"This article was originally published in Polish in issue 4/2013" — a lot of excellent advice. Sad to see C++ have moved in last decade in a direction that makes writing efficient, simple low level code harder and harder :(
There are so many things that are expressible in C++ now that could not be without writing much more code or using per-compilation tools back then. The ability to run code at compile time that is not run at runtime is huge, #embed lets us make other tools output available without linker scripts or compiler specific tools that.
Also, most of the code from the past still works(from 10 years ago definitely works)
Shot in the dark, but maybe the OP is referring to the fact that these code conventions are explicitly discouraged by the C++ core guidelines. The SoA example falls afoul of the rule requiring T* to be used only for singular object pointers, for example.
Not a regular C++ programmer but wouldn’t you use std::span here instead? Sure it’ll carry a few redundant lengths but it makes using functions that take spans easier. When I do write C++ it’s usually for speed so I’m often working at the intrinsics level, though AI has gotten good enough at it that I now generally delegate this work to an agent.
eh, yes-ish, but only thanks to compiler writers. at one point committee went all-out enforcing their lifetime model, making bit_cast not an option. If your code accesses same data using different types, you are spelunking ruins with snake pits and lava. more and more stuff needs magic support code in std::, making no-lib code less and less possible (it used to be that with no-rtti and no-exceptions, you could use all c++ features and only needed cxa_at_exit, operator delete, and few other little things. NOT ANY MORE).
On one hand, you have consteval and stuff, letting you FINALLY initialize data at compile time (hey, 20 years late but still!)
on other hand, it is done in most non-debuggable way possible. try setting breakpoint or adding print to constexpr function that causes your requires clause to fail...
so no, newer C++ the language is not possible to use for low level work. The dialects that compiler makers support are. We will see for how long
I think it has become EASIER: for instance, since C++23 Rust-like move semantics can be used, which provides the compiler with extra information that can be leveraged for the generation of better code.
Or take constexpr - it permits to move computations to compile time that are complex and in older versions either had to be done at runtime, or an ugly workaround had to be used (e.g. assigning a mysterious literal pre-computed in another run or by hand).
The closest thing I can think of is trivial relocation [0] which matches the bitwise copy + no destructor on moved-from bits of Rust's moves, but that was only added to the draft for C++26 and was removed late in the process anyways [1].
Writing clear, concise, and efficient code in C++ has never been simpler or easier. The improvements in C++ over the last 15 years have been qualitative.
So many complex, esoteric, and difficult to maintain incantations that used to be required for efficient code generation are no longer necessary.
It is probably very domain specific. In robotics for example everything is a zero sum game: CPU, memory bandwidth, GPU, battery life etc ... So it is really a topic, probably true for anything embedded actually. Some other offline applications: HFT, Telco etc..
I wish the GUI apps devs respect more the laptop resources they are running on, don't get me started on the 4 instances of chrome I need to run just for discord, signal etc ...
True, but moving from a list of unique polymorphic pointers to a std::variant gains you at least a 2-3x speed up in terms of TLB and cacheline locality. From there, swapping to SOA will net you another 4-8x, so you're looking at nearly 25x improvement by going data first. That may not matter in the unique case of say, games, where rendering a million entities will dwarf the cost of SIMD processing a million entities, but in something like numerical simulations (fluids) or quant it will be warmly welcomed
That, and it frees the compiler from reasoning about virtual inlining, and that the std::variant approach can pack potentially more than one object into a single cacheline. TLBs also work with 4096 byte pages, so 32 polymorphic 128 byte entities may (at the absolute worst case) use 32 distinct pages which requires 32 TLB virtual translations, while the std::variant one uses 1.
The next step of going SOA benefits from all of the above, it just further unlocks you packed quad and oct instructions (AVX256 and 512 depending if you buy AMD or not).
If you take the linked benchmark and use the latest compiler version, the std::variant version is faster. The annoying thing about std::variant (and with some other features of modern C++) is that it generates a bunch of code that the compiler has to optimize away.
Somehow and for some reason rust enums don’t have this problem and are far more ergonomic and easier to work with (not to mention compile times are amazing). I don’t know exactly why it’s better to have it as a first class language primitive and why the compiler has such a problem with std::visit, but clearly c++ meta programming slows down things in a super linear way such that the compiler has problems both from code gen and then optimization.
When writing code for end-user applications, I think it's mostly true. When it's writing code for a database engine, a game engine, a 3d renderer, or anything else that involves heavy data processing, optimization is the core "thing" often and it might not even be a good enough solution without it. Although, a lot of time even then C++ is good enough even then when picking reasonable data structures to represent the data.
I started writing a [CPU-only] 3-D rendering library in C++ recently, after having written the equivalent in C as a proof-of-concept and an experiment. The reason I decided to write it in C++ after C, is not only because I wanted to tap into meta-programming which is facilitated much better with C++, or that I wanted niceties like procedure overloading, but because some things with C or C++ aren't automagically optimised -- like if you want to leverage struct-of-array (SoA) memory layouts because it allows fewer SIMD (AVX in my case) instructions in the rendering pipeline. You do _not_ get that "for free" just writing a single procedure in C++, much less with C. Both languages are layout-sensitive, I mean this is in part what gives you the speed -- optimising with memory layout for cache locality etc. But you have to do it yourself. Meaning that if you need array-of-struct (AoS) or in fact don't know which path the CPU would prefer, there's no other way than roll up your sleeves and one way or another implement both.
The kicker is, in my case I chose C++ because templates allow me to reuse most of the code in the rendering pipeline _regardless_ of whether I go for AoS or SoA layout. I leverage operator overloading to do vector by matrix multiplication which is implemented in both variants. I do have to specify the desired variant during building, but I've profiled and for Intel x86 AVX in my case SoA is something like twice as efficient because I process ("shade") 8 vertices with 4-5 instructions instead of 1 vertex at a time (still shaded with vectorisation -- just "rotated", i.e in the pipeline axis and not vertex buffer axis).
TL;DR; C++ gives you plenty fast by default, but it's not always enough. The difference between 5 and 15 frames per second, well, makes all the difference -- our eyes are only fooled once the frames-per-second rate goes sufficiently up, anything below an acceptable threshold and it's completely different experience. You then either sacrifice resolution or level of detail etc, or decide to squeeze more from the language by helping the compiler.
Definitely need to optimize a bit for games and huge scale web apps. We've been finding big optimizations in our app recently. App works without them because we can scale horizontally but cutting CPU usage by 30% by eliminating redundant work and reducing copies of big objects? Why wouldn't we want to do that? This isn't even fancy algorithm stuff, mostly just shoddy initial implementations by 100s of eng working on a codebase over 7 years (not even that old). Stuff like that creeps in.
And the code might be faster, too, like in that 2005 series of articles by Raymond Chen and Rico Mariani (in which one of them wrote a program in C++ and the other wrote the same program in C#).
There's no mention of branch prediction, or context switching, or synchronisation. Depending on what you're doing, they could be very consequential. There's only very brief mention of parallelisation with threads and with SIMD.
High-performance programming is a big topic. The scope is far too broad for a single blog post, which naturally gives only cursory discussion of C++ and computer architecture. The article isn't bad considering, but I do think it's the wrong format. A blog series, or even a book, would be more fitting.
Learn which instructions SIMD nicely (sqrt / fabs, etc). Use ternaries in loops for masking. Use trig identities and lookup tables (don't recompute sin(3t) when you can use two vector multiples using a table of sin(t) eg. sin(t) * sin(t) * sin(t)). Use divisible constexpr constants in loops to eliminate the SIMD tail. Be careful with type casts and floats. `float x; x += 0.5` will introduce *cvt instructions even if the compiler statically knew better otherwise (use 0.5f). Compile with --fast-math and friends so errno doesn't invalidate your SIMD pipeline.
That has a similar problem to the article, it's trying to fit far too much into too small a format.
What you've written mostly makes sense to someone who already has a solid understanding of SIMD and of C++ (although I can't say I follow all of it), but the target audience is people who don't. For them, each point needs a much lengthier explanation.
Likely the best tip would to `objdump -d` and inspect the assembly then checking performance counters. Prepending (__attribute__((used)) will allow you to inspect your functions.
A quick restrict example:
#define fn __attribute__((used))
fn void copy1(int* to, const int* from, const int size)
{
for(int i = 0; i < size; i++)
to[i] = from[i];
}
fn void copy2(int* to, const int* from)
{
constexpr int size = 1024;
for(int i = 0; i < size; i++)
to[i] = from[i];
}
fn void copy3(int* restrict to, const int* restrict from)
{
constexpr int size = 1024;
for(int i = 0; i < size; i++)
to[i] = from[i];
}
gcc test.c -c -O3 && objdump -d ./test.o
copy1 is 52 lines, copy2 is 28 lines, copy3 is 2 lines (just a call to memcpy).
This is a good starting point for self teaching. The impact of your TLB, L1, and overall instruction count (with IPC) can further be measured with `./perf stat -d -d -d ./a.out`. If you want a quick rule of thumb, no instructions are fast instructions.
I've not read Fog's Optimizing software in C++ but I see it's freely available there as a PDF (182 pages). Looks like a great resource on these topics.
creata's comment [0] mentions the works of Agner Fog, which seem very good, and are freely available.
I haven't read C++ High Performance[1] but it looks like it covers the sorts of topics you'd expect, although it looks like it doesn't cover computer architecture in detail e.g. branch prediction. There are books on that too, of course.
My latest C++ project is assessment engine covering various actuarial type things like calculates risk for insurance etc. Typical performance for bulk calculation reaches millions to 10s of millions assessments per second on 16 core server. Well there is a trick there that inside it JIT compiles rules from a DSL to an executable code. interpreter mode (used mainly for audit mode) is about 3-5 times slower which is still insanely fast
This article reminds me of performance advice I was starting to see in the 2000s decade. Basically it was to not introduce a bunch of pointer heavy data structures to get lower algorithmic complexity. Stuff it all into a vector. You will use some algorithms that the computer science textbook will say it's slower, but if it fits all in cache it doesn't matter. The cache misses following pointers all over town hurts you more.
"This article was originally published in Polish in issue 4/2013" — a lot of excellent advice. Sad to see C++ have moved in last decade in a direction that makes writing efficient, simple low level code harder and harder :(
How? you can write exactly the same low level code today.
My thought too.
There are so many things that are expressible in C++ now that could not be without writing much more code or using per-compilation tools back then. The ability to run code at compile time that is not run at runtime is huge, #embed lets us make other tools output available without linker scripts or compiler specific tools that.
Also, most of the code from the past still works(from 10 years ago definitely works)
Shot in the dark, but maybe the OP is referring to the fact that these code conventions are explicitly discouraged by the C++ core guidelines. The SoA example falls afoul of the rule requiring T* to be used only for singular object pointers, for example.
Not a regular C++ programmer but wouldn’t you use std::span here instead? Sure it’ll carry a few redundant lengths but it makes using functions that take spans easier. When I do write C++ it’s usually for speed so I’m often working at the intrinsics level, though AI has gotten good enough at it that I now generally delegate this work to an agent.
Depending on the specific code, the compiler may even eliminate the redundant lengths.
eh, yes-ish, but only thanks to compiler writers. at one point committee went all-out enforcing their lifetime model, making bit_cast not an option. If your code accesses same data using different types, you are spelunking ruins with snake pits and lava. more and more stuff needs magic support code in std::, making no-lib code less and less possible (it used to be that with no-rtti and no-exceptions, you could use all c++ features and only needed cxa_at_exit, operator delete, and few other little things. NOT ANY MORE).
On one hand, you have consteval and stuff, letting you FINALLY initialize data at compile time (hey, 20 years late but still!)
on other hand, it is done in most non-debuggable way possible. try setting breakpoint or adding print to constexpr function that causes your requires clause to fail...
so no, newer C++ the language is not possible to use for low level work. The dialects that compiler makers support are. We will see for how long
I think it has become EASIER: for instance, since C++23 Rust-like move semantics can be used, which provides the compiler with extra information that can be leveraged for the generation of better code.
Or take constexpr - it permits to move computations to compile time that are complex and in older versions either had to be done at runtime, or an ugly workaround had to be used (e.g. assigning a mysterious literal pre-computed in another run or by hand).
What C++23 feature allows that?
Maybe this one?
https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2022/p22...
I think that's basically clarifying the conditions under which C++11-style moves can be performed.
The closest thing I can think of is trivial relocation [0] which matches the bitwise copy + no destructor on moved-from bits of Rust's moves, but that was only added to the draft for C++26 and was removed late in the process anyways [1].
[0]: https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2025/p27...
[1]: https://herbsutter.com/2025/11/10/trip-report-november-2025-...
Writing clear, concise, and efficient code in C++ has never been simpler or easier. The improvements in C++ over the last 15 years have been qualitative.
So many complex, esoteric, and difficult to maintain incantations that used to be required for efficient code generation are no longer necessary.
How do you process read-only mmaped data in C++, in accordance with the language rules? As an example.
I write in C++ almost every day but never have the need to optimize for speed. Even when you write straightforward code it's already blazingly fast.
I rarely use C++ but when I do it is for speed. It’s not uncommon that carefully crafted intrinsics can 10x the straightforward naive implementation.
It is probably very domain specific. In robotics for example everything is a zero sum game: CPU, memory bandwidth, GPU, battery life etc ... So it is really a topic, probably true for anything embedded actually. Some other offline applications: HFT, Telco etc.. I wish the GUI apps devs respect more the laptop resources they are running on, don't get me started on the 4 instances of chrome I need to run just for discord, signal etc ...
True, but moving from a list of unique polymorphic pointers to a std::variant gains you at least a 2-3x speed up in terms of TLB and cacheline locality. From there, swapping to SOA will net you another 4-8x, so you're looking at nearly 25x improvement by going data first. That may not matter in the unique case of say, games, where rendering a million entities will dwarf the cost of SIMD processing a million entities, but in something like numerical simulations (fluids) or quant it will be warmly welcomed
Is the improvement from using std:variant vs polymorphism just due to the indirection you save on?
That, and it frees the compiler from reasoning about virtual inlining, and that the std::variant approach can pack potentially more than one object into a single cacheline. TLBs also work with 4096 byte pages, so 32 polymorphic 128 byte entities may (at the absolute worst case) use 32 distinct pages which requires 32 TLB virtual translations, while the std::variant one uses 1.
The next step of going SOA benefits from all of the above, it just further unlocks you packed quad and oct instructions (AVX256 and 512 depending if you buy AMD or not).
are you sure?
https://stackoverflow.com/questions/69444641/c17-stdvariant-...
If you take the linked benchmark and use the latest compiler version, the std::variant version is faster. The annoying thing about std::variant (and with some other features of modern C++) is that it generates a bunch of code that the compiler has to optimize away.
Somehow and for some reason rust enums don’t have this problem and are far more ergonomic and easier to work with (not to mention compile times are amazing). I don’t know exactly why it’s better to have it as a first class language primitive and why the compiler has such a problem with std::visit, but clearly c++ meta programming slows down things in a super linear way such that the compiler has problems both from code gen and then optimization.
To add to sibling comment...
GCC 11 (2021) std::visit was slower than virtual dispatch.
GCC 12 (2022) optimized std::visit so it can be faster than virtual dispatch.
https://shubhankar-gambhir.github.io/posts/your-stdlib-imple...
When writing code for end-user applications, I think it's mostly true. When it's writing code for a database engine, a game engine, a 3d renderer, or anything else that involves heavy data processing, optimization is the core "thing" often and it might not even be a good enough solution without it. Although, a lot of time even then C++ is good enough even then when picking reasonable data structures to represent the data.
I started writing a [CPU-only] 3-D rendering library in C++ recently, after having written the equivalent in C as a proof-of-concept and an experiment. The reason I decided to write it in C++ after C, is not only because I wanted to tap into meta-programming which is facilitated much better with C++, or that I wanted niceties like procedure overloading, but because some things with C or C++ aren't automagically optimised -- like if you want to leverage struct-of-array (SoA) memory layouts because it allows fewer SIMD (AVX in my case) instructions in the rendering pipeline. You do _not_ get that "for free" just writing a single procedure in C++, much less with C. Both languages are layout-sensitive, I mean this is in part what gives you the speed -- optimising with memory layout for cache locality etc. But you have to do it yourself. Meaning that if you need array-of-struct (AoS) or in fact don't know which path the CPU would prefer, there's no other way than roll up your sleeves and one way or another implement both.
The kicker is, in my case I chose C++ because templates allow me to reuse most of the code in the rendering pipeline _regardless_ of whether I go for AoS or SoA layout. I leverage operator overloading to do vector by matrix multiplication which is implemented in both variants. I do have to specify the desired variant during building, but I've profiled and for Intel x86 AVX in my case SoA is something like twice as efficient because I process ("shade") 8 vertices with 4-5 instructions instead of 1 vertex at a time (still shaded with vectorisation -- just "rotated", i.e in the pipeline axis and not vertex buffer axis).
TL;DR; C++ gives you plenty fast by default, but it's not always enough. The difference between 5 and 15 frames per second, well, makes all the difference -- our eyes are only fooled once the frames-per-second rate goes sufficiently up, anything below an acceptable threshold and it's completely different experience. You then either sacrifice resolution or level of detail etc, or decide to squeeze more from the language by helping the compiler.
Then you don't work on a product that has any scale <shrug>.
Definitely need to optimize a bit for games and huge scale web apps. We've been finding big optimizations in our app recently. App works without them because we can scale horizontally but cutting CPU usage by 30% by eliminating redundant work and reducing copies of big objects? Why wouldn't we want to do that? This isn't even fancy algorithm stuff, mostly just shoddy initial implementations by 100s of eng working on a codebase over 7 years (not even that old). Stuff like that creeps in.
This is a sign you might be more productive in a higher-level language.
And the code might be faster, too, like in that 2005 series of articles by Raymond Chen and Rico Mariani (in which one of them wrote a program in C++ and the other wrote the same program in C#).
https://devblogs.microsoft.com/oldnewthing/20060731-15/?p=30...
https://learn.microsoft.com/en-us/archive/blogs/ricom/perfor...
This follows Rob Pike’s 5 Rules for Programming
https://web.archive.org/web/20250201145327/https://users.ece...
There's no mention of branch prediction, or context switching, or synchronisation. Depending on what you're doing, they could be very consequential. There's only very brief mention of parallelisation with threads and with SIMD.
High-performance programming is a big topic. The scope is far too broad for a single blog post, which naturally gives only cursory discussion of C++ and computer architecture. The article isn't bad considering, but I do think it's the wrong format. A blog series, or even a book, would be more fitting.
Learn which instructions SIMD nicely (sqrt / fabs, etc). Use ternaries in loops for masking. Use trig identities and lookup tables (don't recompute sin(3t) when you can use two vector multiples using a table of sin(t) eg. sin(t) * sin(t) * sin(t)). Use divisible constexpr constants in loops to eliminate the SIMD tail. Be careful with type casts and floats. `float x; x += 0.5` will introduce *cvt instructions even if the compiler statically knew better otherwise (use 0.5f). Compile with --fast-math and friends so errno doesn't invalidate your SIMD pipeline.
Most applications (including most applications that care about numerical performance) should not use -ffast-math.
That has a similar problem to the article, it's trying to fit far too much into too small a format.
What you've written mostly makes sense to someone who already has a solid understanding of SIMD and of C++ (although I can't say I follow all of it), but the target audience is people who don't. For them, each point needs a much lengthier explanation.
Likely the best tip would to `objdump -d` and inspect the assembly then checking performance counters. Prepending (__attribute__((used)) will allow you to inspect your functions.
A quick restrict example:
copy1 is 52 lines, copy2 is 28 lines, copy3 is 2 lines (just a call to memcpy).
This is a good starting point for self teaching. The impact of your TLB, L1, and overall instruction count (with IPC) can further be measured with `./perf stat -d -d -d ./a.out`. If you want a quick rule of thumb, no instructions are fast instructions.
This seems like domain specific advice.
They're a bit old and missing some details, but I like Agner Fog's manuals.
https://www.agner.org/optimize/
I've not read Fog's Optimizing software in C++ but I see it's freely available there as a PDF (182 pages). Looks like a great resource on these topics.
https://www.agner.org/optimize/optimizing_cpp.pdf
Do you have any recommended essential reading for this?
I'm no expert in this stuff but:
creata's comment [0] mentions the works of Agner Fog, which seem very good, and are freely available.
I haven't read C++ High Performance [1] but it looks like it covers the sorts of topics you'd expect, although it looks like it doesn't cover computer architecture in detail e.g. branch prediction. There are books on that too, of course.
[0] https://news.ycombinator.com/item?id=49868657
[1] https://www.packtpub.com/en-us/product/c-high-performance-97...
My latest C++ project is assessment engine covering various actuarial type things like calculates risk for insurance etc. Typical performance for bulk calculation reaches millions to 10s of millions assessments per second on 16 core server. Well there is a trick there that inside it JIT compiles rules from a DSL to an executable code. interpreter mode (used mainly for audit mode) is about 3-5 times slower which is still insanely fast
i heard a lot of AI and LLM is in python?
You heard correctly. This discussion is about C++ though.
Yes, but the numerical backend of the python libraries are written in lower-level languages. For example, pytorch uses a C++ backend.
Python is often used as (very useful!) glue for calling CUDA, C, C++, Fortran, etc… codes.
So, if you are thinking about the sort of “business logic” that’s often Python, but the performance comes from the parts that are usually not.
This article reminds me of performance advice I was starting to see in the 2000s decade. Basically it was to not introduce a bunch of pointer heavy data structures to get lower algorithmic complexity. Stuff it all into a vector. You will use some algorithms that the computer science textbook will say it's slower, but if it fits all in cache it doesn't matter. The cache misses following pointers all over town hurts you more.