Skip to content

Latest commit

 

History

History
265 lines (193 loc) · 13.4 KB

README.md

File metadata and controls

265 lines (193 loc) · 13.4 KB
Raw

Chapter 5

In this chapter we extend the language grammar with the following features:

  • We introduce the if statement.
  • To support splitting of control flow and merging, we introduce new nodes: Region and Phi.
  • Since we can now have multiple return points, we also introduce the Stop node as the termination.

Here is the complete language grammar for this chapter.

New Nodes

The following new nodes are introduced in this chapter:

Node Name Type Chapter Description Inputs Value
If Control 5 A branching test, sub type of MultiNode A control node and a data predicate node A tuple of two values: one for true, one for false
Region Control 5 A merge point for multiple control flows An input for each control flow that is merging Merged control
Phi Data 5 A phi function picks a value based on control flow A Region, and data nodes for each control path Depends on control flow path taken
Stop Control 5 Termination of the program All return nodes of the function None

Recap

Here is a recap of the nodes introduced in previous chapters:

Node Name Type Chapter Description Inputs Value
Multi Abstract class 4 A node that has a tuple result A tuple
Start Control 1 Start of function, now a MultiNode A tuple with a ctrl token and an arg data node
Proj Data 4 Projection nodes extract values from MultiNode A MultiNode and index Result is the extracted value from the input MultiNode at offset index
Bool Data 4 Represents results of a comparison operator Two data nodes Result is a comparison, represented as integer value where 1=true, 0=false
Not Data 4 Logical not One data node Result converts 0 to 1 and vice versa
Return Control 1 End of function Predecessor control node and a data node for the return value of the function Return value of the function
Constant Data 1 Represents constants such as integer literals None, however Start node is set as input to enable graph walking Value of the constant
Add Data 2 Add two values Two data nodes without restrictions on the order Result of the add operation
Sub Data 2 Subtract a value from another Two data nodes, the first one is subtracted by the second one Result of the subtraction
Mul Data 2 Multiply two values Two data nodes without restrictions on the order Result of the multiplication
Div Data 2 Divide a value by another Two data nodes, the first one is divided by the second one Result of the division
Minus Data 2 Negate a value One data node which value is negated Result of the negation
Scope Symbol Table 3 Represents scopes in the graph Nodes that represent the current value of variables None

If Nodes

If node takes in both control and data (predicate expression) and routes the control token to one of the two control flows, represented by true and false Proj nodes.

Phi Nodes

A Phi reads in both data and control, and outputs a data value. The control input to the Phi points to a Region node. The data inputs to the Phi are one for each of the control inputs to that Region. The result computed by a Phi depends both on the data and the matching control input. At most one control input to the Region can be active at a time, and the Phi passes through the data value from the matching input.1

Region Nodes

Every instruction has a control input from a basic block. If the control input is an edge in our abstract graph, then the basic block must be a node in the abstract graph. So we define a REGION instruction to replace a basic block. A REGION instruction takes control from each predecessor block as input and produces a merged control as an output.2

However:

We can remove the control dependence for any given Node simply by replacing the pointer value with the special value NULL. This operation only makes sense for Nodes that represent data computations. PHI, IF, JUMP and STOP Nodes all require the control input for semantic correctness. REGION Nodes require several control inputs (one per CFG input to the basic block they represent). Almost all other Nodes can live without a control dependence. A data computation without a control dependence does not exactly reside in any particular basic block. Its correct behavior depends solely on the remaining data dependences. It, and the Nodes that depend on it or on which it depends, exists in a “sea” of Nodes, with little control structure.

The “sea” of Nodes is useful for optimization, but does not represent any traditional intermediate representation such as a CFG. We need a way to serialize the graph and get back the control dependences. We do this with a simple global code motion algorithm.3

Thus, we do not associate a control edge on every data node in the graph.

We insert a Region node at a merge point where it takes control from each predecessor's control edge, and produces a merged control as output. Data flows via Phi nodes at these merge points.

Parsing an if Statement

When we parse an if statement, the control flow splits at that point. We must track the names being updated in each part of the if statement, and then merge them at the end. The implementation follows the description in Combining Analyses, Combining Optimizations4.

This involves following:

  1. We create an IfNode with the current control token, i.e. the node mapped to $ctrl, and the if predicate expression as inputs.
  2. We add two ProjNodes - one for the True branch (call it ifT), and the other for the False branch (call it ifF) - these extract values from the tuple result of the IfNode.
  3. We duplicate the current ScopeNode. The duplicated ScopeNode contains all the same symbol tables as the original, and has the same edges.
  4. We set the control token to the True projection node ifT, and parse the true branch of the if statement.
  5. We set the duplicated ScopeNode as the current one.
  6. The control token is set to the False projection node ifF, and if there is an else statement we parse it.
  7. At this point we have two ScopeNodes; the original one, potentially updated by the True branch, and the duplicate one, potentially updated by the False branch.
  8. We create a Region node to represent a merge point.
  9. We merge the two ScopeNodes. We create Phi nodes for any names whose bindings differ. The Phi nodes take the Region node as the control input and the two data nodes from each of the ScopeNodes. The name binding is updated to point to the Phi node.
  10. Finally, we set the Region node as the control token, and discard the duplicated ScopeNode.

Implementation is in parseIf method in Parser.

Operations on ScopeNodes

As explained above, we duplicate ScopeNodes and merge them at a later point. There are some subtleties in how this is implemented that is worth going over.

Duplicating a ScopeNode

Below is the code for duplicating a ScopeNode.

Our goals are:

  1. Duplicate the name bindings across all stack levels
  2. Make the new ScopeNode a user of all the bound nodes
  3. Ensure that the order of defs in the duplicate is the same to allow easy merging

For implementation see scopeNode.dup()

Merging two ScopeNodes

At the merge point we merge two ScopeNodes. The goals are:

  1. Merge names whose bindings have changed between the two nodes. For each such name, a Phi node is created, referencing the two original data nodes.
  2. A new Region node is created representing the merged control flow. The phis have this region node as the first input.
  3. After the merge is completed, the duplicate is discarded, and its use of each of the nodes is also deleted.

The merging logic takes advantage of that fact that the two ScopeNodes have the bound nodes in the same order in the list of inputs. This was ensured during duplicating the ScopeNode. Although only the innermost occurrence of a name can have its binding changed, we scan all the nodes in our input list, and simply ignore ones where the binding has not changed.

For implementation see ScopeNode.mergeScopes()

Example

We show the graph for the following code snippet:

int a = 1; 
if (arg == 1) 
	a = arg+2; 
else
	a = arg-3;
return a;

Before Merging

Following shows the graph just before we merge the two branches of the if statement in a Region node.

Graph1

Note the two ScopeNodes in the graph. One has its $ctrl pointing to the True projection, while the other has $ctrl pointing to False projection. The variable a is bound to the Add node in the True branch, whereas it is bound to the Sub node in the False branch. Thus a will need a Phi node when merging the two scopes.

After Merging

Below is the graph after we created a Region node and merged the two definitions of a in a Phi node.

Graph2

The duplicate ScopeNode has been discarded and was merged into the other one. Since a had two different definitions in both scopes a Phi node was created and is now referenced from a. The Phi node's inputs are the Region node and the Add node from the True branch, and Sub node from the False branch.

Finally

Here is the graph after the return statement was parsed and processed.

Graph3

More Peepholes

Phi's implement a peephole illustrated in the example:

int a=arg==2;
if( arg==1 )
{
    a=arg==3;
}
return a;

Pre-peephole we have:

Graph4

Post-peephole:

Graph5

The implementation is in PhiNode.idealize()

More examples

int c = 3;
int b = 2;
if (arg == 1) {
    b = 3;
    c = 4;
}
return c;

Graph6

int a=arg+1;
int b=arg+2;
if( arg==1 )
    b=b+a;
else
    a=b+1;
return a+b;

Graph7

int a=1;
if( arg==1 )
    if( arg==2 )
        a=2;
    else
        a=3;
else if( arg==3 )
    a=4;
else
    a=5;
return a;

Graph8

Footnotes

  1. Click, C. (1995). Combining Analyses, Combining Optimizations, 132. ↩

  2. Click, C. (1995). Combining Analyses, Combining Optimizations, 129. ↩

  3. Click, C. (1995). Combining Analyses, Combining Optimizations, 86. ↩

  4. Click, C. (1995). Combining Analyses, Combining Optimizations, 102-103. ↩