brintos

brintos / llvm-project-archived public Read only

0
0
Text · 30.9 KiB · ed1dea2 Raw
887 lines · plain
1=======================================================2Kaleidoscope: Extending the Language: Mutable Variables3=======================================================4 5.. contents::6   :local:7 8Chapter 7 Introduction9======================10 11Welcome to Chapter 7 of the "`Implementing a language with12LLVM <index.html>`_" tutorial. In chapters 1 through 6, we've built a13very respectable, albeit simple, `functional programming14language <http://en.wikipedia.org/wiki/Functional_programming>`_. In our15journey, we learned some parsing techniques, how to build and represent16an AST, how to build LLVM IR, and how to optimize the resultant code as17well as JIT compile it.18 19While Kaleidoscope is interesting as a functional language, the fact20that it is functional makes it "too easy" to generate LLVM IR for it. In21particular, a functional language makes it very easy to build LLVM IR22directly in `SSA23form <http://en.wikipedia.org/wiki/Static_single_assignment_form>`_.24Since LLVM requires that the input code be in SSA form, this is a very25nice property and it is often unclear to newcomers how to generate code26for an imperative language with mutable variables.27 28The short (and happy) summary of this chapter is that there is no need29for your front-end to build SSA form: LLVM provides highly tuned and30well tested support for this, though the way it works is a bit31unexpected for some.32 33Why is this a hard problem?34===========================35 36To understand why mutable variables cause complexities in SSA37construction, consider this extremely simple C example:38 39.. code-block:: c40 41    int G, H;42    int test(_Bool Condition) {43      int X;44      if (Condition)45        X = G;46      else47        X = H;48      return X;49    }50 51In this case, we have the variable "X", whose value depends on the path52executed in the program. Because there are two different possible values53for X before the return instruction, a PHI node is inserted to merge the54two values. The LLVM IR that we want for this example looks like this:55 56.. code-block:: llvm57 58    @G = weak global i32 0   ; type of @G is i32*59    @H = weak global i32 0   ; type of @H is i32*60 61    define i32 @test(i1 %Condition) {62    entry:63      br i1 %Condition, label %cond_true, label %cond_false64 65    cond_true:66      %X.0 = load i32, i32* @G67      br label %cond_next68 69    cond_false:70      %X.1 = load i32, i32* @H71      br label %cond_next72 73    cond_next:74      %X.2 = phi i32 [ %X.1, %cond_false ], [ %X.0, %cond_true ]75      ret i32 %X.276    }77 78In this example, the loads from the G and H global variables are79explicit in the LLVM IR, and they live in the then/else branches of the80if statement (cond\_true/cond\_false). In order to merge the incoming81values, the X.2 phi node in the cond\_next block selects the right value82to use based on where control flow is coming from: if control flow comes83from the cond\_false block, X.2 gets the value of X.1. Alternatively, if84control flow comes from cond\_true, it gets the value of X.0. The intent85of this chapter is not to explain the details of SSA form. For more86information, see one of the many `online87references <http://en.wikipedia.org/wiki/Static_single_assignment_form>`_.88 89The question for this article is "who places the phi nodes when lowering90assignments to mutable variables?". The issue here is that LLVM91*requires* that its IR be in SSA form: there is no "non-ssa" mode for92it. However, SSA construction requires non-trivial algorithms and data93structures, so it is inconvenient and wasteful for every front-end to94have to reproduce this logic.95 96Memory in LLVM97==============98 99The 'trick' here is that while LLVM does require all register values to100be in SSA form, it does not require (or permit) memory objects to be in101SSA form. In the example above, note that the loads from G and H are102direct accesses to G and H: they are not renamed or versioned. This103differs from some other compiler systems, which do try to version memory104objects. In LLVM, instead of encoding dataflow analysis of memory into105the LLVM IR, it is handled with `Analysis106Passes <../../WritingAnLLVMPass.html>`_ which are computed on demand.107 108With this in mind, the high-level idea is that we want to make a stack109variable (which lives in memory, because it is on the stack) for each110mutable object in a function. To take advantage of this trick, we need111to talk about how LLVM represents stack variables.112 113In LLVM, all memory accesses are explicit with load/store instructions,114and it is carefully designed not to have (or need) an "address-of"115operator. Notice how the type of the @G/@H global variables is actually116"i32\*" even though the variable is defined as "i32". What this means is117that @G defines *space* for an i32 in the global data area, but its118*name* actually refers to the address for that space. Stack variables119work the same way, except that instead of being declared with global120variable definitions, they are declared with the `LLVM alloca121instruction <../../LangRef.html#alloca-instruction>`_:122 123.. code-block:: llvm124 125    define i32 @example() {126    entry:127      %X = alloca i32           ; type of %X is i32*.128      ...129      %tmp = load i32, i32* %X  ; load the stack value %X from the stack.130      %tmp2 = add i32 %tmp, 1   ; increment it131      store i32 %tmp2, i32* %X  ; store it back132      ...133 134This code shows an example of how you can declare and manipulate a stack135variable in the LLVM IR. Stack memory allocated with the alloca136instruction is fully general: you can pass the address of the stack slot137to functions, you can store it in other variables, etc. In our example138above, we could rewrite the example to use the alloca technique to avoid139using a PHI node:140 141.. code-block:: llvm142 143    @G = weak global i32 0   ; type of @G is i32*144    @H = weak global i32 0   ; type of @H is i32*145 146    define i32 @test(i1 %Condition) {147    entry:148      %X = alloca i32           ; type of %X is i32*.149      br i1 %Condition, label %cond_true, label %cond_false150 151    cond_true:152      %X.0 = load i32, i32* @G153      store i32 %X.0, i32* %X   ; Update X154      br label %cond_next155 156    cond_false:157      %X.1 = load i32, i32* @H158      store i32 %X.1, i32* %X   ; Update X159      br label %cond_next160 161    cond_next:162      %X.2 = load i32, i32* %X  ; Read X163      ret i32 %X.2164    }165 166With this, we have discovered a way to handle arbitrary mutable167variables without the need to create Phi nodes at all:168 169#. Each mutable variable becomes a stack allocation.170#. Each read of the variable becomes a load from the stack.171#. Each update of the variable becomes a store to the stack.172#. Taking the address of a variable just uses the stack address173   directly.174 175While this solution has solved our immediate problem, it introduced176another one: we have now apparently introduced a lot of stack traffic177for very simple and common operations, a major performance problem.178Fortunately for us, the LLVM optimizer has a highly-tuned optimization179pass named "mem2reg" that handles this case, promoting allocas like this180into SSA registers, inserting Phi nodes as appropriate. If you run this181example through the pass, for example, you'll get:182 183.. code-block:: bash184 185    $ llvm-as < example.ll | opt -passes=mem2reg | llvm-dis186    @G = weak global i32 0187    @H = weak global i32 0188 189    define i32 @test(i1 %Condition) {190    entry:191      br i1 %Condition, label %cond_true, label %cond_false192 193    cond_true:194      %X.0 = load i32, i32* @G195      br label %cond_next196 197    cond_false:198      %X.1 = load i32, i32* @H199      br label %cond_next200 201    cond_next:202      %X.01 = phi i32 [ %X.1, %cond_false ], [ %X.0, %cond_true ]203      ret i32 %X.01204    }205 206The mem2reg pass implements the standard "iterated dominance frontier"207algorithm for constructing SSA form and has a number of optimizations208that speed up (very common) degenerate cases. The mem2reg optimization209pass is the answer to dealing with mutable variables, and we highly210recommend that you depend on it. Note that mem2reg only works on211variables in certain circumstances:212 213#. mem2reg is alloca-driven: it looks for allocas and if it can handle214   them, it promotes them. It does not apply to global variables or heap215   allocations.216#. mem2reg only looks for alloca instructions in the entry block of the217   function. Being in the entry block guarantees that the alloca is only218   executed once, which makes analysis simpler.219#. mem2reg only promotes allocas whose uses are direct loads and stores.220   If the address of the stack object is passed to a function, or if any221   funny pointer arithmetic is involved, the alloca will not be222   promoted.223#. mem2reg only works on allocas of `first224   class <../../LangRef.html#first-class-types>`_ values (such as pointers,225   scalars and vectors), and only if the array size of the allocation is226   1 (or missing in the .ll file). mem2reg is not capable of promoting227   structs or arrays to registers. Note that the "sroa" pass is228   more powerful and can promote structs, "unions", and arrays in many229   cases.230 231All of these properties are easy to satisfy for most imperative232languages, and we'll illustrate it below with Kaleidoscope. The final233question you may be asking is: should I bother with this nonsense for my234front-end? Wouldn't it be better if I just did SSA construction235directly, avoiding use of the mem2reg optimization pass? In short, we236strongly recommend that you use this technique for building SSA form,237unless there is an extremely good reason not to. Using this technique238is:239 240-  Proven and well tested: clang uses this technique241   for local mutable variables. As such, the most common clients of LLVM242   are using this to handle a bulk of their variables. You can be sure243   that bugs are found fast and fixed early.244-  Extremely Fast: mem2reg has a number of special cases that make it245   fast in common cases as well as fully general. For example, it has246   fast-paths for variables that are only used in a single block,247   variables that only have one assignment point, good heuristics to248   avoid insertion of unneeded phi nodes, etc.249-  Needed for debug info generation: `Debug information in250   LLVM <../../SourceLevelDebugging.html>`_ relies on having the address of251   the variable exposed so that debug info can be attached to it. This252   technique dovetails very naturally with this style of debug info.253 254If nothing else, this makes it much easier to get your front-end up and255running, and is very simple to implement. Let's extend Kaleidoscope with256mutable variables now!257 258Mutable Variables in Kaleidoscope259=================================260 261Now that we know the sort of problem we want to tackle, let's see what262this looks like in the context of our little Kaleidoscope language.263We're going to add two features:264 265#. The ability to mutate variables with the '=' operator.266#. The ability to define new variables.267 268While the first item is really what this is about, we only have269variables for incoming arguments as well as for induction variables, and270redefining those only goes so far :). Also, the ability to define new271variables is a useful thing regardless of whether you will be mutating272them. Here's a motivating example that shows how we could use these:273 274::275 276    # Define ':' for sequencing: as a low-precedence operator that ignores operands277    # and just returns the RHS.278    def binary : 1 (x y) y;279 280    # Recursive fib, we could do this before.281    def fib(x)282      if (x < 3) then283        1284      else285        fib(x-1)+fib(x-2);286 287    # Iterative fib.288    def fibi(x)289      var a = 1, b = 1, c in290      (for i = 3, i < x in291         c = a + b :292         a = b :293         b = c) :294      b;295 296    # Call it.297    fibi(10);298 299In order to mutate variables, we have to change our existing variables300to use the "alloca trick". Once we have that, we'll add our new301operator, then extend Kaleidoscope to support new variable definitions.302 303Adjusting Existing Variables for Mutation304=========================================305 306The symbol table in Kaleidoscope is managed at code generation time by307the '``NamedValues``' map. This map currently keeps track of the LLVM308"Value\*" that holds the double value for the named variable. In order309to support mutation, we need to change this slightly, so that310``NamedValues`` holds the *memory location* of the variable in question.311Note that this change is a refactoring: it changes the structure of the312code, but does not (by itself) change the behavior of the compiler. All313of these changes are isolated in the Kaleidoscope code generator.314 315At this point in Kaleidoscope's development, it only supports variables316for two things: incoming arguments to functions and the induction317variable of 'for' loops. For consistency, we'll allow mutation of these318variables in addition to other user-defined variables. This means that319these will both need memory locations.320 321To start our transformation of Kaleidoscope, we'll change the322``NamedValues`` map so that it maps to AllocaInst\* instead of Value\*. Once323we do this, the C++ compiler will tell us what parts of the code we need324to update:325 326.. code-block:: c++327 328    static std::map<std::string, AllocaInst*> NamedValues;329 330Also, since we will need to create these allocas, we'll use a helper331function that ensures that the allocas are created in the entry block of332the function:333 334.. code-block:: c++335 336    /// CreateEntryBlockAlloca - Create an alloca instruction in the entry block of337    /// the function.  This is used for mutable variables etc.338    static AllocaInst *CreateEntryBlockAlloca(Function *TheFunction,339                                              StringRef VarName) {340      IRBuilder<> TmpB(&TheFunction->getEntryBlock(),341                     TheFunction->getEntryBlock().begin());342      return TmpB.CreateAlloca(Type::getDoubleTy(*TheContext), nullptr,343                               VarName);344    }345 346This funny looking code creates an IRBuilder object that is pointing at347the first instruction (.begin()) of the entry block. It then creates an348alloca with the expected name and returns it. Because all values in349Kaleidoscope are doubles, there is no need to pass in a type to use.350 351With this in place, the first functionality change we want to make belongs to352variable references. In our new scheme, variables live on the stack, so353code generating a reference to them actually needs to produce a load354from the stack slot:355 356.. code-block:: c++357 358    Value *VariableExprAST::codegen() {359      // Look this variable up in the function.360      AllocaInst *A = NamedValues[Name];361      if (!A)362        return LogErrorV("Unknown variable name");363 364      // Load the value.365      return Builder->CreateLoad(A->getAllocatedType(), A, Name.c_str());366    }367 368As you can see, this is pretty straightforward. Now we need to update369the things that define the variables to set up the alloca. We'll start370with ``ForExprAST::codegen()`` (see the `full code listing <#id1>`_ for371the unabridged code):372 373.. code-block:: c++374 375      Function *TheFunction = Builder->GetInsertBlock()->getParent();376 377      // Create an alloca for the variable in the entry block.378      AllocaInst *Alloca = CreateEntryBlockAlloca(TheFunction, VarName);379 380      // Emit the start code first, without 'variable' in scope.381      Value *StartVal = Start->codegen();382      if (!StartVal)383        return nullptr;384 385      // Store the value into the alloca.386      Builder->CreateStore(StartVal, Alloca);387      ...388 389      // Compute the end condition.390      Value *EndCond = End->codegen();391      if (!EndCond)392        return nullptr;393 394      // Reload, increment, and restore the alloca.  This handles the case where395      // the body of the loop mutates the variable.396      Value *CurVar = Builder->CreateLoad(Alloca->getAllocatedType(), Alloca,397                                          VarName.c_str());398      Value *NextVar = Builder->CreateFAdd(CurVar, StepVal, "nextvar");399      Builder->CreateStore(NextVar, Alloca);400      ...401 402This code is virtually identical to the code `before we allowed mutable403variables <LangImpl05.html#code-generation-for-the-for-loop>`_. The big difference is that we404no longer have to construct a PHI node, and we use load/store to access405the variable as needed.406 407To support mutable argument variables, we need to also make allocas for408them. The code for this is also pretty simple:409 410.. code-block:: c++411 412    Function *FunctionAST::codegen() {413      ...414      Builder->SetInsertPoint(BB);415 416      // Record the function arguments in the NamedValues map.417      NamedValues.clear();418      for (auto &Arg : TheFunction->args()) {419        // Create an alloca for this variable.420        AllocaInst *Alloca = CreateEntryBlockAlloca(TheFunction, Arg.getName());421 422        // Store the initial value into the alloca.423        Builder->CreateStore(&Arg, Alloca);424 425        // Add arguments to variable symbol table.426        NamedValues[std::string(Arg.getName())] = Alloca;427      }428 429      if (Value *RetVal = Body->codegen()) {430        ...431 432For each argument, we make an alloca, store the input value to the433function into the alloca, and register the alloca as the memory location434for the argument. This method gets invoked by ``FunctionAST::codegen()``435right after it sets up the entry block for the function.436 437The final missing piece is adding the mem2reg pass, which allows us to438get good codegen once again:439 440.. code-block:: c++441 442        // Promote allocas to registers.443        TheFPM->addPass(PromotePass());444        // Do simple "peephole" optimizations and bit-twiddling optzns.445        TheFPM->addPass(InstCombinePass());446        // Reassociate expressions.447        TheFPM->addPass(ReassociatePass());448        ...449 450It is interesting to see what the code looks like before and after the451mem2reg optimization runs. For example, this is the before/after code452for our recursive fib function. Before the optimization:453 454.. code-block:: llvm455 456    define double @fib(double %x) {457    entry:458      %x1 = alloca double459      store double %x, double* %x1460      %x2 = load double, double* %x1461      %cmptmp = fcmp ult double %x2, 3.000000e+00462      %booltmp = uitofp i1 %cmptmp to double463      %ifcond = fcmp one double %booltmp, 0.000000e+00464      br i1 %ifcond, label %then, label %else465 466    then:       ; preds = %entry467      br label %ifcont468 469    else:       ; preds = %entry470      %x3 = load double, double* %x1471      %subtmp = fsub double %x3, 1.000000e+00472      %calltmp = call double @fib(double %subtmp)473      %x4 = load double, double* %x1474      %subtmp5 = fsub double %x4, 2.000000e+00475      %calltmp6 = call double @fib(double %subtmp5)476      %addtmp = fadd double %calltmp, %calltmp6477      br label %ifcont478 479    ifcont:     ; preds = %else, %then480      %iftmp = phi double [ 1.000000e+00, %then ], [ %addtmp, %else ]481      ret double %iftmp482    }483 484Here there is only one variable (x, the input argument) but you can485still see the extremely simple-minded code generation strategy we are486using. In the entry block, an alloca is created, and the initial input487value is stored into it. Each reference to the variable does a reload488from the stack. Also, note that we didn't modify the if/then/else489expression, so it still inserts a PHI node. While we could make an490alloca for it, it is actually easier to create a PHI node for it, so we491still just make the PHI.492 493Here is the code after the mem2reg pass runs:494 495.. code-block:: llvm496 497    define double @fib(double %x) {498    entry:499      %cmptmp = fcmp ult double %x, 3.000000e+00500      %booltmp = uitofp i1 %cmptmp to double501      %ifcond = fcmp one double %booltmp, 0.000000e+00502      br i1 %ifcond, label %then, label %else503 504    then:505      br label %ifcont506 507    else:508      %subtmp = fsub double %x, 1.000000e+00509      %calltmp = call double @fib(double %subtmp)510      %subtmp5 = fsub double %x, 2.000000e+00511      %calltmp6 = call double @fib(double %subtmp5)512      %addtmp = fadd double %calltmp, %calltmp6513      br label %ifcont514 515    ifcont:     ; preds = %else, %then516      %iftmp = phi double [ 1.000000e+00, %then ], [ %addtmp, %else ]517      ret double %iftmp518    }519 520This is a trivial case for mem2reg, since there are no redefinitions of521the variable. The point of showing this is to calm your tension about522inserting such blatant inefficiencies :).523 524After the rest of the optimizers run, we get:525 526.. code-block:: llvm527 528    define double @fib(double %x) {529    entry:530      %cmptmp = fcmp ult double %x, 3.000000e+00531      %booltmp = uitofp i1 %cmptmp to double532      %ifcond = fcmp ueq double %booltmp, 0.000000e+00533      br i1 %ifcond, label %else, label %ifcont534 535    else:536      %subtmp = fsub double %x, 1.000000e+00537      %calltmp = call double @fib(double %subtmp)538      %subtmp5 = fsub double %x, 2.000000e+00539      %calltmp6 = call double @fib(double %subtmp5)540      %addtmp = fadd double %calltmp, %calltmp6541      ret double %addtmp542 543    ifcont:544      ret double 1.000000e+00545    }546 547Here we see that the simplifycfg pass decided to clone the return548instruction into the end of the 'else' block. This allowed it to549eliminate some branches and the PHI node.550 551Now that all symbol table references are updated to use stack variables,552we'll add the assignment operator.553 554New Assignment Operator555=======================556 557With our current framework, adding a new assignment operator is really558simple. We will parse it just like any other binary operator, but handle559it internally (instead of allowing the user to define it). The first560step is to set a precedence:561 562.. code-block:: c++563 564     int main() {565       // Install standard binary operators.566       // 1 is lowest precedence.567       BinopPrecedence['='] = 2;568       BinopPrecedence['<'] = 10;569       BinopPrecedence['+'] = 20;570       BinopPrecedence['-'] = 20;571 572Now that the parser knows the precedence of the binary operator, it573takes care of all the parsing and AST generation. We just need to574implement codegen for the assignment operator. This looks like:575 576.. code-block:: c++577 578    Value *BinaryExprAST::codegen() {579      // Special case '=' because we don't want to emit the LHS as an expression.580      if (Op == '=') {581        // This assume we're building without RTTI because LLVM builds that way by582        // default. If you build LLVM with RTTI this can be changed to a583        // dynamic_cast for automatic error checking.584        VariableExprAST *LHSE = static_cast<VariableExprAST*>(LHS.get());585        if (!LHSE)586          return LogErrorV("destination of '=' must be a variable");587 588Unlike the rest of the binary operators, our assignment operator doesn't589follow the "emit LHS, emit RHS, do computation" model. As such, it is590handled as a special case before the other binary operators are handled.591The other strange thing is that it requires the LHS to be a variable. It592is invalid to have "(x+1) = expr" - only things like "x = expr" are593allowed.594 595.. code-block:: c++596 597        // Codegen the RHS.598        Value *Val = RHS->codegen();599        if (!Val)600          return nullptr;601 602        // Look up the name.603        Value *Variable = NamedValues[LHSE->getName()];604        if (!Variable)605          return LogErrorV("Unknown variable name");606 607        Builder->CreateStore(Val, Variable);608        return Val;609      }610      ...611 612Once we have the variable, codegen'ing the assignment is613straightforward: we emit the RHS of the assignment, create a store, and614return the computed value. Returning a value allows for chained615assignments like "X = (Y = Z)".616 617Now that we have an assignment operator, we can mutate loop variables618and arguments. For example, we can now run code like this:619 620::621 622    # Function to print a double.623    extern printd(x);624 625    # Define ':' for sequencing: as a low-precedence operator that ignores operands626    # and just returns the RHS.627    def binary : 1 (x y) y;628 629    def test(x)630      printd(x) :631      x = 4 :632      printd(x);633 634    test(123);635 636When run, this example prints "123" and then "4", showing that we did637actually mutate the value! Okay, we have now officially implemented our638goal: getting this to work requires SSA construction in the general639case. However, to be really useful, we want the ability to define our640own local variables, let's add this next!641 642User-defined Local Variables643============================644 645Adding var/in is just like any other extension we made to646Kaleidoscope: we extend the lexer, the parser, the AST and the code647generator. The first step for adding our new 'var/in' construct is to648extend the lexer. As before, this is pretty trivial, the code looks like649this:650 651.. code-block:: c++652 653    enum Token {654      ...655      // var definition656      tok_var = -13657    ...658    }659    ...660    static int gettok() {661    ...662        if (IdentifierStr == "in")663          return tok_in;664        if (IdentifierStr == "binary")665          return tok_binary;666        if (IdentifierStr == "unary")667          return tok_unary;668        if (IdentifierStr == "var")669          return tok_var;670        return tok_identifier;671    ...672 673The next step is to define the AST node that we will construct. For674var/in, it looks like this:675 676.. code-block:: c++677 678    /// VarExprAST - Expression class for var/in679    class VarExprAST : public ExprAST {680      std::vector<std::pair<std::string, std::unique_ptr<ExprAST>>> VarNames;681      std::unique_ptr<ExprAST> Body;682 683    public:684      VarExprAST(std::vector<std::pair<std::string, std::unique_ptr<ExprAST>>> VarNames,685                 std::unique_ptr<ExprAST> Body)686        : VarNames(std::move(VarNames)), Body(std::move(Body)) {}687 688      Value *codegen() override;689    };690 691var/in allows a list of names to be defined all at once, and each name692can optionally have an initializer value. As such, we capture this693information in the VarNames vector. Also, var/in has a body, this body694is allowed to access the variables defined by the var/in.695 696With this in place, we can define the parser pieces. The first thing we697do is add it as a primary expression:698 699.. code-block:: c++700 701    /// primary702    ///   ::= identifierexpr703    ///   ::= numberexpr704    ///   ::= parenexpr705    ///   ::= ifexpr706    ///   ::= forexpr707    ///   ::= varexpr708    static std::unique_ptr<ExprAST> ParsePrimary() {709      switch (CurTok) {710      default:711        return LogError("unknown token when expecting an expression");712      case tok_identifier:713        return ParseIdentifierExpr();714      case tok_number:715        return ParseNumberExpr();716      case '(':717        return ParseParenExpr();718      case tok_if:719        return ParseIfExpr();720      case tok_for:721        return ParseForExpr();722      case tok_var:723        return ParseVarExpr();724      }725    }726 727Next we define ParseVarExpr:728 729.. code-block:: c++730 731    /// varexpr ::= 'var' identifier ('=' expression)?732    //                    (',' identifier ('=' expression)?)* 'in' expression733    static std::unique_ptr<ExprAST> ParseVarExpr() {734      getNextToken();  // eat the var.735 736      std::vector<std::pair<std::string, std::unique_ptr<ExprAST>>> VarNames;737 738      // At least one variable name is required.739      if (CurTok != tok_identifier)740        return LogError("expected identifier after var");741 742The first part of this code parses the list of identifier/expr pairs743into the local ``VarNames`` vector.744 745.. code-block:: c++746 747      while (true) {748        std::string Name = IdentifierStr;749        getNextToken();  // eat identifier.750 751        // Read the optional initializer.752        std::unique_ptr<ExprAST> Init;753        if (CurTok == '=') {754          getNextToken(); // eat the '='.755 756          Init = ParseExpression();757          if (!Init) return nullptr;758        }759 760        VarNames.push_back(std::make_pair(Name, std::move(Init)));761 762        // End of var list, exit loop.763        if (CurTok != ',') break;764        getNextToken(); // eat the ','.765 766        if (CurTok != tok_identifier)767          return LogError("expected identifier list after var");768      }769 770Once all the variables are parsed, we then parse the body and create the771AST node:772 773.. code-block:: c++774 775      // At this point, we have to have 'in'.776      if (CurTok != tok_in)777        return LogError("expected 'in' keyword after 'var'");778      getNextToken();  // eat 'in'.779 780      auto Body = ParseExpression();781      if (!Body)782        return nullptr;783 784      return std::make_unique<VarExprAST>(std::move(VarNames),785                                           std::move(Body));786    }787 788Now that we can parse and represent the code, we need to support789emission of LLVM IR for it. This code starts out with:790 791.. code-block:: c++792 793    Value *VarExprAST::codegen() {794      std::vector<AllocaInst *> OldBindings;795 796      Function *TheFunction = Builder->GetInsertBlock()->getParent();797 798      // Register all variables and emit their initializer.799      for (unsigned i = 0, e = VarNames.size(); i != e; ++i) {800        const std::string &VarName = VarNames[i].first;801        ExprAST *Init = VarNames[i].second.get();802 803Basically it loops over all the variables, installing them one at a804time. For each variable we put into the symbol table, we remember the805previous value that we replace in OldBindings.806 807.. code-block:: c++808 809        // Emit the initializer before adding the variable to scope, this prevents810        // the initializer from referencing the variable itself, and permits stuff811        // like this:812        //  var a = 1 in813        //    var a = a in ...   # refers to outer 'a'.814        Value *InitVal;815        if (Init) {816          InitVal = Init->codegen();817          if (!InitVal)818            return nullptr;819        } else { // If not specified, use 0.0.820          InitVal = ConstantFP::get(*TheContext, APFloat(0.0));821        }822 823        AllocaInst *Alloca = CreateEntryBlockAlloca(TheFunction, VarName);824        Builder->CreateStore(InitVal, Alloca);825 826        // Remember the old variable binding so that we can restore the binding when827        // we unrecurse.828        OldBindings.push_back(NamedValues[VarName]);829 830        // Remember this binding.831        NamedValues[VarName] = Alloca;832      }833 834There are more comments here than code. The basic idea is that we emit835the initializer, create the alloca, then update the symbol table to836point to it. Once all the variables are installed in the symbol table,837we evaluate the body of the var/in expression:838 839.. code-block:: c++840 841      // Codegen the body, now that all vars are in scope.842      Value *BodyVal = Body->codegen();843      if (!BodyVal)844        return nullptr;845 846Finally, before returning, we restore the previous variable bindings:847 848.. code-block:: c++849 850      // Pop all our variables from scope.851      for (unsigned i = 0, e = VarNames.size(); i != e; ++i)852        NamedValues[VarNames[i].first] = OldBindings[i];853 854      // Return the body computation.855      return BodyVal;856    }857 858The end result of all of this is that we get properly scoped variable859definitions, and we even (trivially) allow mutation of them :).860 861With this, we completed what we set out to do. Our nice iterative fib862example from the intro compiles and runs just fine. The mem2reg pass863optimizes all of our stack variables into SSA registers, inserting PHI864nodes where needed, and our front-end remains simple: no "iterated865dominance frontier" computation anywhere in sight.866 867Full Code Listing868=================869 870Here is the complete code listing for our running example, enhanced with871mutable variables and var/in support. To build this example, use:872 873.. code-block:: bash874 875    # Compile876    clang++ -g toy.cpp `llvm-config --cxxflags --ldflags --system-libs --libs core orcjit native` -O3 -o toy877    # Run878    ./toy879 880Here is the code:881 882.. literalinclude:: ../../../examples/Kaleidoscope/Chapter7/toy.cpp883   :language: c++884 885`Next: Compiling to Object Code <LangImpl08.html>`_886 887