Peephole optimization

In compiler theory, peephole optimization is a kind of optimization performed over a very small set of instructions in a segment of generated code. The set is called a "peephole" or a "window". It works by recognising sets of instructions that can be replaced by shorter or faster set of instructions.

Replacement rules

Common techniques applied in peephole optimization:^[1]

Constant folding - Evaluate constant subexpressions in advance.
Strength reduction - Replace slow operations with faster equivalents.
Null sequences - Delete useless operations
Combine Operations: Replace several operations with one equivalent.
Algebraic Laws: Use algebraic laws to simplify or reorder instructions.
Special Case Instructions: Use instructions designed for special operand cases.
Address Mode Operations: Use address modes to simplify code.

There can, of course, be other types of peephole optimizations involving simplifying the target machine instructions, assuming that the target machine is known in advance. Advantages of a given architecture and instruction sets can be exploited in this case.

Examples

Replacing slow instructions with faster ones

The following Java bytecode

...
aload 1
aload 1
mul
...

can be replaced by

...
aload 1
dup
mul
...

This kind of optimization, like most peephole optimizations, makes certain assumptions about the efficiency of instructions. For instance, in this case, it is assumed that the dup operation (which duplicates and pushes the top of the stack) is more efficient than the aload X operation (which loads a local variable identified as X and pushes it on the stack).

Removing redundant code

Another example is to eliminate redundant load stores.

 a = b + c; 
 d = a + e;

is straightforwardly implemented as

MOV b, R0  # Copy b to the register
ADD c, R0  # Add  c to the register, the register is now b+c
MOV R0, a  # Copy the register to a
MOV a, R0  # Copy a to the register
ADD e, R0  # Add  e to the register, the register is now a+e [(b+c)+e]
MOV R0, d  # Copy the register to d

but can be optimised to

MOV b, R0 # Copy b to the register 
ADD c, R0 # Add c to the register, which is now b+c (a)
MOV R0, a # Copy the register to a
ADD e, R0 # Add e to the register, which is now b+c+e [(a)+e]
MOV R0, d # Copy the register to d

Furthermore, if the compiler knew that the variable a was not used again, the middle operation could be omitted.

Removing redundant stack instructions

If the compiler saves registers on the stack before calling a subroutine and restores them when returning, consecutive calls to subroutines may have redundant stack instructions.

Suppose the compiler generates the following Z80 instructions for each procedure call:

PUSH AF
PUSH BC
PUSH DE
PUSH HL
CALL _ADDR
POP HL
POP DE
POP BC
POP AF

If there were two consecutive subroutine calls, they would look like this:

PUSH AF
PUSH BC
PUSH DE
PUSH HL
CALL _ADDR1
POP HL
POP DE
POP BC
POP AF
PUSH AF
PUSH BC
PUSH DE
PUSH HL
CALL _ADDR2
POP HL
POP DE
POP BC
POP AF

The sequence POP regs followed by PUSH for the same registers is generally redundant. In cases where it is redundant, a peephole optimization would remove these instructions. In the example, this would cause another redundant POP/PUSH pair to appear in the peephole, and these would be removed in turn. Removing all of the redundant code in the example above would eventually leave the following code:

PUSH AF
PUSH BC
PUSH DE
PUSH HL
CALL _ADDR1
CALL _ADDR2
POP HL
POP DE
POP BC
POP AF

Implementation

Modern architectures typically allow for many hundreds of different kinds of peephole optimizations, and it is therefore often appropriate for compiler programmers to implement them using a pattern matching algorithm. ^[2]

References

^ Crafting a Compiler with C++, Fischer/LeBlanc
^ Compilers - Principles, Techniques, and Tools 2e, p560

External links

The dictionary definition of peephole optimization at Wiktionary

[1] Crafting a Compiler with C++, Fischer/LeBlanc

[2] Compilers - Principles, Techniques, and Tools 2e, p560

[1]

[2]