Cross-compiling is taking a computer program and compiling it for a machine that isn’t the one hosting the compilation. Although historically compilers would only compile for the host machine, this is considered an anachronism: all serious native compilers are now cross-compilers.ers.

After all, you don’t want to be building your iPhone app on literal iPhone hardware.are.

Many different compilers have different mechanisms for classifying and identifying targets. A target is a platform that the compiler can produce executable code for. However, due to the runaway popularity of LLVM, virtually all compilers now use target triples. You may have already encountered one, such as the venerable x86_64-unknown-linux, or the evil x86_64-pc-windows. This system is convoluted and almost self-consistent.

But what is a target triple, and where did they come from?

Stupid GCC Conventions

So if you go poking around the Target Triplet page on OSDev, you will learn both true and false things about target triples, because this page is about GCC, not native compilers in general.

warning

Generally, there is no “ground truth” for what a target triple is. There isn’t some standards body that assigns these names. But as we’ll see, LLVM is the trendsetter.the trendsetter.

If you run the following command you can learn the target triple for your machine:

$ gcc -dumpmachine
x86_64-linux-gnu

Now if you’re at all familiar with any system that makes pervasive use of target triples, you will know that this is is not a target triple, because this target’s name is is x86_64-unknown-linux-gnu, which is what both clang and rustc call-

$ clang -dumpmachine
x86_64-pc-linux-gnu
$ rustc -vV | grep host
host: x86_64-unknown-linux-gnu

Oh no.

Well, GCC is missing the the pc or unknown component, and that’s specifically a GCC thing; it allows omitting parts of the triple in such a way that is unambiguous. And they are a GCC invention, so perhaps it’s best to start by assessing GCC’s beliefs.o;s beliefs.

According to GCC, a target triple is a string of the form --dor>-<os>. The “machine” part unambiguously identifies the architecture of the system. Practically speaking, this is the assembly language that the compiler will output at the end. The “vendor” part is essentially irrelevant, and mostly is of benefit for sorting related operating systems together. Finally, the “os” part identifies the operating system that this code is being compiled for. The main thing this identifies for a compiler is the executable format: COFF/PE for Windows, Mach-O for Apple’s operating systems, ELF for Linux and friends, and so on (this, however, is an oversimplification)., is an oversimplification).

But you may notice that x86_64-unknown-linux-gnu has an extra, fourth entry¹, which plays many roles but is most often called the target’s “ABI”. For rdquo;. For linux, it identifies the target’s libc, which has consequences for code generation of some language features, such as thread locals and unwinding. It is optional, since many targets only have one ABI.ABI.

Cross Compiling with GCC

A critical piece of history here is to understand the really stupid way in which GCC does cross compiling. Traditionally, each GCC binary would be built for one target triple. The full name of a GCC binary would include the triple, so when cross-compiling, you would compile with x86_64-unknown-linux-gcc, link with x86_64-unknown-linux-ld, and so on (here, gcc is not the fourth ABI component of a triple; it’s just one of the tools in the the x86_64-unknown-linux toolchain).

Nobody with a brain does this². LLVM and all cross compilers that follow it instead put all of the backends in one binary, and use a compiler flag like --target to select the backend.

But regardless, this is where target triples come from, and why they look the way they look: they began as prefixes for the names of binaries in autoconf scripts.

But GCC is ancient technology. In the 21st century, LLVM rules all native compilers.

Names in the Ancient Language

LLVM’s target triple list is the one that should be regarded as “most official”, for a few reasons:few reasons:

1. Inertia. Everyone and their mother uses LLVM as a middleend and backend, so its naming conventions bubble up into language frontends like clang, rustc swiftc, icc, and nvcc.

2. Upstream work by silicon and operating system vendors. LLVM is what people get hired to work on for the most part, not GCC, so its platform-specific conventions often reflect the preferences of vendors.

These are in no small part because Apple, Google, and Nvidia have armies of compiler engineers contributing to LLVM.

The sources for “official” target triples are many. Generally, I would describe a target triple as “official” when:ial” when:

1. A major compiler (so, clang or rustc) uses it. Rust does a way better job than LLVM of documenting their targets, so I prefer to give it deference. You can find Rust’s official triples les here.

2. A platform developer (e.g., a hardware manufacturer, OS vendor) distributes a toolchain with a target triple in the arch-vendor-os format.

So, what are the names in class (1)? LLVM does not really go out of its way to provide such a list. But we gotta start somewhere, so source-diving it is.

We can dig into Triple.cpp in LLVM’s target triple parser. It lists all of the names LLVM recognizes for each part of a triple. Looking at at Triple::parseArch(), we have the following names, including many, many aliases. The first item on the right column is LLVM’s preferred name for the architecture, as indicated by by Triple::getArchTypeName().

Here we begin to see that target triples are not a neat system. They are hell. Where a list of architecture names contains a “…”, it means that LLVM accepts many more names.y more names.

The problem is that architectures often have versions and features, which subtly change how the compiler generates code. For example, when compiling for an x86_64, we may want to specify that we want AVX512 instructions to be used. On LLVM, you might do that with -mattr=+avx512. Every architecture has a subtly-different way of doing this, because every architecture had a different GCC! Each variant of GCC would put different things behind -mXXX flags (-m for “machine”), meaning that the interface is not actually that uniform. The meanings of ings of -march, -mcpu, -mtune, and -mattr thus vary wildly for this reason.

Because LLVM is supposed to replace GCC (for the most part), it replicates a lot of this wacky behavior.

So uh, we gotta talk about 32-bit ARM architecture names.

ARMTargetParser.cpp

There is a hellish file in LLVM dedicated to parsing ARM architecture names. Although members of the ARM family have many configurable features (which you can discover with llc -march aarch64 -mattr help¹⁰), the name of the architecture is somewhat meaningful, and can hav many options, mostly relating to the many versions of ARM that exist.

How bad is it? Well, we can look at all of the various ARM targets that rustc supports with rustc --print target-list:

$ rustc --print target-list | grep -P 'arm|aarch|thumb'umb' \
  | cut -d- -f1 | sort | uniq
aarch64
aarch64_be
arm
arm64_32
arm64e
arm64ec
armeb
armebv7r
armv4t
armv5te
armv6
armv6k
armv7
armv7a
armv7k
armv7r
armv7s
armv8r
thumbv4t
thumbv5te
thumbv6m
thumbv7a
thumbv7em
thumbv7m
thumbv7neon
thumbv8m.base
thumbv8m.main

Most of these are 32-bit ARM versions, with profile information attached. These correspond to the names given here. Why does ARM stick version numbers in the architecture name, instead of using -mcpu like you would on x86 (e.g. -mcpu alderlake)? I have no idea, because ARM is not my strong suit. It’s likely because of how early ARM support was added to GCC.GCC.

Internally, LLVM calls these “subarchitectures”, although ARM gets special handling because there’s so many variants. SPIR-V, Direct X, and MIPS all have subarchitectures, so you might see something like ething like dxilv1.7 if you’re having a bad day.day.

Of course, LLVM’s ARM support also sports some naughty subarchitectures not part of this system, with naughty made up names.mes.

• arm64e is an Apple thing, which is an enhancement of aarch64 present on some Apple hardware, which adds their own flavor of pointer authentication and some other features.

• arm64ec is a completely unrelated Microsoft invention that is essentially “quo;aarch64 but with an x86_64-ey ABI” to make ake x86_64 emulation on what would otherwise be aarch64-pc-windows-msvc target somewhat more amenable.

note

Why the Windows people invented a whole other ABI instead of making things clean and simple like Apple did with Rosetta on ARM MacBooks? I have no idea, but http://www.emulators.com/docs/abc_arm64ec_explained.htm contains various excuses, none of which I am impressed by. My read is that their compiler org was just worse at life than Apple’s, which is not surprising, since Apple does compilers better than anyone else in the business.ess.

Actually, since we’re on the topic of the names of architectures, I have a few things I need to straighten out.out.

Made Up Names of Architectures

x86 and ARM both seem to attract a lot of people making up nicknames for them, which leads to a lot of confusion in:

1. What the “real” name is.name is.

2. What name a particular toolchain wants.

3. What name you should use in your own cosmopolitan tooling.

Let’s talk about the incorrect names people like to make up for them. Please consider the following a relatively normative reference on what people call these architectures, based on my own experience with many tools.ols.

When we say “x86” unqualified, in 2025, we almost always mean ys mean x86_64, because 32-bit x86 is dead. If you need to talk about 32-bit x86, you should either say “32-bit x86”, “protected mode”cted mode”¹¹, or “i386” (the first Intel microarchitecture that implemented protected mode)ed mode)¹². You should not call it x86_32 or just x86.

You might also call it IA-32 for Intel Architecture 32, (or ia32), but nobody calls it that and you risk confusing people with ia64, or IA-64, the official name of Intel’s failed general-purpose VLIW architecture, re, Itanium, which is in no way compatible with x86. ia64 was what GCC and LLVM named Itanium triples with. Itanium support was drowned in a bathtub during the Obama administration, so it’s not really relevant anymore. Rust has never had official Itanium support.ort.

32-bit x86 is extremely not called “x32”; this is what Linux used to call its x86 ILP3286 ILP32⁴ variant before it was removed (which, following the ARM names, would have been called x86_6432).

There are also many ficticious names for 64-bit x86, which you should avoid unless you want the younger generation to make fun of you. amd64 refers to AMD’s original implementation of long mode in their K8 microarchitecture, first shipped in their eir Athlon 64 product. AMD still makes the best x86 chips (I am writing this on a machine socketed with a Zen2 Threadripper), sure, but calling it amd64 is silly and also looks a lot like arm64, and I am honestly kinda annoyed at how much Go code I’ve seen with files named med fast_arm64.s and fast_amd64.s. Debian also uses amd64/arm64, which makes browsing packages kind of annoying.

On that topic, you should absolutely not call 64-bit mode k8, after the AMD K8. Nobody except for weird computer taxonomists like me know what that is. But Bazel calls it that, and it’s really irritatingting¹³.

You should also not call it x64. Although LLVM does accept amd64 for historical purposes, no one calls it x64 except for Microsoft. And even though it is fairly prevalent on Windows, I absolutely give my gamedev friends a hard time when they write x64.

On the ARM side, well. Arm¹⁴ has a bad habit of not using consistent naming for 64-bit ARM, since they used both AArch64 and ARM64 for it. However, in compiler land, aarch64 appears to be somewhat more popular.

You should also probably stick to the LLVM names for the various architectures, instead of picking your favorite Arm Cortex name (like cortex_m0).

Vendors and Operating Systems

The worst is over. Let’s now move onto examinining the rest of the triple: the platform vendor, and the operating system.tem.

The vendor is intended to identify who is responsible for the ABI definition for that target. Although provides little to no value to the compiler itself, but it does help to sort related targets together. Sort of.

Returning to llvm::Triple, we can examine Triple::VendorType. Vendors almost always correspond to companies which develop operating systems or other platforms that code runs on, with some exceptions.

We can also get the vendors that rustc knows about with a handy dandy command:

rustc --print target-list | grep -P '\w+-\w+-' | cut -d- -f2 | sort | uniq
 | uniq

The result is this. This is just a representative list; I have left off a few that are not going to be especially recognizeable.

Most vendors are the names of organizations that produce hardware or operating systems. For example suse and redhat are used for those organizations’ Linux distributions, as a funny branding thing. Some vendors are projects, like the the mesa vendor used with the Mesa3D OpenGL implementation’s triples.les.

The unknown vendor is used for cases where the vendor is not specified or just not important. For example, the canonical Linux triple is x86_64-unknown-linux… although one could argue it should be d be x86_64-torvalds-linux. It is not uncommon for companies that sell/distribute Linux distributions to have their own target triples, as do SUSE and sometimes RedHat. Notably, there are no triples with a google vendor, even though aarch64-linux-android and aarch64-unknown-fuchsia should really be called aarch64-google-linux-android and aarch64-google-fuchsia. The target triple system begins to show cracks here.

The pc vendor is a bit weirder, and is mostly used by Windows targets. The standard Windows target is x86_64-pc-windows-msvc, but really it should have been x86_64-microsoft-windows-msvc. This is likely complicated by the fact that there is also a x86_64-pc-windows-gnu triple, which is for MinGW code. This platform, despite running on Windows, is not provided by Microsoft, so it would probably make more sense to be called x86_64-unknown-windows-gnu.

But not all Windows targets are pc! UWP apps use a different triple, that replaces the pc with uwp. rustc provides targets for Windows 7 backports that use a win7 “vendor”.”.

Beyond Operating Systems

The third (or sometimes second, ugh) component of a triple is the operating system, or just “system”, since it’s much more general than that. The main thing that compilers get from this component relates to generating code to interact with the operating system (e.g. SEH on Windows) and various details related to linking, such as object file format and relocations.relocations.

It’s also used for setting defines like ike __linux__ in C, which user code can use to determine what to do based on the target.

We’ve seen een linux and windows, but you may have also seen x86_64-apple-darwin. Darwin?

The operating system formerly known as Mac OS X (now macOS¹⁶) is a POSIX operating system. The POSIX substrate that all the Apple-specific things are built on top of is called Darwin. Darwin is a free and open source operating system based on Mach, a research kernel whose name survives in Mach-O, the object file format used by all Apple products.

All of the little doodads Apple sells use the actual official names of their OSes, like aarch64-apple-ios. For, you know, iOS. On your iPhone. Built with Xcode on your iMac.

none is a common value for this entry, which usually means a free-standing environment with no operating system. The object file format is usually specified in the fourth entry of the triple, so you might see something like riscv32imc-unknown-none-elf.

Sometimes the triple refers not to an operating system, but to a complete hardware product. This is common with game console triples, which have “operating system” names like es like ps4, psvita, 3ds, and switch. (Both Sony and Nintendo use LLVM as the basis for their internal toolchains; the Xbox toolchain is just MSVC).

ABI! ABI!

The fourth entry of the triple (and I repeat myself, yes, it’s still a triple) represents the binary interface for the target, when it is ambiguous.ous.

For example, Apple targets never have this, because on an Apple platform, you just shut up and use CoreFoundation.framework as your libc. Except this isn’t true, because of things like ike x86_64-apple-ios-sim, the iOS simulator running on an x86 host.

On the other hand, Windows targets will usually specify -msvc or -gnu, to indicate whether they are built to match MSVC’s ABI or MinGW. Linux targets will usually specify the libc vendor in this position: on: -gnu for glibc, -musl for musl, -newlib for newlib, and so on.

This doesn’t just influence the calling convention; it also influences how language features, such as thread locals and dynamic linking, are handled. This usually requires coordination with the target libc.ibc.

On ARM free-standing (armxxx-unknown-none) targets, -eabi specifies the ARM EABI, which is a standard embeded ABI for ARM. -eabihf is similar, but indicates that no soft float support is necessary (hf stands for hardfloat). (Note that Rust does not include a vendor with these architectures, so they’re more like ike armv7r-none-eabi).

A lot of jankier targets use the ABI portion to specify the object file, such as the aforementioned riscv32imc-unknown-none-elf.

WASM Targets

One last thing to note are the various WebAssembly targets, which completely ignore all of the above conventions. Their triples often only have two components (they are still called triples, hopefully I’ve made that clear by now). Rust is a little bit more on the forefront here than han clang (and anyways I don’t want to get into Emscripten) so I’ll stick to what’s going on in going on in rustc.

There’s a few variants. ts. wasm32-unknown-unknown (here using unknown instead of none as the system, oops) is a completely bare WebAssebly runtime where none of the standard library that needs to interact with the outside world works. This is essentially for building WebAssembly modules to deploy in a browser.

There are also the WASI targets, which provide a standard ABI for talking to the host operating system. These are less meant for browsers and more for people who are using WASI as a security boundary. These have names like wasm32-wasip1, which, unusually, lack a vendor! A “more correct” formulation would have been ve been wasm32-unknown-wasip1.

Aside on Go

Go does the correct thing and distributes a cross compiler. This is well and good.

Unfortunately, they decided to be different and special and do not use the target triple system for naming their targets. Instead, you set the GOARCH and GOOS environment variables before invoking gc. This will sometimes be shown printed with a slash between, such as linux/amd64.

Thankfully, they at least provide documentation for a relevant internal package here, which offers the names of various GOARCH and GOOS values.

They use completely different names from everyone else for a few things, which is guaranteed to trip you up. They use call the 32- and 64-bit variants of x86 386 (note the lack of leading i) and amd64. They call 64-bit ARM arm64, instead of aarch64. They call little-endian MIPSes mipsle instead of mipsel.

They also call 32-bit WebAssembly wasm instead of wasm32, which is a bit silly, and they use js/wasm as their equivalent of wasm32-unknown-unknown, which is very silly.

Android is treated as its own operating system, android, rather than being linux with a particular ABI; their system also can’t account for ABI variants in general, since Go originally wanted to not have to link any system libraries, something that does not actually work.ork.

If you are building a new toolchain, don’t be clever by inventing a cute target triple convention. All you’ll do is annoy people who need to work with a lot of different toolchains by being different and special.special.

Inventing Your Own Triples

Realistically, you probably shouldn’t. But if you must, you should probably figure out what you want out of the triple.ple.

Odds are there isn’t anything interesting to put in the vendor field, so you will avoid people a lot of pain by picking ing unknown. Just include a vendor to avoid pain for people in the future.

You should also avoid inventing a new name for an existing architecture. Don’t name your hobby operating system’s triple triple amd64-unknown-whatever, please. And you definitely don’t want to have an ABI component. One ABI is enough.ugh.

If you’re inventing a triple for a free-standing environment, but want to specify something about the hardware configuration, you’re probably gonna want to use to use -none-bi> for your system. For some firmware use-cases, though, the system entry is a better place, such as for the UEFI triples. Although, I have unforunately seen both x86_64-unknown-uefi and x86_64-pc-none-uefi in the wild.

And most imporantly: this sytem was built up organically. Disabuse yourself now of the idea that the system is consistent and that target triples are easy to parse. Trying to parse them will make you very sad.