<table border="1" cellspacing="0" cellpadding="8">
    <tr>
        <th>Issue</th>
        <td>
            <a href=https://github.com/llvm/llvm-project/issues/55397>55397</a>
        </td>
    </tr>

    <tr>
        <th>Summary</th>
        <td>
            [IROutliner] MultiSource/Benchmarks/Prolangs-C++/simul larger with IR Outliner @ -Oz on AArch64
        </td>
    </tr>

    <tr>
      <th>Labels</th>
      <td>
            llvm:optimizations,
            llvm:codesize
      </td>
    </tr>

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

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

<pre>
    At -Oz, simul is 6.59% larger with the IR outliner than without. This is a teensy tiny benchmark, but it does show a regression.

This was found by building the LLVM test suite with + without the IR outliner enabled using LNT, and comparing `size.__text`.

# Reproducing

IR for simul: https://godbolt.org/z/9P16365hG

```
# There are a bunch of functions, so let's do some horrible awk goop
~/llvm-project/build/bin/clang -Oz /tmp/simul-stripped.ll -Rpass-analysis=asm-printer  -S  -o /dev/null 2>&1| rg "\d+ instructions in function" -o | awk ' BEGIN { x=0 } { x+=$1 } END { print x}' 
177

~/llvm-project/build/bin/clang -Oz /tmp/simul-stripped.ll -Rpass-analysis=asm-printer -mllvm -ir-outliner -S  -o /dev/null 2>&1| rg "\d+ instructions in function" -o | awk ' BEGIN { x=0 } { x+=$1 } END { print x}'
188
```

Potential reduction: https://godbolt.org/z/xxr96zhTW

```
# No outliner
$ ~/llvm-project/build/bin/clang -Oz /tmp/test.ll -Rpass-analysis=asm-printer  -S  -o /dev/null
remark: <unknown>:0:0: 7 instructions in function [-Rpass-analysis=asm-printer]
remark: <unknown>:0:0: 7 instructions in function [-Rpass-analysis=asm-printer]
# -> 14 instructions

# Outliner
$ ~/llvm-project/build/bin/clang -Oz /tmp/test.ll -mllvm -ir-outliner -Rpass-analysis=asm-printer  -S  -o /dev/null
remark: <unknown>:0:0: 10 instructions in function [-Rpass-analysis=asm-printer]
remark: <unknown>:0:0: 10 instructions in function [-Rpass-analysis=asm-printer]
remark: <unknown>:0:0: 6 instructions in function [-Rpass-analysis=asm-printer]
# -> 26 instructions
```


# Analysis 

I compiled everything with remarks + `-save-temps`.

Via `llvm-size-remark-diff`:

```
++ > outlined_ir_func_0, 6 instrs, 0 stack B
== < _GLOBAL__sub_I_simulate.cpp, -1 instrs, 0 stack B
== > _ZN17screen_controllerC1Eii, 3 instrs, 0 stack B
== > _ZN17screen_controllerC2Eii, 3 instrs, 0 stack B

### Summary ###
Total change: 
  instruction count: 11 (6.21%)
  stack byte usage: None
```

>From there I took a look at `_ZN17screen_controllerC1Eii` and saw it was pretty much entirely being outlined:

<img width="1343" alt="Screen Shot 2022-05-11 at 6 11 25 PM" src="https://user-images.githubusercontent.com/4722725/167972261-5b3dc9a2-7619-4bba-9677-b0c98a07a3fd.png">

I extracted that function from the temporary bitcode and ran it through clang again with `-Oz -S -emit-llvm`. Then I used `llvm-extract` to yank out the function.

Since it was being outlined almost entirely, I just pasted it twice into a new IR file, resulting in a much smaller test case which seems to exhibit a related size increase.

cc: @AndrewLitteken 
</pre>
<img width="1px" height="1px" alt="" src="http://email.email.llvm.org/o/eJzNV1tvozgU_jXk5YgITIDmIQ9J04wqdTpVW81K-xIZcIKnYEe2aZr--j3HJOmk0-2OZmerlYAE43P7ztWFrnaTqYPwy3PAzsHKtmtAWsiG6ThgKTTcrIWBrXQ1uFrA5S3ozjVS4aKrufJfcGUI9zWS4cXBCaHsDpxUOyiEKuuWmwfiXnQOpINKCwu21lvca8TaCGulVsMgmgfRtH96ZltuYaU7VUGBjDrZVFKtvRZXV18_oxjrwHbSiV69gM0O2vygqlC8aEQFnSUWV9f3pA5HzqVuN9zQYpBFVj6L4XLpxJPDtxOFApbArdgYXXUl7v7-E8pZadNDFyRTqJ3bWPwTsAVea10VunFDbdb4hiAvxjdxlmRp_emEfxbtr6O4-1oYAZxutB5hBL2CFf5xiJb13tLQCBew3CKm-NYKqLUxEk0Fvn2AtdabPb_8AiU3zWMbog3fRIlUCw8p_UqFz7LhiALGAQK5cO0Gn96k0DojNxtRDZsGwtsNtzbkijc7K9HKObfEUiqHKEN4h7cmBpV4xKfqkIYFCQrP4iA_B4M4Mxak5xV5Syrk3fX24MvRONzi2SABmYEGwuzi0-U1rszgCYVG-G--f2MzXAjYKPZrF9f9ulcJv-Zzou5BiPP8BPIPwSRsSQKE0oTHaPyfwrRH6ezs7ZD0zxvthHKSN5i5Va_Tz8T805MZZ8_1_R__EPPX-pizh8UR_JqjqDz8Ysj2oo3wdQutC5LzTj0ovVXko2Qa7W_I_9Y3EKSzdyUH6fxDxRC6IfKFeHTC7XWR-_K74X8r_v9Dl8TRx_jko-Rkv9P3LHvL92_l-QvhdM8bTnqeb5ySWqp4FGbnamqhvg33NlnfjpFpaPmjCJ1osT686qlfJacdPrKo9YY9aVjJ1YrUSabvVouZl4BW7eOqWkqzJHCWETXHvam-UUZgHS8fYLanxUqYzAlzWH66-jKbXi2XtiuWl0tf3bkTw3KzIcIw_gkuF7D88zrObWlw8FmWWjmjm0aY8_hCSiJM_g0T9hNMDs7qL7jrWgRyB8eV_VClHdbtEqe2tfAx51fh-5BAt3bK-fjGRsHOsiGLcQwM2PiwuRdc7HDq6izvGV1rJd6JpYXRLU1kOMlcgtP6AQeaxv848v974GWRn9Is39LgSAPhxgjndtB2OA9RJzKioSmTwu8QB68DJzmXLQVn5WrfA1mcjBJqnrxx_cKdlw53tXbAIsbCKA0RANQvIyBwDr75TATWlD3Bab_rrDChbBENO1xjDnQFrZAtqOEQMwX3jHLGckJyEWf5GF-yOEyLpCrHnIV5Fo_DUVHwcJzleVhE5fiMRzlPVtVwg9Mm88PBSf7hjGp46TABcQx3L0VhtQcbKOW0oTAopCt1JTySBkd2SfOx0d26hr5w8zWXaj9EY8ZiGccSHIpWupCSk_KWplGF7kPDqmPS7nUgLzkNO64e4DB8H_Q5Sfg7qUpxcOSpz9AXrcZ5_uBSivJL-NbhElY4MpO03kqiVyiMgxJbGvFXWIRoM54iusYRSzSF9_FhW06h1J8USm7xoFBLWheitaSyeKolouPPIZT2GGlYh5ADxgNuP1G-LH3OjKKpqozYXknnxANiMqgmSTVOxnzgpGvEBKvy5e2xi6Zz-Ex63enOlKjoYnY4EGEiL24w1tEBNjzvy9lhvjw5dqGRB3Yk3rdZdPR0aso6Gw0600xejV8-BPdx5x34Q_uW1naCNEjTZJwP6kmeVVnJVmKU8iiNR2laVqOzPB6dibOUpyUbNLwQjSXrMBo902SqN0628pnvDySM_PDylWKO4PST7HwgJ5RZyJxFcZzFbBiPokpwxsejaBxH1Qptw_ovmyHR0-w4MBOvdtGtLX5spHX25SP2PblWwgNO_HmHRz8z0UZxxwfewIm37i_hIJxk">