Hi Muneem,

Can you post how the code would look like with the feature you are
proposing?

You don't need to try to teach everyone about how compilers work. Some of
the things you say are just incorrect, so this will drag on.

You showed the benchmark showing different implementations have different
times. Great.

How would the new improved version look like?
It doesn't need to compile.

Thanks!

Em qui., 2 de abr. de 2026, 16:32, Muneem via Std-Proposals <
std-proposals_at_[hidden]> escreveu:

> >what? Did you confuse it with Link Time Optimization? instantiation is
> >in the Transition Unit not at link time as this would be too late.
> >We have Extern Templates but then the compiler in TU becomes blind to that
> >code as it instantiates in other TU, LTO can help there but
> >then we have exactly the same case if we use normal functions.
> It's different for templates because the linker has to plug in some type,
> which again gives it more flexibility since it knows the INTENT IS TO
> PLUGIN types. current std::variants do rely on templates, but it's not good
> enough for subscripting, like my benchmarking clearly shows this
> consistently. LIke how am I supposed to subscript homogenous lists when the
> speed of branching is not fast enough. Why is assembly code outperformed by
> C++ code (for large code bases)? because C++ code has constructs that give
> compiler intent and context, which helps the compiler do automated
> optimizations). There is no current tool, like the current tools can be
> used for this until there is a construct to allow for runtime indexing of
> homogenous half open sets.
>
> One last detail:
> tools do exist and I tried them, but they dont convey intent, like again,
> I can make my own tools but they won't convey intent, hence won't be fast
> enough. Like why do I use the inline keyword, when the compiler should in
> theory inline any function? Why do I use the constexpr keyword, if the
> compiler can evaluate any thing that It can at compile time if the
> observable behaviour does not change. It's a very simple point that I am
> trying to build on.
>
>
> On Fri, Apr 3, 2026 at 12:28 AM Muneem <itfllow123_at_[hidden]> wrote:
>
>> >what? Did you confuse it with Link Time Optimization? instantiation is
>> >in the Transition Unit not at link time as this would be too late.
>> >We have Extern Templates but then the compiler in TU becomes blind to
>> that
>> >code as it instantiates in other TU, LTO can help there but
>> >then we have exactly the same case if we use normal functions.
>> It's different for templates because the linker has to plug in some type,
>> which again gives it more flexibility since it knows the INTENT IS TO
>> PLUGIN types. current std::variants do rely on templates, but it's not good
>> enough for subscripting, like my benchmarking clearly shows this
>> consistently. LIke how am I supposed to subscript homogenous lists when the
>> speed of branching is not fast enough. Why is assembly code outperformed by
>> C++ code (for large code bases)? because C++ code has constructs that give
>> compiler intent and context, which helps the compiler do automated
>> optimizations). There is no current tool, like the current tools can be
>> used for this until there is a construct to allow for runtime indexing of
>> homogenous half open sets.
>>
>>
>> On Thu, Apr 2, 2026 at 1:11 PM Marcin Jaczewski <
>> marcinjaczewski86_at_[hidden]> wrote:
>>
>>> czw., 2 kwi 2026 o 09:05 Muneem via Std-Proposals
>>> <std-proposals_at_[hidden]> napisał(a):
>>> >
>>> > hi!
>>> > Your point is partly correct, but this issue is quite prevalent, below
>>> are my branches from multiple sources:
>>> > this is the updated code:
>>> > #include <variant>
>>> > #include <iostream>
>>> > #include <chrono>
>>> > #include <ctime>
>>> > #include <iomanip>
>>> > #include<array>
>>> > std::array<int, 3> array_1={1,2,3};
>>> >
>>> > struct A { int get() { return array_1[0]; } };
>>> > struct B { int get() { return array_1[1]; } };
>>> > struct C { int get() { return array_1[2]; } };
>>> >
>>> > struct data_accessed_through_visit {
>>> > static std::variant<A, B, C> obj;
>>> >
>>> > inline int operator()(int) {
>>> > return std::visit([](auto&& arg) {
>>> > return arg.get();
>>> > }, obj);
>>> > }
>>> > };
>>> > std::variant<A, B, C> data_accessed_through_visit::obj=C{};
>>> > int user_index = 0;
>>>
>>> I can hold you right now, you use a static object for the
>>> `std::variant`, this means you test different thing than in other
>>> cases.
>>> This can affect code gen and what optimizer can see and guarantee.
>>>
>>> Besides, why use `variant` here? It was not designed for this but to
>>> store different types in the same memory location.
>>> And none of this is used there, only making it harder to compiler to
>>> see what is going on.
>>>
>>> Why not use simple pointers there? And if you need to do some
>>> calculations function pointers?
>>>
>>> >
>>> > struct data_ternary {
>>> > inline int operator()(int index) {
>>> > return (index == 0) ? array_1[0] : (index == 1) ? array_1[1] :
>>> (index == 1) ? array_1[2] : -1;
>>> > }
>>> > };
>>> >
>>> > struct data_switched {
>>> > inline int operator()(int index) {
>>> > switch(index) {
>>> > case 0: return array_1[0];
>>> > case 1: return array_1[1];
>>> > case 2: return array_1[2];
>>> > default: return -1;
>>> > }
>>> > }
>>> > };
>>> >
>>> > struct data_indexing {
>>> > inline int operator()(int index) {
>>> > return array_1[index];
>>> > }
>>> > };
>>> >
>>> >
>>> >
>>> > volatile int x = 0;
>>> > constexpr uint64_t loop_count=10000;
>>> > static void measure_switch() {
>>> > data_switched obj;
>>> > for (int i=0; i++<loop_count;) {
>>> > x = obj(user_index);
>>> > }
>>> > }
>>> >
>>> > static void measure_visit() {
>>> > data_accessed_through_visit obj;
>>> > for (int i=0; i++<loop_count;) {
>>> > x = obj(user_index);
>>> > }
>>> > }
>>>
>>> Why not use:
>>> ```
>>> static void measure_visit() {
>>> data_accessed_through_visit obj
>>> return std::visit([](auto&& arg) {
>>> for (int i=0; i++<loop_count;) {
>>> x = arg.get();
>>> }
>>> }, obj);
>>> }
>>> ```
>>> And this is more similar to what the compiler sees in other test cases.
>>>
>>> >
>>> > static void measure_ternary() {
>>> > data_ternary obj;
>>> > for (int i=0; i++<loop_count;) {
>>> > x = obj(user_index);
>>> > }
>>> > }
>>> > static void measure_indexing() {
>>> > data_indexing obj;
>>> > for (int i=0; i++<loop_count;) {
>>> > x = obj(user_index);
>>> > }
>>> > }
>>> >
>>> > template<typename func_t>
>>> > void call_func(func_t callable_obj, int arg){
>>> > const auto start = std::chrono::steady_clock::now();
>>> >
>>> > constexpr int how_much_to_loop=1000;
>>> > for(int i=0; i++<how_much_to_loop;){
>>> > callable_obj();
>>> > }
>>> > const auto end = std::chrono::steady_clock::now();
>>> > auto result=
>>> std::chrono::duration_cast<std::chrono::nanoseconds>(end -
>>> start).count()/how_much_to_loop;
>>> > std::cout<<result/how_much_to_loop<<std::endl;
>>> >
>>> > }
>>> >
>>> > int main() {
>>> > std::cout << "Enter index (0 for A, 1 for B, 2 for C): ";
>>> > if (!(std::cin >> user_index)) return 1;
>>> >
>>> > // Set the variant state
>>> > if (user_index == 0) data_accessed_through_visit::obj = A{};
>>> > else if (user_index == 1) data_accessed_through_visit::obj = B{};
>>> > else if (user_index == 2) data_accessed_through_visit::obj = C{};
>>> >
>>> > std::cout << "Time (ns) for switch: ";
>>> > call_func(measure_switch, user_index);
>>> >
>>> > std::cout << "Time (ns) for visit: ";
>>> > call_func(measure_visit, user_index);
>>> >
>>> > std::cout << "Time (ns) for ternary: ";
>>> > call_func(measure_ternary, user_index);
>>> >
>>> > std::cout << "Time (ns) for subscript: ";
>>> > call_func(measure_indexing, user_index);
>>> >
>>> > return 0;
>>> > }
>>> > the bench marks consistently show that these syntax constructs do
>>> matter (the smaller the index range is, the more the compiler can flatten
>>> it and know how to branch), notice how ternary is outperforming them all
>>> even though its nesting, This means that adding new syntax with the sole
>>> purpose to give compilers as much information as possible is actually
>>> useful.
>>>
>>> And what do you propose here exactly? diffrent structures behave
>>> differently.
>>> How exactly it will look and work, right now you use lot of vague
>>> statement that do not show anything.
>>>
>>> > Consider how templates and instantiation give the compiler extra
>>> insight. why? because templates are instantiated at the point of
>>> instantiation which can be delayed upto link time. these are the benchmarks:
>>>
>>> what? did you confuse it with Link Time Optimization? instantiation is
>>> in the Transition Unit not at link time as this would be too late.
>>> We have Extern Templates but then compiler in TU becomes blind to that
>>> code as it instantiationed in other TU, LTO can help there but
>>> then we have exactly the same case if we use normal functions.
>>>
>>> > benchmarks for g++:
>>> > Enter index (0 for A, 1 for B, 2 for C): 2
>>> > Time (ns) for switch: 33
>>> > Time (ns) for visit: 278
>>> > Time (ns) for ternary: 19
>>> > Time (ns) for subscript: 34
>>> > PS C:\Users\drnoo\Downloads> .\a.exe
>>> > Enter index (0 for A, 1 for B, 2 for C): 2
>>> > Time (ns) for switch: 33
>>> > Time (ns) for visit: 296
>>> > Time (ns) for ternary: 20
>>> > Time (ns) for subscript: 35
>>> > PS C:\Users\drnoo\Downloads> .\a.exe
>>> > Enter index (0 for A, 1 for B, 2 for C): 2
>>> > Time (ns) for switch: 34
>>> > Time (ns) for visit: 271
>>> > Time (ns) for ternary: 17
>>> > Time (ns) for subscript: 33
>>> > PS C:\Users\drnoo\Downloads> .\a.exe
>>> > Enter index (0 for A, 1 for B, 2 for C): 2
>>> > Time (ns) for switch: 34
>>> > Time (ns) for visit: 281
>>> > Time (ns) for ternary: 19
>>> > Time (ns) for subscript: 32
>>> > PS C:\Users\drnoo\Downloads> .\a.exe
>>> > Enter index (0 for A, 1 for B, 2 for C): 2
>>> > Time (ns) for switch: 34
>>> > Time (ns) for visit: 282
>>> > Time (ns) for ternary: 20
>>> > Time (ns) for subscript: 34
>>> > I really have to go to sleep now ( I am having some issues with visual
>>> studio 2026), I Hope, it would be acceptable for me to send the benchmarks
>>> for that tomorrow.
>>> >
>>> > regards, Muneem
>>> >
>>> >
>>> > On Thu, Apr 2, 2026 at 10:54 AM Thiago Macieira via Std-Proposals <
>>> std-proposals_at_[hidden]> wrote:
>>> >>
>>> >> On Wednesday, 1 April 2026 21:56:44 Pacific Daylight Time Muneem via
>>> Std-
>>> >> Proposals wrote:
>>> >> > /*
>>> >> > Time (ns) for switch: 168100
>>> >> > Time (ns) for visit: 3664100
>>> >> > Time (ns) for ternary: 190900
>>> >> > It keeps on getting worse!
>>> >> > */
>>> >>
>>> >> So far you've maybe shown that one implementation is generating bad
>>> code. Have
>>> >> you tried others?
>>> >>
>>> >> You need to prove that this is an inherent and unavoidable problem of
>>> the
>>> >> requirements, not that it just happened to be bad for this
>>> implementation.
>>> >> Just quickly reading the proposed benchmark code, it would seem
>>> there's no
>>> >> such inherent reason and you're making an unfounded and probably
>>> incorrect
>>> >> assumption about how things actually work.
>>> >>
>>> >> In fact, I pasted a portion of your code into godbolt just to see
>>> what the
>>> >> variant visit code, which you claim to be unnecessarily slow, would
>>> look like:
>>> >> https://gcc.godbolt.org/z/WK5bMzcae
>>> >>
>>> >> The first thing to note in the GCC/libstdc++ pane is that it does not
>>> use
>>> >> user_index. The compiler thinks it's a constant, meaning this
>>> benchmark is
>>> >> faulty. And thus it has constant-propagated this value and is
>>> *incredibly*
>>> >> efficient in doing nothing useful. MSVC did likewise.
>>> >>
>>> >> Since MSVC outputs the out-of-line copy of inlined functions, we can
>>> see the
>>> >> operator() expansion without the proapagation of the user_index
>>> constant. And
>>> >> it's no different than what a ternary or switch would look like.
>>> >>
>>> >> In the Clang/libc++ pane, we see indirect function calls. I don't
>>> know why
>>> >> libc++ std::variant is implemented this way, but it could be why it
>>> is slow
>>> >> for you if you're using this implementation. If you tell Clang to
>>> instead use
>>> >> libstdc++ (remove the last argument of the command-line), the indirect
>>> >> function call disappears and we see an unrolled loop of loading the
>>> value 10.
>>> >> That would mean Clang is even more efficient at doing nothing.
>>> >>
>>> >> Conclusion: it looks like your assumption that there is a problem to
>>> be solved
>>> >> is faulty. There is no problem.
>>> >>
>>> >> --
>>> >> Thiago Macieira - thiago (AT) macieira.info - thiago (AT) kde.org
>>> >> Principal Engineer - Intel Data Center - Platform & Sys. Eng.
>>> >> --
>>> >> Std-Proposals mailing list
>>> >> Std-Proposals_at_[hidden]
>>> >> https://lists.isocpp.org/mailman/listinfo.cgi/std-proposals
>>> >
>>> > --
>>> > Std-Proposals mailing list
>>> > Std-Proposals_at_[hidden]
>>> > https://lists.isocpp.org/mailman/listinfo.cgi/std-proposals
>>>
>> --
> Std-Proposals mailing list
> Std-Proposals_at_[hidden]
> https://lists.isocpp.org/mailman/listinfo.cgi/std-proposals
>