source: icGREP/icgrep-devel/icgrep/pablo/pablo_compiler.cpp @ 4685

Last change on this file since 4685 was 4684, checked in by nmedfort, 4 years ago

First attempt to intergrate 'generate_predefined_ucd_functions' into build process.

File size: 34.9 KB
Line 
1/*
2 *  Copyright (c) 2014-15 International Characters.
3 *  This software is licensed to the public under the Open Software License 3.0.
4 *  icgrep is a trademark of International Characters.
5 */
6
7#include <pablo/pablo_compiler.h>
8#include <pablo/codegenstate.h>
9#include <pablo/carry_data.h>
10#include <pablo/carry_manager.h>
11#include <pablo/printer_pablos.h>
12#include <pablo/function.h>
13#include <cc/cc_namemap.hpp>
14#include <re/re_name.h>
15#include <stdexcept>
16#include <include/simd-lib/bitblock.hpp>
17#include <sstream>
18#include <IDISA/idisa_builder.h>
19#include <llvm/IR/Verifier.h>
20#include <llvm/Pass.h>
21#include <llvm/PassManager.h>
22#include <llvm/ADT/SmallVector.h>
23#include <llvm/Analysis/Passes.h>
24#include <llvm/IR/BasicBlock.h>
25#include <llvm/IR/CallingConv.h>
26#include <llvm/IR/Constants.h>
27#include <llvm/IR/DataLayout.h>
28#include <llvm/IR/DerivedTypes.h>
29#include <llvm/IR/Function.h>
30#include <llvm/IR/GlobalVariable.h>
31#include <llvm/IR/InlineAsm.h>
32#include <llvm/IR/Instructions.h>
33#include <llvm/IR/LLVMContext.h>
34#include <llvm/IR/Module.h>
35#include <llvm/Support/FormattedStream.h>
36#include <llvm/Support/MathExtras.h>
37#include <llvm/Support/Casting.h>
38#include <llvm/Support/Compiler.h>
39#include <llvm/Support/Debug.h>
40#include <llvm/Support/TargetSelect.h>
41#include <llvm/Support/Host.h>
42#include <llvm/Transforms/Scalar.h>
43#include <llvm/ExecutionEngine/ExecutionEngine.h>
44#include <llvm/ExecutionEngine/MCJIT.h>
45#include <llvm/IRReader/IRReader.h>
46#include <llvm/Bitcode/ReaderWriter.h>
47#include <llvm/Support/MemoryBuffer.h>
48#include <llvm/IR/IRBuilder.h>
49#include <llvm/Support/CommandLine.h>
50#include <llvm/ADT/Twine.h>
51#include <iostream>
52
53static cl::OptionCategory eIRDumpOptions("LLVM IR Dump Options", "These options control dumping of LLVM IR.");
54static cl::opt<bool> DumpGeneratedIR("dump-generated-IR", cl::init(false), cl::desc("print LLVM IR generated by RE compilation"), cl::cat(eIRDumpOptions));
55
56static cl::OptionCategory fTracingOptions("Run-time Tracing Options", "These options control execution traces.");
57static cl::opt<bool> TraceNext("trace-next-nodes", cl::init(false), cl::desc("Generate dynamic traces of executed Next nodes (while control variables)."), cl::cat(fTracingOptions));
58static cl::opt<bool> DumpTrace("dump-trace", cl::init(false), cl::desc("Generate dynamic traces of executed assignments."), cl::cat(fTracingOptions));
59
60extern "C" {
61  void wrapped_print_register(char * regName, BitBlock bit_block) {
62      print_register<BitBlock>(regName, bit_block);
63  }
64}
65
66namespace pablo {
67
68PabloCompiler::PabloCompiler()
69: mMod(nullptr)
70, mExecutionEngine(nullptr)
71, mBuilder(nullptr)
72, mCarryManager(nullptr)
73, mCarryOffset(0)
74, mBitBlockType(VectorType::get(IntegerType::get(getGlobalContext(), 64), BLOCK_SIZE / 64))
75, iBuilder(mBitBlockType)
76, mInputType(nullptr)
77, mCarryDataPtr(nullptr)
78, mWhileDepth(0)
79, mIfDepth(0)
80, mZeroInitializer(ConstantAggregateZero::get(mBitBlockType))
81, mOneInitializer(ConstantVector::getAllOnesValue(mBitBlockType))
82, mFunction(nullptr)
83, mInputAddressPtr(nullptr)
84, mOutputAddressPtr(nullptr)
85, mMaxWhileDepth(0)
86, mPrintRegisterFunction(nullptr) {
87
88}
89
90PabloCompiler::~PabloCompiler() {
91}
92   
93
94void PabloCompiler::genPrintRegister(std::string regName, Value * bitblockValue) {
95    Constant * regNameData = ConstantDataArray::getString(mMod->getContext(), regName);
96    GlobalVariable *regStrVar = new GlobalVariable(*mMod,
97                                                   ArrayType::get(IntegerType::get(mMod->getContext(), 8), regName.length()+1),
98                                                   /*isConstant=*/ true,
99                                                   /*Linkage=*/ GlobalValue::PrivateLinkage,
100                                                   /*Initializer=*/ regNameData);
101    Value * regStrPtr = mBuilder->CreateGEP(regStrVar, {mBuilder->getInt64(0), mBuilder->getInt32(0)});
102    mBuilder->CreateCall(mPrintRegisterFunction, {regStrPtr, bitblockValue});
103}
104
105CompiledPabloFunction PabloCompiler::compile(PabloFunction & function) {
106
107    Examine(function);
108
109    InitializeNativeTarget();
110    InitializeNativeTargetAsmPrinter();
111    InitializeNativeTargetAsmParser();
112
113    Module * module = new Module("", getGlobalContext());
114
115    mMod = module;
116
117    std::string errMessage;
118    #ifdef USE_LLVM_3_5
119    EngineBuilder builder(mMod);
120    #else
121    EngineBuilder builder(std::move(std::unique_ptr<Module>(mMod)));
122    #endif
123    builder.setErrorStr(&errMessage);
124    builder.setMCPU(sys::getHostCPUName());
125    #ifdef USE_LLVM_3_5
126    builder.setUseMCJIT(true);
127    #endif
128    builder.setOptLevel(mMaxWhileDepth ? CodeGenOpt::Level::Less : CodeGenOpt::Level::None);
129    mExecutionEngine = builder.create();
130    if (mExecutionEngine == nullptr) {
131        throw std::runtime_error("Could not create ExecutionEngine: " + errMessage);
132    }
133    DeclareFunctions();
134
135    auto func = compile(function, mMod);
136
137    //Display the IR that has been generated by this module.
138    if (LLVM_UNLIKELY(DumpGeneratedIR)) {
139        module->dump();
140    }
141    //Create a verifier.  The verifier will print an error message if our module is malformed in any way.
142    verifyModule(*module, &dbgs());
143
144    mExecutionEngine->finalizeObject();
145    ExecutionEngine * engine = mExecutionEngine;
146    mExecutionEngine = nullptr; // <-- pass ownership of the execution engine to the caller
147
148    return CompiledPabloFunction(func.second, func.first, engine);
149}
150
151std::pair<llvm::Function *, size_t> PabloCompiler::compile(PabloFunction & function, Module * module) {
152
153    Examine(function);
154
155    mMod = module;
156
157    mBuilder = new IRBuilder<>(mMod->getContext());
158
159    iBuilder.initialize(mMod, mBuilder);
160
161    mCarryManager = new CarryManager(mBuilder, mBitBlockType, mZeroInitializer, mOneInitializer, &iBuilder);
162
163    GenerateFunction(function);
164
165    mBuilder->SetInsertPoint(BasicBlock::Create(mMod->getContext(), "entry", mFunction,0));
166
167    //The basis bits structure
168    for (unsigned i = 0; i != function.getNumOfParameters(); ++i) {
169        Value* indices[] = {mBuilder->getInt64(0), mBuilder->getInt32(i)};
170        Value * gep = mBuilder->CreateGEP(mInputAddressPtr, indices);
171        LoadInst * basisBit = mBuilder->CreateAlignedLoad(gep, BLOCK_SIZE/8, false, function.getParameter(i)->getName()->to_string());
172        mMarkerMap.insert(std::make_pair(function.getParameter(i), basisBit));
173        if (DumpTrace) {
174            genPrintRegister(function.getParameter(i)->getName()->to_string(), basisBit);
175        }
176    }
177       
178    mCarryOffset = mCarryManager->initialize(&(function.getEntryBlock()), mCarryDataPtr);
179   
180    //Generate the IR instructions for the function.
181    compileBlock(function.getEntryBlock());
182   
183    mCarryManager->generateBlockNoIncrement();
184
185    if (DumpTrace || TraceNext) {
186        genPrintRegister("mBlockNo", mBuilder->CreateAlignedLoad(mBuilder->CreateBitCast(mCarryManager->getBlockNoPtr(), PointerType::get(mBitBlockType, 0)), BLOCK_SIZE/8, false));
187    }
188   
189    // Write the output values out
190    for (unsigned i = 0; i != function.getNumOfResults(); ++i) {
191        assert (function.getResult(i));
192        SetOutputValue(mMarkerMap[function.getResult(i)], i);
193    }
194
195    //Terminate the block
196    ReturnInst::Create(mMod->getContext(), mBuilder->GetInsertBlock());
197
198    // Clean up
199    delete mCarryManager; mCarryManager = nullptr;
200    delete mBuilder; mBuilder = nullptr;
201    mMod = nullptr; // don't delete this. It's either owned by the ExecutionEngine or the calling function.
202
203    //Return the required size of the carry data area to the process_block function.
204    return std::make_pair(mFunction, mCarryOffset * sizeof(BitBlock));
205}
206
207inline void PabloCompiler::GenerateFunction(PabloFunction & function) {
208    mInputType = PointerType::get(StructType::get(mMod->getContext(), std::vector<Type *>(function.getNumOfParameters(), mBitBlockType)), 0);
209    Type * carryType = PointerType::get(mBitBlockType, 0);
210    Type * outputType = PointerType::get(StructType::get(mMod->getContext(), std::vector<Type *>(function.getNumOfResults(), mBitBlockType)), 0);
211    FunctionType * functionType = FunctionType::get(Type::getVoidTy(mMod->getContext()), {{mInputType, carryType, outputType}}, false);
212
213#ifdef USE_UADD_OVERFLOW
214#ifdef USE_TWO_UADD_OVERFLOW
215    // Type Definitions for llvm.uadd.with.overflow.carryin.i128 or .i256
216    std::vector<Type*>StructTy_0_fields;
217    StructTy_0_fields.push_back(IntegerType::get(mMod->getContext(), BLOCK_SIZE));
218    StructTy_0_fields.push_back(IntegerType::get(mMod->getContext(), 1));
219    StructType *StructTy_0 = StructType::get(mMod->getContext(), StructTy_0_fields, /*isPacked=*/false);
220
221    std::vector<Type*>FuncTy_1_args;
222    FuncTy_1_args.push_back(IntegerType::get(mMod->getContext(), BLOCK_SIZE));
223    FuncTy_1_args.push_back(IntegerType::get(mMod->getContext(), BLOCK_SIZE));
224    FunctionType* FuncTy_1 = FunctionType::get(
225                                              /*Result=*/StructTy_0,
226                                              /*Params=*/FuncTy_1_args,
227                                              /*isVarArg=*/false);
228
229    mFunctionUaddOverflow = mMod->getFunction("llvm.uadd.with.overflow.i" +
230                                              std::to_string(BLOCK_SIZE));
231    if (!mFunctionUaddOverflow) {
232        mFunctionUaddOverflow= Function::Create(
233          /*Type=*/ FuncTy_1,
234          /*Linkage=*/ GlobalValue::ExternalLinkage,
235          /*Name=*/ "llvm.uadd.with.overflow.i" + std::to_string(BLOCK_SIZE), mMod); // (external, no body)
236        mFunctionUaddOverflow->setCallingConv(CallingConv::C);
237    }
238    AttributeSet mFunctionUaddOverflowPAL;
239    {
240        SmallVector<AttributeSet, 4> Attrs;
241        AttributeSet PAS;
242        {
243          AttrBuilder B;
244          B.addAttribute(Attribute::NoUnwind);
245          B.addAttribute(Attribute::ReadNone);
246          PAS = AttributeSet::get(mMod->getContext(), ~0U, B);
247        }
248
249        Attrs.push_back(PAS);
250        mFunctionUaddOverflowPAL = AttributeSet::get(mMod->getContext(), Attrs);
251    }
252    mFunctionUaddOverflow->setAttributes(mFunctionUaddOverflowPAL);
253#else
254    // Type Definitions for llvm.uadd.with.overflow.carryin.i128 or .i256
255    std::vector<Type*>StructTy_0_fields;
256    StructTy_0_fields.push_back(IntegerType::get(mMod->getContext(), BLOCK_SIZE));
257    StructTy_0_fields.push_back(IntegerType::get(mMod->getContext(), 1));
258    StructType *StructTy_0 = StructType::get(mMod->getContext(), StructTy_0_fields, /*isPacked=*/false);
259
260    std::vector<Type*>FuncTy_1_args;
261    FuncTy_1_args.push_back(IntegerType::get(mMod->getContext(), BLOCK_SIZE));
262    FuncTy_1_args.push_back(IntegerType::get(mMod->getContext(), BLOCK_SIZE));
263    FuncTy_1_args.push_back(IntegerType::get(mMod->getContext(), 1));
264    FunctionType* FuncTy_1 = FunctionType::get(
265                                              /*Result=*/StructTy_0,
266                                              /*Params=*/FuncTy_1_args,
267                                              /*isVarArg=*/false);
268
269    mFunctionUaddOverflowCarryin = mMod->getFunction("llvm.uadd.with.overflow.carryin.i" +
270                                              std::to_string(BLOCK_SIZE));
271    if (!mFunctionUaddOverflowCarryin) {
272        mFunctionUaddOverflowCarryin = Function::Create(
273          /*Type=*/ FuncTy_1,
274          /*Linkage=*/ GlobalValue::ExternalLinkage,
275          /*Name=*/ "llvm.uadd.with.overflow.carryin.i" + std::to_string(BLOCK_SIZE), mMod); // (external, no body)
276        mFunctionUaddOverflowCarryin->setCallingConv(CallingConv::C);
277    }
278    AttributeSet mFunctionUaddOverflowCarryinPAL;
279    {
280        SmallVector<AttributeSet, 4> Attrs;
281        AttributeSet PAS;
282        {
283          AttrBuilder B;
284          B.addAttribute(Attribute::NoUnwind);
285          B.addAttribute(Attribute::ReadNone);
286          PAS = AttributeSet::get(mMod->getContext(), ~0U, B);
287        }
288
289        Attrs.push_back(PAS);
290        mFunctionUaddOverflowCarryinPAL = AttributeSet::get(mMod->getContext(), Attrs);
291    }
292    mFunctionUaddOverflowCarryin->setAttributes(mFunctionUaddOverflowCarryinPAL);
293#endif
294#endif
295
296    //Starts on process_block
297    SmallVector<AttributeSet, 4> Attrs;
298    Attrs.push_back(AttributeSet::get(mMod->getContext(), ~0U, { Attribute::NoUnwind, Attribute::UWTable }));
299    Attrs.push_back(AttributeSet::get(mMod->getContext(), 1U, { Attribute::ReadOnly, Attribute::NoCapture }));
300    Attrs.push_back(AttributeSet::get(mMod->getContext(), 2U, { Attribute::NoCapture }));
301    Attrs.push_back(AttributeSet::get(mMod->getContext(), 3U, { Attribute::ReadNone, Attribute::NoCapture }));
302    AttributeSet AttrSet = AttributeSet::get(mMod->getContext(), Attrs);
303
304    // Create the function that will be generated.
305    mFunction = Function::Create(functionType, GlobalValue::ExternalLinkage, function.getName()->value(), mMod);
306    mFunction->setCallingConv(CallingConv::C);
307    mFunction->setAttributes(AttrSet);
308
309    Function::arg_iterator args = mFunction->arg_begin();
310    mInputAddressPtr = args++;
311    mInputAddressPtr->setName("input");
312    mCarryDataPtr = args++;
313    mCarryDataPtr->setName("carry");
314    mOutputAddressPtr = args++;
315    mOutputAddressPtr->setName("output");
316}
317
318inline void PabloCompiler::Examine(PabloFunction & function) {
319    if (mMod == nullptr) {
320
321        mWhileDepth = 0;
322        mIfDepth = 0;
323        mMaxWhileDepth = 0;
324
325        Examine(function.getEntryBlock());
326
327        if (LLVM_UNLIKELY(mWhileDepth != 0 || mIfDepth != 0)) {
328            throw std::runtime_error("Malformed Pablo AST: Unbalanced If or While nesting depth!");
329        }
330    }
331}
332
333
334void PabloCompiler::Examine(PabloBlock & block) {
335    for (Statement * stmt : block) {
336        if (Call * call = dyn_cast<Call>(stmt)) {
337            mCalleeMap.insert(std::make_pair(call->getCallee(), nullptr));
338        }
339        else if (If * ifStatement = dyn_cast<If>(stmt)) {
340            Examine(ifStatement->getBody());
341        }
342        else if (While * whileStatement = dyn_cast<While>(stmt)) {
343            mMaxWhileDepth = std::max(mMaxWhileDepth, ++mWhileDepth);
344            Examine(whileStatement->getBody());
345            --mWhileDepth;
346        }
347    }
348}
349
350inline void PabloCompiler::DeclareFunctions() {
351    if (DumpTrace || TraceNext) {
352        //This function can be used for testing to print the contents of a register from JIT'd code to the terminal window.
353        mPrintRegisterFunction = mMod->getOrInsertFunction("wrapped_print_register", Type::getVoidTy(mMod->getContext()), Type::getInt8PtrTy(mMod->getContext()), mBitBlockType, NULL);
354        if (mExecutionEngine) mExecutionEngine->addGlobalMapping(cast<GlobalValue>(mPrintRegisterFunction), (void *)&wrapped_print_register);
355    }
356}
357
358void PabloCompiler::compileBlock(PabloBlock & block) {
359    mPabloBlock = & block;
360    mCarryManager->ensureCarriesLoadedLocal();
361    for (const Statement * statement : block) {
362        compileStatement(statement);
363    }
364    mCarryManager->ensureCarriesStoredLocal();
365    mPabloBlock = block.getParent();
366}
367
368Value * PabloCompiler::genBitTest2(Value * e1, Value * e2) {
369    Type * t1 = e1->getType();
370    Type * t2 = e2->getType();
371    if (t1 == mBitBlockType) {
372        if (t2 == mBitBlockType) {
373            return iBuilder.bitblock_any(mBuilder->CreateOr(e1, e2));
374        }
375        else {
376            Value * m1 = mBuilder->CreateZExt(iBuilder.hsimd_signmask(16, e1), t2);
377            return mBuilder->CreateICmpNE(mBuilder->CreateOr(m1, e2), ConstantInt::get(t2, 0));
378        }
379    }
380    else if (t2 == mBitBlockType) {
381        Value * m2 = mBuilder->CreateZExt(iBuilder.hsimd_signmask(16, e2), t1);
382        return mBuilder->CreateICmpNE(mBuilder->CreateOr(e1, m2), ConstantInt::get(t1, 0));
383    }
384    else {
385        return mBuilder->CreateICmpNE(mBuilder->CreateOr(e1, e2), ConstantInt::get(t1, 0));
386    }
387}
388
389void PabloCompiler::compileIf(const If * ifStatement) {       
390    //
391    //  The If-ElseZero stmt:
392    //  if <predicate:expr> then <body:stmt>* elsezero <defined:var>* endif
393    //  If the value of the predicate is nonzero, then determine the values of variables
394    //  <var>* by executing the given statements.  Otherwise, the value of the
395    //  variables are all zero.  Requirements: (a) no variable that is defined within
396    //  the body of the if may be accessed outside unless it is explicitly
397    //  listed in the variable list, (b) every variable in the defined list receives
398    //  a value within the body, and (c) the logical consequence of executing
399    //  the statements in the event that the predicate is zero is that the
400    //  values of all defined variables indeed work out to be 0.
401    //
402    //  Simple Implementation with Phi nodes:  a phi node in the if exit block
403    //  is inserted for each variable in the defined variable list.  It receives
404    //  a zero value from the ifentry block and the defined value from the if
405    //  body.
406    //
407
408    BasicBlock * ifEntryBlock = mBuilder->GetInsertBlock();
409    BasicBlock * ifBodyBlock = BasicBlock::Create(mMod->getContext(), "if.body", mFunction, 0);
410    BasicBlock * ifEndBlock = BasicBlock::Create(mMod->getContext(), "if.end", mFunction, 0);
411   
412    PabloBlock & ifBody = ifStatement -> getBody();
413   
414    Value * if_test_value = compileExpression(ifStatement->getCondition());
415   
416    mCarryManager->enterScope(&ifBody);
417    if (mCarryManager->blockHasCarries()) {
418        // load the summary variable
419        Value* last_if_pending_data = mCarryManager->getCarrySummaryExpr();
420        mBuilder->CreateCondBr(genBitTest2(if_test_value, last_if_pending_data), ifBodyBlock, ifEndBlock);
421
422    }
423    else {
424        mBuilder->CreateCondBr(iBuilder.bitblock_any(if_test_value), ifBodyBlock, ifEndBlock);
425    }
426    // Entry processing is complete, now handle the body of the if.
427    mBuilder->SetInsertPoint(ifBodyBlock);
428   
429   
430    ++mIfDepth;
431    compileBlock(ifBody);
432    --mIfDepth;
433    if (mCarryManager->blockHasCarries()) {
434        mCarryManager->generateCarryOutSummaryCodeIfNeeded();
435    }
436    BasicBlock * ifBodyFinalBlock = mBuilder->GetInsertBlock();
437    mBuilder->CreateBr(ifEndBlock);
438    //End Block
439    mBuilder->SetInsertPoint(ifEndBlock);
440    for (const PabloAST * node : ifStatement->getDefined()) {
441        const Assign * assign = cast<Assign>(node);
442        PHINode * phi = mBuilder->CreatePHI(mBitBlockType, 2, assign->getName()->value());
443        auto f = mMarkerMap.find(assign);
444        assert (f != mMarkerMap.end());
445        phi->addIncoming(mZeroInitializer, ifEntryBlock);
446        phi->addIncoming(f->second, ifBodyFinalBlock);
447        mMarkerMap[assign] = phi;
448    }
449    // Create the phi Node for the summary variable, if needed.
450    mCarryManager->addSummaryPhiIfNeeded(ifEntryBlock, ifBodyFinalBlock);
451    mCarryManager->leaveScope();
452}
453
454void PabloCompiler::compileWhile(const While * whileStatement) {
455
456    PabloBlock & whileBody = whileStatement -> getBody();
457   
458    BasicBlock * whileEntryBlock = mBuilder->GetInsertBlock();
459    BasicBlock * whileBodyBlock = BasicBlock::Create(mMod->getContext(), "while.body", mFunction, 0);
460    BasicBlock * whileEndBlock = BasicBlock::Create(mMod->getContext(), "while.end", mFunction, 0);
461
462    mCarryManager->enterScope(&whileBody);
463    mCarryManager->ensureCarriesLoadedRecursive();
464
465    const auto & nextNodes = whileStatement->getVariants();
466    std::vector<PHINode *> nextPhis;
467    nextPhis.reserve(nextNodes.size());
468
469    // On entry to the while structure, proceed to execute the first iteration
470    // of the loop body unconditionally.   The while condition is tested at the end of
471    // the loop.
472
473    mBuilder->CreateBr(whileBodyBlock);
474    mBuilder->SetInsertPoint(whileBodyBlock);
475
476    //
477    // There are 3 sets of Phi nodes for the while loop.
478    // (1) Carry-ins: (a) incoming carry data first iterations, (b) zero thereafter
479    // (2) Carry-out accumulators: (a) zero first iteration, (b) |= carry-out of each iteration
480    // (3) Next nodes: (a) values set up before loop, (b) modified values calculated in loop.
481
482    mCarryManager->initializeCarryDataPhisAtWhileEntry(whileEntryBlock);
483
484    // for any Next nodes in the loop body, initialize to (a) pre-loop value.
485    for (const Next * n : nextNodes) {
486        PHINode * phi = mBuilder->CreatePHI(mBitBlockType, 2, n->getName()->value());
487        auto f = mMarkerMap.find(n->getInitial());
488        assert (f != mMarkerMap.end());
489        phi->addIncoming(f->second, whileEntryBlock);
490        mMarkerMap[n->getInitial()] = phi;
491        nextPhis.push_back(phi);
492    }
493
494    //
495    // Now compile the loop body proper.  Carry-out accumulated values
496    // and iterated values of Next nodes will be computed.
497    ++mWhileDepth;
498    compileBlock(whileBody);
499
500    BasicBlock * whileBodyFinalBlock = mBuilder->GetInsertBlock();
501
502    mCarryManager->extendCarryDataPhisAtWhileBodyFinalBlock(whileBodyFinalBlock);
503
504    // Terminate the while loop body with a conditional branch back.
505    mBuilder->CreateCondBr(iBuilder.bitblock_any(compileExpression(whileStatement->getCondition())), whileBodyBlock, whileEndBlock);
506
507    // and for any Next nodes in the loop body
508    for (unsigned i = 0; i < nextNodes.size(); i++) {
509        const Next * n = nextNodes[i];
510        auto f = mMarkerMap.find(n->getExpr());
511        if (LLVM_UNLIKELY(f == mMarkerMap.end())) {
512            throw std::runtime_error("Next node expression was not compiled!");
513        }
514        nextPhis[i]->addIncoming(f->second, whileBodyFinalBlock);
515    }
516
517    mBuilder->SetInsertPoint(whileEndBlock);
518    if (mCarryManager->blockHasCarries()) {
519        mCarryManager->generateCarryOutSummaryCodeIfNeeded();
520    }
521    --mWhileDepth;
522
523    mCarryManager->ensureCarriesStoredRecursive();
524    mCarryManager->leaveScope();
525}
526
527
528void PabloCompiler::compileStatement(const Statement * stmt) {
529    Value * expr = nullptr;
530    if (const Assign * assign = dyn_cast<const Assign>(stmt)) {
531        expr = compileExpression(assign->getExpression());
532    }
533    else if (const Next * next = dyn_cast<const Next>(stmt)) {
534        expr = compileExpression(next->getExpr());
535        if (TraceNext) {
536            genPrintRegister(next->getName()->to_string(), expr);
537        }
538    }
539    else if (const If * ifStatement = dyn_cast<const If>(stmt)) {
540        compileIf(ifStatement);
541        return;
542    }
543    else if (const While * whileStatement = dyn_cast<const While>(stmt)) {
544        compileWhile(whileStatement);
545        return;
546    }
547    else if (const Call* call = dyn_cast<Call>(stmt)) {
548        //Call the callee once and store the result in the marker map.
549        if (mMarkerMap.count(call) != 0) {
550            return;
551        }
552
553        const Prototype * proto = call->getPrototype();
554        const String * callee = proto->getName();
555
556        Type * inputType = StructType::get(mMod->getContext(), std::vector<Type *>{proto->getNumOfParameters(), mBitBlockType});
557        Type * carryType = mBitBlockType;
558        Type * outputType = StructType::get(mMod->getContext(), std::vector<Type *>{proto->getNumOfResults(), mBitBlockType});
559        FunctionType * functionType = FunctionType::get(Type::getVoidTy(mMod->getContext()), std::vector<Type *>{PointerType::get(inputType, 0), PointerType::get(carryType, 0), PointerType::get(outputType, 0)}, false);
560
561        //Starts on process_block
562        SmallVector<AttributeSet, 3> Attrs;
563        Attrs.push_back(AttributeSet::get(mMod->getContext(), 1U, { Attribute::ReadOnly, Attribute::NoCapture }));
564        Attrs.push_back(AttributeSet::get(mMod->getContext(), 2U, { Attribute::NoCapture }));
565        Attrs.push_back(AttributeSet::get(mMod->getContext(), 3U, { Attribute::ReadNone, Attribute::NoCapture }));
566        AttributeSet AttrSet = AttributeSet::get(mMod->getContext(), Attrs);
567
568        Function * externalFunction = cast<Function>(mMod->getOrInsertFunction(callee->value(), functionType, AttrSet));
569        if (LLVM_UNLIKELY(externalFunction == nullptr)) {
570            throw std::runtime_error("Could not create static method call for external function \"" + callee->to_string() + "\"");
571        }
572        externalFunction->setCallingConv(llvm::CallingConv::C);
573
574        if (mExecutionEngine) mExecutionEngine->addGlobalMapping(externalFunction, proto->getFunctionPtr());
575
576        // add mCarryOffset to mCarryDataPtr
577        Value * carryFramePtr = mBuilder->CreateGEP(mCarryDataPtr, mBuilder->getInt64(mCarryOffset));
578        AllocaInst * outputStruct = mBuilder->CreateAlloca(outputType);
579        mBuilder->CreateCall3(externalFunction, mInputAddressPtr, carryFramePtr, outputStruct);
580        Value * outputPtr = mBuilder->CreateGEP(outputStruct, { mBuilder->getInt32(0), mBuilder->getInt32(0) });
581        expr = mBuilder->CreateAlignedLoad(outputPtr, BLOCK_SIZE / 8, false);
582
583        mCarryOffset += (proto->getRequiredStateSpace() + (BLOCK_SIZE / 8) - 1) / (BLOCK_SIZE / 8);
584    }
585    else if (const And * pablo_and = dyn_cast<And>(stmt)) {
586        expr = mBuilder->CreateAnd(compileExpression(pablo_and->getExpr1()), compileExpression(pablo_and->getExpr2()), "and");
587    }
588    else if (const Or * pablo_or = dyn_cast<Or>(stmt)) {
589        expr = mBuilder->CreateOr(compileExpression(pablo_or->getExpr1()), compileExpression(pablo_or->getExpr2()), "or");
590    }
591    else if (const Xor * pablo_xor = dyn_cast<Xor>(stmt)) {
592        expr = mBuilder->CreateXor(compileExpression(pablo_xor->getExpr1()), compileExpression(pablo_xor->getExpr2()), "xor");
593    }
594    else if (const Sel * sel = dyn_cast<Sel>(stmt)) {
595        Value* ifMask = compileExpression(sel->getCondition());
596        Value* ifTrue = mBuilder->CreateAnd(ifMask, compileExpression(sel->getTrueExpr()));
597        Value* ifFalse = mBuilder->CreateAnd(genNot(ifMask), compileExpression(sel->getFalseExpr()));
598        expr = mBuilder->CreateOr(ifTrue, ifFalse);
599    }
600    else if (const Not * pablo_not = dyn_cast<Not>(stmt)) {
601        expr = genNot(compileExpression(pablo_not->getExpr()));
602    }
603    else if (const Advance * adv = dyn_cast<Advance>(stmt)) {
604        Value* strm_value = compileExpression(adv->getExpr());
605        int shift = adv->getAdvanceAmount();
606        unsigned advance_index = adv->getLocalAdvanceIndex();
607        expr = mCarryManager->advanceCarryInCarryOut(advance_index, shift, strm_value);
608    }
609    else if (const MatchStar * mstar = dyn_cast<MatchStar>(stmt)) {
610        Value * marker = compileExpression(mstar->getMarker());
611        Value * cc = compileExpression(mstar->getCharClass());
612        Value * marker_and_cc = mBuilder->CreateAnd(marker, cc);
613        unsigned carry_index = mstar->getLocalCarryIndex();
614        expr = mBuilder->CreateOr(mBuilder->CreateXor(genAddWithCarry(marker_and_cc, cc, carry_index), cc), marker, "matchstar");
615    }
616    else if (const ScanThru * sthru = dyn_cast<ScanThru>(stmt)) {
617        Value * marker_expr = compileExpression(sthru->getScanFrom());
618        Value * cc_expr = compileExpression(sthru->getScanThru());
619        unsigned carry_index = sthru->getLocalCarryIndex();
620        expr = mBuilder->CreateAnd(genAddWithCarry(marker_expr, cc_expr, carry_index), genNot(cc_expr), "scanthru");
621    }
622    else {
623        llvm::raw_os_ostream cerr(std::cerr);
624        PabloPrinter::print(stmt, cerr);
625        throw std::runtime_error("Unrecognized Pablo Statement! can't compile.");
626    }
627    mMarkerMap[stmt] = expr;
628    if (DumpTrace) {
629        genPrintRegister(stmt->getName()->to_string(), expr);
630    }
631   
632}
633
634Value * PabloCompiler::compileExpression(const PabloAST * expr) {
635    if (isa<Ones>(expr)) {
636        return mOneInitializer;
637    }
638    else if (isa<Zeroes>(expr)) {
639        return mZeroInitializer;
640    }
641    auto f = mMarkerMap.find(expr);
642    if (LLVM_UNLIKELY(f == mMarkerMap.end())) {
643        std::string o;
644        llvm::raw_string_ostream str(o);
645        str << "\"";
646        PabloPrinter::print(expr, str);
647        str << "\" was used before definition!";
648        throw std::runtime_error(str.str());
649    }
650    return f->second;
651}
652
653
654#ifdef USE_UADD_OVERFLOW
655#ifdef USE_TWO_UADD_OVERFLOW
656PabloCompiler::SumWithOverflowPack PabloCompiler::callUaddOverflow(Value* int128_e1, Value* int128_e2) {
657    std::vector<Value*> struct_res_params;
658    struct_res_params.push_back(int128_e1);
659    struct_res_params.push_back(int128_e2);
660    CallInst* struct_res = CallInst::Create(mFunctionUaddOverflow, struct_res_params, "uadd_overflow_res", mBasicBlock);
661    struct_res->setCallingConv(CallingConv::C);
662    struct_res->setTailCall(false);
663    AttributeSet struct_res_PAL;
664    struct_res->setAttributes(struct_res_PAL);
665
666    SumWithOverflowPack ret;
667
668    std::vector<unsigned> int128_sum_indices;
669    int128_sum_indices.push_back(0);
670    ret.sum = ExtractValueInst::Create(struct_res, int128_sum_indices, "sum", mBasicBlock);
671
672    std::vector<unsigned> int1_obit_indices;
673    int1_obit_indices.push_back(1);
674    ret.obit = ExtractValueInst::Create(struct_res, int1_obit_indices, "obit", mBasicBlock);
675
676    return ret;
677}
678#else
679PabloCompiler::SumWithOverflowPack PabloCompiler::callUaddOverflow(Value* int128_e1, Value* int128_e2, Value* int1_cin) {
680    std::vector<Value*> struct_res_params;
681    struct_res_params.push_back(int128_e1);
682    struct_res_params.push_back(int128_e2);
683    struct_res_params.push_back(int1_cin);
684    CallInst* struct_res = CallInst::Create(mFunctionUaddOverflowCarryin, struct_res_params, "uadd_overflow_res", mBasicBlock);
685    struct_res->setCallingConv(CallingConv::C);
686    struct_res->setTailCall(false);
687    AttributeSet struct_res_PAL;
688    struct_res->setAttributes(struct_res_PAL);
689
690    SumWithOverflowPack ret;
691
692    std::vector<unsigned> int128_sum_indices;
693    int128_sum_indices.push_back(0);
694    ret.sum = ExtractValueInst::Create(struct_res, int128_sum_indices, "sum", mBasicBlock);
695
696    std::vector<unsigned> int1_obit_indices;
697    int1_obit_indices.push_back(1);
698    ret.obit = ExtractValueInst::Create(struct_res, int1_obit_indices, "obit", mBasicBlock);
699
700    return ret;
701}
702#endif
703#endif
704
705
706Value* PabloCompiler::genAddWithCarry(Value* e1, Value* e2, unsigned localIndex) {
707    Value * carryq_value = mCarryManager->getCarryOpCarryIn(localIndex);
708#ifdef USE_TWO_UADD_OVERFLOW
709    //This is the ideal implementation, which uses two uadd.with.overflow
710    //The back end should be able to recognize this pattern and combine it into uadd.with.overflow.carryin
711    CastInst* int128_e1 = new BitCastInst(e1, mBuilder->getIntNTy(BLOCK_SIZE), "e1_128", mBasicBlock);
712    CastInst* int128_e2 = new BitCastInst(e2, mBuilder->getIntNTy(BLOCK_SIZE), "e2_128", mBasicBlock);
713    CastInst* int128_carryq_value = new BitCastInst(carryq_value, mBuilder->getIntNTy(BLOCK_SIZE), "carryq_128", mBasicBlock);
714
715    SumWithOverflowPack sumpack0, sumpack1;
716
717    sumpack0 = callUaddOverflow(int128_e1, int128_e2);
718    sumpack1 = callUaddOverflow(sumpack0.sum, int128_carryq_value);
719
720    Value* obit = mBuilder->CreateOr(sumpack0.obit, sumpack1.obit, "carry_bit");
721    Value* sum = mBuilder->CreateBitCast(sumpack1.sum, mBitBlockType, "ret_sum");
722
723    /*obit is the i1 carryout, zero extend and insert it into a v2i64 or v4i64 vector.*/
724    ConstantAggregateZero* const_packed_5 = ConstantAggregateZero::get(mBitBlockType);
725    ConstantInt* const_int32_6 = ConstantInt::get(mMod->getContext(), APInt(32, StringRef("0"), 10));
726    CastInst* int64_o0 = new ZExtInst(obit, IntegerType::get(mMod->getContext(), 64), "o0", mBasicBlock);
727    InsertElementInst* carry_out = InsertElementInst::Create(const_packed_5, int64_o0, const_int32_6, "carry_out", mBasicBlock);
728
729#elif defined USE_UADD_OVERFLOW
730    //use llvm.uadd.with.overflow.i128 or i256
731    CastInst* int128_e1 = new BitCastInst(e1, mBuilder->getIntNTy(BLOCK_SIZE), "e1_128", mBasicBlock);
732    CastInst* int128_e2 = new BitCastInst(e2, mBuilder->getIntNTy(BLOCK_SIZE), "e2_128", mBasicBlock);
733
734    //get i1 carryin from iBLOCK_SIZE
735    ConstantInt* const_int32_6 = ConstantInt::get(mMod->getContext(), APInt(32, StringRef("0"), 10));
736    ExtractElementInst * int64_carryq_value = ExtractElementInst::Create(carryq_value, const_int32_6, "carryq_64", mBasicBlock);
737    CastInst* int1_carryq_value = new TruncInst(int64_carryq_value, IntegerType::get(mMod->getContext(), 1), "carryq_1", mBasicBlock);
738
739    SumWithOverflowPack sumpack0;
740    sumpack0 = callUaddOverflow(int128_e1, int128_e2, int1_carryq_value);
741    Value* obit = sumpack0.obit;
742    Value* sum = mBuilder->CreateBitCast(sumpack0.sum, mBitBlockType, "sum");
743
744    /*obit is the i1 carryout, zero extend and insert it into a v2i64 or v4i64 vector.*/
745    ConstantAggregateZero* const_packed_5 = ConstantAggregateZero::get(mBitBlockType);
746    CastInst* int64_o0 = new ZExtInst(obit, IntegerType::get(mMod->getContext(), 64), "o0", mBasicBlock);
747    InsertElementInst* carry_out = InsertElementInst::Create(const_packed_5, int64_o0, const_int32_6, "carry_out", mBasicBlock);
748#elif (BLOCK_SIZE == 128)
749    //calculate carry through logical ops
750    Value* carrygen = mBuilder->CreateAnd(e1, e2, "carrygen");
751    Value* carryprop = mBuilder->CreateOr(e1, e2, "carryprop");
752    Value* digitsum = mBuilder->CreateAdd(e1, e2, "digitsum");
753    Value* partial = mBuilder->CreateAdd(digitsum, carryq_value, "partial");
754    Value* digitcarry = mBuilder->CreateOr(carrygen, mBuilder->CreateAnd(carryprop, genNot(partial)));
755    Value* mid_carry_in = genShiftLeft64(mBuilder->CreateLShr(digitcarry, 63), "mid_carry_in");
756
757    Value* sum = mBuilder->CreateAdd(partial, mid_carry_in, "sum");
758    Value* carry_out = genShiftHighbitToLow(BLOCK_SIZE, mBuilder->CreateOr(carrygen, mBuilder->CreateAnd(carryprop, genNot(sum))));
759#else
760    //BLOCK_SIZE == 256, there is no other implementation
761    static_assert(false, "Add with carry for 256-bit bitblock requires USE_UADD_OVERFLOW");
762#endif //USE_TWO_UADD_OVERFLOW
763
764    mCarryManager->setCarryOpCarryOut(localIndex, carry_out);
765    return sum;
766}
767
768Value * PabloCompiler::genShiftHighbitToLow(unsigned FieldWidth, Value * op) {
769    unsigned FieldCount = BLOCK_SIZE/FieldWidth;
770    VectorType * vType = VectorType::get(IntegerType::get(mMod->getContext(), FieldWidth), FieldCount);
771    Value * v = mBuilder->CreateBitCast(op, vType);
772    return mBuilder->CreateBitCast(mBuilder->CreateLShr(v, FieldWidth - 1), mBitBlockType);
773}
774
775Value* PabloCompiler::genShiftLeft64(Value* e, const Twine &namehint) {
776    Value* i128_val = mBuilder->CreateBitCast(e, mBuilder->getIntNTy(BLOCK_SIZE));
777    return mBuilder->CreateBitCast(mBuilder->CreateShl(i128_val, 64, namehint), mBitBlockType);
778}
779
780inline Value* PabloCompiler::genNot(Value* expr) {
781    return mBuilder->CreateXor(expr, mOneInitializer, "not");
782}
783   
784void PabloCompiler::SetOutputValue(Value * marker, const unsigned index) {
785    if (LLVM_UNLIKELY(marker == nullptr)) {
786        throw std::runtime_error("Cannot set result " + std::to_string(index) + " to Null");
787    }
788    if (LLVM_UNLIKELY(marker->getType()->isPointerTy())) {
789        marker = mBuilder->CreateAlignedLoad(marker, BLOCK_SIZE/8, false);
790    }
791    Value* indices[] = {mBuilder->getInt64(0), mBuilder->getInt32(index)};
792    Value* gep = mBuilder->CreateGEP(mOutputAddressPtr, indices);
793    mBuilder->CreateAlignedStore(marker, gep, BLOCK_SIZE/8, false);
794}
795
796CompiledPabloFunction::CompiledPabloFunction(size_t carryDataSize, Function * function, ExecutionEngine * executionEngine)
797: CarryDataSize(carryDataSize)
798, FunctionPointer(executionEngine->getPointerToFunction(function))
799, mFunction(function)
800, mExecutionEngine(executionEngine)
801{
802
803}
804
805// Clean up the memory for the compiled function once we're finished using it.
806CompiledPabloFunction::~CompiledPabloFunction() {
807    if (mExecutionEngine) {
808        assert (mFunction);
809        // mExecutionEngine->freeMachineCodeForFunction(mFunction); // This function only prints a "not supported" message. Reevaluate with LLVM 3.6.
810        delete mExecutionEngine;
811    }
812}
813
814}
Note: See TracBrowser for help on using the repository browser.