imported>Jochen Burghardt: /* Invariant code detection */ the authors are active, not their paper (and shorter); rm 2nd wikilink (kept 1st one since a draft seems to be forthcoming)

2025-09-04T14:41:18Z

Invariant code detection: the authors are active, not their paper (and shorter); rm 2nd wikilink (kept 1st one since a draft seems to be forthcoming)

New page

{{short description|Type of compiler optimization}}
{{More citations needed|date=January 2021}}

In [[computer programming]], [[loop-invariant code]] consists of statements or expressions (in an [[imperative programming|imperative]] [[programming language]]) that can be moved outside the body of a loop without affecting the semantics of the program. '''Loop-invariant code motion''' (also called '''hoisting''' or '''scalar promotion''') is a [[compiler optimization]] that performs this movement automatically.

==Example==
In the following code sample, two optimizations can be applied.

<syntaxhighlight lang="C">
int i = 0;
while (i < n) {
x = y + z;
a[i] = 6 * i + x * x;
++i;
}
</syntaxhighlight>

Although the calculation <code>x = y + z</code> and <code>x * x</code> is loop-invariant, precautions must be taken before moving the code outside the loop. It is possible that the loop condition is <code>false</code> (for example, if <code>n</code> holds a negative value), and in such case, the loop body should not be executed at all. One way of guaranteeing correct behaviour is using a conditional branch outside of the loop. Evaluating the loop condition can have [[Side effect (computer science)|side effects]], so an additional evaluation by the <code>if</code> construct should be compensated by replacing the <code>while</code> loop with a <code>[[Do while loop|do {} while]]</code>. If the code used <code>do {} while</code> in the first place, the whole guarding process is not needed, as the loop body is guaranteed to execute at least once.

<syntaxhighlight lang="C">
int i = 0;
if (i < n) {
x = y + z;
int const t1 = x * x;
do {
a[i] = 6 * i + t1;
++i;
} while (i < n);
}
</syntaxhighlight>

This code can be optimized further. For example, [[strength reduction]] could remove the two multiplications inside the loop (<code>6*i</code> and <code>a[i]</code>), and [[induction variable]] elimination could then elide <code>i</code> completely. Since <code>6 * i</code> must be in lock step with <code>i</code> itself, there is no need to have both.

==Invariant code detection==
Usually, a [[Reaching definition|reaching definitions analysis]] is used to detect whether a statement or expression is loop invariant.

For example, if all reaching definitions for the operands of some simple expression are outside of the loop, the expression can be moved out of the loop.

Moyen, Rubiano and [[Thomas Seiller|Seiller]] used data-flow dependence analysis<ref>{{cite book |last1=Moyen |first1=Jean-Yves |last2=Rubiano |first2=Thomas |last3=Seiller |first3=Thomas |chapter=Loop Quasi-Invariant Chunk Detection |title=Automated Technology for Verification and Analysis |series=Lecture Notes in Computer Science |date=2017 |volume=10482 |pages=91–108 |doi=10.1007/978-3-319-68167-2_7|isbn=978-3-319-68166-5 }}</ref> to detect not only invariant commands but larger code fragments such as an inner loop. The analysis also detects quasi-invariants of arbitrary degrees, that is commands or code fragments that become invariant after a fixed number of iterations of the loop body. This technique was later used by Aubert, Rubiano, Rusch, and Seiller to automatically parallelise loops.<ref>{{cite book |last1=Aubert |first1=Clément |last2=Rubiano |first2=Thomas
|last3=Rusch |first3=Neea |last4=Seiller |first4=Thomas |chapter= Distributing and Parallelizing Non-canonical Loops |title= Verification, Model Checking, and Abstract Interpretation |series=Lecture Notes in Computer Science |date=2023 |volume=13881 |pages=91–108 |doi=10.1007/978-3-031-24950-1_1 }}</ref>

==Benefits==
Loop-invariant code which has been hoisted out of a loop is executed less often, providing a speedup. Another effect of this transformation is allowing constants to be stored in registers and not having to calculate the address and access the memory (or cache line) at each iteration.

However, if too many variables are created, there will be high [[register pressure]], especially on processors with few registers, like the 32-bit [[x86]]. If the compiler runs out of registers, some variables will be [[register spilling|spilled]]. To counteract this, the inverse optimization can be performed, [[rematerialization]].

==See also==
* [[Code motion]]
* [[Loop invariant]]

==Further reading==
* Aho, Alfred V.; Sethi, Ravi; & Ullman, Jeffrey D. (1986). Compilers: Principles, Techniques, and Tools. Addison Wesley. {{ISBN|0-201-10088-6}}.

==References==
{{Reflist}}

==External links==
* [https://web.archive.org/web/20170806083023/http://www.compileroptimizations.com/category/hoisting.htm Compiler Optimizations — Hoisting]

{{Compiler optimizations}}

[[Category:Compiler optimizations]]

Loop-invariant code motion - Revision history

imported>Jochen Burghardt: /* Invariant code detection */ the authors are active, not their paper (and shorter); rm 2nd wikilink (kept 1st one since a draft seems to be forthcoming)