Main Page | See live article | Alphabetical index

Compiler optimization

Compiler optimization is used to improve the efficiency (in terms of running time or resource usage) of the code output by a compiler. These techniques allow programmers to write code in a straightforward manner, expressing their intentions clearly, while allowing the computer to make choices about implementation details that lead to efficient code. Contrary to what the term might imply, this rarely results in code that is perfectly "optimal" by any measure, only code that is much improved compared to direct translation of the programmer's original code.

Table of contents
1 Problems with Optimization
2 Types of Optimizations
3 Factors affecting optimization
4 Intended use of the generated code
5 Optimization techniques
6 Links

Problems with Optimization

Further problems with optimizing compilers are:

This is where so-called "post pass" optimizers come in. These tools take the program code output by an "optimizing" compiler and optimize it even further (see http://www.absint.com/aipop for an example of such tool). As opposed to compilers which optimize intermediate representations of programs, post pass optimizers work on the Assembly language level. Post pass compilers are also limited by this, however, in that much of the information found in the original source is lost.

In the early times of computer science, compiler optimizations were not as good as hand-written ones. As compiler technologies have improved, good compilers can often generate better code than human programmers — and good post pass optimizers can improve highly hand-optimized code even further. In the RISC CPU architecture, compiler optimization is the key for obtaining an optimal code, because the RISC instruction set is so compact that it is hard for a human to manually schedule or combine small instructions to get efficient results. Indeed, these architectures were designed to be easy for compiler writers to target.

Types of Optimizations

Techniques in optimization can be broken up along various dimensions:

; scope of the optimization: The effect of a particular compiler optimization can affect anywhere from a single statement to an entire program. Some examples of optimization scopes:

Local techniques tend to be easier to implement, but result in lesser gains, while global techniques make the opposite tradeoff.

; programming language-independent vs. language-dependent: Most high-level languages share common programming constructs and abstractions--decision (if, switch, case), looping (for, while, repeat .. until, do .. while), encapsulation (structures, objects). Thus similar optimization techniques can be used across languages. However certain language features make some kinds of optimizations possible and/or difficult. For instance, the existence of pointers in C and C++ makes certain optmizations of array accesses difficult. Conversely, in some languages functions are not permitted to have "side effects". Therefore, if repeated calls to the same function with the same arguments are made, the compiler can immediately infer that results need only be computed once and the result referred to repeatedly.

; machine independent vs. machine dependent: A lot of optimizations that operate on abstract programming concepts (loops, objects, structures) are independent of the machine targetted by the compiler. But many of the most effective optimizations are those that best exploit special features of the target platform.

These dimensions aren't completely orthogonal. It is often the case that machine dependent optimizations are local, for example.

An instance of a local machine dependent optimization: to set a register to 0, the obvious way is to use the constant 0 with the instruction that sets a register value to a constant. A less obvious way is to XOR a register with itself. It is up to the compiler to know which instruction variant to use. On many RISC machines, both instructions would be equally appropriate, since they would both be the same length and take the same time. On many other microprocessors such as the Intel x86 family, it turns out that the XOR variant is shorter and maybe faster (no need to decode an immediate operand nor use the internal "immediate operand register").

Factors affecting optimization

; The machine itself

Many of the choices about which optimizations can and should be done depend on the characteristics of the target machine. It is sometimes possible to parameterize some of these machine dependent factors, so that a single piece of compiler code can be used to optimize different machines just by altering the machine description parameters. See for example GCC for a compiler that exemplifies this approach.

; The architecture of the target CPU

Compilers can schedule, or reorder, instructions so that pipeline stalls occur less frequently. Here again, instructions have to be scheduled so that the various functional units are fully fed with instructions to execute.

; The architecture of the machine

Intended use of the generated code

; Debugging : When a programmer is still writing an application, they will recompile and test often, and so compilation must be fast. This is one reason most optimizations are avoided while debugging. Also, debugging code is usually stepped through in a symbolic debugger, and optimizing transformations, particularly those that reorder code, can make it difficult to identify the output code with the line numbers in the source code from which the code was generated. This can confuse the debugging tools and the programmer using them.

; General purpose use : Prepackaged software is very often expected to be executed on a variety of machines and CPUs that may share the same instruction set, but have different timing, cache or memory characteristics. So, the code may not be tuned to any particular CPU, or may be tuned to work well on the most popular CPU and work reasonably on other CPUs.

; Special purpose use : If the software is compiled to be used on one or a few very similar machines, with known characteristics, then the compiler can heavily tune the generated code to those machines alone. Important special cases include code designed for parallel and vector processors, for which we have parallelizing compilers.

Optimization techniques

Common Themes

To a large extent, optimization techniques have the following themes, which sometime conflict

; Avoid redundancy : If something has already been computed, it's generally better to store it and reuse it later, instead of recomputing it.

; Less code : There is less work for the CPU, cache, and memory. So, likely to be faster.

; Straight line code, fewer jumps : Less complicated code. Jumps interfere with the prefetching of instructions, thus slowing down code.

; Code locality : Pieces of code executed close together in time should be placed close together in memory, which increases spatial locality of reference.

; Extract more information from code : The more information the compiler has, the better it can optimize.

Optimization techniques

Loop Optimizations

Some optimization techniques primarily designed to operate on loops include:

; induction variable analysis : Roughly, if a variable in a loop is a simple function of the index variable, such as j := 4*i+1, it can be updated appropriately each time the loop variable is changed. This is a strength reduction, and also may allow the index variable's definitions to become dead code. This information is also useful for bounds-checking elimination and dependence analysis, among other things.

; loop fission : Loop fission attempts to break a loop into multiple loops over the same index range but each taking only a part of the loop's body. This can improve locality of reference, both of the data being accessed in the loop and the code in the loop's body.

; loop fusion or loop combining : Another technique which attempts to reduce loop overhead. When two adjacent loops would iterate the same number of times (whether or not that number is known at compile time), their bodies can be combined as long as they make no reference to each other's data.

; loop interchange : These optimizations exchange inner loops with outer loops. When the loop variables index into an array, such a transformation can improve locality of reference, depending on the array's layout.

; loop-invariant code motion : If a quantity is computed inside a loop during every iteration, and its value is the same for each iteration, it can vastly improve efficiency to hoist it outside the loop and compute its value just once before the loop begins. This is particularly important with the address-calculation expressions generated by loops over arrays.

; loop reversal : Loop reversal reverses the order in which values are assigned to the index variable. This is a subtle optimization which can help eliminate dependencies and thus enable other optimizations.

; loop unrolling : Duplicates the body of the loop multiple times, in order to decrease the number of times the loop condition is tested and the number of jumps, which hurt performance by impairing the instruction pipeline. Completely unrolling a loop eliminates all overhead, but requires that the number of iterations be known at compile time.

; loop splitting : Loop splitting attempts to simplify a loop or eliminate dependencies by breaking it into multiple loops which have the same bodies but iterate over different contiguous portions of the index range. A useful special case is loop peeling, which can simplify a loop with a problematic first iteration by performing that iteration separately before entering the loop.

; unswitching : Unswitching moves a conditional inside a loop outside of it by duplicating the loop's body, and placing a version of it inside each of the if and else clauses of the conditional.

Data-flow Optimizations

Data flow optimizations primarily depend on how certain properties of data are propagated by control edges in the control flow graph. Some of these include:

; common subexpression elimination : In the expression "(a+b)-(a+b)/4", "common subexpression" refers to the duplicated "(a+b)". Compilers implementing this technique realize that "(a+b)" won't change, and as such, only calculate its value once.

; constant folding and propagation : replacing expressions consisting of constants (e.g. "3 + 5") which their final value ("8") rather than doing the calculation in run-time. Used in most modern languages.

SSA-based Optimizations

These optimizations are intended to be done after transforming the program into a special form called SSA form, in which every variable is assigned in only one place. Although some function without SSA, they are most effective with SSA. Many optimizations listed in other sections also benefit with no special changes, such as register allocation.

; global value numbering: GVN eliminates redundancy by constructing a value graph of the program, and then determining which values are computed by equivalent expressions. GVN is able to identify some redundancy that common subexpression elimination cannot, and vice versa.

; sparse conditional constant propagation: Effectively equivalent to iteratively performing constant propagation, constant folding, and dead code elimination until there is no change, but is much more efficient. This optimization symbolically executes the program, simultaneously propagating constant values and eliminating portions of the control flow graph that this makes unreachable.

Back End Optimizations

; register allocation: The most frequently used variables should be kept in processor registers for fastest access. To find which variables to put in registers a interference-graph is created. Each variable is a vertex and when two variables are used at the same time (have an intersecting liverange) they have an edge between them. This graph is colored using for example Chaitin's algorithm using the same number of colors as there are registers. If the coloring fails one variable is "spilled" to memory and the coloring is retried.

; instruction selection: Most architectures, particularly CISC architectures and those with many addressing modes, offer several different ways of performing a particular operation, using entirely different sequences of instructions. The job of the instruction selector is to do a good job overall of choosing which instructions to implement which operators in the low-level intermediate representation with. For example, on many processors in the 68000 family, complex addressing modes can be used in statements like "lea 25(a1,d5*4),a0", allowing a single instruction to perform a significant amount of arithmetic with less storage.

; instruction scheduling : Instruction scheduling is an important optimization for modern pipelined processors, which avoids stalls or bubbles in the pipeline by clustering instructions with no dependencies together, while being careful to preserve the original semantics.

Functional Language Optimizations

Although many of these also apply to non-functional languages, they either originate in, are most easily implemented in, or are particular critical in functional languages such as Lisp and ML.

; removing recursion : Recursion is often expensive, as a function call consumes stack space and involves some overhead related to parameter passing and flushing the instruction cache. Tail recursive algorithms can be converted to iteration, which does not have call overhead and uses a constant amount of stack space, through a process called tail recursion elimination.

Other Optimizations

Please help separate and categorize these further and create detailed pages for them, especially the more complex ones, or link to one where one exists.

; bounds-checking elimination : Several modern languages, most notably Java, enforce bounds-checking of all array accesses. This is a severe performance bottleneck on certain applications such as scientific code. Bounds-checking elimination allows the compiler to safely remove bounds-checking in many situations where it can determine that the index must fall within valid bounds, for example if it is a simple loop variable.

; Branch offset optimization (machine independent): Choose the shortest branch displacement that reaches target

; code-block reordering: Code-block reordering alters the order of the basic blocks in a program in order to reduce conditional branches and improve locality of reference.

; dead code elimination: Removes instructions that will not affect the behaviour of the program, for example definitions which have no uses, called dead code. This reduces code size and eliminates unnecessary computation.

; factoring out of invariants : If an expression is carried out both when a condition is met and otherwise, it can be written just once outside of the conditional statement. Similarly, if certain types of expressions (e.g. the assignment of a constant into a variable) appear inside a loop, they can be moved out of it because their effect will be the same no matter if they're executed many times or just once. Also known as total redundancy elimination. A more powerful optimization is partial redundancy elimination (PRE).

; in-line expansion : When some code invokes a procedure, it is possible to directly insert the body of the procedure inside the calling code rather than transferring control to it. This saves the overhead related to procedure calls, as well as providing great opportunity for many different parameter-specific optimizations, but comes at the cost of space; the procedure body is duplicated each time the procedure is called inline. Generally, inlining is useful in performance-critical code that makes a large number of calls to small procedures.

; jump threading: In this pass, conditional jumps in the code that branch to identical or inverse tests are detected, and can be "threaded" through a second conditional test.

; strength reduction : A general term encompassing optimizations that replace complex or difficult or expensive operations with simpler ones. Induction variable analysis can accomplish this with variables inside a loop which depend on the loop variable. Several peephole optimizations also fall into this category, such as replacing division by a constant with multiplication by its reciprocal, converting multiplies into a series of bit-shifts and adds, and replacing large instructions with equivalent smaller ones that load more quickly.

; reduction of cache collisions : (e.g. by disrupting alignment within a page)

; Stack height reduction: Rearrange expression tree to minimize resources needed for expression evaluation.

; test reordering : If we have two tests that are the condition for something, we can first deal with the simpler tests (e.g. comparing a variable to something) and only then with the complex tests (e.g. those that require a function call). This technique complements lazy evaluation, but can be used only when the tests are not dependent on one another. Short-circuiting semantics can make this difficult.

; unreachable code elimination: Unreachable code refers to code which can never be executed. Unreachable code elimination prevents the compiler from emitting instructions for such code, decreasing code size.

Links

redhat.com | GCC Optimization

A summary of the optimization methods used by the popular free compiler GCC.