Optimizations: the speed size tradeoff

Everyone wants their program to be super fast and super small but it's usually not possible to have both characteristics. This section discusses the different optimization levels that rustc provides and how they affect the execution time and binary size of a program.

No optimizations

This is the default. When you call cargo build you use the development (AKA dev) profile. This profile is optimized for debugging so it enables debug information and does not enable any optimizations, i.e. it uses -C opt-level = 0.

At least for bare metal development, debuginfo is zero cost in the sense that it won't occupy space in Flash / ROM so we actually recommend that you enable debuginfo in the release profile -- it is disabled by default. That will let you use breakpoints when debugging release builds.

[profile.release]
# symbols are nice and they don't increase the size on Flash
debug = true

No optimizations is great for debugging because stepping through the code feels like you are executing the program statement by statement, plus you can print stack variables and function arguments in GDB. When the code is optimized, trying to print variables results in $0 = <value optimized out> being printed.

The biggest downside of the dev profile is that the resulting binary will be huge and slow. The size is usually more of a problem because unoptimized binaries can occupy dozens of KiB of Flash, which your target device may not have -- the result: your unoptimized binary doesn't fit in your device!

Can we have smaller, debugger friendly binaries? Yes, there's a trick.

Optimizing dependencies

There's a Cargo feature named profile-overrides that lets you override the optimization level of dependencies. You can use that feature to optimize all dependencies for size while keeping the top crate unoptimized and debugger friendly.

Beware that generic code can sometimes be optimized alongside the crate where it is instantiated, rather than the crate where it is defined. If you create an instance of a generic struct in your application and find that it pulls in code with a large footprint, it may be that increasing the optimisation level of the relevant dependencies has no effect.

Here's an example:

# Cargo.toml
[package]
name = "app"
# ..

[profile.dev.package."*"] # +
opt-level = "z" # +

Without the override:

$ cargo size --bin app -- -A
app  :
section               size        addr
.vector_table         1024   0x8000000
.text                 9060   0x8000400
.rodata               1708   0x8002780
.data                    0  0x20000000
.bss                     4  0x20000000

With the override:

$ cargo size --bin app -- -A
app  :
section               size        addr
.vector_table         1024   0x8000000
.text                 3490   0x8000400
.rodata               1100   0x80011c0
.data                    0  0x20000000
.bss                     4  0x20000000

That's a 6 KiB reduction in Flash usage without any loss in the debuggability of the top crate. If you step into a dependency then you'll start seeing those <value optimized out> messages again but it's usually the case that you want to debug the top crate and not the dependencies. And if you do need to debug a dependency then you can use the profile-overrides feature to exclude a particular dependency from being optimized. See example below:

# ..

# don't optimize the `cortex-m-rt` crate
[profile.dev.package.cortex-m-rt] # +
opt-level = 0 # +

# but do optimize all the other dependencies
[profile.dev.package."*"]
codegen-units = 1 # better optimizations
opt-level = "z"

Now the top crate and cortex-m-rt are debugger friendly!

Optimize for speed

As of 2018-09-18 rustc supports three "optimize for speed" levels: opt-level = 1, 2 and 3. When you run cargo build --release you are using the release profile which defaults to opt-level = 3.

Both opt-level = 2 and 3 optimize for speed at the expense of binary size, but level 3 does more vectorization and inlining than level 2. In particular, you'll see that at opt-level equal to or greater than 2 LLVM will unroll loops. Loop unrolling has a rather high cost in terms of Flash / ROM (e.g. from 26 bytes to 194 for a zero this array loop) but can also halve the execution time given the right conditions (e.g. number of iterations is big enough).

Currently there's no way to disable loop unrolling in opt-level = 2 and 3 so if you can't afford its cost you should optimize your program for size.

Optimize for size

As of 2018-09-18 rustc supports two "optimize for size" levels: opt-level = "s" and "z". These names were inherited from clang / LLVM and are not too descriptive but "z" is meant to give the idea that it produces smaller binaries than "s".

If you want your release binaries to be optimized for size then change the profile.release.opt-level setting in Cargo.toml as shown below.

[profile.release]
# or "z"
opt-level = "s"

These two optimization levels greatly reduce LLVM's inline threshold, a metric used to decide whether to inline a function or not. One of Rust principles are zero cost abstractions; these abstractions tend to use a lot of newtypes and small functions to hold invariants (e.g. functions that borrow an inner value like deref, as_ref) so a low inline threshold can make LLVM miss optimization opportunities (e.g. eliminate dead branches, inline calls to closures).

When optimizing for size you may want to try increasing the inline threshold to see if that has any effect on the binary size. The recommended way to change the inline threshold is to append the -C inline-threshold flag to the other rustflags in .cargo/config.toml.

# .cargo/config.toml
# this assumes that you are using the cortex-m-quickstart template
[target.'cfg(all(target_arch = "arm", target_os = "none"))']
rustflags = [
  # ..
  "-C", "inline-threshold=123", # +
]

What value to use? As of 1.29.0 these are the inline thresholds that the different optimization levels use:

  • opt-level = 3 uses 275
  • opt-level = 2 uses 225
  • opt-level = "s" uses 75
  • opt-level = "z" uses 25

You should try 225 and 275 when optimizing for size.