[llvm:optimizations] [missed-optimization] Unable to hoist out call to readonly function in a loop

[mattgodbolt]

In clang 21.1.0 (and earlier), in the following code, which is about as reduced as I can get:

#include <cstddef>

// [[gnu::noinline]]
// [[gnu::const]]
// [[gnu::noinline, gnu::const]]
static std::size_t count_ints(const int *ints) {
  std::size_t num{};
  while (*ints) {
    num++;
    ints++;
  }
  return num;
}

std::size_t num_compares{};

// Returns if a zero-termed list of ints has 1234
bool has_1234(const int *ints) {
  for (std::size_t index = 0; index < count_ints(ints); ++index) {
    ++num_compares;
    if (ints[index] == 1234) {
      return true;
    }
  }
  return false;
}

CE link: https://godbolt.org/z/raPrvq9aT

Here I would expect:

• the count_ints function to be noted as const (or similar) Certainly "pure". TBAA should note that it only reads through an int* if at all. Maybe argmemonly (I think?)
• as a result the count_ints is called inside the loop (even though it's cleverly been turned into a wcslen. I'd hope LICM could yank it out
• that results in poor performance (O(n2)) though naturally this code is awful (on purpose!)

Interestingly, marking it gnu::const (not strictly true) still causes the wcslen to be inlined and called every iteration. Making it noinline and const "fixes" this but I think is not strictly true (as I understand it the read through *ints makes it non-const which is to say "readnone" in llvm).

Similar reported to gcc.

I discussed this over Discord with a few folks more familiar with LLVM and they said to file this with the tags in the title. Sorry if this is not the correct way to do this.

[vidsinghal]

Hello, i'd like to work on this if this is planned to be implemented?