LLVM每日谈之十六 LLVM的学习感悟

Although I knew that LLVM uses a Single Static Assignment (SSA) format, it wasn't obvious to me how prevalent this idea was in LLVM until I reallystarted using it. Reading the Programmer's Manual and Language Reference,I noted that most of the important LLVM IR (Intermediate Representation) C++ classes were derived from the Value class. The full power of that simpledesign only became fully understood once I started constructing executableexpressions for Stacker.

This really makes your programming go faster. Think about compiling codefor the following C/C++ expression: (a|b)*((x+1)/(y+1)). Assumingthe values are on the stack in the order a, b, x, y, this could beexpressed in stacker as: 1 + SWAP 1 + / ROT2 OR *.You could write a function using LLVM that computes this expression like this:

Value* expression(BasicBlock*bb, Value* a, Value* b, Value* x, Value* y ){    
 Instruction* tail = bb-<getTerminator();    
 ConstantSInt* one = ConstantSInt::get( Type::IntTy, 1);    
 BinaryOperator* or1 = 	BinaryOperator::create( Instruction::Or, a, b, "", tail );
 BinaryOperator* add1 = 	BinaryOperator::create( Instruction::Add, x, one, "", tail );
 BinaryOperator* add2 =	BinaryOperator::create( Instruction::Add, y, one, "", tail ); 
 BinaryOperator* div1 = 	BinaryOperator::create( Instruction::Div, add1, add2, "", tail);
 BinaryOperator* mult1 = 	BinaryOperator::create( Instruction::Mul, or1, div1, "", tail );    return mult1;}

"Okay, big deal," you say. It is a big deal. Here's why. Note that I didn'thave to tell this function which kinds of Values are being passed in. They could beInstructions, Constants, GlobalVariables, etc. Furthermore, if you specify Values that are incorrect for this sequence of operations, LLVM will either notice right away (at compilation time) or the LLVM Verifier will pick up the inconsistency when the compiler runs. In no case will you make a type error that gets passed through to the generated program. This really helps you write a compiler that always generates correct code!

The second point is that we don't have to worry about branching, registers,stack variables, saving partial results, etc. The instructions we create are the values we use. Note that all that was created in the abovecode is a Constant value and five operators.Each of the instructions is the resulting value of that instruction. This saves a lot of time.

The lesson is this: SSA form is very powerful: there is no differencebetween a value and the instruction that created it. This is fullyenforced by the LLVM IR. Use it to your best advantage.

I had to learn about terminating blocks the hard way: using the debugger to figure out what the LLVM verifier was trying to tell me and begging forhelp on the LLVMdev mailing list. I hope you avoid this experience.

Emblazon this rule in your mind:

• AllBasicBlocks in your compiler must be terminated with a terminating instruction (branch, return, etc.).

Terminating instructions are a semantic requirement of the LLVM IR. Thereis no facility for implicitly chaining together blocks placed into a functionin the order they occur. Indeed, in the general case, blocks will not beadded to the function in the order of execution because of the recursiveway compilers are written.

Furthermore, if you don't terminate your blocks, your compiler code will compile just fine. You won't find out about the problem until you're running the compiler and the module you just created fails on the LLVM Verifier.

After a little initial fumbling around, I quickly caught on to how blocksshould be constructed. In general, here's what I learned:

1. Create your blocks early. While writing your compiler, you will encounter several situations where you know apriori that you will need several blocks. For example, if-then-else, switch, while, and for statements in C/C++ all need multiple blocks for expression in LVVM. The rule is, create them early.
2. Terminate your blocks early. This just reduces the chances that you forget to terminate your blocks which is required (go here for more).
3. Use getTerminator() for instruction insertion. I noticed early on that many of the constructors for the Instruction classes take an optional insert_before argument. At first, I thought this was a mistake because clearly the normal mode of inserting instructions would be one at a time after some other instruction, not before. However, if you hold on to your terminating instruction (or use the handy dandy getTerminator() method on a BasicBlock), it can always be used as the insert_before argument to your instruction constructors. This causes the instruction to automatically be inserted in the RightPlace™ place, just before the terminating instruction. The nice thing about this design is that you can pass blocks around and insert new instructions into them without ever knowing what instructions came before. This makes for some very clean compiler design.

The foregoing is such an important principal, its worth making an idiom:

BasicBlock* bb = new BasicBlock();
bb-<getInstList().push_back( new Branch( ... ) );
new Instruction(..., bb-<getTerminator() );

To make this clear, consider the typical if-then-else statement(see StackerCompiler::handle_if() method). We can set this upin a single function using LLVM in the following way:

using namespace llvm;
BasicBlock*MyCompiler::handle_if( BasicBlock* bb, SetCondInst* condition ){
 // Create the blocks to contain code in the structure of if/then/else    
 BasicBlock* then = new BasicBlock();
 BasicBlock* else = new BasicBlock();
 BasicBlock* exit = new BasicBlock();    // Insert the branch instruction for the "if"    
bb-<getInstList().push_back( new BranchInst( then, else, condition ) );    // Set up the terminating instructions    then-<getInstList().push_back( new BranchInst( exit ) );    else-<getInstList().push_back( new BranchInst( exit ) );    // Fill in the then part .. details excised for brevity    this-<fill_in( then );    // Fill in the else part .. details excised for brevity    this-<fill_in( else );    // Return a block to the caller that can be filled in with the code    // that follows the if/then/else construct.    return exit;}

Presumably in the foregoing, the calls to the "fill_in" method would add the instructions for the "then" and "else" parts. They would use the third partof the idiom almost exclusively (inserting new instructions before the terminator). Furthermore, they could even recurse back to handle_if should they encounter another if/then/else statement, and it will just work.

Note how cleanly this all works out. In particular, the push_back methods onthe BasicBlock's instruction list. These are lists of type Instruction which also happen to be Values. To create the "if" branch, we merely instantiate a BranchInst that takes as arguments the blocks to branch to and the condition to branch on. The blocksact like branch labels! This new BranchInst terminatesthe BasicBlock provided as an argument. To give the caller a wayto keep inserting after calling handle_if, we create an "exit" blockwhich is returned to the caller. Note that the "exit" block is used as the terminator for both the "then" and the "else" blocks. This guarantees that nomatter what else "handle_if" or "fill_in" does, they end up at the "exit" block.

One of the first things I noticed is the frequent use of the "push_back"method on the various lists. This is so common that it is worth mentioning.The "push_back" inserts a value into an STL list, vector, array, etc. at theend. The method might have also been named "insert_tail" or "append".Although I've used STL quite frequently, my use of push_back wasn't veryhigh in other programs. In LLVM, you'll use it all the time.

It took a little getting used to and several rounds of postings to the LLVMmailing list to wrap my head around this instruction correctly. Even though I hadread the Language Reference and Programmer's Manual a couple times each, I stillmissed a few very key points:

• GetElementPtrInst gives you back a Value for the last thing indexed.
• All global variables in LLVM are pointers.
• Pointers must also be dereferenced with the GetElementPtrInst instruction.

This means that when you look up an element in the global variable (assumingit's a struct or array), you must deference the pointer first! For manythings, this leads to the idiom:

std::vector index_vector;
index_vector.push_back( ConstantSInt::get( Type::LongTy, 0 );// ... push other indices ...
GetElementPtrInst* gep = new GetElementPtrInst( ptr, index_vector );

For example, suppose we have a global variable whose type is [24 x int]. Thevariable itself represents a pointer to that array. To subscript thearray, we need two indices, not just one. The first index (0) dereferences thepointer. The second index subscripts the array. If you're a "C" programmer, thiswill run against your grain because you'll naturally think of the global arrayvariable and the address of its first element as the same. That tripped me upfor a while until I realized that they really do differ .. by type.Remember that LLVM is a strongly typed language itself. Everythinghas a type. The "type" of the global variable is [24 x int]*. That is, it'sa pointer to an array of 24 ints. When you dereference that global variable witha single (0) index, you now have a "[24 x int]" type. Althoughthe pointer value of the dereferenced global and the address of the zero'th elementin the array will be the same, they differ in their type. The zero'th element hastype "int" while the pointer value has type "[24 x int]".

Get this one aspect of LLVM right in your head, and you'll save yourselfa lot of compiler writing headaches down the road.

Linkage types in LLVM can be a little confusing, especially if your compilerwriting mind has affixed very hard concepts to particular words like "weak","external", "global", "linkonce", etc. LLVM does not use the precisedefinitions of, say, ELF or GCC, even though they share common terms. To be fair,the concepts are related and similar but not precisely the same. This can leadyou to think you know what a linkage type represents but in fact it is slightlydifferent. I recommend you read the Language Reference on this topic very carefully. Then, read it again.

Here are some handy tips that I discovered along the way:

• Uninitialized means external. That is, the symbol is declared in the current module and can be used by that module but it is not defined by that module.
• Setting an initializer changes a global's linkage type from whatever it was to a normal, defined global (not external). You'll need to call the setLinkage() method to reset it if you specify the initializer after the GlobalValue has been constructed. This is important for LinkOnce and Weak linkage types.
• Appending linkage can be used to keep track of compilation information at runtime. It could be used, for example, to build a full table of all the C++ virtual tables or hold the C++ RTTI data, or whatever. Appending linkage can only be applied to arrays. The arrays are concatenated together at link time.

Constants in LLVM took a little getting used to until I discovered a few utilityfunctions in the LLVM IR that make things easier. Here's what I learned:

• Constants are Values like anything else and can be operands of instructions
• Integer constants, frequently needed, can be created using the static "get" methods of the ConstantInt, ConstantSInt, and ConstantUInt classes. The nice thing about these is that you can "get" any kind of integer quickly.
• There's a special method on Constant class which allows you to get the null constant for any type. This is really handy for initializing large arrays or structures, etc.