<table border="1" cellspacing="0" cellpadding="8">
    <tr>
        <th>Issue</th>
        <td>
            <a href=https://github.com/llvm/llvm-project/issues/136275>136275</a>
        </td>
    </tr>

    <tr>
        <th>Summary</th>
        <td>
            Trivial comparison of pointer defeats store->load forwarding
        </td>
    </tr>

    <tr>
      <th>Labels</th>
      <td>
            llvm:analysis
      </td>
    </tr>

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

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

<pre>
    LLVM version 20.1.1
target: x86_64-apple-darwin20.6.0

Given the following IR:

```llvm
declare ptr @calloc(i64, i64)
declare void @free(ptr)
declare void @bar(ptr noalias nocapture readonly %0, i1 %info)

define void @foo(i32 %argc, ptr nocapture readnone %argv) local_unnamed_addr {
entry:
  %0 = tail call ptr @calloc(i64 1, i64 64)
 %1 = icmp ult ptr %0, inttoptr (i64 u0xFFFFFFFFFFFFFFFF to ptr)
  store i64 0, ptr %0, align 4

  ; NOTE: the %1 here causes %0 to be considered captured and therefore potentially aliased.
  ; So the load and branch below the call can't be optimized away.
  tail call void @bar(ptr %0, i1 %1)
  %2 = load i64, ptr %0, align 4
  %3 = add nsw i64 %2, -1
  store i64 %3, ptr %0, align 4
  %4 = icmp eq i64 %2, 0
  br i1 %4, label %free_it, label %end

free_it:
  tail call void @free(ptr %0)
  ret void

end:
  ret void
}
```

I would expect the load at `%2`, the add at `%3`, the icmp at `%4`, and the branch to be optimized away.  But they aren't, because `%0` is considered captured by `llvm::DetermineUseCaptureKind()`:

```c++
  case Instruction::ICmp: {
    unsigned Idx = U.getOperandNo();
    unsigned OtherIdx = 1 - Idx;
    if (auto *CPN = dyn_cast<ConstantPointerNull>(I->getOperand(OtherIdx))) { ... }

    // Otherwise, be conservative. There are crazy ways to capture pointers
    // using comparisons.
    return UseCaptureKind::MAY_CAPTURE;
```

And therefore `%0` is assumed potentially aliased.

Even passing `%1` to `@llvm.assume` (which is how I initially hit this condition) prevents the optimization.
</pre>
<img width="1" height="1" alt="" src="http://email.email.llvm.org/o/eJyEVt1u4zgPfRrlhohhy07iXOTCTdsPwTfTGcx2FtirgraYRANF8kpy0szTLyQ5bfqz2MJAGpE8JA-P6KBzcqeJVmx2w2a3Exz83tjVYf8LO9M6oyetEefVly9_foUjWSeNBp5nRVawvPFod-RZ2cBzPX-aV1Pse0VTgfYkNc-zeZazvGF58z95JA1-T7A1SpmT1DvY_GBlk8xsnqdHqeOB5Y2gTqEl6L0FVuUdKmU6xms5rxhfQ_xYXvkdjRTBcWuJGK97bz-3t2iTGbRBJdGBNh32frAEllAYrc7A-CyPWYrwr9Rbk8Ai3lbqq3TGhKJKHhzR7roQltCvUbXRNHocGV-CMh2qp0FrPJB4QiEssMUNyxvS3p4TKRDLAFbegkepIFDwGR1QjITAhZMQWMRA2R16GJRPcZeutPcmHcT4IX--f_cH3sALhQDOG0sxRX7p7wKGSu40VIkcAFbewMO3x7sgiDDrWMmeLEGHgyOXevIGWoLOaCcFWRIwciUAtQhxlrYhY288aS9RqTPEYZHIXtL8YWIGZTCFtRZ1t4eWlDlFSySsQ834wod8pvfyIH-HLCc8J6BXZj8K5K0KigsZjM94JDdmHvX4OSXRuYzOKARod4okBoTgOC3esRu8_wOtep0r_f0GLY8urR3rjWUpbEmFb-FaPEn_5oy0SGO7GEfZfeTk9VKNZSUmLPnoklAC3IhwbVjcXt_u5LqBkxmUAHruqfNXY_QQfEJD89h8sATmXgzllSGS8GKpRsuooIscktTejR7gZohpz4CWokBCaEtRpCNgKBek-1Sm7Rkuq6psWNnckid7kJp-Olonp_9LLRivA1fz_OOa6xi_CU-gq0NHsNHO26Hz0uiEuVkf-nCN0mIAABh0XNQCNuI5yuBntiP_rSeLWjyYMVv5wf1buFGXmAKmIf7VTW7DIsDBG2C8WX9_iG7irJ86dJ6V67XRzqP2343UnuzDoBQr7xivN1NW3r1WwHh9yRTqiE-oHrIsg1EHKSXj94zfp7pO0lEiPzJN9oheHimDx7g2wu7uLP4-wwnPLozzslf7VI57izm48GLpzKFHK53RLhvtlvxgNbwbUGT6a_PX07r5_vjzx12i5Z1cmzdL6a060LnhQOLfVhXLm7vw3uvDS1bvxuAiBAe-5zmrooyyhBPOGa9Pe9ntA_renGADUssReS-DbJMohYxa4UvoLR1Jexd1P0odgzGbiFUpluUSJ7QqFlW1nJe8KCb7VU7LmmO1qLGdCSyXHMvlIqcSxbwSZbecyBXP-Syvijqv87wosoXIW1HMttVigbyuW1bldECpsli-sbuJdG6gVVHO-WI2iVvGxZ8UnI8XBTWqs5OOcR5-Z9hVOJ-2w84FFqTz7hXMS69o9WjlUaK6mieY7WXyIGhL6F1aoEGLcYlsjT2hFVLvJoNVq733vQtzjvrYSb8f2qwzB8bvY1npY9pb84s6z_h9bMMxfj92clzxfwIAAP__kE7D4g">