<table border="1" cellspacing="0" cellpadding="8">
    <tr>
        <th>Issue</th>
        <td>
            <a href=https://github.com/llvm/llvm-project/issues/63498>63498</a>
        </td>
    </tr>

    <tr>
        <th>Summary</th>
        <td>
            [flang] [mlir] Why is `-mmlir --math-runtime=precise` not enabled by default when the optimization level is -O2?
        </td>
    </tr>

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

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

    <tr>
      <th>Reporter</th>
      <td>
          k-arrows
      </td>
    </tr>
</table>

<pre>
    This issue is not a bug report, just a question.  My question is, why is `-mmlir --math-runtime=precise` not enabled by default when the optimization level is -O2?

Here's the background to this question. 

I checked if flang-new passes the Netlib LAPACK tests. In LAPACK's Makefile, the optimization level of the Fortran compiler is -O2, and if you use this as it is, you can run the test as follows.

```shell
$ diff make.inc.example make.inc
9c9
< CC = clang
---
> CC = gcc
20,21c20,21
< FC = flang-new
< FFLAGS = -O2
---
> FC = gfortran
> FFLAGS = -O2 -frecursive
23c23
< FFLAGS_NOOPT = -O0
---
> FFLAGS_NOOPT = -O0 -frecursive
46c46
< #TIMER = INT_ETIME
---
> TIMER = INT_ETIME
50c50
< TIMER = INT_CPU_TIME
---
> #TIMER = INT_CPU_TIME
$ make
```

I used LAPACK 3.11.0 and got the following result:

```shell
 -->   LAPACK TESTING SUMMARY  <--
                Processing LAPACK Testing output found in the TESTING directory
SUMMARY nb test run     numerical error         other error
================ ===========     =================       ================
REAL 1327023         0       (0.000%)        0       (0.000%)
DOUBLE PRECISION        1327845         0       (0.000%)        0 (0.000%)
COMPLEX                 282008          27027   (9.584%) 1855    (0.658%)
COMPLEX16               282791          27170   (9.608%) 1856    (0.656%)

--> ALL PRECISIONS      3219667         54197 (1.683%)        3711    (0.115%)
```

You can see errors in COMPLEX and COMPLEX16 tests. Then, after I enabled `-mmlir --math-runtime=precise`, I ran the test again as follows and got the following result:
```shell
$ diff make.inc.example make.inc
9c9
< CC = gcc
---
> CC = clang
20,21c20,21
< FC = gfortran
< FFLAGS = -O2 -frecursive
---
> FC = flang-new
> FFLAGS = -O2 -mmlir --math-runtime=precise
23c23
< FFLAGS_NOOPT = -O0 -frecursive
---
> FFLAGS_NOOPT = -O0
46c46
< TIMER = INT_ETIME
---
> #TIMER = INT_ETIME
50c50
< #TIMER = INT_CPU_TIME
---
> TIMER = INT_CPU_TIME
$ make
```

```shell
                        -->   LAPACK TESTING SUMMARY  <--
                Processing LAPACK Testing output found in the TESTING directory
SUMMARY                 nb test run     numerical error other error
================        =========== =================       ================
REAL                    1327023 0       (0.000%)        0       (0.000%)
DOUBLE PRECISION 1327845         0       (0.000%)        0       (0.000%)
COMPLEX 786775          0       (0.000%)        0       (0.000%)
COMPLEX16 787842          0       (0.000%)        0 (0.000%)

--> ALL PRECISIONS      4229485         0       (0.000%) 0       (0.000%)
```

As you can see, in this case, no errors were observed at all.

Personally, I don't like the default state where it gives errors in tests of very famous apps like LAPACK. Also in my opinion, if the optimization level is -O2, it is desirable to turn off optimizations that cause numerical errors. I think that's why LAPACK's Makefile also specifies -O2 instead of -Ofast.

This is a matter of policy and may not be an issue that I should insist on further. However, I would like to know if some decisions have already been made in Flang community. I have noticed that the simillar issue (https://github.com/llvm/llvm-project/issues/58799) is still open.
</pre>
<img width="1px" height="1px" alt="" src="http://email.email.llvm.org/o/eJzMWFtz4joS_jXmpQuXLd8feGBIOIfa3GqSqd3zNCXsNuhEllhJhmV__ZZkzC1kSM7Mbq2KCrHU_fVFan1uqNZsIRBHXvLFS24GtDVLqUavQ6qU3OjBXFbb0cuSaWBatwhMg5AGKMzbBShcSWU8MoE_W20n_9miNkwKH-B-u38Cpq3MZrm16l4aDJuGMwXDYUPNcqhaYViDXnSzUlgyjV4aOCMo6JxjBfMtVFjTlhvYLFGAWSLIlWEN-zd1-BzXyC328JF40dQLbrxg3P39HRV6JNNOaU7L14WSrajASDA2qoPHx1ozKJdYvmIFrIaaU7EYCtzAimqNHdQDGs7mcDd-Gk_-Bga10T7MxG7CWbynr1gzjjb2d1yWtVuZSmUUFVDKZsU4qj4UMgEqnA9b2UKrsfOZamBml1S7UFIBqu3yYj2xArXkXG60fxyVlwbdRy-R890ciaFidQ0NfUWfidLHf9FmxXE_0ckVZbFTiCYwmYAX3UBpE9PNDofDfvm2X16UO10SeGRCwnL3fcCZdoL7BB-tTO_Gvz27VZuItzZ2qou6S93RwokmDGuFZas0W-POmagk0bmh7w-Pj08vO6XgkrkLYm-x47SM0wO2R6KX2f3tV6cwe3j5fmsfL6C_K5UEZRIcAE_lJk_fvr8D-MbyqazddLu9Z8fitAJajVV_viM_DP3AHcaFNO6gdQeMCXsL6JYbLxpfO2swHFrnoEd9uX1-mT38Bs_f7u_HX_8A8KJJHwWcjSclS9Ta2uu1beGKBcjWrFoDtStr1lVBj1wxhaWRatuB9obEvKsTWzR2iLZBxUrKAZWSam9UmiWqbq7P7c2nPnBdwo7PwvZ6f0HThfH1dnwHYUSygET7YIMekuSBHwSBRxKPFD9e7eBuHr99ubuFp6-3k9nz7PGh17EW8jj5sIVL2JPH-6e723-cHwcgOQmC_Og5C0jWYRd-ksc77DBPkr3JNMkvQIfpW-isCI-hwyzoodMgP0CnR9DpMXRfkva4j-_uDql57iAjEhZpmu1NJHFYZBYp9NM8Os1LlIXh3k4YJid2LpXuHztK0Ijd2dW2LPpE2ho-RL4jrpclCsc2tUEFsz3zfoirreIMLH8dCGhBmTiioQ_eHL-Yn_YEdImdjsjrGj-dk8w5Pb0lgktU9Ybl3nLVlVR_lL-u-fMu5Z3x18fI60csd8ZfV2jpB5z4af66yD_vjP8XWjof12jq5-mpv5B_nsf-Gwx1YfSk9QvJ6rMs9T50f8VmeZplB8CfRwxTyPIsj8nHMS-hXWOlmJAizq9m4kcOX6zFsd63KRpdO-TqgWkoqXbPQvZMtUGFIOca1RoroAYo5ydNzBMqLQXlfNvxTiWFRzIDnL2iK7K-VdSGGrQNo0LbLi3YGvURHzrmsw3YGtUWatrIVgNdrXSH1FWzD2OupRVvtiBXTDDpeJLVV5pQK2NbNKhQM2XJ1HWcrRIg6_pE0faT1EBJbX93VuK2p7SZEq9OyLWVto2-0GUCtZ7qFZasZui8ACa0QVrZKIePNdXmJJO7rh4oNNRY3pc1rCRn5dbRdUO3rgufI1Cx6_6dpzPQS9lye6tppg1IAXWr7F3kw-9yg2tU3dZsnFS3MRJehdzYxGnZ2E0qmXbBL-na-q6QVluYIwpoaIU25VNLmLYnblrBzNamwgkLaViJVeeM3QbNGsY5VTsnPZIvjVlp-15Bph6ZLphZtnO_lI1Hppyv-6_hSsk_sTQemTpN7ZFpkmdFYU8506AN4xzkCoU_qEZRVUQFHeAoTPMsTeM8DwbLEWY1VmWRU4ySuK6xKGIsq7zEFINsHsUDNiIBiYKUJEEU50HgY1WleRSTlAQpyTLqxQE2lHHfuuRLtRg4Z0ZpFBf5gNM5cu1-nSFE4KaL0SPES24GauTCmLcL7cUBZ_Zdbo9imOHuZx334uElN-AlX-wLhv337_-r32IGreKjv74dLgn_CQAA___OEQ0i">