Genetic Improvement of LLVM Intermediate Representation

Abstract

Evolving LLVM IR is widely applicable, with LLVM Clang offering support for an increasing range of computer hardware and programming languages. Local search mutations are used to hill climb industry C code released to support geographic open standards: Open Location Code (OLC) from Google and Uber's Hexagonal Hierarchical Spatial Index (H3), giving up to two percent speed up on compiler optimised code.

Genetic Improvement of OLC and H3 with Magpie

“…We use four open source C/C++ examples. Two industrial geospatial programs, both written in C [8]. One from Google's OLC and the other from Uber's H3 (see Section 3.1).…”

“…We used the same test cases as before [8] when optimising OLC and H3. Both Google's Open Location Code (OLC) https://github.com/ google/open-location-code (downloaded 4 August 2022) and Uber's Hexagonal Hierarchical Geospatial Indexing System (H3) https: //github.com/uber/h3 (downloaded the previous day) are open industry standards (total sizes OLC 14 024 and H3 15 015 lines of source code).…”

“…For OLC we used Google's 16 character coding and for H3 we used Uber's highest resolution (-r 15) which uses 15 characters. Following our earlier work [8], we use as test cases the position of actual addresses.…”

“…For Google's OLC we used the [8] dataset which was the location of the first ten thousand GB postcodes downloaded from https:// www.getthedata.com/downloads/open_postcode_geo.csv.zip (dated 16 March 2022). The addresses are alphabetically sorted starting with AB1 0AA, which is in Aberdeen.…”

“…Magpie generated 700 OLC, 19 077 H3, 904 Blue and 3151 SEEDS patches (see Table 1). The OLC and H3 (700 and 19077) values are taken from our previous work [8]. As before [8], we use a coupon collector [3] argument to calculate how many random samples would be needed to be almost certain of visiting every of the C/C++ source code at least once.…”