 
	 * larger. But a) why would this API be used for frames that large b)
	 * it isn't worth the complexity to support. */
	assert(SIZE_MAX >= PY_SSIZE_T_MAX);
	if (frameHeader.frameContentSize > PY_SSIZE_T_MAX) {
		PyErr_SetString(PyExc_ValueError,
			"chunk 0 is too large to decompress on this platform");
		return NULL;
	}

	if (ensure_dctx(self, 0)) {
		goto finally;
	}

	buffer1Size = (size_t)frameHeader.frameContentSize;
	buffer1 = PyMem_Malloc(buffer1Size);
	if (!buffer1) {
		goto finally;
	}

	outBuffer.dst = buffer1;
	outBuffer.size = buffer1Size;
	outBuffer.pos = 0;

	inBuffer.src = chunkData;
	inBuffer.size = chunkSize;
	inBuffer.pos = 0;

	Py_BEGIN_ALLOW_THREADS
	zresult = ZSTD_decompress_generic(self->dctx, &outBuffer, &inBuffer);
	Py_END_ALLOW_THREADS
	if (ZSTD_isError(zresult)) {
		PyErr_Format(ZstdError, "could not decompress chunk 0: %s", ZSTD_getErrorName(zresult));
		goto finally;
	}
	else if (zresult) {
		PyErr_Format(ZstdError, "chunk 0 did not decompress full frame");
		goto finally;
	}

	buffer1ContentSize = outBuffer.pos;

	/* Special case of a simple chain. */
	if (1 == chunksLen) {
		result = PyBytes_FromStringAndSize(buffer1, buffer1Size);
		goto finally;
	}

	/* This should ideally look at next chunk. But this is slightly simpler. */
	buffer2Size = (size_t)frameHeader.frameContentSize;
	buffer2 = PyMem_Malloc(buffer2Size);
	if (!buffer2) {
		goto finally;
	}

			PyErr_Format(PyExc_ValueError, "chunk %zd must be bytes", chunkIndex);
			goto finally;
		}

		PyBytes_AsStringAndSize(chunk, &chunkData, &chunkSize);
		zresult = ZSTD_getFrameHeader(&frameHeader, (void*)chunkData, chunkSize);
		if (ZSTD_isError(zresult)) {
			PyErr_Format(PyExc_ValueError, "chunk %zd is not a valid zstd frame", chunkIndex);
			goto finally;
		}
		else if (zresult) {
			PyErr_Format(PyExc_ValueError, "chunk %zd is too small to contain a zstd frame", chunkIndex);
			goto finally;
		}

		if (ZSTD_CONTENTSIZE_UNKNOWN == frameHeader.frameContentSize) {
			PyErr_Format(PyExc_ValueError, "chunk %zd missing content size in frame", chunkIndex);
			goto finally;
		}

		assert(ZSTD_CONTENTSIZE_ERROR != frameHeader.frameContentSize);

		if (frameHeader.frameContentSize > PY_SSIZE_T_MAX) {
			PyErr_Format(PyExc_ValueError,
				"chunk %zd is too large to decompress on this platform", chunkIndex);
			goto finally;
		}

		inBuffer.src = chunkData;
		inBuffer.size = chunkSize;
		inBuffer.pos = 0;

		parity = chunkIndex % 2;

		/* This could definitely be abstracted to reduce code duplication. */
		if (parity) {
			/* Resize destination buffer to hold larger content. */
			if (buffer2Size < frameHeader.frameContentSize) {
				buffer2Size = (size_t)frameHeader.frameContentSize;
				destBuffer = PyMem_Realloc(buffer2, buffer2Size);
				if (!destBuffer) {
					goto finally;
				}
				buffer2 = destBuffer;
			}

			Py_BEGIN_ALLOW_THREADS
			zresult = ZSTD_DCtx_refPrefix_advanced(self->dctx,
				buffer1, buffer1ContentSize, ZSTD_dct_rawContent);
			Py_END_ALLOW_THREADS
			if (ZSTD_isError(zresult)) {
				PyErr_Format(ZstdError,
					"failed to load prefix dictionary at chunk %zd", chunkIndex);
				goto finally;
			}

			outBuffer.dst = buffer2;
			outBuffer.size = buffer2Size;
			outBuffer.pos = 0;

			Py_BEGIN_ALLOW_THREADS
			zresult = ZSTD_decompress_generic(self->dctx, &outBuffer, &inBuffer);
			Py_END_ALLOW_THREADS
			if (ZSTD_isError(zresult)) {
				PyErr_Format(ZstdError, "could not decompress chunk %zd: %s",
					chunkIndex, ZSTD_getErrorName(zresult));
				goto finally;
			}
			else if (zresult) {
				PyErr_Format(ZstdError, "chunk %zd did not decompress full frame",
					chunkIndex);
				goto finally;
			}

			buffer2ContentSize = outBuffer.pos;
		}
		else {
			if (buffer1Size < frameHeader.frameContentSize) {
				buffer1Size = (size_t)frameHeader.frameContentSize;
				destBuffer = PyMem_Realloc(buffer1, buffer1Size);
				if (!destBuffer) {
					goto finally;
				}
				buffer1 = destBuffer;
			}

			Py_BEGIN_ALLOW_THREADS
			zresult = ZSTD_DCtx_refPrefix_advanced(self->dctx,
				buffer2, buffer2ContentSize, ZSTD_dct_rawContent);
			Py_END_ALLOW_THREADS
			if (ZSTD_isError(zresult)) {
				PyErr_Format(ZstdError,
					"failed to load prefix dictionary at chunk %zd", chunkIndex);
				goto finally;
			}

			outBuffer.dst = buffer1;
			outBuffer.size = buffer1Size;
			outBuffer.pos = 0;

			Py_BEGIN_ALLOW_THREADS
			zresult = ZSTD_decompress_generic(self->dctx, &outBuffer, &inBuffer);
			Py_END_ALLOW_THREADS
			if (ZSTD_isError(zresult)) {
				PyErr_Format(ZstdError, "could not decompress chunk %zd: %s",
					chunkIndex, ZSTD_getErrorName(zresult));
				goto finally;
			}
			else if (zresult) {
				PyErr_Format(ZstdError, "chunk %zd did not decompress full frame",
					chunkIndex);
				goto finally;
			}

			buffer1ContentSize = outBuffer.pos;
		}
	}

	result = PyBytes_FromStringAndSize(parity ? buffer2 : buffer1,
		parity ? buffer2ContentSize : buffer1ContentSize);

	}
	if (buffer1) {
		PyMem_Free(buffer1);
	}

	return result;
}

typedef struct {
	void* sourceData;
	size_t sourceSize;
	size_t destSize;
} FramePointer;

typedef struct {
	FramePointer* frames;
	Py_ssize_t framesSize;

	Py_ssize_t endOffset;
	unsigned long long totalSourceSize;

	/* Compression state and settings. */
	ZSTD_DCtx* dctx;
	int requireOutputSizes;

	/* Output storage. */
	DestBuffer* destBuffers;
	Py_ssize_t destCount;

	assert(NULL == state->destBuffers);
	assert(0 == state->destCount);
	assert(state->endOffset - state->startOffset >= 0);

	/* We could get here due to the way work is allocated. Ideally we wouldn't
	   get here. But that would require a bit of a refactor in the caller. */
	if (state->totalSourceSize > SIZE_MAX) {
		state->error = WorkerError_memory;
		state->errorOffset = 0;
		return;
	}

	/*
	 * We need to allocate a buffer to hold decompressed data. How we do this
	 * depends on what we know about the output. The following scenarios are
	 * possible:
	 *

	 */

	/* Resolve ouput segments. */
	for (frameIndex = state->startOffset; frameIndex <= state->endOffset; frameIndex++) {
		FramePointer* fp = &framePointers[frameIndex];
		unsigned long long decompressedSize;

		if (0 == fp->destSize) {
			decompressedSize = ZSTD_getFrameContentSize(fp->sourceData, fp->sourceSize);

			if (ZSTD_CONTENTSIZE_ERROR == decompressedSize) {
				state->error = WorkerError_unknownSize;
				state->errorOffset = frameIndex;
				return;
			}
			else if (ZSTD_CONTENTSIZE_UNKNOWN == decompressedSize) {
				if (state->requireOutputSizes) {
					state->error = WorkerError_unknownSize;
					state->errorOffset = frameIndex;
					return;
				}

				/* This will fail the assert for .destSize > 0 below. */
				decompressedSize = 0;
			}

			if (decompressedSize > SIZE_MAX) {
				state->error = WorkerError_memory;
				state->errorOffset = frameIndex;
				return;
			}

			fp->destSize = (size_t)decompressedSize;
		}

		totalOutputSize += fp->destSize;
	}

	destBuffer = &state->destBuffers[state->destCount - 1];

	assert(framePointers[state->startOffset].destSize > 0); /* For now. */

	allocationSize = roundpow2((size_t)state->totalSourceSize);

	if (framePointers[state->startOffset].destSize > allocationSize) {
		allocationSize = roundpow2(framePointers[state->startOffset].destSize);
	}

	}

	destBuffer->segmentsSize = remainingItems;

	for (frameIndex = state->startOffset; frameIndex <= state->endOffset; frameIndex++) {
		ZSTD_outBuffer outBuffer;
		ZSTD_inBuffer inBuffer;
		const void* source = framePointers[frameIndex].sourceData;
		const size_t sourceSize = framePointers[frameIndex].sourceSize;
		void* dest;
		const size_t decompressedSize = framePointers[frameIndex].destSize;
		size_t destAvailable = destBuffer->destSize - destOffset;

			destBuffer = &state->destBuffers[state->destCount - 1];

			/* Don't take any chances will non-NULL pointers. */
			memset(destBuffer, 0, sizeof(DestBuffer));

			allocationSize = roundpow2((size_t)state->totalSourceSize);

			if (decompressedSize > allocationSize) {
				allocationSize = roundpow2(decompressedSize);
			}

			currentBufferStartIndex = frameIndex;
		}

		dest = (char*)destBuffer->dest + destOffset;

		outBuffer.dst = dest;
		outBuffer.size = decompressedSize;
		outBuffer.pos = 0;

		inBuffer.src = source;
		inBuffer.size = sourceSize;
		inBuffer.pos = 0;

		zresult = ZSTD_decompress_generic(state->dctx, &outBuffer, &inBuffer);
		if (ZSTD_isError(zresult)) {
			state->error = WorkerError_zstd;
			state->zresult = zresult;
			state->errorOffset = frameIndex;
			return;
		}
		else if (zresult || outBuffer.pos != decompressedSize) {
			state->error = WorkerError_sizeMismatch;
			state->zresult = outBuffer.pos;
			state->errorOffset = frameIndex;
			return;
		}

		destBuffer->segments[localOffset].offset = destOffset;
		destBuffer->segments[localOffset].length = outBuffer.pos;
		destOffset += outBuffer.pos;
		localOffset++;
		remainingItems--;
	}

	if (destBuffer->destSize > destOffset) {

		destBuffer->destSize = destOffset;
	}
}

ZstdBufferWithSegmentsCollection* decompress_from_framesources(ZstdDecompressor* decompressor, FrameSources* frames,
	Py_ssize_t threadCount) {
	Py_ssize_t i = 0;
	int errored = 0;
	Py_ssize_t segmentsCount;
	ZstdBufferWithSegments* bws = NULL;
	PyObject* resultArg = NULL;
	Py_ssize_t resultIndex;
	ZstdBufferWithSegmentsCollection* result = NULL;
	FramePointer* framePointers = frames->frames;
	unsigned long long workerBytes = 0;
	Py_ssize_t currentThread = 0;
	Py_ssize_t workerStartOffset = 0;
	POOL_ctx* pool = NULL;
	WorkerState* workerStates = NULL;
	unsigned long long bytesPerWorker;

	/* Caller should normalize 0 and negative values to 1 or larger. */
	assert(threadCount >= 1);

	/* More threads than inputs makes no sense under any conditions. */
	threadCount = frames->framesSize < threadCount ? frames->framesSize
												   : threadCount;

	/* TODO lower thread count if input size is too small and threads would just
	   add overhead. */

	if (decompressor->dict) {
		if (ensure_ddict(decompressor->dict)) {
			return NULL;
		}
	}

	/* If threadCount==1, we don't start a thread pool. But we do leverage the

		}
	}

	bytesPerWorker = frames->compressedSize / threadCount;

	if (bytesPerWorker > SIZE_MAX) {
		PyErr_SetString(ZstdError, "too much data per worker for this platform");
		goto finally;
	}

	for (i = 0; i < threadCount; i++) {
		size_t zresult;

		workerStates[i].dctx = ZSTD_createDCtx();
		if (NULL == workerStates[i].dctx) {
			PyErr_NoMemory();
			goto finally;
		}

		ZSTD_copyDCtx(workerStates[i].dctx, decompressor->dctx);

		if (decompressor->dict) {
			zresult = ZSTD_DCtx_refDDict(workerStates[i].dctx, decompressor->dict->ddict);
			if (zresult) {
				PyErr_Format(ZstdError, "unable to reference prepared dictionary: %s",
					ZSTD_getErrorName(zresult));
				goto finally;
			}
		}

		workerStates[i].framePointers = framePointers;
		workerStates[i].requireOutputSizes = 1;
	}

	Py_BEGIN_ALLOW_THREADS

			PyErr_NoMemory();
			errored = 1;
			break;

		case WorkerError_sizeMismatch:
			PyErr_Format(ZstdError, "error decompressing item %zd: decompressed %zu bytes; expected %zu",
				workerStates[i].errorOffset, workerStates[i].zresult,
				framePointers[workerStates[i].errorOffset].destSize);
			errored = 1;
			break;

			if (frameSizesP) {
				decompressedSize = frameSizesP[i];
			}

			if (sourceSize > SIZE_MAX) {
				PyErr_Format(PyExc_ValueError,
					"item %zd is too large for this platform", i);
				goto finally;
			}

			if (decompressedSize > SIZE_MAX) {
				PyErr_Format(PyExc_ValueError,
					"decompressed size of item %zd is too large for this platform", i);
				goto finally;
			}

			framePointers[i].sourceData = sourceData;
			framePointers[i].sourceSize = (size_t)sourceSize;
			framePointers[i].destSize = (size_t)decompressedSize;
		}
	}
	else if (PyObject_TypeCheck(frames, &ZstdBufferWithSegmentsCollectionType)) {
		Py_ssize_t offset = 0;
		ZstdBufferWithSegments* buffer;

		for (i = 0; i < collection->bufferCount; i++) {
			Py_ssize_t segmentIndex;
			buffer = collection->buffers[i];

			for (segmentIndex = 0; segmentIndex < buffer->segmentCount; segmentIndex++) {
				unsigned long long decompressedSize = frameSizesP ? frameSizesP[offset] : 0;

				if (buffer->segments[segmentIndex].offset + buffer->segments[segmentIndex].length > buffer->dataSize) {
					PyErr_Format(PyExc_ValueError, "item %zd has offset outside memory area",
						offset);
					goto finally;
				}

				if (buffer->segments[segmentIndex].length > SIZE_MAX) {
					PyErr_Format(PyExc_ValueError,
						"item %zd in buffer %zd is too large for this platform",
						segmentIndex, i);
					goto finally;
				}

				if (decompressedSize > SIZE_MAX) {
					PyErr_Format(PyExc_ValueError,
						"decompressed size of item %zd in buffer %zd is too large for this platform",
						segmentIndex, i);
					goto finally;
				}

				totalInputSize += buffer->segments[segmentIndex].length;

				framePointers[offset].sourceData = (char*)buffer->data + buffer->segments[segmentIndex].offset;
				framePointers[offset].sourceSize = (size_t)buffer->segments[segmentIndex].length;
				framePointers[offset].destSize = (size_t)decompressedSize;

				offset++;
			}
		}
	}

		if (!framePointers) {
			PyErr_NoMemory();
			goto finally;
		}

		frameBuffers = PyMem_Malloc(frameCount * sizeof(Py_buffer));
		if (NULL == frameBuffers) {
			PyErr_NoMemory();
			goto finally;
		}

		memset(frameBuffers, 0, frameCount * sizeof(Py_buffer));

		/* Do a pass to assemble info about our input buffers and output sizes. */
		for (i = 0; i < frameCount; i++) {
			unsigned long long decompressedSize = frameSizesP ? frameSizesP[i] : 0;

			if (0 != PyObject_GetBuffer(PyList_GET_ITEM(frames, i),
				&frameBuffers[i], PyBUF_CONTIG_RO)) {
				PyErr_Clear();
				PyErr_Format(PyExc_TypeError, "item %zd not a bytes like object", i);
				goto finally;
			}

			if (decompressedSize > SIZE_MAX) {
				PyErr_Format(PyExc_ValueError,
					"decompressed size of item %zd is too large for this platform", i);
				goto finally;
			}

			totalInputSize += frameBuffers[i].len;

			framePointers[i].sourceData = frameBuffers[i].buf;
			framePointers[i].sourceSize = frameBuffers[i].len;
			framePointers[i].destSize = (size_t)decompressedSize;
		}
	}
	else {
		PyErr_SetString(PyExc_TypeError, "argument must be list or BufferWithSegments");
		goto finally;

static PyMethodDef Decompressor_methods[] = {
	{ "copy_stream", (PyCFunction)Decompressor_copy_stream, METH_VARARGS | METH_KEYWORDS,
	Decompressor_copy_stream__doc__ },
	{ "decompress", (PyCFunction)Decompressor_decompress, METH_VARARGS | METH_KEYWORDS,
	Decompressor_decompress__doc__ },
	{ "decompressobj", (PyCFunction)Decompressor_decompressobj, METH_VARARGS | METH_KEYWORDS,
	Decompressor_decompressobj__doc__ },
	{ "read_to_iter", (PyCFunction)Decompressor_read_to_iter, METH_VARARGS | METH_KEYWORDS,
	Decompressor_read_to_iter__doc__ },
	/* TODO Remove deprecated API */
	{ "read_from", (PyCFunction)Decompressor_read_to_iter, METH_VARARGS | METH_KEYWORDS,
	Decompressor_read_to_iter__doc__ },
	{ "stream_reader", (PyCFunction)Decompressor_stream_reader,
	METH_VARARGS | METH_KEYWORDS, Decompressor_stream_reader__doc__ },
	{ "stream_writer", (PyCFunction)Decompressor_stream_writer, METH_VARARGS | METH_KEYWORDS,
	Decompressor_stream_writer__doc__ },
	/* TODO remove deprecated API */
	{ "write_to", (PyCFunction)Decompressor_stream_writer, METH_VARARGS | METH_KEYWORDS,
	Decompressor_stream_writer__doc__ },
	{ "decompress_content_dict_chain", (PyCFunction)Decompressor_decompress_content_dict_chain,
	  METH_VARARGS | METH_KEYWORDS, Decompressor_decompress_content_dict_chain__doc__ },
	{ "multi_decompress_to_buffer", (PyCFunction)Decompressor_multi_decompress_to_buffer,
	  METH_VARARGS | METH_KEYWORDS, Decompressor_multi_decompress_to_buffer__doc__ },
	{ "memory_size", (PyCFunction)Decompressor_memory_size, METH_NOARGS,
	Decompressor_memory_size__doc__ },
	{ NULL, NULL }
};

PyTypeObject ZstdDecompressorType = {
	PyVarObject_HEAD_INIT(NULL, 0)