Written by Reid Spencer

This document is another way to learn about LLVM. Unlike the LLVM Reference Manual or LLVM Programmer's Manual, this document walks you through the implementation of a programming language named Stacker. Stacker was invented specifically as a demonstration of LLVM. The emphasis in this document is not on describing the intricacies of LLVM itself, but on how to use it to build your own compiler system.

Amongst other things, LLVM is a platform for compiler writers. Because of its exceptionally clean and small IR (intermediate representation), compiler writing with LLVM is much easier than with other system. As proof, the author of Stacker wrote the entire compiler (language definition, lexer, parser, code generator, etc.) in about four days! That's important to know because it shows how quickly you can get a new language up when using LLVM. Furthermore, this was the first language the author ever created using LLVM. The learning curve is included in that four days.

The language described here, Stacker, is Forth-like. Programs are simple collections of word definitions and the only thing definitions can do is manipulate a stack or generate I/O. Stacker is not a "real" programming language; its very simple. Although it is computationally complete, you wouldn't use it for your next big project. However, the fact that it is complete, its simple, and it doesn't have a C-like syntax make it useful for demonstration purposes. It shows that LLVM could be applied to a wide variety of language syntaxes.

The basic notions behind stacker is very simple. There's a stack of integers (or character pointers) that the program manipulates. Pretty much the only thing the program can do is manipulate the stack and do some limited I/O operations. The language provides you with several built-in words that manipulate the stack in interesting ways. To get your feet wet, here's how you write the traditional "Hello, World" program in Stacker:

: hello_world "Hello, World!" >s DROP CR ;
: MAIN hello_world ;


This has two "definitions" (Stacker manipulates words, not functions and words have definitions): MAIN and hello_world. The MAIN definition is standard, it tells Stacker where to start. Here, MAIN is defined to simply invoke the word hello_world. The hello_world definition tells stacker to push the "Hello, World!" string onto the stack, print it out (>s), pop it off the stack (DROP), and finally print a carriage return (CR). Although hello_world uses the stack, its net effect is null. Well written Stacker definitions have that characteristic.

Exercise for the reader: how could you make this a one line program?

Stacker was written for two purposes: (a) to get the author over the learning curve and (b) to provide a simple example of how to write a compiler using LLVM. During the development of Stacker, many lessons about LLVM were learned. Those lessons are described in the following subsections.

Although I knew that LLVM used a Single Static Assignment (SSA) format, it wasn't obvious to me how prevalent this idea was in LLVM until I really started 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 simple design only became fully understood once I started constructing executable expressions for Stacker.

This really makes your programming go faster. Think about compiling code for the following C/C++ expression: (a|b)*((x+1)/(y+1)). You could write a function using LLVM that does exactly that, this way:

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 = 
	new BinaryOperator::create( Instruction::Or, a, b, "", tail );
    BinaryOperator* add1 = 
	new BinaryOperator::create( Instruction::Add, x, one, "", tail );
    BinaryOperator* add2 =
	new BinaryOperator::create( Instruction::Add, y, one, "", tail );
    BinaryOperator* div1 = 
	new BinaryOperator::create( Instruction::Div, add1, add2, "", tail);
    BinaryOperator* mult1 = 
	new 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't have to tell this function which kinds of Values are being passed in. They could be instructions, Constants, Global Variables, 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 above code is a Constant value and five operators. Each of the instructions is the resulting value of that instruction.

The lesson is this: SSA form is very powerful: there is no difference between a value and the instruction that created it. This is fully enforced 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 for help on the LLVMdev mailing list. I hope you avoid this experience.

Emblazon this rule in your mind:

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

Terminating instructions are a semantic requirement of the LLVM IR. There is no facility for implicitly chaining together blocks placed into a function in the order they occur. Indeed, in the general case, blocks will not be added to the function in the order of execution because of the recursive way 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 blocks should be constructed. The use of the standard template library really helps simply the interface. 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&tm; 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 known 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 up in 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 part of 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 all "just work".

Note how cleanly this all works out. In particular, the push_back methods on the 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 blocks act like branch labels! This new BranchInst terminates the BasicBlock provided as an argument. To give the caller a way to keep inserting after calling handle_if we create an "exit" block which is returned to the caller. Note that the "exit" block is used as the terminator for both the "then" and the "else" blocks. This gaurantees that no matter 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 the end. The method might have also been named "insert_tail" or "append". Althought I've used STL quite frequently, my use of push_back wasn't very high 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 LLVM mail list to wrap my head around this instruction correctly. Even though I had read the Language Reference and Programmer's Manual a couple times each, I still missed 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 (assuming its a struct or array), you must deference the pointer first! For many things, 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]. The variable itself represents a pointer to that array. To subscript the array, we need two indices, not just one. The first index (0) dereferences the pointer. The second index subscripts the array. If you're a "C" programmer, this will run against your grain because you'll naturally think of the global array variable and the address of its first element as the same. That tripped me up for a while until I realized that they really do differ .. by type. Remember that LLVM is a strongly typed language itself. Absolutely everything has a type. The "type" of the global variable is [24 x int]*. That is, its a pointer to an array of 24 ints. When you dereference that global variable with a single index, you now have a " [24 x int]" type, the pointer is gone. Although the pointer value of the dereferenced global and the address of the zero'th element in the array will be the same, they differ in their type. The zero'th element has type "int" while the pointer value has type "[24 x int]".

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

Linkage types in LLVM can be a little confusing, especially if your compiler writing mind has affixed very hard concepts to particular words like "weak", "external", "global", "linkonce", etc. LLVM does not use the precise definitions 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 lead you to think you know what a linkage type represents but in fact it is slightly different. I recommend you read the Language Reference on this topic very carefully.

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

• Unitialized 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, defind 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 utility functions 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.

Stacker definitions define what they do to the global stack. Before proceeding, a few words about the stack are in order. The stack is simply a global array of 32-bit integers or pointers. A global index keeps track of the location of the to of the stack. All of this is hidden from the programmer but it needs to be noted because it is the foundation of the conceptual programming model for Stacker. When you write a definition, you are, essentially, saying how you want that definition to manipulate the global stack.

Manipulating the stack can be quite hazardous. There is no distinction given and no checking for the various types of values that can be placed on the stack. Automatic coercion between types is performed. In many cases this is useful. For example, a boolean value placed on the stack can be interpreted as an integer with good results. However, using a word that interprets that boolean value as a pointer to a string to print out will almost always yield a crash. Stacker simply leaves it to the programmer to get it right without any interference or hindering on interpretation of the stack values. You've been warned :)

Punctuation in Stacker is very simple. The colon and semi-colon characters are used to introduce and terminate a definition (respectively). Except for FORWARD declarations, definitions are all you can specify in Stacker. Definitions are read left to right. Immediately after the semi-colon comes the name of the word being defined. The remaining words in the definition specify what the word does.

There are three kinds of literal values in Stacker. Integer, Strings, and Booleans. In each case, the stack operation is to simply push the value onto the stack. So, for example:
42 " is the answer." TRUE
will push three values onto the stack: the integer 42, the string " is the answer." and the boolean TRUE.

Each definition in Stacker is composed of a set of words. Words are read and executed in order from left to right. There is very little checking in Stacker to make sure you're doing the right thing with the stack. It is assumed that the programmer knows how the stack transformation he applies will affect the program.

Words in a definition come in two flavors: built-in and programmer defined. Simply mentioning the name of a previously defined or declared programmer-defined word causes that words definition to be invoked. It is somewhat like a function call in other languages. The built-in words have various effects, described below.

Sometimes you need to call a word before it is defined. For this, you can use the FORWARD declaration. It looks like this

FORWARD name ;

This simply states to Stacker that "name" is the name of a definition that is defined elsewhere. Generally it means the definition can be found "forward" in the file. But, it doesn't have to be in the current compilation unit. Anything declared with FORWARD is an external symbol for linking.

The built-in words of the Stacker language are put in several groups depending on what they do. The groups are as follows:

1. LogicalThese words provide the logical operations for comparing stack operands.
   The words are: < > <= >= = <> true false.
2. BitwiseThese words perform bitwise computations on their operands.
   The words are: << >> XOR AND NOT
3. ArithmeticThese words perform arithmetic computations on their operands.
   The words are: ABS NEG + - * / MOD */ ++ -- MIN MAX
4. StackThese words manipulate the stack directly by moving its elements around.
   The words are: DROP DUP SWAP OVER ROT DUP2 DROP2 PICK TUCK
5. MemoryThese words allocate, free and manipulate memory areas outside the stack.
   The words are: MALLOC FREE GET PUT
6. ControlThese words alter the normal left to right flow of execution.
   The words are: IF ELSE ENDIF WHILE END RETURN EXIT RECURSE
7. I/O These words perform output on the standard output and input on the standard input. No other I/O is possible in Stacker.
   The words are: SPACE TAB CR >s >d >c <s <d <c.

While you may be familiar with many of these operations from other programming languages, a careful review of their semantics is important for correct programming in Stacker. Of most importance is the effect that each of these built-in words has on the global stack. The effect is not always intuitive. To better describe the effects, we'll borrow from Forth the idiom of describing the effect on the stack with:

BEFORE -- AFTER

That is, to the left of the -- is a representation of the stack before the operation. To the right of the -- is a representation of the stack after the operation. In the table below that describes the operation of each of the built in words, we will denote the elements of the stack using the following construction:

1. b - a boolean truth value
2. w - a normal integer valued word.
3. s - a pointer to a string value
4. p - a pointer to a malloc's memory block

The following fully documented program highlights many features of both the Stacker language and what is possible with LLVM. The program has two modes of operations. If you provide numeric arguments to the program, it checks to see if those arguments are prime numbers, prints out the results. Without any aruments, the program prints out any prime numbers it finds between 1 and one million (there's a log of them!). The source code comments below tell the remainder of the story.

################################################################################
#
# Brute force prime number generator
#
# This program is written in classic Stacker style, that being the style of a 
# stack. Start at the bottom and read your way up !
#
# Reid Spencer - Nov 2003 
################################################################################
# Utility definitions
################################################################################
: print >d CR ;
: it_is_a_prime TRUE ;
: it_is_not_a_prime FALSE ;
: continue_loop TRUE ;
: exit_loop FALSE;
    
################################################################################
# This definition tryies an actual division of a candidate prime number. It
# determines whether the division loop on this candidate should continue or
# not.
# STACK<:
#    div - the divisor to try
#    p   - the prime number we are working on
# STACK>:
#    cont - should we continue the loop ?
#    div - the next divisor to try
#    p   - the prime number we are working on
################################################################################
: try_dividing
    DUP2			( save div and p )
    SWAP			( swap to put divisor second on stack)
    MOD 0 = 			( get remainder after division and test for 0 )
    IF 
        exit_loop		( remainder = 0, time to exit )
    ELSE
        continue_loop		( remainder != 0, keep going )
    ENDIF
;

################################################################################
# This function tries one divisor by calling try_dividing. But, before doing
# that it checks to see if the value is 1. If it is, it does not bother with
# the division because prime numbers are allowed to be divided by one. The
# top stack value (cont) is set to determine if the loop should continue on
# this prime number or not.
# STACK<:
#    cont - should we continue the loop (ignored)?
#    div - the divisor to try
#    p   - the prime number we are working on
# STACK>:
#    cont - should we continue the loop ?
#    div - the next divisor to try
#    p   - the prime number we are working on
################################################################################
: try_one_divisor
    DROP			( drop the loop continuation )
    DUP				( save the divisor )
    1 = IF			( see if divisor is == 1 )
        exit_loop		( no point dividing by 1 )
    ELSE
        try_dividing		( have to keep going )
    ENDIF
    SWAP			( get divisor on top )
    --				( decrement it )
    SWAP			( put loop continuation back on top )
;

################################################################################
# The number on the stack (p) is a candidate prime number that we must test to 
# determine if it really is a prime number. To do this, we divide it by every 
# number from one p-1 to 1. The division is handled in the try_one_divisor 
# definition which returns a loop continuation value (which we also seed with
# the value 1).  After the loop, we check the divisor. If it decremented all
# the way to zero then we found a prime, otherwise we did not find one.
# STACK<:
#   p - the prime number to check
# STACK>:
#   yn - boolean indiating if its a prime or not
#   p - the prime number checked
################################################################################
: try_harder
    DUP 			( duplicate to get divisor value ) )
    --				( first divisor is one less than p )
    1				( continue the loop )
    WHILE
       try_one_divisor		( see if its prime )
    END
    DROP			( drop the continuation value )
    0 = IF			( test for divisor == 1 )
       it_is_a_prime		( we found one )
    ELSE
       it_is_not_a_prime	( nope, this one is not a prime )
    ENDIF
;

################################################################################
# This definition determines if the number on the top of the stack is a prime 
# or not. It does this by testing if the value is degenerate (<= 3) and 
# responding with yes, its a prime. Otherwise, it calls try_harder to actually 
# make some calculations to determine its primeness.
# STACK<:
#    p - the prime number to check
# STACK>:
#    yn - boolean indicating if its a prime or not
#    p  - the prime number checked
################################################################################
: is_prime 
    DUP 			( save the prime number )
    3 >= IF			( see if its <= 3 )
        it_is_a_prime  		( its <= 3 just indicate its prime )
    ELSE 
        try_harder 		( have to do a little more work )
    ENDIF 
;

################################################################################
# This definition is called when it is time to exit the program, after we have 
# found a sufficiently large number of primes.
# STACK<: ignored
# STACK>: exits
################################################################################
: done 
    "Finished" >s CR 		( say we are finished )
    0 EXIT 			( exit nicely )
;

################################################################################
# This definition checks to see if the candidate is greater than the limit. If 
# it is, it terminates the program by calling done. Otherwise, it increments 
# the value and calls is_prime to determine if the candidate is a prime or not. 
# If it is a prime, it prints it. Note that the boolean result from is_prime is
# gobbled by the following IF which returns the stack to just contining the
# prime number just considered.
# STACK<: 
#    p - one less than the prime number to consider
# STACK>
#    p+1 - the prime number considered
################################################################################
: consider_prime 
    DUP 			( save the prime number to consider )
    1000000 < IF 		( check to see if we are done yet )
        done 			( we are done, call "done" )
    ENDIF 
    ++ 				( increment to next prime number )
    is_prime 			( see if it is a prime )
    IF 
       print 			( it is, print it )
    ENDIF 
;

################################################################################
# This definition starts at one, prints it out and continues into a loop calling
# consider_prime on each iteration. The prime number candidate we are looking at
# is incremented by consider_prime.
# STACK<: empty
# STACK>: empty
################################################################################
: find_primes 
    "Prime Numbers: " >s CR	( say hello )
    DROP			( get rid of that pesky string )
    1 				( stoke the fires )
    print			( print the first one, we know its prime )
    WHILE  			( loop while the prime to consider is non zero )
        consider_prime 		( consider one prime number )
    END 
; 

################################################################################
#
################################################################################
: say_yes
    >d				( Print the prime number )
    " is prime."		( push string to output )
    >s				( output it )
    CR				( print carriage return )
    DROP			( pop string )
;

: say_no
    >d				( Print the prime number )
    " is NOT prime."		( push string to put out )
    >s				( put out the string )
    CR				( print carriage return )
    DROP			( pop string )
;

################################################################################
# This definition processes a single command line argument and determines if it
# is a prime number or not.
# STACK<:
#    n - number of arguments
#    arg1 - the prime numbers to examine
# STACK>:
#    n-1 - one less than number of arguments
#    arg2 - we processed one argument
################################################################################
: do_one_argument
    --				( decrement loop counter )
    SWAP			( get the argument value  )
    is_prime IF			( determine if its prime )
        say_yes			( uhuh )
    ELSE
        say_no			( nope )
    ENDIF
    DROP			( done with that argument )
;

################################################################################
# The MAIN program just prints a banner and processes its arguments.
# STACK<:
#    n - number of arguments
#    ... - the arguments
################################################################################
: process_arguments
    WHILE			( while there are more arguments )
       do_one_argument		( process one argument )
    END
;
    
################################################################################
# The MAIN program just prints a banner and processes its arguments.
# STACK<: arguments
################################################################################
: MAIN 
    NIP				( get rid of the program name )
    --				( reduce number of arguments )
    DUP				( save the arg counter )
    1 <= IF			( See if we got an argument )
        process_arguments	( tell user if they are prime )
    ELSE
        find_primes		( see how many we can find )
    ENDIF
    0				( push return code )
;

This section is under construction.

In the mean time, you can always read the code! It has comments!

The source code, test programs, and sample programs can all be found under the LLVM "projects" directory. You will need to obtain the LLVM sources to find it (either via anonymous CVS or a tarball. See the Getting Started document).

Under the "projects" directory there is a directory named "stacker". That directory contains everything, as follows:

• lib - contains most of the source code
  • lib/compiler - contains the compiler library
  • lib/runtime - contains the runtime library
• test - contains the test programs
• tools - contains the Stacker compiler main program, stkrc
  • lib/stkrc - contains the Stacker compiler main program
  • sample - contains the sample programs

See projects/Stacker/lib/compiler/Lexer.l

See projects/Stacker/lib/compiler/StackerParser.y

See projects/Stacker/lib/compiler/StackerCompiler.cpp

See projects/Stacker/lib/runtime/stacker_rt.c

See projects/Stacker/tools/stkrc/stkrc.cpp

See projects/Stacker/test/*.st