<table border="1" cellspacing="0" cellpadding="8">
    <tr>
        <th>Issue</th>
        <td>
            <a href=https://github.com/llvm/llvm-project/issues/118152>118152</a>
        </td>
    </tr>

    <tr>
        <th>Summary</th>
        <td>
            [clang] -ffast-math in 19.1.0 prevents function from returning intended __m128 bitmask
        </td>
    </tr>

    <tr>
      <th>Labels</th>
      <td>
            clang
      </td>
    </tr>

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

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

<pre>
    Hello.

A function like below may return an unexpected `__m128` value when compiled using clang 19.1.0 with `--fast-math -O2`.

```
auto maskFunc = []() {
 return _mm_castsi128_ps(_mm_set_epi32( 0x7FFFFFFF, 0x00000000, 0x00000000, 0xFFFFFFFF ));
};
```

Sample code at Compiler Explorer https://godbolt.org/z/xG5r1n7xY , demonstrating different result from lambda return and immediate assignment.

Workaround:

- Remove `-ffast-math` or `-O2`
- Add `-fno-finite-math-only`
- Use clang 18

---

I can speculate that this is because the i32 values set in the sample function are NaN when interpreted as float, and `-ffast-math` optimizer discarded such return value as invalid. However, the `__m128` value was intended for use in bitwise operation intrinsics like [_mm_and_ps()](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_and_ps), and not as a value in floating point calculation.

This was permitted in clang 18, but no longer as of clang 19.1.0. Was this behavior change intended?

If so, is guaranteeing assignment and returning of an arbitrary `__m128` value (which may contain NaN or INFINITY) no longer possible with with `-ffast-math` enabled for clang 19 onward? In addition to bitwise intrinsics, comparison intrinsics like [_mm_cmpgt_ps()](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_cmpgt_ps) will use `0xFFFFFFFF` as truth values, so handling non-finite value are necessary for multiple intrinsics by definition.

Or, is it simply not recommended to use these intrinsics with `-ffast-math` (or more specifically `-ffinite-math-only`)? Using `-fno-finite-math-only` is a valid workaround, but will affect the entire compilation, despite the affected code having no intention of introducing NaN or INFINITY in float calculations elsewhere.

Regards,
</pre>
<img width="1px" height="1px" alt="" src="http://email.email.llvm.org/o/eJzEVk9v27gT_TT0ZWBDpuJ_Bx_c5qdfc0mBbouip4AiR9ZsKVIgKdvpp18MpdhJke51DQNxKIoz896bN1Qx0tEh7sXqg1jdz9SQWh_2sSXXmFntzfP-E1rrF6I4iOJwgGZwOpF3YOknQo3Wn6FTzxAwDcGBcjA4vPSoExoQ6-LpqVvKrVgXcFJ2QDi36ED7rieLBoZI7gjaKneE5W6xXBRwptTyi_N5o2Kadyq1MP8sxbqYchDrYvoWBzUkD52KP6vBaRDlPYx1CLkVcgdi80EUh5fcnrruSauYIi3l9qmPQm55KWJ6wp5KKeQWisumGj9CfoTiUkyf9_6bNlYg5I6_JQcTm_vpx6s0RXH4S3W9RdDeIKgEH0cEAvzv0lsfMECbUh9FeRCyErI6elN7mxY-HIWsfglZXf6_Cku3ufwAjm6w8y6moBIDaKhpMKBLEDAONkETfAdWdbVRN2YMUNehIZUQRt47dGlC9bsPP1XwgzOcQ16awxfs_AkzG82VDebSh7yYack7D8aM25yfN-QoYd47984-v-z5FvGF6u0UYT4ffzyAVg5ij3qwnF5qVYLUUgSKUKNWQ-RFBCrlKKQIEROQy6txBPeqTRUQHtXjKDZyCUMfkAWpIjTWq8QQMiDvVNYn6ugXBjAUtQoGDcRBty8wjipWEcidlCWzgE_-jCcMfCTn8p7m8_aEjg9rfACuhhzUlM4UEXyPTKTPuQZykXQc-0usPrBElTOjXllmWd1v1XI-nxccwC6074SstOdoaXwiZDVEISt0QlbGa_59izM_DmQwLxm8LNrUWSHLhJckyvvXsXcvmDmfuH41FUduhJR12HtyCbSymUbybhLXV2aSUegxdJSYCXI3LciPUA8JnAfr3REDH--bN7awgO8qjoqosVUn8gF0q9wRr8iKsprE1ED0fChFOA4qKJcQOb2b6HMhI6P8wDdsXCrUlIIKz-9xKOT23JJus9kxvopc1pgP8PBYPTw-fP3BlnMrovcxUm1xdLSrrb2VGzpV20kVL_WCd2cVuB54cKCMoayN5K-CudHHZbKbqkDxz_rRXX9M_5mCbtF3cCZrs_zFurg5KCPB9IYhtVN7c2HRQ6ucsUyR824ylpcWDAgONcbIjDF-3WATsQ-8QqF-BoP5vZsaP4dJHJQgUtfb56zpgNp33dikycPkOG_Q_hOLQm45vg-YPYwa0sra52nnO3bI06KCb3n6_YttcpK5z8jA-WbQU79kLFXToE7ZedAlCjgN19x-46SIPaXRPcfNaMYxxG2UkR1bKIvMN7lcbwbNz34T-LXZX_d4BLQRzy0GnAD-gkcVDFM4M_vS7MqdmuF-uSnlrlytZTFr9-tipxTutN41qtGyQKO2xW5T3Jn1rqnVekZ7Wci75bIsirIoltvFpl4Wsiw3q6282xXlUtwV2CmyC2tPHc_JGcU44H653C5XcmZVjTbma42UubOElHzDCXt-YV4PxyjuCksxxdsRiZLNd6HxjdU9vGKaq58uKX3AE7oUbyMnT9ybo1ztfvQR7ly-psyGYPe_TXpK7VBPbceJTH_mffB_o-YWzIVx0021nfbynwAAAP__C7A4Zw">