1 | // |
---|
2 | // Created by wxy325 on 2018/3/16. |
---|
3 | // |
---|
4 | |
---|
5 | #include "lz4_index_builder.h" |
---|
6 | |
---|
7 | |
---|
8 | #include <kernels/kernel_builder.h> |
---|
9 | #include <iostream> |
---|
10 | #include <string> |
---|
11 | #include <llvm/Support/raw_ostream.h> |
---|
12 | #include <kernels/streamset.h> |
---|
13 | |
---|
14 | using namespace llvm; |
---|
15 | using namespace kernel; |
---|
16 | using namespace std; |
---|
17 | |
---|
18 | namespace kernel{ |
---|
19 | |
---|
20 | LZ4IndexBuilderKernel::LZ4IndexBuilderKernel(const std::unique_ptr<kernel::KernelBuilder> &iBuilder) |
---|
21 | : SegmentOrientedKernel("LZ4IndexBuilderKernel", |
---|
22 | // Inputs |
---|
23 | { |
---|
24 | Binding{iBuilder->getStreamSetTy(1, 8), "byteStream", BoundedRate(0, 1)}, |
---|
25 | Binding{iBuilder->getStreamSetTy(1, 1), "extender", RateEqualTo("byteStream")}, |
---|
26 | |
---|
27 | // block data |
---|
28 | Binding{iBuilder->getStreamSetTy(1, 1), "isCompressed", BoundedRate(0, 1), |
---|
29 | AlwaysConsume()}, |
---|
30 | Binding{iBuilder->getStreamSetTy(1, 64), "blockStart", BoundedRate(0, 1), |
---|
31 | AlwaysConsume()}, |
---|
32 | Binding{iBuilder->getStreamSetTy(1, 64), "blockEnd", BoundedRate(0, 1), |
---|
33 | AlwaysConsume()} |
---|
34 | |
---|
35 | }, |
---|
36 | //Outputs |
---|
37 | { |
---|
38 | // Uncompressed_data |
---|
39 | Binding{iBuilder->getStreamSetTy(1, 64), "uncompressedStartPos", |
---|
40 | BoundedRate(0, 1)}, |
---|
41 | Binding{iBuilder->getStreamSetTy(1, 64), "uncompressedLength", |
---|
42 | BoundedRate(0, 1)}, |
---|
43 | Binding{iBuilder->getStreamSetTy(1, 64), "uncompressedOutputPos", |
---|
44 | BoundedRate(0, 1)}, |
---|
45 | |
---|
46 | Binding{iBuilder->getStreamSetTy(1, 1), "deletionMarker", BoundedRate(0, 1)}, |
---|
47 | Binding{iBuilder->getStreamSetTy(1, 64), "m0Start", BoundedRate(0, 1)}, |
---|
48 | Binding{iBuilder->getStreamSetTy(1, 64), "m0End", BoundedRate(0, 1)}, |
---|
49 | Binding{iBuilder->getStreamSetTy(1, 64), "matchOffset", BoundedRate(0, 1)}, |
---|
50 | Binding{iBuilder->getStreamSetTy(1, 1), "M0Marker", BoundedRate(0, 1)} |
---|
51 | }, |
---|
52 | //Arguments |
---|
53 | { |
---|
54 | Binding{iBuilder->getSizeTy(), "fileSize"} |
---|
55 | }, |
---|
56 | {}, |
---|
57 | //Internal states: |
---|
58 | { |
---|
59 | Binding{iBuilder->getSizeTy(), "blockDataIndex"}, |
---|
60 | Binding{iBuilder->getInt64Ty(), "m0OutputPos"} |
---|
61 | }) { |
---|
62 | this->setStride(4 * 1024 * 1024); |
---|
63 | addAttribute(MustExplicitlyTerminate()); |
---|
64 | } |
---|
65 | |
---|
66 | void LZ4IndexBuilderKernel::generateDoSegmentMethod(const std::unique_ptr<KernelBuilder> &iBuilder) { |
---|
67 | |
---|
68 | BasicBlock* exitBlock = iBuilder->CreateBasicBlock("exitBlock"); |
---|
69 | BasicBlock* blockEndConBlock = iBuilder->CreateBasicBlock("blockEndConBlock"); |
---|
70 | |
---|
71 | Value * blockDataIndex = iBuilder->getScalarField("blockDataIndex"); |
---|
72 | |
---|
73 | Value * totalNumber = iBuilder->CreateAdd(iBuilder->getAvailableItemCount("blockEnd"), iBuilder->getProcessedItemCount("blockEnd")); |
---|
74 | Value * totalExtender = iBuilder->CreateAdd(iBuilder->getAvailableItemCount("extender"), iBuilder->getProcessedItemCount("extender")); |
---|
75 | |
---|
76 | Value * blockEnd = this->generateLoadInt64NumberInput(iBuilder, "blockEnd", blockDataIndex); |
---|
77 | |
---|
78 | iBuilder->CreateCondBr(iBuilder->CreateICmpULT(blockDataIndex, totalNumber), blockEndConBlock, exitBlock); |
---|
79 | |
---|
80 | iBuilder->SetInsertPoint(blockEndConBlock); |
---|
81 | Value * blockStart = this->generateLoadInt64NumberInput(iBuilder, "blockStart", blockDataIndex); |
---|
82 | BasicBlock * processBlock = iBuilder->CreateBasicBlock("processBlock"); |
---|
83 | iBuilder->CreateCondBr(iBuilder->CreateICmpULE(blockEnd, totalExtender), processBlock, exitBlock); |
---|
84 | |
---|
85 | iBuilder->SetInsertPoint(processBlock); |
---|
86 | |
---|
87 | //TODO handle uncompressed block |
---|
88 | |
---|
89 | this->generateProcessCompressedBlock(iBuilder, blockStart, blockEnd); |
---|
90 | |
---|
91 | Value * newBlockDataIndex = iBuilder->CreateAdd(blockDataIndex, iBuilder->getInt64(1)); |
---|
92 | iBuilder->setScalarField("blockDataIndex", newBlockDataIndex); |
---|
93 | iBuilder->setProcessedItemCount("blockEnd", newBlockDataIndex); |
---|
94 | iBuilder->setProcessedItemCount("blockStart", newBlockDataIndex); |
---|
95 | iBuilder->setProcessedItemCount("isCompressed", newBlockDataIndex); |
---|
96 | |
---|
97 | iBuilder->setProcessedItemCount("byteStream", blockEnd); |
---|
98 | iBuilder->CreateBr(exitBlock); |
---|
99 | |
---|
100 | iBuilder->SetInsertPoint(exitBlock); |
---|
101 | } |
---|
102 | |
---|
103 | Value* LZ4IndexBuilderKernel::processLiteral(const std::unique_ptr<KernelBuilder> &iBuilder, Value* token, Value* tokenPos, Value* blockEnd) { |
---|
104 | BasicBlock* entryBlock = iBuilder->GetInsertBlock(); |
---|
105 | |
---|
106 | Value * extendedLiteralValue = iBuilder->CreateICmpEQ(iBuilder->CreateAnd(token, iBuilder->getInt8(0xf0)), iBuilder->getInt8(0xf0)); |
---|
107 | |
---|
108 | BasicBlock* extendLiteralLengthBody = iBuilder->CreateBasicBlock("block_data_loop_handle_compressed_block_extend_literal_length_body"); |
---|
109 | BasicBlock* extendLiteralLengthExit = iBuilder->CreateBasicBlock("block_data_loop_handle_compressed_block_extend_literal_length_exit"); |
---|
110 | |
---|
111 | iBuilder->CreateCondBr(extendedLiteralValue, extendLiteralLengthBody, extendLiteralLengthExit); |
---|
112 | |
---|
113 | iBuilder->SetInsertPoint(extendLiteralLengthBody); |
---|
114 | Value* newCursorPos = this->advanceUntilNextZero(iBuilder, "extender", iBuilder->CreateAdd(tokenPos, iBuilder->getInt64(1)), blockEnd); |
---|
115 | BasicBlock* advanceFinishBlock = iBuilder->GetInsertBlock(); |
---|
116 | |
---|
117 | iBuilder->CreateBr(extendLiteralLengthExit); |
---|
118 | |
---|
119 | iBuilder->SetInsertPoint(extendLiteralLengthExit); |
---|
120 | |
---|
121 | PHINode* phiCursorPosAfterLiteral = iBuilder->CreatePHI(iBuilder->getInt64Ty(), 2); |
---|
122 | phiCursorPosAfterLiteral->addIncoming(newCursorPos, advanceFinishBlock); |
---|
123 | phiCursorPosAfterLiteral->addIncoming(tokenPos, entryBlock); |
---|
124 | |
---|
125 | Value * literalExtensionSize = iBuilder->CreateSub(phiCursorPosAfterLiteral, tokenPos); |
---|
126 | Value * finalLengthByte = this->generateLoadSourceInputByte(iBuilder, phiCursorPosAfterLiteral); |
---|
127 | finalLengthByte = iBuilder->CreateZExt(finalLengthByte, iBuilder->getInt64Ty()); |
---|
128 | Value * literalLengthExtendValue = iBuilder->CreateSelect( |
---|
129 | iBuilder->CreateICmpUGT(literalExtensionSize, iBuilder->getSize(0)), |
---|
130 | iBuilder->CreateAdd( |
---|
131 | iBuilder->CreateMul( |
---|
132 | iBuilder->CreateSub(literalExtensionSize, iBuilder->getSize(1)), |
---|
133 | iBuilder->getSize(255) |
---|
134 | ), |
---|
135 | finalLengthByte |
---|
136 | ), |
---|
137 | iBuilder->getSize(0) |
---|
138 | ); |
---|
139 | literalLengthExtendValue = iBuilder->CreateZExt(literalLengthExtendValue, iBuilder->getInt64Ty()); |
---|
140 | Value* literalLengthBase = iBuilder->CreateLShr(iBuilder->CreateZExt(token, iBuilder->getInt64Ty()), iBuilder->getInt64(4)); |
---|
141 | Value* literalLength = iBuilder->CreateAdd(literalLengthBase, literalLengthExtendValue); |
---|
142 | |
---|
143 | Value* offsetPos = iBuilder->CreateAdd( |
---|
144 | iBuilder->CreateAdd( |
---|
145 | phiCursorPosAfterLiteral, |
---|
146 | literalLength), |
---|
147 | iBuilder->getSize(1)); |
---|
148 | |
---|
149 | // TODO Clear Output Buffer at the beginning instead of marking 0 |
---|
150 | this->markCircularOutputBitstream(iBuilder, "deletionMarker", iBuilder->getProducedItemCount("deletionMarker"), iBuilder->CreateAdd(phiCursorPosAfterLiteral, iBuilder->getSize(1)), true); |
---|
151 | this->markCircularOutputBitstream(iBuilder, "deletionMarker", iBuilder->CreateAdd(phiCursorPosAfterLiteral, iBuilder->getSize(1)), offsetPos, false); |
---|
152 | this->increaseScalarField(iBuilder, "m0OutputPos", literalLength); //TODO m0OutputPos may be removed from scalar fields |
---|
153 | return offsetPos; |
---|
154 | } |
---|
155 | |
---|
156 | Value* LZ4IndexBuilderKernel::processMatch(const std::unique_ptr<KernelBuilder> &iBuilder, Value* offsetPos, Value* token, Value* blockEnd) { |
---|
157 | Constant* INT64_ONE = iBuilder->getInt64(1); |
---|
158 | |
---|
159 | BasicBlock* entryBlock = iBuilder->GetInsertBlock(); |
---|
160 | |
---|
161 | Value* matchLengthStartPos = iBuilder->CreateAdd(offsetPos, INT64_ONE); |
---|
162 | Value* extendedMatchValue = iBuilder->CreateICmpEQ(iBuilder->CreateAnd(token, iBuilder->getInt8(0xf)), iBuilder->getInt8(0xf)); |
---|
163 | |
---|
164 | BasicBlock* extendMatchBodyBlock = iBuilder->CreateBasicBlock("block_data_loop_handle_compressed_block_loop_extend_match_body"); |
---|
165 | BasicBlock* extendMatchExitBlock = iBuilder->CreateBasicBlock("block_data_loop_handle_compressed_block_loop_extend_match_exit"); |
---|
166 | |
---|
167 | iBuilder->CreateCondBr(extendedMatchValue, extendMatchBodyBlock, extendMatchExitBlock); |
---|
168 | |
---|
169 | iBuilder->SetInsertPoint(extendMatchBodyBlock); |
---|
170 | |
---|
171 | //ExtendMatchBodyBlock |
---|
172 | Value* newCursorPos = this->advanceUntilNextZero(iBuilder, "extender", iBuilder->CreateAdd(matchLengthStartPos, INT64_ONE), blockEnd); |
---|
173 | BasicBlock* advanceFinishBlock = iBuilder->GetInsertBlock(); |
---|
174 | |
---|
175 | // ----May be in a different segment now |
---|
176 | iBuilder->CreateBr(extendMatchExitBlock); |
---|
177 | |
---|
178 | //ExtendMatchExitBlock |
---|
179 | iBuilder->SetInsertPoint(extendMatchExitBlock); |
---|
180 | PHINode* phiCursorPosAfterMatch = iBuilder->CreatePHI(iBuilder->getInt64Ty(), 2); |
---|
181 | phiCursorPosAfterMatch->addIncoming(newCursorPos, advanceFinishBlock); |
---|
182 | phiCursorPosAfterMatch->addIncoming(matchLengthStartPos, entryBlock); |
---|
183 | |
---|
184 | Value* oldMatchExtensionSize = iBuilder->CreateSub(phiCursorPosAfterMatch, matchLengthStartPos); |
---|
185 | extendedMatchValue = iBuilder->CreateICmpEQ(iBuilder->CreateAnd(token, iBuilder->getInt8(0xf)), iBuilder->getInt8(0xf)); |
---|
186 | Value* matchExtensionSize = iBuilder->CreateSelect( |
---|
187 | iBuilder->CreateICmpEQ(extendedMatchValue, iBuilder->getInt1(true)), |
---|
188 | oldMatchExtensionSize, |
---|
189 | iBuilder->getSize(0) |
---|
190 | ); |
---|
191 | Value* matchLengthBase = iBuilder->CreateZExt(iBuilder->CreateAnd(token, iBuilder->getInt8(0x0f)), iBuilder->getInt64Ty()); |
---|
192 | Value* matchLength = iBuilder->CreateAdd(matchLengthBase, iBuilder->getInt64(4)); |
---|
193 | |
---|
194 | |
---|
195 | Value* extensionLastBitPos = iBuilder->CreateAdd(offsetPos, iBuilder->getSize(1)); |
---|
196 | extensionLastBitPos = iBuilder->CreateAdd(extensionLastBitPos, matchExtensionSize); |
---|
197 | |
---|
198 | Value* extensionLastBitValue = this->generateLoadSourceInputByte(iBuilder, extensionLastBitPos); |
---|
199 | extensionLastBitValue = iBuilder->CreateZExt(extensionLastBitValue, iBuilder->getSizeTy()); |
---|
200 | |
---|
201 | |
---|
202 | Value* matchLengthAddValue = iBuilder->CreateSelect( |
---|
203 | iBuilder->CreateICmpUGT(matchExtensionSize, iBuilder->getSize(0)), |
---|
204 | iBuilder->CreateAdd( |
---|
205 | iBuilder->CreateMul( |
---|
206 | iBuilder->CreateSub(matchExtensionSize, iBuilder->getSize(1)), |
---|
207 | iBuilder->getSize(255) |
---|
208 | ), |
---|
209 | extensionLastBitValue |
---|
210 | ) |
---|
211 | , |
---|
212 | iBuilder->getSize(0) |
---|
213 | ); |
---|
214 | matchLengthAddValue = iBuilder->CreateZExt(matchLengthAddValue, iBuilder->getInt64Ty()); |
---|
215 | |
---|
216 | matchLength = iBuilder->CreateAdd(matchLength, matchLengthAddValue); |
---|
217 | |
---|
218 | Value* outputPos = iBuilder->getScalarField("m0OutputPos"); |
---|
219 | |
---|
220 | Value* outputEndPos = iBuilder->CreateSub( |
---|
221 | iBuilder->CreateAdd(outputPos, matchLength), |
---|
222 | iBuilder->getInt64(1) |
---|
223 | ); |
---|
224 | |
---|
225 | Value* matchOffset = iBuilder->CreateAdd( |
---|
226 | iBuilder->CreateZExt(this->generateLoadSourceInputByte(iBuilder, offsetPos), iBuilder->getSizeTy()), |
---|
227 | iBuilder->CreateShl(iBuilder->CreateZExt(this->generateLoadSourceInputByte(iBuilder, iBuilder->CreateAdd(offsetPos, iBuilder->getSize(1))), iBuilder->getSizeTy()), iBuilder->getSize(8)) |
---|
228 | ); |
---|
229 | this->generateStoreNumberOutput(iBuilder, "m0Start", outputPos); |
---|
230 | this->generateStoreNumberOutput(iBuilder, "m0End", outputEndPos); |
---|
231 | this->generateStoreNumberOutput(iBuilder, "matchOffset", matchOffset); |
---|
232 | this->increaseScalarField(iBuilder, "m0OutputPos", matchLength); |
---|
233 | this->markCircularOutputBitstream(iBuilder, "M0Marker", outputPos, outputEndPos, true, false); |
---|
234 | |
---|
235 | return iBuilder->CreateAdd(phiCursorPosAfterMatch, INT64_ONE); |
---|
236 | } |
---|
237 | |
---|
238 | void LZ4IndexBuilderKernel::generateProcessCompressedBlock(const std::unique_ptr<KernelBuilder> &iBuilder, Value* blockStart, Value* blockEnd) { |
---|
239 | // Constant |
---|
240 | BasicBlock* entryBlock = iBuilder->GetInsertBlock(); |
---|
241 | |
---|
242 | Value* m0OutputBlockPtr = iBuilder->getOutputStreamBlockPtr("M0Marker", iBuilder->getSize(0)); |
---|
243 | iBuilder->CreateMemSet(m0OutputBlockPtr, iBuilder->getInt8(0), 4 * 1024 * 1024 / 8, true); |
---|
244 | |
---|
245 | |
---|
246 | Value* isTerminal = iBuilder->CreateICmpEQ(blockEnd, iBuilder->getScalarField("fileSize")); |
---|
247 | iBuilder->setTerminationSignal(isTerminal); |
---|
248 | |
---|
249 | //TODO use memset to clear output buffer for extract marker |
---|
250 | |
---|
251 | BasicBlock* exitBlock = iBuilder->CreateBasicBlock("processCompressedExitBlock"); |
---|
252 | |
---|
253 | BasicBlock* processCon = iBuilder->CreateBasicBlock("processCompressedConBlock"); |
---|
254 | BasicBlock* processBody = iBuilder->CreateBasicBlock("processCompressedBodyBlock"); |
---|
255 | |
---|
256 | iBuilder->CreateBr(processCon); |
---|
257 | iBuilder->SetInsertPoint(processCon); |
---|
258 | |
---|
259 | PHINode* phiCursorValue = iBuilder->CreatePHI(iBuilder->getInt64Ty(), 3); // phiCursorValue should always be the position of next token except for the final sequence |
---|
260 | phiCursorValue->addIncoming(blockStart, entryBlock); |
---|
261 | |
---|
262 | iBuilder->CreateCondBr(iBuilder->CreateICmpULT(phiCursorValue, blockEnd), processBody, exitBlock); |
---|
263 | |
---|
264 | // Process Body |
---|
265 | iBuilder->SetInsertPoint(processBody); |
---|
266 | |
---|
267 | //TODO add acceleration here |
---|
268 | Value* token = this->generateLoadSourceInputByte(iBuilder, phiCursorValue); |
---|
269 | // Process Literal |
---|
270 | BasicBlock* processLiteralBlock = iBuilder->CreateBasicBlock("processLiteralBlock"); |
---|
271 | iBuilder->CreateBr(processLiteralBlock); |
---|
272 | iBuilder->SetInsertPoint(processLiteralBlock); |
---|
273 | |
---|
274 | Value* offsetPos = this->processLiteral(iBuilder, token, phiCursorValue, blockEnd); |
---|
275 | // Process Match |
---|
276 | BasicBlock* handleM0BodyBlock = iBuilder->CreateBasicBlock("block_data_loop_handle_compressed_block_loop_handle_m0_body"); |
---|
277 | BasicBlock* handleM0ElseBlock = iBuilder->CreateBasicBlock("block_data_loop_handle_compressed_block_loop_handle_m0_else"); |
---|
278 | |
---|
279 | iBuilder->CreateCondBr( |
---|
280 | iBuilder->CreateICmpULT(offsetPos, blockEnd), |
---|
281 | handleM0BodyBlock, |
---|
282 | handleM0ElseBlock |
---|
283 | ); |
---|
284 | |
---|
285 | // HandleM0Body |
---|
286 | iBuilder->SetInsertPoint(handleM0BodyBlock); |
---|
287 | Value* nextTokenPos = this->processMatch(iBuilder, offsetPos, token, blockEnd); |
---|
288 | phiCursorValue->addIncoming(nextTokenPos, iBuilder->GetInsertBlock()); |
---|
289 | |
---|
290 | iBuilder->CreateBr(processCon); |
---|
291 | |
---|
292 | |
---|
293 | // HandleM0Else |
---|
294 | iBuilder->SetInsertPoint(handleM0ElseBlock); |
---|
295 | |
---|
296 | phiCursorValue->addIncoming(offsetPos, handleM0ElseBlock); |
---|
297 | // Store final M0 pos to make sure the bit stream will be long enough |
---|
298 | Value* finalM0OutputPos = iBuilder->getScalarField("m0OutputPos"); |
---|
299 | this->generateStoreNumberOutput(iBuilder, "m0Start", finalM0OutputPos); |
---|
300 | this->generateStoreNumberOutput(iBuilder, "m0End", finalM0OutputPos); |
---|
301 | this->generateStoreNumberOutput(iBuilder, "matchOffset", iBuilder->getInt64(0)); |
---|
302 | iBuilder->setProducedItemCount("M0Marker", finalM0OutputPos); |
---|
303 | // finalM0OutputPos should always be 4MB * n except for the final block |
---|
304 | |
---|
305 | iBuilder->CreateBr(processCon); |
---|
306 | |
---|
307 | |
---|
308 | iBuilder->SetInsertPoint(exitBlock); |
---|
309 | } |
---|
310 | |
---|
311 | Value * LZ4IndexBuilderKernel::advanceUntilNextZero(const unique_ptr<KernelBuilder> &iBuilder, string inputName, Value * startPos, Value * maxPos) { |
---|
312 | |
---|
313 | unsigned int bitBlockWidth = iBuilder->getBitBlockWidth(); |
---|
314 | Constant* INT64_BIT_BLOCK_WIDTH = iBuilder->getInt64(bitBlockWidth); |
---|
315 | Type* bitBlockType = iBuilder->getBitBlockType(); |
---|
316 | Type* bitBlockWidthIntTy = iBuilder->getIntNTy(bitBlockWidth); |
---|
317 | |
---|
318 | BasicBlock* entryBlock = iBuilder->GetInsertBlock(); |
---|
319 | |
---|
320 | BasicBlock* advanceConBlock = iBuilder->CreateBasicBlock("advanceConBlock"); |
---|
321 | BasicBlock* advanceBodyBlock = iBuilder->CreateBasicBlock("advanceBodyBlock"); |
---|
322 | BasicBlock* advanceExitBlock = iBuilder->CreateBasicBlock("advanceExitBlock"); |
---|
323 | |
---|
324 | iBuilder->CreateBr(advanceConBlock); |
---|
325 | // TODO special handling for the first advance may have better performance |
---|
326 | iBuilder->SetInsertPoint(advanceConBlock); |
---|
327 | |
---|
328 | PHINode* phiCurrentPos = iBuilder->CreatePHI(iBuilder->getInt64Ty(), 2); |
---|
329 | phiCurrentPos->addIncoming(startPos, entryBlock); |
---|
330 | PHINode* phiIsFinish = iBuilder->CreatePHI(iBuilder->getInt1Ty(), 2); |
---|
331 | phiIsFinish->addIncoming(iBuilder->getInt1(false), entryBlock); |
---|
332 | iBuilder->CreateCondBr(iBuilder->CreateNot(phiIsFinish), advanceBodyBlock, advanceExitBlock); |
---|
333 | |
---|
334 | iBuilder->SetInsertPoint(advanceBodyBlock); |
---|
335 | |
---|
336 | Value * currentBlockGlobalPos = iBuilder->CreateAnd(phiCurrentPos, ConstantExpr::getNeg(INT64_BIT_BLOCK_WIDTH)); |
---|
337 | Value * currentPosBitBlockOffset = iBuilder->CreateURem(phiCurrentPos, INT64_BIT_BLOCK_WIDTH); |
---|
338 | |
---|
339 | Value * ptr = iBuilder->CreatePointerCast(iBuilder->getRawInputPointer(inputName, currentBlockGlobalPos), bitBlockType->getPointerTo()); |
---|
340 | |
---|
341 | Value * currentBitValue = iBuilder->CreateBitCast(iBuilder->CreateLoad(ptr), bitBlockWidthIntTy); |
---|
342 | currentBitValue = iBuilder->CreateLShr(currentBitValue, iBuilder->CreateZExt(currentPosBitBlockOffset, bitBlockWidthIntTy)); |
---|
343 | currentBitValue = iBuilder->CreateNot(currentBitValue); |
---|
344 | |
---|
345 | Value * forwardZeroCount = iBuilder->CreateTrunc(iBuilder->CreateCountForwardZeroes(currentBitValue), iBuilder->getInt64Ty()); |
---|
346 | Value * newOffset = iBuilder->CreateAdd(currentPosBitBlockOffset, forwardZeroCount); |
---|
347 | newOffset = iBuilder->CreateUMin(newOffset, INT64_BIT_BLOCK_WIDTH); |
---|
348 | |
---|
349 | Value * actualAdvanceValue = iBuilder->CreateSub(newOffset, currentPosBitBlockOffset); |
---|
350 | Value * newPos = iBuilder->CreateAdd(phiCurrentPos, actualAdvanceValue); |
---|
351 | if (maxPos) { |
---|
352 | newPos = iBuilder->CreateUMin(maxPos, newPos); |
---|
353 | actualAdvanceValue = iBuilder->CreateSub(newPos, phiCurrentPos); |
---|
354 | newOffset = iBuilder->CreateAdd(currentPosBitBlockOffset, actualAdvanceValue); |
---|
355 | } |
---|
356 | |
---|
357 | phiIsFinish->addIncoming(iBuilder->CreateICmpNE(newOffset, INT64_BIT_BLOCK_WIDTH), iBuilder->GetInsertBlock()); |
---|
358 | phiCurrentPos->addIncoming(newPos, iBuilder->GetInsertBlock()); |
---|
359 | iBuilder->CreateBr(advanceConBlock); |
---|
360 | |
---|
361 | iBuilder->SetInsertPoint(advanceExitBlock); |
---|
362 | return phiCurrentPos; |
---|
363 | } |
---|
364 | |
---|
365 | Value * LZ4IndexBuilderKernel::generateLoadInt64NumberInput(const unique_ptr<KernelBuilder> &iBuilder, string inputBufferName, Value * globalOffset) { |
---|
366 | Constant* SIZE_STRIDE_SIZE = iBuilder->getSize(getStride()); |
---|
367 | Value * processed = iBuilder->getProcessedItemCount(inputBufferName); |
---|
368 | processed = iBuilder->CreateAnd(processed, ConstantExpr::getNeg(SIZE_STRIDE_SIZE)); |
---|
369 | Value * offset = iBuilder->CreateSub(globalOffset, processed); |
---|
370 | Value * valuePtr = iBuilder->getRawInputPointer(inputBufferName, offset); |
---|
371 | return iBuilder->CreateLoad(valuePtr); |
---|
372 | } |
---|
373 | |
---|
374 | Value *LZ4IndexBuilderKernel::generateLoadSourceInputByte(const std::unique_ptr<KernelBuilder> &iBuilder, Value * offset) { |
---|
375 | Value * ptr = iBuilder->getRawInputPointer("byteStream", offset); |
---|
376 | return iBuilder->CreateLoad(ptr); |
---|
377 | } |
---|
378 | |
---|
379 | void LZ4IndexBuilderKernel::increaseScalarField(const unique_ptr<KernelBuilder> &iBuilder, const string &fieldName, Value *value) { |
---|
380 | Value *fieldValue = iBuilder->getScalarField(fieldName); |
---|
381 | fieldValue = iBuilder->CreateAdd(fieldValue, value); |
---|
382 | iBuilder->setScalarField(fieldName, fieldValue); |
---|
383 | } |
---|
384 | |
---|
385 | size_t LZ4IndexBuilderKernel::getOutputBufferSize(const unique_ptr<KernelBuilder> &iBuilder, string bufferName) { |
---|
386 | return this->getOutputStreamSetBuffer(bufferName)->getBufferBlocks() * iBuilder->getStride(); |
---|
387 | } |
---|
388 | |
---|
389 | // Assume we have enough output buffer |
---|
390 | llvm::BasicBlock *LZ4IndexBuilderKernel::markCircularOutputBitstream(const std::unique_ptr<KernelBuilder> &iBuilder, |
---|
391 | const std::string &bitstreamName, |
---|
392 | llvm::Value *start, llvm::Value *end, bool isOne, |
---|
393 | bool setProduced) { |
---|
394 | const unsigned int bitBlockWidth = iBuilder->getBitBlockWidth(); |
---|
395 | Value* SIZE_BIT_BLOCK_WIDTH = iBuilder->getSize(bitBlockWidth); |
---|
396 | Value* SIZE_ONE = iBuilder->getSize(1); |
---|
397 | Type * const INT_BIT_BLOCK_TY = iBuilder->getIntNTy(bitBlockWidth); |
---|
398 | Type * const BIT_BLOCK_TY = iBuilder->getBitBlockType(); |
---|
399 | Constant* INT_BIT_BLOCK_ONE = ConstantInt::get(INT_BIT_BLOCK_TY, 1); |
---|
400 | Constant* INT_BIT_BLOCK_ZERO = ConstantInt::get(INT_BIT_BLOCK_TY, 0); |
---|
401 | |
---|
402 | BasicBlock *entryBlock = iBuilder->GetInsertBlock(); |
---|
403 | BasicBlock *conBlock = iBuilder->CreateBasicBlock("mark_bit_one_con"); |
---|
404 | BasicBlock *bodyBlock = iBuilder->CreateBasicBlock("mark_bit_one_body"); |
---|
405 | BasicBlock *exitBlock = iBuilder->CreateBasicBlock("mark_bit_one_exit"); |
---|
406 | |
---|
407 | Value * startBlockLocalIndex = iBuilder->CreateUDiv(start, SIZE_BIT_BLOCK_WIDTH); |
---|
408 | |
---|
409 | iBuilder->CreateBr(conBlock); |
---|
410 | |
---|
411 | // Con |
---|
412 | iBuilder->SetInsertPoint(conBlock); |
---|
413 | |
---|
414 | PHINode *curBlockLocalIndex = iBuilder->CreatePHI(iBuilder->getSizeTy(), 2); |
---|
415 | curBlockLocalIndex->addIncoming(startBlockLocalIndex, entryBlock); |
---|
416 | |
---|
417 | |
---|
418 | iBuilder->CreateCondBr( |
---|
419 | iBuilder->CreateICmpULT(iBuilder->CreateMul(curBlockLocalIndex, SIZE_BIT_BLOCK_WIDTH), end), |
---|
420 | bodyBlock, |
---|
421 | exitBlock |
---|
422 | ); |
---|
423 | |
---|
424 | // Body |
---|
425 | iBuilder->SetInsertPoint(bodyBlock); |
---|
426 | |
---|
427 | Value * const currentPosition = iBuilder->CreateMul(curBlockLocalIndex, SIZE_BIT_BLOCK_WIDTH); |
---|
428 | Value * lowestBitPosition = iBuilder->CreateURem(start, SIZE_BIT_BLOCK_WIDTH); |
---|
429 | lowestBitPosition = iBuilder->CreateZExt(lowestBitPosition, INT_BIT_BLOCK_TY); |
---|
430 | Value * outputLowestBitValue = iBuilder->CreateShl(INT_BIT_BLOCK_ONE, lowestBitPosition); |
---|
431 | Value * const hasNotReachedStart = iBuilder->CreateICmpULE(currentPosition, start); |
---|
432 | outputLowestBitValue = iBuilder->CreateSelect(hasNotReachedStart, outputLowestBitValue, INT_BIT_BLOCK_ONE); |
---|
433 | |
---|
434 | Value * const nextPosition = iBuilder->CreateMul(iBuilder->CreateAdd(curBlockLocalIndex, SIZE_ONE), SIZE_BIT_BLOCK_WIDTH); |
---|
435 | Value * const hasNotReachEnd = iBuilder->CreateICmpULE(nextPosition, end); |
---|
436 | Value * producedItemsCount = iBuilder->CreateSelect(hasNotReachEnd, nextPosition, end); |
---|
437 | Value * highestBitPosition = iBuilder->CreateURem(end, SIZE_BIT_BLOCK_WIDTH); |
---|
438 | highestBitPosition = iBuilder->CreateZExt(highestBitPosition, INT_BIT_BLOCK_TY); |
---|
439 | Value * outputHighestBitValue = iBuilder->CreateShl(INT_BIT_BLOCK_ONE, highestBitPosition); |
---|
440 | outputHighestBitValue = iBuilder->CreateSelect(hasNotReachEnd, INT_BIT_BLOCK_ZERO, outputHighestBitValue); |
---|
441 | Value * bitMask = iBuilder->CreateSub(outputHighestBitValue, outputLowestBitValue); |
---|
442 | bitMask = iBuilder->CreateBitCast(bitMask, BIT_BLOCK_TY); |
---|
443 | |
---|
444 | Value * targetPtr = iBuilder->CreatePointerCast(iBuilder->getRawOutputPointer(bitstreamName, currentPosition), iBuilder->getBitBlockType()->getPointerTo()); |
---|
445 | Value * oldValue = iBuilder->CreateBlockAlignedLoad(targetPtr); |
---|
446 | Value * newValue = nullptr; |
---|
447 | if (isOne) { |
---|
448 | newValue = iBuilder->CreateOr(oldValue, bitMask); |
---|
449 | } else { |
---|
450 | newValue = iBuilder->CreateAnd(oldValue, iBuilder->CreateNot(bitMask)); |
---|
451 | } |
---|
452 | iBuilder->CreateStore(newValue, targetPtr); |
---|
453 | |
---|
454 | if (setProduced) { |
---|
455 | iBuilder->setProducedItemCount(bitstreamName, producedItemsCount); |
---|
456 | } |
---|
457 | |
---|
458 | curBlockLocalIndex->addIncoming(iBuilder->CreateAdd(curBlockLocalIndex, SIZE_ONE), bodyBlock); |
---|
459 | iBuilder->CreateBr(conBlock); |
---|
460 | |
---|
461 | // Exit |
---|
462 | iBuilder->SetInsertPoint(exitBlock); |
---|
463 | return exitBlock; |
---|
464 | } |
---|
465 | |
---|
466 | |
---|
467 | void LZ4IndexBuilderKernel::generateStoreNumberOutput(const unique_ptr<KernelBuilder> &iBuilder, |
---|
468 | const string & outputBufferName, |
---|
469 | Value * value) { |
---|
470 | |
---|
471 | Value * outputOffset = iBuilder->getProducedItemCount(outputBufferName); |
---|
472 | Value * outputRawPtr = iBuilder->getRawOutputPointer(outputBufferName, outputOffset); |
---|
473 | iBuilder->CreateStore(value, outputRawPtr); |
---|
474 | iBuilder->setProducedItemCount(outputBufferName, iBuilder->CreateAdd(outputOffset, iBuilder->getSize(1))); |
---|
475 | } |
---|
476 | |
---|
477 | } |
---|