[C++20] [Modules] Generating LLVM IR takes time proportional to unused code in imported modules

[davidstone]

Given the following two translation units:

export module a;

constexpr bool f() {
	for (unsigned n = 0; n != 1'000'000; ++n) {
	}
	return true;
}

template<typename>
struct s {
	static constexpr auto a = f();
	static constexpr auto b = f();
	static constexpr auto c = f();
	static constexpr auto d = f();
};

template struct s<int>;
template struct s<long>;

export module b;

import a;

After calling

clang++ -std=c++20 --precompile -x c++-module -c a.cpp -o a.pcm

I would expect module b should be compiled very quickly. Instead, I see

$ time clang++ -std=c++20 -fmodule-file=a=a.pcm --precompile -x c++-module -c b.cpp -o b.pcm

________________________________________________________
Executed in   13.09 millis    fish           external
   usr time    3.69 millis  594.00 micros    3.09 millis
   sys time    9.42 millis  157.00 micros    9.26 millis

which is reasonable, but

$ time clang++ -c b.pcm -o b.o

________________________________________________________
Executed in    5.45 secs    fish           external
   usr time    5.44 secs  677.00 micros    5.44 secs
   sys time    0.01 secs  177.00 micros    0.01 secs

which is not reasonable. This time scales linearly with the number of explicit template instantiations in module a.

This is as of 9f20142, but I have experienced this slowness in any version that supports modules.

[llvmbot]

@llvm/issue-subscribers-clang-modules

[ChuanqiXu9]

For

// c.cpp
import a;

I get

$ time clang++ -std=c++20 -fprebuilt-module-path=. c.cpp -Xclang -disable-llvm-passes -S -emit-llvm -o c.ll

real	0m8.117s
user	0m8.018s
sys	0m0.012s

So the problem occurs during IR generation and this is not limited to module units.