Lab for Automated Reasoning and Analysis LARA

Register Allocation using Liveness Information

static int x;
static int y;
static int z;
static int res;
 
int f() {
  int xy;
  int yz;
  int xz; 
  int res1; 
  xy = x * y;
  yz = y * z;
  xz = x * z;
  res1 = xy + yz;
  res = res1 + xz;
}

How to generate code that we saw in Register Machine Model in Scala

      Move(Register(1), Memory(xAddr)), 
      Move(Register(2), Memory(yAddr)),
      Move(Register(3), Memory(zAddr)),
      Quad(Register(4), Mul, Register(1), Register(2)),
      Quad(Register(2), Mul, Register(2), Register(3)),
      Quad(Register(4), Add, Register(4), Register(2)),
      Quad(Register(1), Mul, Register(1), Register(3)),
      Quad(Register(4), Add, Register(4), Register(1)),
      Move(Memory(rAddr), Register(4))

One solution:

  1. keep variables in registers when possible
  2. reuse registers if variable not used any more
  3. if no available register, free a register by saving it to memory ('spilling')

A lot of research was done and better techiques were found

Liveness in the Example

Variable is not live at a given program point if its current value will not be used in any future execution.

code live variables
{}
x = xAddr
{x}
y = yAddr
{x,y}
xy = x * y
{xy,x,y}
z = zAddr
{xy,x,y,z}
yz = y * z
{xy,yz,x,z}
xz = x * z
{xy, yz,xz}
res1 = xy + yz
{res1,xz}
res = res1 + xz
{}

Slightly differently ordered program:

code live variables
{}
x = xAddr
{x}
y = yAddr
{x,y}
xy = x * y
{xy,x,y}
z = zAddr
{xy,x,y,z}
yz = y * z
{xy,yz,x,z}
res1 = xy + yz
{res1,x,z}
xz = x * z
{res1,xz}
res = res1 + xz
{}

Interference Graph

Undirected graph

Nodes are variables: x,y,z,xy,yz,xz,res1

Edges: there is an edge ${p,q}$ if at some program point $u$, we have ${p,q} \subseteq live(u)$

  • both variables are live at the same time at some point

A Register Allocation Algorithm

Starting point:

  • assign variables to memory
  • assume temporary (intermediate) values will be in registers
  • introduce explicit instructions to move between memory and registers

Goal is to assign temporaries to a limited number of K registers

If two temporary variables are both live at same point, we cannot keep them in same register

Algorithm phases:

Build: construct interference graph

  • nodes are temporary variables
  • there is edge if two variables both live at some program point

Goal: color the graph using K colors such that adjecent nodes have different color

  • colors are registers
  • if we need values of two temporaries, we can store them in different registers

Simplify

If there is a node with less than K neighbors, we will always be able to color it!

  • so we can remove it from the graph

This reduces graph size

  • it is incomplete
  • e.g. every planar can be colored by at most 4 colors (and can have nodes with 100 neighbors)

Spill

If every node has K or more neighbors, we remove one of them

  • we mark it as node for potential spilling
  • then remove it and continue

Select

Assign colors backwards, adding nodes that were removed

If we find a node that was spilled, we check if we are lucky that we can color it

  • if yes, continue
  • if no, insert instructions to save and load values from memory
  • restart (now we have graph that is easier to color)

For the first of the examples, we can remove the nodes e.g. in this order:

res1
xz
xy
y
x
z
yz

This gives a 4-coloring

Suppose we try to do it with 3 registers only. We get potential spill trying to color xy. So we change the program to introduce a location in memory (e.g. on stack frame) to store the value of xy right after we assign to it and reload it immediately afterwards. We get this program:

code live variables
{}
x = xAddr
{x}
y = yAddr
{x,y}
xy = x * y
{x,y,xy}
xyAddr = xy
{x,y}
z = zAddr
{x,y,z}
yz = y * z
{yz,x,z}
xz = x * z
{yz,xz}
xy1 = xyAddr
{xy1,yz,xz}
res1 = xy1 + yz
{res1,xz}
res = res1 + xz
{}

Note that xy was split into xy and xy1, each of which has shorter life span. It “hibernates” in memory and is woken up when the value is loaded.

We succeed in coloring it, e.g. with these colors:

res1 -> 1
xz   -> 3
xy1  -> 2
yz   -> 1
xy   -> 1
y    -> 3
x    -> 2
z    -> 1

Conservative Coalescing

Suppose that variables tmp1 and tmp2 are both assigned to the same register R and the program has an instruction:

tmp2 = tmp1

which moves the value of tmp1 into tmp2. This instruction then becomes

R = R

which can be simply omitted.

How to force a register allocator to assign tmp1 and tmp2 to same register?

  • merge the nodes for tmp1 and tmp2 in the interference graph
  • this is called coalescing

If we always coalesce non-interfering nodes when there are assignments, then our graph may become more difficult to color, and we may in fact need more registers!

Conservative coalescing: coalesce only if merged node of tmp1 and tmp2 will have a small degree so that we are sure that we will be able to color it.

Example 1

i = 0
s = s + i
i = i + b
j = i
s = s + j + b
j = j + 1

Live variables:

code live variables
{s,b}
i = 0
{s,b,i}
s = s + i
{s,b,i}
i = i + b
{s,b,i}
j = i
{s,j,b}
s = s + j +b
{j}
j = j + 1
{}

Example 2

i = 0
while (i < a) {
  s = s + i
  i = i + b
}
j = i
while (j < b) {
  s = s + j + b
  j = j + 1
}

References

 
cc09/register_allocation_using_liveness_information.txt · Last modified: 2012/12/10 01:56 by vkuncak
 
© EPFL 2018 - Legal notice