• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1=======================================================
2Kaleidoscope: Extending the Language: Mutable Variables
3=======================================================
4
5.. contents::
6   :local:
7
8Chapter 7 Introduction
9======================
10
11Welcome to Chapter 7 of the "`Implementing a language with
12LLVM <index.html>`_" tutorial. In chapters 1 through 6, we've built a
13very respectable, albeit simple, `functional programming
14language <http://en.wikipedia.org/wiki/Functional_programming>`_. In our
15journey, we learned some parsing techniques, how to build and represent
16an AST, how to build LLVM IR, and how to optimize the resultant code as
17well as JIT compile it.
18
19While Kaleidoscope is interesting as a functional language, the fact
20that it is functional makes it "too easy" to generate LLVM IR for it. In
21particular, a functional language makes it very easy to build LLVM IR
22directly in `SSA
23form <http://en.wikipedia.org/wiki/Static_single_assignment_form>`_.
24Since LLVM requires that the input code be in SSA form, this is a very
25nice property and it is often unclear to newcomers how to generate code
26for an imperative language with mutable variables.
27
28The short (and happy) summary of this chapter is that there is no need
29for your front-end to build SSA form: LLVM provides highly tuned and
30well tested support for this, though the way it works is a bit
31unexpected for some.
32
33Why is this a hard problem?
34===========================
35
36To understand why mutable variables cause complexities in SSA
37construction, consider this extremely simple C example:
38
39.. code-block:: c
40
41    int G, H;
42    int test(_Bool Condition) {
43      int X;
44      if (Condition)
45        X = G;
46      else
47        X = H;
48      return X;
49    }
50
51In this case, we have the variable "X", whose value depends on the path
52executed in the program. Because there are two different possible values
53for X before the return instruction, a PHI node is inserted to merge the
54two values. The LLVM IR that we want for this example looks like this:
55
56.. code-block:: llvm
57
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_false
64
65    cond_true:
66      %X.0 = load i32* @G
67      br label %cond_next
68
69    cond_false:
70      %X.1 = load i32* @H
71      br label %cond_next
72
73    cond_next:
74      %X.2 = phi i32 [ %X.1, %cond_false ], [ %X.0, %cond_true ]
75      ret i32 %X.2
76    }
77
78In this example, the loads from the G and H global variables are
79explicit in the LLVM IR, and they live in the then/else branches of the
80if statement (cond\_true/cond\_false). In order to merge the incoming
81values, the X.2 phi node in the cond\_next block selects the right value
82to use based on where control flow is coming from: if control flow comes
83from the cond\_false block, X.2 gets the value of X.1. Alternatively, if
84control flow comes from cond\_true, it gets the value of X.0. The intent
85of this chapter is not to explain the details of SSA form. For more
86information, see one of the many `online
87references <http://en.wikipedia.org/wiki/Static_single_assignment_form>`_.
88
89The question for this article is "who places the phi nodes when lowering
90assignments to mutable variables?". The issue here is that LLVM
91*requires* that its IR be in SSA form: there is no "non-ssa" mode for
92it. However, SSA construction requires non-trivial algorithms and data
93structures, so it is inconvenient and wasteful for every front-end to
94have to reproduce this logic.
95
96Memory in LLVM
97==============
98
99The 'trick' here is that while LLVM does require all register values to
100be in SSA form, it does not require (or permit) memory objects to be in
101SSA form. In the example above, note that the loads from G and H are
102direct accesses to G and H: they are not renamed or versioned. This
103differs from some other compiler systems, which do try to version memory
104objects. In LLVM, instead of encoding dataflow analysis of memory into
105the LLVM IR, it is handled with `Analysis
106Passes <../WritingAnLLVMPass.html>`_ which are computed on demand.
107
108With this in mind, the high-level idea is that we want to make a stack
109variable (which lives in memory, because it is on the stack) for each
110mutable object in a function. To take advantage of this trick, we need
111to 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 actually
116"i32\*" even though the variable is defined as "i32". What this means is
117that @G defines *space* for an i32 in the global data area, but its
118*name* actually refers to the address for that space. Stack variables
119work the same way, except that instead of being declared with global
120variable definitions, they are declared with the `LLVM alloca
121instruction <../LangRef.html#alloca-instruction>`_:
122
123.. code-block:: llvm
124
125    define i32 @example() {
126    entry:
127      %X = alloca i32           ; type of %X is i32*.
128      ...
129      %tmp = load i32* %X       ; load the stack value %X from the stack.
130      %tmp2 = add i32 %tmp, 1   ; increment it
131      store i32 %tmp2, i32* %X  ; store it back
132      ...
133
134This code shows an example of how you can declare and manipulate a stack
135variable in the LLVM IR. Stack memory allocated with the alloca
136instruction is fully general: you can pass the address of the stack slot
137to functions, you can store it in other variables, etc. In our example
138above, we could rewrite the example to use the alloca technique to avoid
139using a PHI node:
140
141.. code-block:: llvm
142
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_false
150
151    cond_true:
152      %X.0 = load i32* @G
153      store i32 %X.0, i32* %X   ; Update X
154      br label %cond_next
155
156    cond_false:
157      %X.1 = load i32* @H
158      store i32 %X.1, i32* %X   ; Update X
159      br label %cond_next
160
161    cond_next:
162      %X.2 = load i32* %X       ; Read X
163      ret i32 %X.2
164    }
165
166With this, we have discovered a way to handle arbitrary mutable
167variables 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 address
173   directly.
174
175While this solution has solved our immediate problem, it introduced
176another one: we have now apparently introduced a lot of stack traffic
177for very simple and common operations, a major performance problem.
178Fortunately for us, the LLVM optimizer has a highly-tuned optimization
179pass named "mem2reg" that handles this case, promoting allocas like this
180into SSA registers, inserting Phi nodes as appropriate. If you run this
181example through the pass, for example, you'll get:
182
183.. code-block:: bash
184
185    $ llvm-as < example.ll | opt -mem2reg | llvm-dis
186    @G = weak global i32 0
187    @H = weak global i32 0
188
189    define i32 @test(i1 %Condition) {
190    entry:
191      br i1 %Condition, label %cond_true, label %cond_false
192
193    cond_true:
194      %X.0 = load i32* @G
195      br label %cond_next
196
197    cond_false:
198      %X.1 = load i32* @H
199      br label %cond_next
200
201    cond_next:
202      %X.01 = phi i32 [ %X.1, %cond_false ], [ %X.0, %cond_true ]
203      ret i32 %X.01
204    }
205
206The mem2reg pass implements the standard "iterated dominance frontier"
207algorithm for constructing SSA form and has a number of optimizations
208that speed up (very common) degenerate cases. The mem2reg optimization
209pass is the answer to dealing with mutable variables, and we highly
210recommend that you depend on it. Note that mem2reg only works on
211variables in certain circumstances:
212
213#. mem2reg is alloca-driven: it looks for allocas and if it can handle
214   them, it promotes them. It does not apply to global variables or heap
215   allocations.
216#. mem2reg only looks for alloca instructions in the entry block of the
217   function. Being in the entry block guarantees that the alloca is only
218   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 any
221   funny pointer arithmetic is involved, the alloca will not be
222   promoted.
223#. mem2reg only works on allocas of `first
224   class <../LangRef.html#first-class-types>`_ values (such as pointers,
225   scalars and vectors), and only if the array size of the allocation is
226   1 (or missing in the .ll file). mem2reg is not capable of promoting
227   structs or arrays to registers. Note that the "sroa" pass is
228   more powerful and can promote structs, "unions", and arrays in many
229   cases.
230
231All of these properties are easy to satisfy for most imperative
232languages, and we'll illustrate it below with Kaleidoscope. The final
233question you may be asking is: should I bother with this nonsense for my
234front-end? Wouldn't it be better if I just did SSA construction
235directly, avoiding use of the mem2reg optimization pass? In short, we
236strongly recommend that you use this technique for building SSA form,
237unless there is an extremely good reason not to. Using this technique
238is:
239
240-  Proven and well tested: clang uses this technique
241   for local mutable variables. As such, the most common clients of LLVM
242   are using this to handle a bulk of their variables. You can be sure
243   that bugs are found fast and fixed early.
244-  Extremely Fast: mem2reg has a number of special cases that make it
245   fast in common cases as well as fully general. For example, it has
246   fast-paths for variables that are only used in a single block,
247   variables that only have one assignment point, good heuristics to
248   avoid insertion of unneeded phi nodes, etc.
249-  Needed for debug info generation: `Debug information in
250   LLVM <../SourceLevelDebugging.html>`_ relies on having the address of
251   the variable exposed so that debug info can be attached to it. This
252   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 and
255running, and is very simple to implement. Let's extend Kaleidoscope with
256mutable variables now!
257
258Mutable Variables in Kaleidoscope
259=================================
260
261Now that we know the sort of problem we want to tackle, let's see what
262this 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 have
269variables for incoming arguments as well as for induction variables, and
270redefining those only goes so far :). Also, the ability to define new
271variables is a useful thing regardless of whether you will be mutating
272them. 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 operands
277    # 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) then
283        1
284      else
285        fib(x-1)+fib(x-2);
286
287    # Iterative fib.
288    def fibi(x)
289      var a = 1, b = 1, c in
290      (for i = 3, i < x in
291         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 variables
300to use the "alloca trick". Once we have that, we'll add our new
301operator, then extend Kaleidoscope to support new variable definitions.
302
303Adjusting Existing Variables for Mutation
304=========================================
305
306The symbol table in Kaleidoscope is managed at code generation time by
307the '``NamedValues``' map. This map currently keeps track of the LLVM
308"Value\*" that holds the double value for the named variable. In order
309to support mutation, we need to change this slightly, so that
310``NamedValues`` holds the *memory location* of the variable in question.
311Note that this change is a refactoring: it changes the structure of the
312code, but does not (by itself) change the behavior of the compiler. All
313of these changes are isolated in the Kaleidoscope code generator.
314
315At this point in Kaleidoscope's development, it only supports variables
316for two things: incoming arguments to functions and the induction
317variable of 'for' loops. For consistency, we'll allow mutation of these
318variables in addition to other user-defined variables. This means that
319these will both need memory locations.
320
321To start our transformation of Kaleidoscope, we'll change the
322NamedValues map so that it maps to AllocaInst\* instead of Value\*. Once
323we do this, the C++ compiler will tell us what parts of the code we need
324to 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 helper
331function that ensures that the allocas are created in the entry block of
332the function:
333
334.. code-block:: c++
335
336    /// CreateEntryBlockAlloca - Create an alloca instruction in the entry block of
337    /// the function.  This is used for mutable variables etc.
338    static AllocaInst *CreateEntryBlockAlloca(Function *TheFunction,
339                                              const std::string &VarName) {
340      IRBuilder<> TmpB(&TheFunction->getEntryBlock(),
341                     TheFunction->getEntryBlock().begin());
342      return TmpB.CreateAlloca(Type::getDoubleTy(TheContext), 0,
343                               VarName.c_str());
344    }
345
346This funny looking code creates an IRBuilder object that is pointing at
347the first instruction (.begin()) of the entry block. It then creates an
348alloca with the expected name and returns it. Because all values in
349Kaleidoscope 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 to
352variable references. In our new scheme, variables live on the stack, so
353code generating a reference to them actually needs to produce a load
354from the stack slot:
355
356.. code-block:: c++
357
358    Value *VariableExprAST::codegen() {
359      // Look this variable up in the function.
360      Value *V = NamedValues[Name];
361      if (!V)
362        return LogErrorV("Unknown variable name");
363
364      // Load the value.
365      return Builder.CreateLoad(V, Name.c_str());
366    }
367
368As you can see, this is pretty straightforward. Now we need to update
369the things that define the variables to set up the alloca. We'll start
370with ``ForExprAST::codegen()`` (see the `full code listing <#id1>`_ for
371the 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 where
395      // the body of the loop mutates the variable.
396      Value *CurVar = Builder.CreateLoad(Alloca);
397      Value *NextVar = Builder.CreateFAdd(CurVar, StepVal, "nextvar");
398      Builder.CreateStore(NextVar, Alloca);
399      ...
400
401This code is virtually identical to the code `before we allowed mutable
402variables <LangImpl5.html#code-generation-for-the-for-loop>`_. The big difference is that we
403no longer have to construct a PHI node, and we use load/store to access
404the variable as needed.
405
406To support mutable argument variables, we need to also make allocas for
407them. The code for this is also pretty simple:
408
409.. code-block:: c++
410
411    Function *FunctionAST::codegen() {
412      ...
413      Builder.SetInsertPoint(BB);
414
415      // Record the function arguments in the NamedValues map.
416      NamedValues.clear();
417      for (auto &Arg : TheFunction->args()) {
418        // Create an alloca for this variable.
419        AllocaInst *Alloca = CreateEntryBlockAlloca(TheFunction, Arg.getName());
420
421        // Store the initial value into the alloca.
422        Builder.CreateStore(&Arg, Alloca);
423
424        // Add arguments to variable symbol table.
425        NamedValues[Arg.getName()] = Alloca;
426      }
427
428      if (Value *RetVal = Body->codegen()) {
429        ...
430
431For each argument, we make an alloca, store the input value to the
432function into the alloca, and register the alloca as the memory location
433for the argument. This method gets invoked by ``FunctionAST::codegen()``
434right after it sets up the entry block for the function.
435
436The final missing piece is adding the mem2reg pass, which allows us to
437get good codegen once again:
438
439.. code-block:: c++
440
441        // Promote allocas to registers.
442        TheFPM->add(createPromoteMemoryToRegisterPass());
443        // Do simple "peephole" optimizations and bit-twiddling optzns.
444        TheFPM->add(createInstructionCombiningPass());
445        // Reassociate expressions.
446        TheFPM->add(createReassociatePass());
447        ...
448
449It is interesting to see what the code looks like before and after the
450mem2reg optimization runs. For example, this is the before/after code
451for our recursive fib function. Before the optimization:
452
453.. code-block:: llvm
454
455    define double @fib(double %x) {
456    entry:
457      %x1 = alloca double
458      store double %x, double* %x1
459      %x2 = load double, double* %x1
460      %cmptmp = fcmp ult double %x2, 3.000000e+00
461      %booltmp = uitofp i1 %cmptmp to double
462      %ifcond = fcmp one double %booltmp, 0.000000e+00
463      br i1 %ifcond, label %then, label %else
464
465    then:       ; preds = %entry
466      br label %ifcont
467
468    else:       ; preds = %entry
469      %x3 = load double, double* %x1
470      %subtmp = fsub double %x3, 1.000000e+00
471      %calltmp = call double @fib(double %subtmp)
472      %x4 = load double, double* %x1
473      %subtmp5 = fsub double %x4, 2.000000e+00
474      %calltmp6 = call double @fib(double %subtmp5)
475      %addtmp = fadd double %calltmp, %calltmp6
476      br label %ifcont
477
478    ifcont:     ; preds = %else, %then
479      %iftmp = phi double [ 1.000000e+00, %then ], [ %addtmp, %else ]
480      ret double %iftmp
481    }
482
483Here there is only one variable (x, the input argument) but you can
484still see the extremely simple-minded code generation strategy we are
485using. In the entry block, an alloca is created, and the initial input
486value is stored into it. Each reference to the variable does a reload
487from the stack. Also, note that we didn't modify the if/then/else
488expression, so it still inserts a PHI node. While we could make an
489alloca for it, it is actually easier to create a PHI node for it, so we
490still just make the PHI.
491
492Here is the code after the mem2reg pass runs:
493
494.. code-block:: llvm
495
496    define double @fib(double %x) {
497    entry:
498      %cmptmp = fcmp ult double %x, 3.000000e+00
499      %booltmp = uitofp i1 %cmptmp to double
500      %ifcond = fcmp one double %booltmp, 0.000000e+00
501      br i1 %ifcond, label %then, label %else
502
503    then:
504      br label %ifcont
505
506    else:
507      %subtmp = fsub double %x, 1.000000e+00
508      %calltmp = call double @fib(double %subtmp)
509      %subtmp5 = fsub double %x, 2.000000e+00
510      %calltmp6 = call double @fib(double %subtmp5)
511      %addtmp = fadd double %calltmp, %calltmp6
512      br label %ifcont
513
514    ifcont:     ; preds = %else, %then
515      %iftmp = phi double [ 1.000000e+00, %then ], [ %addtmp, %else ]
516      ret double %iftmp
517    }
518
519This is a trivial case for mem2reg, since there are no redefinitions of
520the variable. The point of showing this is to calm your tension about
521inserting such blatent inefficiencies :).
522
523After the rest of the optimizers run, we get:
524
525.. code-block:: llvm
526
527    define double @fib(double %x) {
528    entry:
529      %cmptmp = fcmp ult double %x, 3.000000e+00
530      %booltmp = uitofp i1 %cmptmp to double
531      %ifcond = fcmp ueq double %booltmp, 0.000000e+00
532      br i1 %ifcond, label %else, label %ifcont
533
534    else:
535      %subtmp = fsub double %x, 1.000000e+00
536      %calltmp = call double @fib(double %subtmp)
537      %subtmp5 = fsub double %x, 2.000000e+00
538      %calltmp6 = call double @fib(double %subtmp5)
539      %addtmp = fadd double %calltmp, %calltmp6
540      ret double %addtmp
541
542    ifcont:
543      ret double 1.000000e+00
544    }
545
546Here we see that the simplifycfg pass decided to clone the return
547instruction into the end of the 'else' block. This allowed it to
548eliminate some branches and the PHI node.
549
550Now that all symbol table references are updated to use stack variables,
551we'll add the assignment operator.
552
553New Assignment Operator
554=======================
555
556With our current framework, adding a new assignment operator is really
557simple. We will parse it just like any other binary operator, but handle
558it internally (instead of allowing the user to define it). The first
559step is to set a precedence:
560
561.. code-block:: c++
562
563     int main() {
564       // Install standard binary operators.
565       // 1 is lowest precedence.
566       BinopPrecedence['='] = 2;
567       BinopPrecedence['<'] = 10;
568       BinopPrecedence['+'] = 20;
569       BinopPrecedence['-'] = 20;
570
571Now that the parser knows the precedence of the binary operator, it
572takes care of all the parsing and AST generation. We just need to
573implement codegen for the assignment operator. This looks like:
574
575.. code-block:: c++
576
577    Value *BinaryExprAST::codegen() {
578      // Special case '=' because we don't want to emit the LHS as an expression.
579      if (Op == '=') {
580        // Assignment requires the LHS to be an identifier.
581        VariableExprAST *LHSE = dynamic_cast<VariableExprAST*>(LHS.get());
582        if (!LHSE)
583          return LogErrorV("destination of '=' must be a variable");
584
585Unlike the rest of the binary operators, our assignment operator doesn't
586follow the "emit LHS, emit RHS, do computation" model. As such, it is
587handled as a special case before the other binary operators are handled.
588The other strange thing is that it requires the LHS to be a variable. It
589is invalid to have "(x+1) = expr" - only things like "x = expr" are
590allowed.
591
592.. code-block:: c++
593
594        // Codegen the RHS.
595        Value *Val = RHS->codegen();
596        if (!Val)
597          return nullptr;
598
599        // Look up the name.
600        Value *Variable = NamedValues[LHSE->getName()];
601        if (!Variable)
602          return LogErrorV("Unknown variable name");
603
604        Builder.CreateStore(Val, Variable);
605        return Val;
606      }
607      ...
608
609Once we have the variable, codegen'ing the assignment is
610straightforward: we emit the RHS of the assignment, create a store, and
611return the computed value. Returning a value allows for chained
612assignments like "X = (Y = Z)".
613
614Now that we have an assignment operator, we can mutate loop variables
615and arguments. For example, we can now run code like this:
616
617::
618
619    # Function to print a double.
620    extern printd(x);
621
622    # Define ':' for sequencing: as a low-precedence operator that ignores operands
623    # and just returns the RHS.
624    def binary : 1 (x y) y;
625
626    def test(x)
627      printd(x) :
628      x = 4 :
629      printd(x);
630
631    test(123);
632
633When run, this example prints "123" and then "4", showing that we did
634actually mutate the value! Okay, we have now officially implemented our
635goal: getting this to work requires SSA construction in the general
636case. However, to be really useful, we want the ability to define our
637own local variables, let's add this next!
638
639User-defined Local Variables
640============================
641
642Adding var/in is just like any other extension we made to
643Kaleidoscope: we extend the lexer, the parser, the AST and the code
644generator. The first step for adding our new 'var/in' construct is to
645extend the lexer. As before, this is pretty trivial, the code looks like
646this:
647
648.. code-block:: c++
649
650    enum Token {
651      ...
652      // var definition
653      tok_var = -13
654    ...
655    }
656    ...
657    static int gettok() {
658    ...
659        if (IdentifierStr == "in")
660          return tok_in;
661        if (IdentifierStr == "binary")
662          return tok_binary;
663        if (IdentifierStr == "unary")
664          return tok_unary;
665        if (IdentifierStr == "var")
666          return tok_var;
667        return tok_identifier;
668    ...
669
670The next step is to define the AST node that we will construct. For
671var/in, it looks like this:
672
673.. code-block:: c++
674
675    /// VarExprAST - Expression class for var/in
676    class VarExprAST : public ExprAST {
677      std::vector<std::pair<std::string, std::unique_ptr<ExprAST>>> VarNames;
678      std::unique_ptr<ExprAST> Body;
679
680    public:
681      VarExprAST(std::vector<std::pair<std::string, std::unique_ptr<ExprAST>>> VarNames,
682                 std::unique_ptr<ExprAST> Body)
683        : VarNames(std::move(VarNames)), Body(std::move(Body)) {}
684
685      Value *codegen() override;
686    };
687
688var/in allows a list of names to be defined all at once, and each name
689can optionally have an initializer value. As such, we capture this
690information in the VarNames vector. Also, var/in has a body, this body
691is allowed to access the variables defined by the var/in.
692
693With this in place, we can define the parser pieces. The first thing we
694do is add it as a primary expression:
695
696.. code-block:: c++
697
698    /// primary
699    ///   ::= identifierexpr
700    ///   ::= numberexpr
701    ///   ::= parenexpr
702    ///   ::= ifexpr
703    ///   ::= forexpr
704    ///   ::= varexpr
705    static std::unique_ptr<ExprAST> ParsePrimary() {
706      switch (CurTok) {
707      default:
708        return LogError("unknown token when expecting an expression");
709      case tok_identifier:
710        return ParseIdentifierExpr();
711      case tok_number:
712        return ParseNumberExpr();
713      case '(':
714        return ParseParenExpr();
715      case tok_if:
716        return ParseIfExpr();
717      case tok_for:
718        return ParseForExpr();
719      case tok_var:
720        return ParseVarExpr();
721      }
722    }
723
724Next we define ParseVarExpr:
725
726.. code-block:: c++
727
728    /// varexpr ::= 'var' identifier ('=' expression)?
729    //                    (',' identifier ('=' expression)?)* 'in' expression
730    static std::unique_ptr<ExprAST> ParseVarExpr() {
731      getNextToken();  // eat the var.
732
733      std::vector<std::pair<std::string, std::unique_ptr<ExprAST>>> VarNames;
734
735      // At least one variable name is required.
736      if (CurTok != tok_identifier)
737        return LogError("expected identifier after var");
738
739The first part of this code parses the list of identifier/expr pairs
740into the local ``VarNames`` vector.
741
742.. code-block:: c++
743
744      while (1) {
745        std::string Name = IdentifierStr;
746        getNextToken();  // eat identifier.
747
748        // Read the optional initializer.
749        std::unique_ptr<ExprAST> Init;
750        if (CurTok == '=') {
751          getNextToken(); // eat the '='.
752
753          Init = ParseExpression();
754          if (!Init) return nullptr;
755        }
756
757        VarNames.push_back(std::make_pair(Name, std::move(Init)));
758
759        // End of var list, exit loop.
760        if (CurTok != ',') break;
761        getNextToken(); // eat the ','.
762
763        if (CurTok != tok_identifier)
764          return LogError("expected identifier list after var");
765      }
766
767Once all the variables are parsed, we then parse the body and create the
768AST node:
769
770.. code-block:: c++
771
772      // At this point, we have to have 'in'.
773      if (CurTok != tok_in)
774        return LogError("expected 'in' keyword after 'var'");
775      getNextToken();  // eat 'in'.
776
777      auto Body = ParseExpression();
778      if (!Body)
779        return nullptr;
780
781      return llvm::make_unique<VarExprAST>(std::move(VarNames),
782                                           std::move(Body));
783    }
784
785Now that we can parse and represent the code, we need to support
786emission of LLVM IR for it. This code starts out with:
787
788.. code-block:: c++
789
790    Value *VarExprAST::codegen() {
791      std::vector<AllocaInst *> OldBindings;
792
793      Function *TheFunction = Builder.GetInsertBlock()->getParent();
794
795      // Register all variables and emit their initializer.
796      for (unsigned i = 0, e = VarNames.size(); i != e; ++i) {
797        const std::string &VarName = VarNames[i].first;
798        ExprAST *Init = VarNames[i].second.get();
799
800Basically it loops over all the variables, installing them one at a
801time. For each variable we put into the symbol table, we remember the
802previous value that we replace in OldBindings.
803
804.. code-block:: c++
805
806        // Emit the initializer before adding the variable to scope, this prevents
807        // the initializer from referencing the variable itself, and permits stuff
808        // like this:
809        //  var a = 1 in
810        //    var a = a in ...   # refers to outer 'a'.
811        Value *InitVal;
812        if (Init) {
813          InitVal = Init->codegen();
814          if (!InitVal)
815            return nullptr;
816        } else { // If not specified, use 0.0.
817          InitVal = ConstantFP::get(TheContext, APFloat(0.0));
818        }
819
820        AllocaInst *Alloca = CreateEntryBlockAlloca(TheFunction, VarName);
821        Builder.CreateStore(InitVal, Alloca);
822
823        // Remember the old variable binding so that we can restore the binding when
824        // we unrecurse.
825        OldBindings.push_back(NamedValues[VarName]);
826
827        // Remember this binding.
828        NamedValues[VarName] = Alloca;
829      }
830
831There are more comments here than code. The basic idea is that we emit
832the initializer, create the alloca, then update the symbol table to
833point to it. Once all the variables are installed in the symbol table,
834we evaluate the body of the var/in expression:
835
836.. code-block:: c++
837
838      // Codegen the body, now that all vars are in scope.
839      Value *BodyVal = Body->codegen();
840      if (!BodyVal)
841        return nullptr;
842
843Finally, before returning, we restore the previous variable bindings:
844
845.. code-block:: c++
846
847      // Pop all our variables from scope.
848      for (unsigned i = 0, e = VarNames.size(); i != e; ++i)
849        NamedValues[VarNames[i].first] = OldBindings[i];
850
851      // Return the body computation.
852      return BodyVal;
853    }
854
855The end result of all of this is that we get properly scoped variable
856definitions, and we even (trivially) allow mutation of them :).
857
858With this, we completed what we set out to do. Our nice iterative fib
859example from the intro compiles and runs just fine. The mem2reg pass
860optimizes all of our stack variables into SSA registers, inserting PHI
861nodes where needed, and our front-end remains simple: no "iterated
862dominance frontier" computation anywhere in sight.
863
864Full Code Listing
865=================
866
867Here is the complete code listing for our running example, enhanced with
868mutable variables and var/in support. To build this example, use:
869
870.. code-block:: bash
871
872    # Compile
873    clang++ -g toy.cpp `llvm-config --cxxflags --ldflags --system-libs --libs core mcjit native` -O3 -o toy
874    # Run
875    ./toy
876
877Here is the code:
878
879.. literalinclude:: ../../examples/Kaleidoscope/Chapter7/toy.cpp
880   :language: c++
881
882`Next: Compiling to Object Code <LangImpl08.html>`_
883
884