source: icGREP/icgrep-devel/icgrep/pablo/pablo_compiler.cpp @ 4676

Last change on this file since 4676 was 4676, checked in by cameron, 4 years ago

Simplify/unify conditional creation of explicit carry summary code and summary phi nodes.

File size: 35.1 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, mBuilder(nullptr)
71, mCarryManager(nullptr)
72, mBitBlockType(VectorType::get(IntegerType::get(getGlobalContext(), 64), BLOCK_SIZE / 64))
73, iBuilder(mBitBlockType)
74, mInputType(nullptr)
75, mCarryDataPtr(nullptr)
76, mWhileDepth(0)
77, mIfDepth(0)
78, mZeroInitializer(ConstantAggregateZero::get(mBitBlockType))
79, mOneInitializer(ConstantVector::getAllOnesValue(mBitBlockType))
80, mFunction(nullptr)
81, mInputAddressPtr(nullptr)
82, mOutputAddressPtr(nullptr)
83, mMaxWhileDepth(0)
84, mPrintRegisterFunction(nullptr) {
85
86}
87
88PabloCompiler::~PabloCompiler() {
89}
90   
91void PabloCompiler::InstallExternalFunction(std::string C_fn_name, void * fn_ptr, const size_t carrySize) {
92    mExternalMap.insert(std::make_pair(C_fn_name, fn_ptr));
93}
94
95void PabloCompiler::genPrintRegister(std::string regName, Value * bitblockValue) {
96    Constant * regNameData = ConstantDataArray::getString(mMod->getContext(), regName);
97    GlobalVariable *regStrVar = new GlobalVariable(*mMod,
98                                                   ArrayType::get(IntegerType::get(mMod->getContext(), 8), regName.length()+1),
99                                                   /*isConstant=*/ true,
100                                                   /*Linkage=*/ GlobalValue::PrivateLinkage,
101                                                   /*Initializer=*/ regNameData);
102    Value * regStrPtr = mBuilder->CreateGEP(regStrVar, {mBuilder->getInt64(0), mBuilder->getInt32(0)});
103    mBuilder->CreateCall(mPrintRegisterFunction, {regStrPtr, bitblockValue});
104}
105
106CompiledPabloFunction PabloCompiler::compile(PabloFunction & function) {
107
108    Examine(function);
109
110    InitializeNativeTarget();
111    InitializeNativeTargetAsmPrinter();
112    InitializeNativeTargetAsmParser();
113
114    Module * module = new Module("", getGlobalContext());
115
116    mMod = module;
117
118    std::string errMessage;
119    #ifdef USE_LLVM_3_5
120    EngineBuilder builder(mMod);
121    #else
122    EngineBuilder builder(std::move(std::unique_ptr<Module>(mMod)));
123    #endif
124    builder.setErrorStr(&errMessage);
125    builder.setMCPU(sys::getHostCPUName());
126    #ifdef USE_LLVM_3_5
127    builder.setUseMCJIT(true);
128    #endif
129    builder.setOptLevel(mMaxWhileDepth ? CodeGenOpt::Level::Less : CodeGenOpt::Level::None);
130    ExecutionEngine * engine = builder.create();
131    if (engine == nullptr) {
132        throw std::runtime_error("Could not create ExecutionEngine: " + errMessage);
133    }
134    DeclareFunctions(engine);
135    DeclareCallFunctions(engine);
136
137    auto func = compile(function, mMod);
138
139    //Display the IR that has been generated by this module.
140    if (LLVM_UNLIKELY(DumpGeneratedIR)) {
141        module->dump();
142    }
143    //Create a verifier.  The verifier will print an error message if our module is malformed in any way.
144    verifyModule(*module, &dbgs());
145
146    engine->finalizeObject();
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.getParameters().size(); ++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    unsigned totalCarryDataSize = 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.getResults().size(); ++i) {
191        SetOutputValue(mMarkerMap[function.getResult(i)], i);
192    }
193
194    //Terminate the block
195    ReturnInst::Create(mMod->getContext(), mBuilder->GetInsertBlock());
196
197    // Clean up
198    delete mCarryManager; mCarryManager = nullptr;
199    delete mBuilder; mBuilder = nullptr;
200    mMod = nullptr; // don't delete this. It's either owned by the ExecutionEngine or the calling function.
201
202    //Return the required size of the carry data area to the process_block function.
203    return std::make_pair(mFunction, totalCarryDataSize * sizeof(BitBlock));
204}
205
206inline void PabloCompiler::GenerateFunction(PabloFunction & function) {
207    mInputType = PointerType::get(StructType::get(mMod->getContext(), std::vector<Type *>(function.getParameters().size(), mBitBlockType)), 0);
208    Type * carryType = PointerType::get(mBitBlockType, 0);
209    Type * outputType = PointerType::get(StructType::get(mMod->getContext(), std::vector<Type *>(function.getResults().size(), mBitBlockType)), 0);
210    FunctionType * functionType = FunctionType::get(Type::getVoidTy(mMod->getContext()), {{mInputType, carryType, outputType}}, false);
211
212#ifdef USE_UADD_OVERFLOW
213#ifdef USE_TWO_UADD_OVERFLOW
214    // Type Definitions for llvm.uadd.with.overflow.carryin.i128 or .i256
215    std::vector<Type*>StructTy_0_fields;
216    StructTy_0_fields.push_back(IntegerType::get(mMod->getContext(), BLOCK_SIZE));
217    StructTy_0_fields.push_back(IntegerType::get(mMod->getContext(), 1));
218    StructType *StructTy_0 = StructType::get(mMod->getContext(), StructTy_0_fields, /*isPacked=*/false);
219
220    std::vector<Type*>FuncTy_1_args;
221    FuncTy_1_args.push_back(IntegerType::get(mMod->getContext(), BLOCK_SIZE));
222    FuncTy_1_args.push_back(IntegerType::get(mMod->getContext(), BLOCK_SIZE));
223    FunctionType* FuncTy_1 = FunctionType::get(
224                                              /*Result=*/StructTy_0,
225                                              /*Params=*/FuncTy_1_args,
226                                              /*isVarArg=*/false);
227
228    mFunctionUaddOverflow = mMod->getFunction("llvm.uadd.with.overflow.i" +
229                                              std::to_string(BLOCK_SIZE));
230    if (!mFunctionUaddOverflow) {
231        mFunctionUaddOverflow= Function::Create(
232          /*Type=*/ FuncTy_1,
233          /*Linkage=*/ GlobalValue::ExternalLinkage,
234          /*Name=*/ "llvm.uadd.with.overflow.i" + std::to_string(BLOCK_SIZE), mMod); // (external, no body)
235        mFunctionUaddOverflow->setCallingConv(CallingConv::C);
236    }
237    AttributeSet mFunctionUaddOverflowPAL;
238    {
239        SmallVector<AttributeSet, 4> Attrs;
240        AttributeSet PAS;
241        {
242          AttrBuilder B;
243          B.addAttribute(Attribute::NoUnwind);
244          B.addAttribute(Attribute::ReadNone);
245          PAS = AttributeSet::get(mMod->getContext(), ~0U, B);
246        }
247
248        Attrs.push_back(PAS);
249        mFunctionUaddOverflowPAL = AttributeSet::get(mMod->getContext(), Attrs);
250    }
251    mFunctionUaddOverflow->setAttributes(mFunctionUaddOverflowPAL);
252#else
253    // Type Definitions for llvm.uadd.with.overflow.carryin.i128 or .i256
254    std::vector<Type*>StructTy_0_fields;
255    StructTy_0_fields.push_back(IntegerType::get(mMod->getContext(), BLOCK_SIZE));
256    StructTy_0_fields.push_back(IntegerType::get(mMod->getContext(), 1));
257    StructType *StructTy_0 = StructType::get(mMod->getContext(), StructTy_0_fields, /*isPacked=*/false);
258
259    std::vector<Type*>FuncTy_1_args;
260    FuncTy_1_args.push_back(IntegerType::get(mMod->getContext(), BLOCK_SIZE));
261    FuncTy_1_args.push_back(IntegerType::get(mMod->getContext(), BLOCK_SIZE));
262    FuncTy_1_args.push_back(IntegerType::get(mMod->getContext(), 1));
263    FunctionType* FuncTy_1 = FunctionType::get(
264                                              /*Result=*/StructTy_0,
265                                              /*Params=*/FuncTy_1_args,
266                                              /*isVarArg=*/false);
267
268    mFunctionUaddOverflowCarryin = mMod->getFunction("llvm.uadd.with.overflow.carryin.i" +
269                                              std::to_string(BLOCK_SIZE));
270    if (!mFunctionUaddOverflowCarryin) {
271        mFunctionUaddOverflowCarryin = Function::Create(
272          /*Type=*/ FuncTy_1,
273          /*Linkage=*/ GlobalValue::ExternalLinkage,
274          /*Name=*/ "llvm.uadd.with.overflow.carryin.i" + std::to_string(BLOCK_SIZE), mMod); // (external, no body)
275        mFunctionUaddOverflowCarryin->setCallingConv(CallingConv::C);
276    }
277    AttributeSet mFunctionUaddOverflowCarryinPAL;
278    {
279        SmallVector<AttributeSet, 4> Attrs;
280        AttributeSet PAS;
281        {
282          AttrBuilder B;
283          B.addAttribute(Attribute::NoUnwind);
284          B.addAttribute(Attribute::ReadNone);
285          PAS = AttributeSet::get(mMod->getContext(), ~0U, B);
286        }
287
288        Attrs.push_back(PAS);
289        mFunctionUaddOverflowCarryinPAL = AttributeSet::get(mMod->getContext(), Attrs);
290    }
291    mFunctionUaddOverflowCarryin->setAttributes(mFunctionUaddOverflowCarryinPAL);
292#endif
293#endif
294
295    //Starts on process_block
296    SmallVector<AttributeSet, 4> Attrs;
297    Attrs.push_back(AttributeSet::get(mMod->getContext(), ~0U, { Attribute::NoUnwind, Attribute::UWTable }));
298    Attrs.push_back(AttributeSet::get(mMod->getContext(), 1U, { Attribute::ReadOnly, Attribute::NoCapture }));
299    Attrs.push_back(AttributeSet::get(mMod->getContext(), 2U, { Attribute::NoCapture }));
300    Attrs.push_back(AttributeSet::get(mMod->getContext(), 3U, { Attribute::ReadNone, Attribute::NoCapture }));
301    AttributeSet AttrSet = AttributeSet::get(mMod->getContext(), Attrs);
302
303    // Create the function that will be generated.
304    mFunction = Function::Create(functionType, GlobalValue::ExternalLinkage, function.getName()->value(), mMod);
305    mFunction->setCallingConv(CallingConv::C);
306    mFunction->setAttributes(AttrSet);
307
308    Function::arg_iterator args = mFunction->arg_begin();
309    mInputAddressPtr = args++;
310    mInputAddressPtr->setName("input");
311    mCarryDataPtr = args++;
312    mCarryDataPtr->setName("carry");
313    mOutputAddressPtr = args++;
314    mOutputAddressPtr->setName("output");
315}
316
317inline void PabloCompiler::Examine(PabloFunction & function) {
318    if (mMod == nullptr) {
319
320        mWhileDepth = 0;
321        mIfDepth = 0;
322        mMaxWhileDepth = 0;
323
324        Examine(function.getEntryBlock());
325
326        if (LLVM_UNLIKELY(mWhileDepth != 0 || mIfDepth != 0)) {
327            throw std::runtime_error("Malformed Pablo AST: Unbalanced If or While nesting depth!");
328        }
329    }
330}
331
332
333void PabloCompiler::Examine(PabloBlock & block) {
334    for (Statement * stmt : block) {
335        if (Call * call = dyn_cast<Call>(stmt)) {
336            mCalleeMap.insert(std::make_pair(call->getCallee(), nullptr));
337        }
338        else if (If * ifStatement = dyn_cast<If>(stmt)) {
339            Examine(ifStatement->getBody());
340        }
341        else if (While * whileStatement = dyn_cast<While>(stmt)) {
342            mMaxWhileDepth = std::max(mMaxWhileDepth, ++mWhileDepth);
343            Examine(whileStatement->getBody());
344            --mWhileDepth;
345        }
346    }
347}
348
349inline void PabloCompiler::DeclareFunctions(ExecutionEngine * const engine) {
350    if (DumpTrace || TraceNext) {
351        //This function can be used for testing to print the contents of a register from JIT'd code to the terminal window.
352        mPrintRegisterFunction = mMod->getOrInsertFunction("wrapped_print_register", Type::getVoidTy(mMod->getContext()), Type::getInt8PtrTy(mMod->getContext()), mBitBlockType, NULL);
353        if (engine) engine->addGlobalMapping(cast<GlobalValue>(mPrintRegisterFunction), (void *)&wrapped_print_register);
354    }
355}
356   
357void PabloCompiler::DeclareCallFunctions(ExecutionEngine * const engine) {
358    for (auto mapping : mCalleeMap) {
359        const String * callee = mapping.first;
360        auto ei = mExternalMap.find(callee->value());
361        if (ei != mExternalMap.end()) {
362
363            Type * inputType = PointerType::get(StructType::get(mMod->getContext(), std::vector<Type *>{8, mBitBlockType}), 0);
364            Type * carryType = PointerType::get(mBitBlockType, 0);
365            Type * outputType = PointerType::get(StructType::get(mMod->getContext(), std::vector<Type *>{1, mBitBlockType}), 0);
366            FunctionType * functionType = FunctionType::get(Type::getVoidTy(mMod->getContext()), std::vector<Type *>{inputType, carryType, outputType}, false);
367
368            //Starts on process_block
369            SmallVector<AttributeSet, 3> Attrs;
370            Attrs.push_back(AttributeSet::get(mMod->getContext(), 1U, { Attribute::ReadOnly, Attribute::NoCapture }));
371            Attrs.push_back(AttributeSet::get(mMod->getContext(), 2U, { Attribute::NoCapture }));
372            Attrs.push_back(AttributeSet::get(mMod->getContext(), 3U, { Attribute::ReadNone, Attribute::NoCapture }));
373            AttributeSet AttrSet = AttributeSet::get(mMod->getContext(), Attrs);
374
375            Function * externalFunction = cast<Function>(mMod->getOrInsertFunction(callee->value(), functionType, AttrSet));
376            if (LLVM_UNLIKELY(externalFunction == nullptr)) {
377                throw std::runtime_error("Could not create static method call for external function \"" + callee->to_string() + "\"");
378            }
379            externalFunction->setCallingConv(llvm::CallingConv::C);
380
381            if (engine) engine->addGlobalMapping(externalFunction, ei->second);
382            mCalleeMap[callee] = externalFunction;
383        }
384        else {
385            throw std::runtime_error("External function \"" + callee->to_string() + "\" not installed");
386        }
387    }
388}
389
390void PabloCompiler::compileBlock(PabloBlock & block) {
391    mPabloBlock = & block;
392    for (const Statement * statement : block) {
393        compileStatement(statement);
394    }
395    mPabloBlock = block.getParent();
396}
397
398Value * PabloCompiler::genBitTest2(Value * e1, Value * e2) {
399    Type * t1 = e1->getType();
400    Type * t2 = e2->getType();
401    if (t1 == mBitBlockType) {
402        if (t2 == mBitBlockType) {
403            return iBuilder.bitblock_any(mBuilder->CreateOr(e1, e2));
404        }
405        else {
406            Value * m1 = mBuilder->CreateZExt(iBuilder.hsimd_signmask(16, e1), t2);
407            return mBuilder->CreateICmpNE(mBuilder->CreateOr(m1, e2), ConstantInt::get(t2, 0));
408        }
409    }
410    else if (t2 == mBitBlockType) {
411        Value * m2 = mBuilder->CreateZExt(iBuilder.hsimd_signmask(16, e2), t1);
412        return mBuilder->CreateICmpNE(mBuilder->CreateOr(e1, m2), ConstantInt::get(t1, 0));
413    }
414    else {
415        return mBuilder->CreateICmpNE(mBuilder->CreateOr(e1, e2), ConstantInt::get(t1, 0));
416    }
417}
418
419void PabloCompiler::compileIf(const If * ifStatement) {       
420    //
421    //  The If-ElseZero stmt:
422    //  if <predicate:expr> then <body:stmt>* elsezero <defined:var>* endif
423    //  If the value of the predicate is nonzero, then determine the values of variables
424    //  <var>* by executing the given statements.  Otherwise, the value of the
425    //  variables are all zero.  Requirements: (a) no variable that is defined within
426    //  the body of the if may be accessed outside unless it is explicitly
427    //  listed in the variable list, (b) every variable in the defined list receives
428    //  a value within the body, and (c) the logical consequence of executing
429    //  the statements in the event that the predicate is zero is that the
430    //  values of all defined variables indeed work out to be 0.
431    //
432    //  Simple Implementation with Phi nodes:  a phi node in the if exit block
433    //  is inserted for each variable in the defined variable list.  It receives
434    //  a zero value from the ifentry block and the defined value from the if
435    //  body.
436    //
437
438    BasicBlock * ifEntryBlock = mBuilder->GetInsertBlock();
439    BasicBlock * ifBodyBlock = BasicBlock::Create(mMod->getContext(), "if.body", mFunction, 0);
440    BasicBlock * ifEndBlock = BasicBlock::Create(mMod->getContext(), "if.end", mFunction, 0);
441   
442    PabloBlock & ifBody = ifStatement -> getBody();
443   
444    Value * if_test_value = compileExpression(ifStatement->getCondition());
445   
446    mCarryManager->enterScope(&ifBody);
447    if (mCarryManager->blockHasCarries()) {
448        // load the summary variable
449        Value* last_if_pending_data = mCarryManager->getCarrySummaryExpr();
450        mBuilder->CreateCondBr(genBitTest2(if_test_value, last_if_pending_data), ifBodyBlock, ifEndBlock);
451
452    }
453    else {
454        mBuilder->CreateCondBr(iBuilder.bitblock_any(if_test_value), ifBodyBlock, ifEndBlock);
455    }
456    // Entry processing is complete, now handle the body of the if.
457    mBuilder->SetInsertPoint(ifBodyBlock);
458   
459   
460    ++mIfDepth;
461    compileBlock(ifBody);
462    --mIfDepth;
463    if (mCarryManager->blockHasCarries()) {
464        mCarryManager->generateCarryOutSummaryCodeIfNeeded();
465    }
466    BasicBlock * ifBodyFinalBlock = mBuilder->GetInsertBlock();
467    mBuilder->CreateBr(ifEndBlock);
468    //End Block
469    mBuilder->SetInsertPoint(ifEndBlock);
470    for (const PabloAST * node : ifStatement->getDefined()) {
471        const Assign * assign = cast<Assign>(node);
472        PHINode * phi = mBuilder->CreatePHI(mBitBlockType, 2, assign->getName()->value());
473        auto f = mMarkerMap.find(assign);
474        assert (f != mMarkerMap.end());
475        phi->addIncoming(mZeroInitializer, ifEntryBlock);
476        phi->addIncoming(f->second, ifBodyFinalBlock);
477        mMarkerMap[assign] = phi;
478    }
479    // Create the phi Node for the summary variable, if needed.
480    mCarryManager->addSummaryPhiIfNeeded(ifEntryBlock, ifBodyFinalBlock);
481    mCarryManager->leaveScope();
482}
483
484void PabloCompiler::compileWhile(const While * whileStatement) {
485
486    PabloBlock & whileBody = whileStatement -> getBody();
487   
488    BasicBlock * whileEntryBlock = mBuilder->GetInsertBlock();
489    BasicBlock * whileBodyBlock = BasicBlock::Create(mMod->getContext(), "while.body", mFunction, 0);
490    BasicBlock * whileEndBlock = BasicBlock::Create(mMod->getContext(), "while.end", mFunction, 0);
491
492    mCarryManager->enterScope(&whileBody);
493    mCarryManager->ensureCarriesLoadedRecursive();
494
495    const auto & nextNodes = whileStatement->getVariants();
496    std::vector<PHINode *> nextPhis;
497    nextPhis.reserve(nextNodes.size());
498
499    // On entry to the while structure, proceed to execute the first iteration
500    // of the loop body unconditionally.   The while condition is tested at the end of
501    // the loop.
502
503    mBuilder->CreateBr(whileBodyBlock);
504    mBuilder->SetInsertPoint(whileBodyBlock);
505
506    //
507    // There are 3 sets of Phi nodes for the while loop.
508    // (1) Carry-ins: (a) incoming carry data first iterations, (b) zero thereafter
509    // (2) Carry-out accumulators: (a) zero first iteration, (b) |= carry-out of each iteration
510    // (3) Next nodes: (a) values set up before loop, (b) modified values calculated in loop.
511
512    mCarryManager->initializeCarryDataPhisAtWhileEntry(whileEntryBlock);
513
514    // for any Next nodes in the loop body, initialize to (a) pre-loop value.
515    for (const Next * n : nextNodes) {
516        PHINode * phi = mBuilder->CreatePHI(mBitBlockType, 2, n->getName()->value());
517        auto f = mMarkerMap.find(n->getInitial());
518        assert (f != mMarkerMap.end());
519        phi->addIncoming(f->second, whileEntryBlock);
520        mMarkerMap[n->getInitial()] = phi;
521        nextPhis.push_back(phi);
522    }
523
524    //
525    // Now compile the loop body proper.  Carry-out accumulated values
526    // and iterated values of Next nodes will be computed.
527    ++mWhileDepth;
528    compileBlock(whileBody);
529
530    BasicBlock * whileBodyFinalBlock = mBuilder->GetInsertBlock();
531
532    mCarryManager->extendCarryDataPhisAtWhileBodyFinalBlock(whileBodyFinalBlock);
533
534    // Terminate the while loop body with a conditional branch back.
535    mBuilder->CreateCondBr(iBuilder.bitblock_any(compileExpression(whileStatement->getCondition())), whileBodyBlock, whileEndBlock);
536
537    // and for any Next nodes in the loop body
538    for (unsigned i = 0; i < nextNodes.size(); i++) {
539        const Next * n = nextNodes[i];
540        auto f = mMarkerMap.find(n->getExpr());
541        if (LLVM_UNLIKELY(f == mMarkerMap.end())) {
542            throw std::runtime_error("Next node expression was not compiled!");
543        }
544        nextPhis[i]->addIncoming(f->second, whileBodyFinalBlock);
545    }
546
547    mBuilder->SetInsertPoint(whileEndBlock);
548    --mWhileDepth;
549
550    mCarryManager->ensureCarriesStoredRecursive();
551    mCarryManager->leaveScope();
552}
553
554
555void PabloCompiler::compileStatement(const Statement * stmt) {
556    Value * expr = nullptr;
557    if (const Assign * assign = dyn_cast<const Assign>(stmt)) {
558        expr = compileExpression(assign->getExpression());
559    }
560    else if (const Next * next = dyn_cast<const Next>(stmt)) {
561        expr = compileExpression(next->getExpr());
562        if (TraceNext) {
563            genPrintRegister(next->getName()->to_string(), expr);
564        }
565    }
566    else if (const If * ifStatement = dyn_cast<const If>(stmt)) {
567        compileIf(ifStatement);
568        return;
569    }
570    else if (const While * whileStatement = dyn_cast<const While>(stmt)) {
571        compileWhile(whileStatement);
572        return;
573    }
574    else if (const Call* call = dyn_cast<Call>(stmt)) {
575        //Call the callee once and store the result in the marker map.
576        if (mMarkerMap.count(call) != 0) {
577            return;
578        }
579        auto ci = mCalleeMap.find(call->getCallee());
580        if (LLVM_UNLIKELY(ci == mCalleeMap.end())) {
581            throw std::runtime_error("Unexpected error locating static function for \"" + call->getCallee()->to_string() + "\"");
582        }       
583
584        Function * function = ci->second;
585        auto arg = function->getArgumentList().begin();
586        Value * carryFramePtr = ConstantPointerNull::get(cast<PointerType>((++arg)->getType()));
587        AllocaInst * outputStruct = mBuilder->CreateAlloca(cast<PointerType>((++arg)->getType())->getElementType());
588        mBuilder->CreateCall3(function, mInputAddressPtr, carryFramePtr, outputStruct);
589        Value * outputPtr = mBuilder->CreateGEP(outputStruct, { mBuilder->getInt32(0), mBuilder->getInt32(0) });
590        expr = mBuilder->CreateAlignedLoad(outputPtr, BLOCK_SIZE / 8, false);
591    }
592    else if (const And * pablo_and = dyn_cast<And>(stmt)) {
593        expr = mBuilder->CreateAnd(compileExpression(pablo_and->getExpr1()), compileExpression(pablo_and->getExpr2()), "and");
594    }
595    else if (const Or * pablo_or = dyn_cast<Or>(stmt)) {
596        expr = mBuilder->CreateOr(compileExpression(pablo_or->getExpr1()), compileExpression(pablo_or->getExpr2()), "or");
597    }
598    else if (const Xor * pablo_xor = dyn_cast<Xor>(stmt)) {
599        expr = mBuilder->CreateXor(compileExpression(pablo_xor->getExpr1()), compileExpression(pablo_xor->getExpr2()), "xor");
600    }
601    else if (const Sel * sel = dyn_cast<Sel>(stmt)) {
602        Value* ifMask = compileExpression(sel->getCondition());
603        Value* ifTrue = mBuilder->CreateAnd(ifMask, compileExpression(sel->getTrueExpr()));
604        Value* ifFalse = mBuilder->CreateAnd(genNot(ifMask), compileExpression(sel->getFalseExpr()));
605        expr = mBuilder->CreateOr(ifTrue, ifFalse);
606    }
607    else if (const Not * pablo_not = dyn_cast<Not>(stmt)) {
608        expr = genNot(compileExpression(pablo_not->getExpr()));
609    }
610    else if (const Advance * adv = dyn_cast<Advance>(stmt)) {
611        Value* strm_value = compileExpression(adv->getExpr());
612        int shift = adv->getAdvanceAmount();
613        unsigned advance_index = adv->getLocalAdvanceIndex();
614        expr = mCarryManager->advanceCarryInCarryOut(advance_index, shift, strm_value);
615    }
616    else if (const MatchStar * mstar = dyn_cast<MatchStar>(stmt)) {
617        Value * marker = compileExpression(mstar->getMarker());
618        Value * cc = compileExpression(mstar->getCharClass());
619        Value * marker_and_cc = mBuilder->CreateAnd(marker, cc);
620        unsigned carry_index = mstar->getLocalCarryIndex();
621        expr = mBuilder->CreateOr(mBuilder->CreateXor(genAddWithCarry(marker_and_cc, cc, carry_index), cc), marker, "matchstar");
622    }
623    else if (const ScanThru * sthru = dyn_cast<ScanThru>(stmt)) {
624        Value * marker_expr = compileExpression(sthru->getScanFrom());
625        Value * cc_expr = compileExpression(sthru->getScanThru());
626        unsigned carry_index = sthru->getLocalCarryIndex();
627        expr = mBuilder->CreateAnd(genAddWithCarry(marker_expr, cc_expr, carry_index), genNot(cc_expr), "scanthru");
628    }
629    else {
630        llvm::raw_os_ostream cerr(std::cerr);
631        PabloPrinter::print(stmt, cerr);
632        throw std::runtime_error("Unrecognized Pablo Statement! can't compile.");
633    }
634    mMarkerMap[stmt] = expr;
635    if (DumpTrace) {
636        genPrintRegister(stmt->getName()->to_string(), expr);
637    }
638   
639}
640
641Value * PabloCompiler::compileExpression(const PabloAST * expr) {
642    if (isa<Ones>(expr)) {
643        return mOneInitializer;
644    }
645    else if (isa<Zeroes>(expr)) {
646        return mZeroInitializer;
647    }
648    auto f = mMarkerMap.find(expr);
649    if (LLVM_UNLIKELY(f == mMarkerMap.end())) {
650        std::string o;
651        llvm::raw_string_ostream str(o);
652        str << "\"";
653        PabloPrinter::print(expr, str);
654        str << "\" was used before definition!";
655        throw std::runtime_error(str.str());
656    }
657    return f->second;
658}
659
660
661#ifdef USE_UADD_OVERFLOW
662#ifdef USE_TWO_UADD_OVERFLOW
663PabloCompiler::SumWithOverflowPack PabloCompiler::callUaddOverflow(Value* int128_e1, Value* int128_e2) {
664    std::vector<Value*> struct_res_params;
665    struct_res_params.push_back(int128_e1);
666    struct_res_params.push_back(int128_e2);
667    CallInst* struct_res = CallInst::Create(mFunctionUaddOverflow, struct_res_params, "uadd_overflow_res", mBasicBlock);
668    struct_res->setCallingConv(CallingConv::C);
669    struct_res->setTailCall(false);
670    AttributeSet struct_res_PAL;
671    struct_res->setAttributes(struct_res_PAL);
672
673    SumWithOverflowPack ret;
674
675    std::vector<unsigned> int128_sum_indices;
676    int128_sum_indices.push_back(0);
677    ret.sum = ExtractValueInst::Create(struct_res, int128_sum_indices, "sum", mBasicBlock);
678
679    std::vector<unsigned> int1_obit_indices;
680    int1_obit_indices.push_back(1);
681    ret.obit = ExtractValueInst::Create(struct_res, int1_obit_indices, "obit", mBasicBlock);
682
683    return ret;
684}
685#else
686PabloCompiler::SumWithOverflowPack PabloCompiler::callUaddOverflow(Value* int128_e1, Value* int128_e2, Value* int1_cin) {
687    std::vector<Value*> struct_res_params;
688    struct_res_params.push_back(int128_e1);
689    struct_res_params.push_back(int128_e2);
690    struct_res_params.push_back(int1_cin);
691    CallInst* struct_res = CallInst::Create(mFunctionUaddOverflowCarryin, struct_res_params, "uadd_overflow_res", mBasicBlock);
692    struct_res->setCallingConv(CallingConv::C);
693    struct_res->setTailCall(false);
694    AttributeSet struct_res_PAL;
695    struct_res->setAttributes(struct_res_PAL);
696
697    SumWithOverflowPack ret;
698
699    std::vector<unsigned> int128_sum_indices;
700    int128_sum_indices.push_back(0);
701    ret.sum = ExtractValueInst::Create(struct_res, int128_sum_indices, "sum", mBasicBlock);
702
703    std::vector<unsigned> int1_obit_indices;
704    int1_obit_indices.push_back(1);
705    ret.obit = ExtractValueInst::Create(struct_res, int1_obit_indices, "obit", mBasicBlock);
706
707    return ret;
708}
709#endif
710#endif
711
712
713Value* PabloCompiler::genAddWithCarry(Value* e1, Value* e2, unsigned localIndex) {
714    Value * carryq_value = mCarryManager->getCarryOpCarryIn(localIndex);
715#ifdef USE_TWO_UADD_OVERFLOW
716    //This is the ideal implementation, which uses two uadd.with.overflow
717    //The back end should be able to recognize this pattern and combine it into uadd.with.overflow.carryin
718    CastInst* int128_e1 = new BitCastInst(e1, mBuilder->getIntNTy(BLOCK_SIZE), "e1_128", mBasicBlock);
719    CastInst* int128_e2 = new BitCastInst(e2, mBuilder->getIntNTy(BLOCK_SIZE), "e2_128", mBasicBlock);
720    CastInst* int128_carryq_value = new BitCastInst(carryq_value, mBuilder->getIntNTy(BLOCK_SIZE), "carryq_128", mBasicBlock);
721
722    SumWithOverflowPack sumpack0, sumpack1;
723
724    sumpack0 = callUaddOverflow(int128_e1, int128_e2);
725    sumpack1 = callUaddOverflow(sumpack0.sum, int128_carryq_value);
726
727    Value* obit = mBuilder->CreateOr(sumpack0.obit, sumpack1.obit, "carry_bit");
728    Value* sum = mBuilder->CreateBitCast(sumpack1.sum, mBitBlockType, "ret_sum");
729
730    /*obit is the i1 carryout, zero extend and insert it into a v2i64 or v4i64 vector.*/
731    ConstantAggregateZero* const_packed_5 = ConstantAggregateZero::get(mBitBlockType);
732    ConstantInt* const_int32_6 = ConstantInt::get(mMod->getContext(), APInt(32, StringRef("0"), 10));
733    CastInst* int64_o0 = new ZExtInst(obit, IntegerType::get(mMod->getContext(), 64), "o0", mBasicBlock);
734    InsertElementInst* carry_out = InsertElementInst::Create(const_packed_5, int64_o0, const_int32_6, "carry_out", mBasicBlock);
735
736#elif defined USE_UADD_OVERFLOW
737    //use llvm.uadd.with.overflow.i128 or i256
738    CastInst* int128_e1 = new BitCastInst(e1, mBuilder->getIntNTy(BLOCK_SIZE), "e1_128", mBasicBlock);
739    CastInst* int128_e2 = new BitCastInst(e2, mBuilder->getIntNTy(BLOCK_SIZE), "e2_128", mBasicBlock);
740
741    //get i1 carryin from iBLOCK_SIZE
742    ConstantInt* const_int32_6 = ConstantInt::get(mMod->getContext(), APInt(32, StringRef("0"), 10));
743    ExtractElementInst * int64_carryq_value = ExtractElementInst::Create(carryq_value, const_int32_6, "carryq_64", mBasicBlock);
744    CastInst* int1_carryq_value = new TruncInst(int64_carryq_value, IntegerType::get(mMod->getContext(), 1), "carryq_1", mBasicBlock);
745
746    SumWithOverflowPack sumpack0;
747    sumpack0 = callUaddOverflow(int128_e1, int128_e2, int1_carryq_value);
748    Value* obit = sumpack0.obit;
749    Value* sum = mBuilder->CreateBitCast(sumpack0.sum, mBitBlockType, "sum");
750
751    /*obit is the i1 carryout, zero extend and insert it into a v2i64 or v4i64 vector.*/
752    ConstantAggregateZero* const_packed_5 = ConstantAggregateZero::get(mBitBlockType);
753    CastInst* int64_o0 = new ZExtInst(obit, IntegerType::get(mMod->getContext(), 64), "o0", mBasicBlock);
754    InsertElementInst* carry_out = InsertElementInst::Create(const_packed_5, int64_o0, const_int32_6, "carry_out", mBasicBlock);
755#elif (BLOCK_SIZE == 128)
756    //calculate carry through logical ops
757    Value* carrygen = mBuilder->CreateAnd(e1, e2, "carrygen");
758    Value* carryprop = mBuilder->CreateOr(e1, e2, "carryprop");
759    Value* digitsum = mBuilder->CreateAdd(e1, e2, "digitsum");
760    Value* partial = mBuilder->CreateAdd(digitsum, carryq_value, "partial");
761    Value* digitcarry = mBuilder->CreateOr(carrygen, mBuilder->CreateAnd(carryprop, genNot(partial)));
762    Value* mid_carry_in = genShiftLeft64(mBuilder->CreateLShr(digitcarry, 63), "mid_carry_in");
763
764    Value* sum = mBuilder->CreateAdd(partial, mid_carry_in, "sum");
765    Value* carry_out = genShiftHighbitToLow(BLOCK_SIZE, mBuilder->CreateOr(carrygen, mBuilder->CreateAnd(carryprop, genNot(sum))));
766#else
767    //BLOCK_SIZE == 256, there is no other implementation
768    static_assert(false, "Add with carry for 256-bit bitblock requires USE_UADD_OVERFLOW");
769#endif //USE_TWO_UADD_OVERFLOW
770
771    mCarryManager->setCarryOpCarryOut(localIndex, carry_out);
772    return sum;
773}
774
775Value * PabloCompiler::genShiftHighbitToLow(unsigned FieldWidth, Value * op) {
776    unsigned FieldCount = BLOCK_SIZE/FieldWidth;
777    VectorType * vType = VectorType::get(IntegerType::get(mMod->getContext(), FieldWidth), FieldCount);
778    Value * v = mBuilder->CreateBitCast(op, vType);
779    return mBuilder->CreateBitCast(mBuilder->CreateLShr(v, FieldWidth - 1), mBitBlockType);
780}
781
782Value* PabloCompiler::genShiftLeft64(Value* e, const Twine &namehint) {
783    Value* i128_val = mBuilder->CreateBitCast(e, mBuilder->getIntNTy(BLOCK_SIZE));
784    return mBuilder->CreateBitCast(mBuilder->CreateShl(i128_val, 64, namehint), mBitBlockType);
785}
786
787inline Value* PabloCompiler::genNot(Value* expr) {
788    return mBuilder->CreateXor(expr, mOneInitializer, "not");
789}
790   
791void PabloCompiler::SetOutputValue(Value * marker, const unsigned index) {
792    if (marker->getType()->isPointerTy()) {
793        marker = mBuilder->CreateAlignedLoad(marker, BLOCK_SIZE/8, false);
794    }
795    Value* indices[] = {mBuilder->getInt64(0), mBuilder->getInt32(index)};
796    Value* gep = mBuilder->CreateGEP(mOutputAddressPtr, indices);
797    mBuilder->CreateAlignedStore(marker, gep, BLOCK_SIZE/8, false);
798}
799
800CompiledPabloFunction::CompiledPabloFunction(size_t carryDataSize, Function * function, ExecutionEngine * executionEngine)
801: CarryDataSize(carryDataSize)
802, FunctionPointer(executionEngine->getPointerToFunction(function))
803, mFunction(function)
804, mExecutionEngine(executionEngine)
805{
806
807}
808
809// Clean up the memory for the compiled function once we're finished using it.
810CompiledPabloFunction::~CompiledPabloFunction() {
811    if (mExecutionEngine) {
812        assert (mFunction);
813        // mExecutionEngine->freeMachineCodeForFunction(mFunction); // This function only prints a "not supported" message. Reevaluate with LLVM 3.6.
814        delete mExecutionEngine;
815    }
816}
817
818}
Note: See TracBrowser for help on using the repository browser.