Describe the bug

_rdtsc() intrinsic somehow get dropped at linkage. Sample:

#include <x86intrin.h>
#include <utility>

long tsc;

[[gnu::noinline]] 
void call_cold(auto&& f) {
    std::forward<decltype(f)>(f)();
}

int main() {
    call_cold([&](){
        tsc = _rdtsc();
    });
    return tsc;
}

Without 'link to binary' output flag the code is correct -- the call_cold<>() instance contains RDTSC instruction (with all expected calculations). However with linking to binary the instruction is dropped:

call_cold<main::{lambda()#1}>(main::{lambda()#1}&&) [clone .constprop.0]::
 mov    %rax,0x2f00(%rip)        # 404030 <tsc>
 ret    

I checked on my local compiler 12.2.1 -- objdump -dC a.out gives expected code (with RDTSC) so it doesn't seem like a compiler issue.

Actually, it might be just cutting off the beginning of the function as the function address is expected to be aligned while the first instruction in the function which I can see is 401129.

Steps to reproduce

1. Copy-paste the snippet to the godbolt editor
2. Select g++ 12.2 for x86_64
3. Make sure the "link-to-binary" flag is chosen
4. Inspect the assembler output for function call_cold<main::{lambda()#1}>(main::{lambda()#1}&&) [clone .constprop.0] -- for me is shows only the above 2 instructions (mov + ret)

Expected behavior

Taken from objdump locally:
0000000000001020 <void call_cold<main::{lambda()#1}>(main::{lambda()#1}&&) [clone .constprop.0]>:
    1020:       0f 31                   rdtsc
    1022:       48 c1 e2 20             shl    $0x20,%rdx
    1026:       48 09 d0                or     %rdx,%rax
    1029:       48 89 05 e8 2f 00 00    mov    %rax,0x2fe8(%rip)        # 4018 <tsc>
    1030:       c3                      ret

Reproduction link

https://godbolt.org/z/xPe6s597e

Screenshots

Not applicable

Operating System

Manjaro Linux

Browser version

Firefox 110.0 (64-bit)