<table border="1" cellspacing="0" cellpadding="8">
    <tr>
        <th>Issue</th>
        <td>
            <a href=https://github.com/llvm/llvm-project/issues/58348>58348</a>
        </td>
    </tr>

    <tr>
        <th>Summary</th>
        <td>
            Regression missed optimization with constant signed division going from version 9 to 10
        </td>
    </tr>

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

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

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

<pre>
    clang 10 and newer produce sub-optimal code when doing a signed floor division for the same llvm IR

I have 3 versions of floor_div that preform division with rounding towards -inf instead of 0
https://godbolt.org/z/dbfWhEdGY

```
long long xfloor_div_16(long long a)
{
     return (a + 0x8000000000000000ULL) / 16 - 0x8000000000000000 / 16;
}

long long yfloor_div_16(long long a)
{ 
    return (long long)((unsigned long long)(a) / 16 * 16) / 16;
}

long long zfloor_div_16(long long a)
{
    unsigned long long q = a; return (long long)(q - q % 16) / 16;
}
```

current versions of clang/llvm compile the first version to a single arithmetic shift right for power of 2 constants, whereas version 9 and older compile all three to optimal code even though both old and new versions produce the same llvm IR
current versions of gcc compile the the last 2 optimally, but not the first :-(

the first version gets optimized directly to optimal llvm IR 
```
%2 = ashr i64 %0, 4, !dbg !17
```
versions 2 & 3 both produce:
```
%2 = and i64 %0, -16, !dbg !23
%3 = sdiv i64 %2, 16, !dbg !24
```

which gets turned into a single x86 sar instruction in version <= 9 
```
sar     rax, 4
```
while it becomes an adjustment + conditional move before the sar in versions >= 10

```
and     rdi, -16
lea     rax, [rdi + 15]
test    rdi, rdi
cmovns  rax, rdi
sar     rax, 4
```



</pre>
<img width="1px" height="1px" alt="" src="http://email.email.llvm.org/o/eJyVVVtvozgU_jXwcpQKDKTwwEMzna5GmqeRVqt9GhlswJXBGdskbX_9HptLSJu5bERMHJ_L58_fOa4Uey1rSYcW4gjowGDgZ67hqBUbaw5mrHbqaEVPJdSKcTh3fACmBDpQMKIdOINGKqWBiZMwQg3Q4MR26Et7DlKeevjyLYgeg-hhGr9AR08cEjhx7RwMqGaK8R1joCu1mJ9jmP4S9CxsB1qNA3OprTpTzQzsxNCAGIzllLko0ZShs_ZoguQhIE_4tIpVSto7pVucveGXVc0_3Wf2179bWME-mh8_lQrz-OFlxfY93gckv6zQgBSz8_1h-gHuo7kd9QBoS3E4QPSSR9efv79-RV9cfIJ4D7sbFvNakByWDI9btBcQr38GDy74LvBWW2dJcnzGYT7Ud0t0AzcgDw7Z-sfvMb79Xwo_woAfECSPaJ8cfr6BH0gl2pHsd_iuz3oa61FrPtgrWfrSwChexrXqj0JyL-5GaLOaoh59NQwtrlKNUu25FTWYTjQWtGg766viqFxtYViCsVC1dLAmIJ9cUWlOzRqu8JWoJEPrJSmVEhNrzl2yq5LkJyxJ26mx7aBSWCbouJTyZTNLRd-uzFt7b-v6asvuKynumiz55atDX40WBmU3tGDp7ZycNuR-5Kzl1kyBxBseNBOa11a-brc3Q4Tbh0YyMknCdBrEPnXnHjlAqRsCErOqda_4_qb_ulWMQvbYjjx3M02uefw6KfK7zblzgttmJcnqkXgP45rb7EKc7QeP9BfiPHei7ibOnPaRMDFsVfeS7_FYtW-GeqytY1gMK9lB8slhKG5T6Rx9Y6AvE3-3jBAB5hEWKo6q4AYpAMqeR2N7pxzX51DUTLjUeHa9wh5fuS6-aE5vABlE9Nkhiq92-S6lI9njYmLheGosnG7xBtkBLTyCOAuyucYtR7FdnN1rkjoiw_yL8_r_H5GwHUNexvt9ti_SNCchKxNWJAUNrbCSl994q7nx3PfCGDyvWerUrtfZ0gOWe3S97Fp_vzZa9ZuWgKcdR-GoZfnuesNQY3WHZzL3qfm1Qyk_Y0nhFAGMHDvNU5YnaR52JcsZrxIW53FBCho3VcaSpCga0tCous9pKGnFpSmR2YAQ10d8CPyN9IaiJBEhcRQnJMrSlNxxnBLCmjSv6X2GOk4j3lMh7xwOd--GuvSQqrE1uCiFseaySI0ngPt0GJ-O2M10-cypECL0mUuP_D-PPZrp">