source: icGREP/icgrep-devel/icgrep/kernels/scanmatchgen.cpp @ 5184

Last change on this file since 5184 was 5104, checked in by cameron, 3 years ago

getStreamSetBlockPtr

File size: 13.3 KB
Line 
1/*
2 *  Copyright (c) 2015 International Characters.
3 *  This software is licensed to the public under the Open Software License 3.0.
4 */
5
6#include "kernel.h"
7#include "scanmatchgen.h"
8#include <llvm/IR/Intrinsics.h>
9#include <IDISA/idisa_builder.h>
10#include <llvm/Support/raw_os_ostream.h>
11
12using namespace llvm;
13
14namespace kernel {
15
16Value * generateForwardZeroesMask(IDISA::IDISA_Builder * iBuilder, Value * bits) {
17    Value * bits_minus1 = iBuilder->CreateSub(bits, ConstantInt::get(bits->getType(), 1));
18    return iBuilder->CreateAnd(bits_minus1, iBuilder->CreateNot(bits));
19}
20
21Value * generatePopcount(IDISA::IDISA_Builder * iBuilder, Value * bits) {
22    Value * ctpopFunc = Intrinsic::getDeclaration(iBuilder->getModule(), Intrinsic::ctpop, bits->getType());
23    return iBuilder->CreateCall(ctpopFunc, std::vector<Value *>({bits}));
24}
25
26Value * generateCountForwardZeroes(IDISA::IDISA_Builder * iBuilder, Value * bits) {
27    Value * cttzFunc = Intrinsic::getDeclaration(iBuilder->getModule(), Intrinsic::cttz, bits->getType());
28    return iBuilder->CreateCall(cttzFunc, std::vector<Value *>({bits, ConstantInt::get(iBuilder->getInt1Ty(), 0)}));
29}
30
31Value * generateCountReverseZeroes(IDISA::IDISA_Builder * iBuilder, Value * bits) {
32    Value * ctlzFunc = Intrinsic::getDeclaration(iBuilder->getModule(), Intrinsic::ctlz, bits->getType());
33    return iBuilder->CreateCall(ctlzFunc, std::vector<Value *>({bits, ConstantInt::get(iBuilder->getInt1Ty(), 0)}));
34}
35
36Value * generateResetLowestBit(IDISA::IDISA_Builder * iBuilder, Value * bits) {
37    Value * bits_minus1 = iBuilder->CreateSub(bits, ConstantInt::get(bits->getType(), 1));
38    return iBuilder->CreateAnd(bits_minus1, bits);
39}
40
41       
42void scanMatchKernel::generateDoBlockMethod() {
43    IDISA::IDISA_Builder::InsertPoint savePoint = iBuilder->saveIP();
44    Module * m = iBuilder->getModule();
45    Function * scanWordFunction = generateScanWordRoutine(m);
46    const unsigned fieldCount = iBuilder->getBitBlockWidth() / mScanwordBitWidth;
47    Type * T = iBuilder->getIntNTy(mScanwordBitWidth);
48    Type * scanwordVectorType =  VectorType::get(T, fieldCount);
49
50    Function * doBlockFunction = m->getFunction(mKernelName + doBlock_suffix);
51
52    iBuilder->SetInsertPoint(BasicBlock::Create(iBuilder->getContext(), "entry", doBlockFunction, 0));
53    Value * kernelStuctParam = getParameter(doBlockFunction, "self");
54    Value * blockNo = getScalarField(kernelStuctParam, blockNoScalar);
55    Value * scanwordPos = iBuilder->CreateMul(blockNo, ConstantInt::get(blockNo->getType(), iBuilder->getBitBlockWidth()));
56   
57    Value * recordStart = getScalarField(kernelStuctParam, "LineStart");
58    Value * recordNum = getScalarField(kernelStuctParam, "LineNum");
59    Value * matchResultsPtr = getStreamSetBlockPtr(kernelStuctParam, "matchResults", blockNo);   
60    Value * matches = iBuilder->CreateBlockAlignedLoad(matchResultsPtr, {iBuilder->getInt32(0), iBuilder->getInt32(0)});
61    Value * linebreaks = iBuilder->CreateBlockAlignedLoad(matchResultsPtr, {iBuilder->getInt32(0), iBuilder->getInt32(1)});
62    Value * matchWordVector = iBuilder->CreateBitCast(matches, scanwordVectorType);
63    Value * breakWordVector = iBuilder->CreateBitCast(linebreaks, scanwordVectorType);
64    for(unsigned i = 0; i < fieldCount; ++i){
65        Value * matchWord = iBuilder->CreateExtractElement(matchWordVector, ConstantInt::get(T, i));
66        Value * recordBreaksWord = iBuilder->CreateExtractElement(breakWordVector, ConstantInt::get(T, i));
67        Value * wordResult = iBuilder->CreateCall(scanWordFunction, {kernelStuctParam, matchWord, recordBreaksWord, scanwordPos, recordStart, recordNum});
68        scanwordPos = iBuilder->CreateAdd(scanwordPos, ConstantInt::get(T, mScanwordBitWidth));
69        recordStart = iBuilder->CreateExtractValue(wordResult, std::vector<unsigned>({0}));
70        recordNum = iBuilder->CreateExtractValue(wordResult, std::vector<unsigned>({1}));
71    }
72    setScalarField(kernelStuctParam, "LineStart", recordStart);
73    setScalarField(kernelStuctParam, "LineNum", recordNum);
74    iBuilder -> CreateRetVoid();
75    iBuilder->restoreIP(savePoint);
76}
77
78   
79Function * scanMatchKernel::generateScanWordRoutine(Module * m) {
80    Function * function = m->getFunction("scan_matches_in_scanword");
81    if (LLVM_UNLIKELY(function != nullptr)) {
82        return function;
83    }
84   
85    LLVMContext & ctxt = m->getContext();
86    Type * T = iBuilder->getIntNTy(mScanwordBitWidth);
87    Type * S = PointerType::get(iBuilder->getIntNTy(8), 0);
88    Type * returnType = StructType::get(ctxt, std::vector<Type *>({T, T}));
89    FunctionType * functionType = FunctionType::get(returnType, std::vector<Type *>({PointerType::get(mKernelStateType, 0), T, T, T, T, T}), false);
90   
91    SmallVector<AttributeSet, 6> Attrs;
92    Attrs.push_back(AttributeSet::get(ctxt, ~0U, std::vector<Attribute::AttrKind>({ Attribute::NoUnwind, Attribute::UWTable })));
93    Attrs.push_back(AttributeSet::get(ctxt, 1, std::vector<Attribute::AttrKind>({})));
94    Attrs.push_back(AttributeSet::get(ctxt, 2, std::vector<Attribute::AttrKind>({})));
95    Attrs.push_back(AttributeSet::get(ctxt, 3, std::vector<Attribute::AttrKind>({})));
96    Attrs.push_back(AttributeSet::get(ctxt, 4, std::vector<Attribute::AttrKind>({})));
97    Attrs.push_back(AttributeSet::get(ctxt, 5, std::vector<Attribute::AttrKind>({})));
98    AttributeSet AttrSet = AttributeSet::get(ctxt, Attrs);
99   
100    function = Function::Create(functionType, GlobalValue::ExternalLinkage, "scan_matches_in_scanword", m);
101    function->setCallingConv(CallingConv::C);
102    function->setAttributes(AttrSet);
103    function->addFnAttr(llvm::Attribute::AlwaysInline);
104   
105    Function::arg_iterator args = function->arg_begin();
106    Value * instance = &*(args++);
107    instance->setName("this");
108    Value * matches_input_parm = &*(args++);
109    matches_input_parm->setName("matches");
110    Value * record_breaks_input_parm = &*(args++);
111    record_breaks_input_parm->setName("breaks");
112    Value * scanwordPos = &*(args++);
113    scanwordPos->setName("scanwordPos");
114    Value * recordStart_input_parm = &*(args++);
115    recordStart_input_parm->setName("pendingLineStart");
116    Value * recordNum_input_parm = &*(args++);
117    recordNum_input_parm->setName("lineNum");
118   
119    Constant * matchProcessor;
120    if (mIsNameExpression) {
121        matchProcessor = m->getOrInsertFunction("insert_codepoints", Type::getVoidTy(ctxt), T, T, T, S, nullptr);
122    } else {
123        matchProcessor = m->getOrInsertFunction("wrapped_report_match", Type::getVoidTy(ctxt), T, T, T, S, T, T, nullptr);
124    }
125    iBuilder->SetInsertPoint(BasicBlock::Create(ctxt, "entry", function,0));
126   
127    BasicBlock * entry_block = iBuilder->GetInsertBlock();
128    BasicBlock * matches_test_block = BasicBlock::Create(ctxt, "matches_test_block", function, 0);
129    BasicBlock * process_matches_loop_entry = BasicBlock::Create(ctxt, "process_matches_loop", function, 0);
130    BasicBlock * prior_breaks_block = BasicBlock::Create(ctxt, "prior_breaks_block", function, 0);
131    BasicBlock * loop_final_block = BasicBlock::Create(ctxt, "loop_final_block", function, 0);
132    BasicBlock * matches_done_block = BasicBlock::Create(ctxt, "matches_done_block", function, 0);
133    BasicBlock * remaining_breaks_block = BasicBlock::Create(ctxt, "remaining_breaks_block", function, 0);
134    BasicBlock * return_block = BasicBlock::Create(ctxt, "return_block", function, 0);
135   
136   
137    // The match scanner works with a loop involving four variables:
138    // (a) the bit stream scanword of matches marking the ends of selected records,
139    // (b) the bit stream scanword of record_breaks marking the ends of all records,
140    // (c) the integer lastRecordNum indicating the number of records processed so far,
141    // (d) the index lastRecordStart indicating the file position of the last record.
142    // We set up a loop structure, in which a set of 4 phi nodes initialize these
143    // variables from either the input to the scanner or the computed values within
144    // the loop body.
145   
146   
147    iBuilder->CreateBr(matches_test_block);
148   
149    // LOOP Test Block
150    iBuilder->SetInsertPoint(matches_test_block);
151    PHINode * matches_phi = iBuilder->CreatePHI(T, 2, "matches");
152    PHINode * record_breaks_phi = iBuilder->CreatePHI(T, 2, "record_breaks");
153    PHINode * recordNum_phi = iBuilder->CreatePHI(T, 2, "recordNum");
154    PHINode * recordStart_phi = iBuilder->CreatePHI(T, 2, "recordStart");
155    matches_phi->addIncoming(matches_input_parm, entry_block);
156    record_breaks_phi->addIncoming(record_breaks_input_parm, entry_block);
157    recordNum_phi->addIncoming(recordNum_input_parm, entry_block);
158    recordStart_phi->addIncoming(recordStart_input_parm, entry_block);
159    Value * have_matches_cond = iBuilder->CreateICmpNE(matches_phi, ConstantInt::get(T, 0));
160    iBuilder->CreateCondBr(have_matches_cond, process_matches_loop_entry, matches_done_block);
161   
162    // LOOP BODY
163    // The loop body is entered if we have more matches to process.
164    iBuilder->SetInsertPoint(process_matches_loop_entry);
165    Value * prior_breaks = iBuilder->CreateAnd(generateForwardZeroesMask(iBuilder, matches_phi), record_breaks_phi);
166    // Within the loop we have a conditional block that is executed if there are any prior
167    // record breaks.
168    Value * prior_breaks_cond = iBuilder->CreateICmpNE(prior_breaks, ConstantInt::get(T, 0));
169    iBuilder->CreateCondBr(prior_breaks_cond, prior_breaks_block, loop_final_block);
170   
171    // PRIOR_BREAKS_BLOCK
172    // If there are prior breaks, we count them and compute the record start position.
173    iBuilder->SetInsertPoint(prior_breaks_block);
174    Value * matchRecordNum = iBuilder->CreateAdd(generatePopcount(iBuilder, prior_breaks), recordNum_phi);
175    Value * reverseDistance = generateCountReverseZeroes(iBuilder, prior_breaks);
176    Value * width = ConstantInt::get(T, mScanwordBitWidth);
177    Value * matchRecordStart = iBuilder->CreateAdd(scanwordPos, iBuilder->CreateSub(width, reverseDistance));
178    iBuilder->CreateBr(loop_final_block);
179   
180    // LOOP FINAL BLOCK
181    // The prior breaks, if any have been counted.  Set up phi nodes for the recordNum
182    // and recortStart depending on whether the conditional execution of prior_breaks_block.
183    iBuilder->SetInsertPoint(loop_final_block);
184    PHINode * matchRecordNum_phi = iBuilder->CreatePHI(T, 2, "matchRecordNum");
185    PHINode * matchRecordStart_phi = iBuilder->CreatePHI(T, 2, "matchRecordStart");
186    matchRecordNum_phi->addIncoming(recordNum_phi, process_matches_loop_entry);
187    matchRecordNum_phi->addIncoming(matchRecordNum, prior_breaks_block);
188    matchRecordStart_phi->addIncoming(recordStart_phi, process_matches_loop_entry);
189    matchRecordStart_phi->addIncoming(matchRecordStart, prior_breaks_block);   
190    Value * matchRecordEnd = iBuilder->CreateAdd(scanwordPos, generateCountForwardZeroes(iBuilder, matches_phi));
191   
192
193    Value * fileBuf = getScalarField(instance, "FileBuf");
194    if (mIsNameExpression) {
195        iBuilder->CreateCall(matchProcessor, std::vector<Value *>({matchRecordNum_phi, matchRecordStart_phi, matchRecordEnd, fileBuf}));
196    } else {
197        Value * fileSize = getScalarField(instance, "FileSize");
198        Value * fileIdx = getScalarField(instance, "FileIdx");
199        iBuilder->CreateCall(matchProcessor, std::vector<Value *>({matchRecordNum_phi, matchRecordStart_phi, matchRecordEnd, fileBuf, fileSize, fileIdx}));
200    }
201   
202    Value * remaining_matches = generateResetLowestBit(iBuilder, matches_phi);
203    Value * remaining_breaks = iBuilder->CreateXor(record_breaks_phi, prior_breaks);
204    matches_phi->addIncoming(remaining_matches, loop_final_block);
205    record_breaks_phi->addIncoming(remaining_breaks, loop_final_block);
206    recordNum_phi->addIncoming(matchRecordNum_phi, loop_final_block);
207    recordStart_phi->addIncoming(matchRecordStart_phi, loop_final_block);
208    iBuilder->CreateBr(matches_test_block);
209   
210   
211    // LOOP EXIT/MATCHES_DONE
212    iBuilder->SetInsertPoint(matches_done_block);
213    // When the matches are done, there may be additional record breaks remaining
214    Value * more_breaks_cond = iBuilder->CreateICmpNE(record_breaks_phi, ConstantInt::get(T, 0));
215    iBuilder->CreateCondBr(more_breaks_cond, remaining_breaks_block, return_block);
216   
217    // REMAINING_BREAKS_BLOCK: process remaining record breaks after all matches are processed
218    iBuilder->SetInsertPoint(remaining_breaks_block);
219    Value * break_count = generatePopcount(iBuilder, record_breaks_phi);
220    Value * final_record_num = iBuilder->CreateAdd(recordNum_phi, break_count);
221    Value * reverseZeroes = generateCountReverseZeroes(iBuilder, record_breaks_phi);
222    Value * pendingLineStart = iBuilder->CreateAdd(scanwordPos, iBuilder->CreateSub(width, reverseZeroes));
223    iBuilder->CreateBr(return_block);
224   
225    // RETURN block
226    iBuilder->SetInsertPoint(return_block);
227    PHINode * finalRecordCount_phi = iBuilder->CreatePHI(T, 2, "finalRecordCount");
228    PHINode * finalRecordStart_phi = iBuilder->CreatePHI(T, 2, "finalRecordStart");
229    finalRecordCount_phi->addIncoming(recordNum_phi, matches_done_block);
230    finalRecordCount_phi->addIncoming(final_record_num, remaining_breaks_block);
231    finalRecordStart_phi->addIncoming(recordStart_phi, matches_done_block);
232    finalRecordStart_phi->addIncoming(pendingLineStart, remaining_breaks_block);
233    Value * retVal = UndefValue::get(returnType);
234    retVal = iBuilder->CreateInsertValue(retVal, finalRecordStart_phi, 0);
235    retVal = iBuilder->CreateInsertValue(retVal, finalRecordCount_phi, 1);
236    iBuilder->CreateRet(retVal);
237   
238    return function;
239}
240
241}
Note: See TracBrowser for help on using the repository browser.