https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116416
Bug ID: 116416
Summary: Missing optimization: compile time evaluation of
prvalue
Product: gcc
Version: 14.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: nuttyrunner at gmail dot com
Target Milestone: ---
gcc -x c++ -Og -std=c++20 yields inconsistent compile time evaluation for
following example:
struct Str {
constexpr Str(){
}
constexpr Str(const char *instr) {
str = instr; length = 0;
for(auto index = 0;instr[index];++index) {
++length;
}
}
const char *str = nullptr;
int length = 0;
};
extern void callback(Str str);
void func1() {
callback(Str{"Test"}); // Does not get constexpr optimized by gcc
}
void func2() {
Str str{"Test"};
callback(str); // Does get constexpr optimized by gcc
}
func2's lvalue str initializer gets compile time evaluated and the call yields
plain mov-s in assembly,
but prvalue Str{"Test"} does not - instead it inlines the constructor and
performs the counting each time func1 is invoked.
constexpr does not guarantee compile time evaluation, however this specific
usecase is extremely important to our code base for not counting the lengths of
strings each time performance-wise, and it seems to be some small semantic
difference?
Please note that the "callback" function call is not important, same happens
without it, it is in the example to illustrate that it is unrelated to dead
code removal.
That optimization is something we could try and look into in the gcc codebase
to propose a patch, but we would appreciate any advice/pointers as to where to
look.
Godbolt: https://godbolt.org/z/W1KK3hqao
