<table border="1" cellspacing="0" cellpadding="8">
    <tr>
        <th>Issue</th>
        <td>
            <a href=https://github.com/llvm/llvm-project/issues/64317>64317</a>
        </td>
    </tr>

    <tr>
        <th>Summary</th>
        <td>
            `__builtin_assume` in destructors leads to worse codegen
        </td>
    </tr>

    <tr>
      <th>Labels</th>
      <td>
            new issue
      </td>
    </tr>

    <tr>
      <th>Assignees</th>
      <td>
      </td>
    </tr>

    <tr>
      <th>Reporter</th>
      <td>
          cor3ntin
      </td>
    </tr>
</table>

<pre>
    Consider the following C++ code

```cpp
struct S {
    S();
    ~S();
    int x = 0;
};

S::~S() { 
    __builtin_assume(x); // #1
}

void f(int n) {
    S s;
    if (n == 0)
        throw;
    else
 f(n - 1);
}
```

https://godbolt.org/z/ddchhPssK

If line \#1 is commented,  llvm is able to see that `S::~S` does nothing and as such it figures out that `f` requires no cleanup and
not call to `_Unwind_Resume` is performed.

If on the other hand `__builtin_assume(x)` is present - and x is not a constant, then:

* Some IR is created to load `x`
* Llvm does not deduce that S::~S() does nothing
* A landing pad / call to `_Unwind_Resume`  is generated.

The outcome of that is that `__builtin_assume` cause significant performance regression compared to identical code without it,
when it appears in cleaning code, because it forces an onerous unwinding in places that may otherwise not need one.

I think I have convinced myself that this is a bug in llvm rather than clang but I'm not completely certain. 



</pre>
<img width="1px" height="1px" alt="" src="http://email.email.llvm.org/o/eJx8Vc9vMjcQ_WvMZRS0eGFhD3sgiZCi9lB9ac-R1x5Yt97x1j9C0kP_9mq85IOk_YqQQWNm5r03bwcVoz0RYic292LzuFA5DT502oeakqVF78179-ApWoMB0oBw9M75s6UTPAh5L-Q9aG9QVI-i2l_OpprfeprmSEwh6wTPILb3cwQA4FnInZCtqG9if_9X0FKCNxD1I1Tf42L7eP1ezmdR70W9_6jAveBa4-Wlz9YlSy8qxjyikLu3uQ8IeRDyAELWq2vtm8Kv3ho4CrljHHQpfUMD4me0RxByR4x3hizb6yW_0hD8-VMKunhRsPQhuIPVrQhXQB_S3uIbUpoiky88Tt703qWlDychD38JeTBGD8MvMf50m_R0BGcJQWwemDjYCNqPI1JCI-QDgHOvI0dV7xCSh4gIaVAJRFNdpW4qMB4jkE8De0KRARUhZj2ATXC0pxwwgs_pe_KRkwL-mW0oiaAdKsoT587YyCfQyjluK5rq5Tc6WzIv37AMrqkY1oTh6MOIZvmFladiU58GDDAwHi7xg-lfagWMSAnuCvw3DjEEBdpTTIoSC5IGJKZ9a3S5h2c_Ijx9K_oFVAkNo3Zelb5v11HJPfzMkn7IBQZN1hdNv3r3VtNr_h6cIsMyT1xeHv5fJcZ0QsLAqD7J9CsLlJNm7P44Q7Dx-4T-pVZTgVY5IvCysEerFaWPCSjSCAFPAWO0nthFkwqzDNYgJauVKzsCzjYN7ATLgs5IzgMSG0VNE6oQwdJsByZZ9op8gB7n5uwnHzRGUASeMPgcIRfW_HNLMDnF14XHqN5nE5xtxCI4IRrO--wYYJH_gCcY1CvyxF8taTQwvkd0F23SYGN5FqDPpVF5OIIqHkuDYtCKTtDnBE9CbsfSj5VwmNC9g8aQlKUlfLJPORemq01bt2qB3appq1autvVqMXRaY7NtEXd1q3dq3bSoN01Vy1YZjb3sF7aTlayrXbWSldxtVku9Xm00HjdGNeu-bbVYVzgq65YMlxfCwsaYsWvW9Wq7cKpHF8vil5LwDOVSSMn_A6HjnLs-n6JYV87GFK9Vkk0Oux8YxRIYnBe-DxEcKhPZC2cfIpaZnpAWObjuy96yacj9UvtRyAO3unzcTcH_jjoJeSgAo5CHQuCfAAAA__8IZRAO">