<table border="1" cellspacing="0" cellpadding="8">
<tr>
<th>Issue</th>
<td>
<a href=https://github.com/llvm/llvm-project/issues/57912>57912</a>
</td>
</tr>
<tr>
<th>Summary</th>
<td>
Broken epilog vector body missing %resume.val contribution in phi
</td>
</tr>
<tr>
<th>Labels</th>
<td>
backend:AArch64,
miscompilation,
vectorization,
SVE
</td>
</tr>
<tr>
<th>Assignees</th>
<td>
</td>
</tr>
<tr>
<th>Reporter</th>
<td>
peterwaller-arm
</td>
</tr>
</table>
<pre>
I have bisected a case of a checksum failure in a non-trivial internal C++ workload to a recent change, 582f8ef19fe791e66c0bbe090edebfd74ba2b55d (`[LV] Keep track of cost-based ScalarAfterVec in VPWidenPointerInd`, committed Mon Sep 19 18:14:35 2022 +0100).
I have made a minimal IR reproducer, which shows the issue when fed to `bin/opt -passes='function(loop-vectorize),default<O2>' -S` (though `default<O2>` is not necessary to induce the issue, that is what is shown below).
```llvm
target triple = "aarch64"
define void @repro_issue_57912(ptr %arg, i64 %arg1) "target-features"="+sve,+sve2" {
entry:
br label %loop.body
loop.body:
%ind = phi i64 [ %ind.next, %loop.body ], [ 0, %entry ]
%gep = phi ptr [ %gep.next, %loop.body ], [ %arg, %entry ]
store i8 0, ptr %gep, align 1
%gep.next = getelementptr inbounds i8, ptr %gep, i64 1
%ind.next = add nuw nsw i64 %ind, 1
%cmp = icmp eq i64 %ind, %arg1
br i1 %cmp, label %exit, label %loop.body
exit:
ret void
}
```
On main this produces the following epilog loop:
```llvm
vec.epilog.vector.body: ; preds = %vec.epilog.vector.body, %vec.epilog.ph
%pointer.phi11 = phi ptr [ %arg, %vec.epilog.ph ], [ %ptr.ind12, %vec.epilog.vector.body ]
%index10 = phi i64 [ %vec.epilog.resume.val, %vec.epilog.ph ], [ %index.next16, %vec.epilog.vector.body ]
store <vscale x 8 x i8> zeroinitializer, ptr %pointer.phi11, align 1
%index.next16 = add nuw i64 %index10, %2
%ptr.ind12 = getelementptr i8, ptr %pointer.phi11, i64 %2
%9 = icmp eq i64 %index.next16, %n.vec6
br i1 %9, label %vec.epilog.middle.block, label %vec.epilog.vector.body, !llvm.loop !2
```
With a revert of 582f8ef19fe791e66c0bbe090edebfd74ba2b55d on main, I get the following, for which the workload does not have a checksum failure:
```llvm
vec.epilog.vector.body: ; preds = %vec.epilog.vector.body, %vec.epilog.ph
%index8 = phi i64 [ %vec.epilog.resume.val, %vec.epilog.ph ], [ %index.next10, %vec.epilog.vector.body ]
%next.gep9 = getelementptr i8, ptr %arg, i64 %index8
store <vscale x 8 x i8> zeroinitializer, ptr %next.gep9, align 1
%index.next10 = add nuw i64 %index8, %2
%9 = icmp eq i64 %index.next10, %n.vec4
br i1 %9, label %vec.epilog.middle.block, label %vec.epilog.vector.body, !llvm.loop !2
```
What I see is that in the broken case, the `%pointer.phi11` begins at `%arg`, and in the working case, it the equivalent store destination `%next.gep9` begins at `%arg + %vec.epilog.resume.val`. So it appears there is a problem with how the phi has been constructed in this epilog body, and it is not commencing where the main vector body has left off as it should.
cc @fhahn - My guess is that the patch is not at fault but has revealed a bug in epilog vectorization, do you concur? Would it be possible to revert your patch until a fix can be found? Thanks in advance.
</pre>
<img width="1px" height="1px" alt="" src="http://email.email.llvm.org/o/eJzNV9tu2zgQ_Rr7hbChi2VbD37IpQWC3aLFZtE-FpQ0sriRSZek7GS_fs9Qcnxp0nSxWKCBY0u8zJy5HHKmMNXT6k40ckeiUI5KT5WQopSOhKn5qaHywXUbUUvVdpaE0hjVRk-8VTslWwx4shoPN6PkGh-xN_ahNbIS3mCppZK0hxyp1zRKbkS2TOol1XFe0yKPaT4vo6KgKI-ooqKuFrNCJkWWVWKULEfzaJRd__55lN2K34i2wltZPjCy0jg_KQCzEvelbKW9qgHjM5UM8POnL6oi_ckEbHe6YjlQXZrNRnk28YPR4h7y4lzEy1F6Fc_wlWYiiZIEiq-jOMKOfDqKbkfRVf89uGkjK4JdG6XVBlbf_QETt9ZUXUmWlewbVTbCNWbvhG_gMOc6wihpUVNwCtAUSo-S92brxWQrnSM3Sm9HyaLudOmVwdyyNWY72SEgxqq_4bgcsiuqZdf6UXrzMRml77BBTO4hjV3lG9OtG5Z9uQrzyiFkXmjEwjlpnxiF0gz5CJGx-0Z6XrwfftkKLQpqzf7SG-zS8Gnb3aYf8tKuySNGatuSgEXAlUhpy2Y-w9PpbmBUmsTOKMR5FgUPfg0wvmbIigQGbb3F_gwyGZmaz4a3GEhYcK9tUpP0yEvHCtiH2HrtdmxN_4B3RHRx3etFKtonhLp_E6KwopWwj2Wzx6cFCNFPHl-Py7EKbgumbRvVg8quh-GppkfPWE9lYf42jGFZNEwGEGHiKHaNZDyIDZb3YjH8ptijk16S7ZBAiPCy1z44FWL5TbZqrUV8DiPoC1jgXmppA5G8TenCdLpykPW9JPZEfO6loxxZVUJ3e6Hd_hBHzPO20y3lpveA4gf6drHyEPtj3FQ87OLp5yDSo_JnAxdRDfPHgFrkKyfhkJuL24vkPs3Zjxrcx-niGzBjYHxP8dq0YIjSa0Fb1Zq1YKXPWl7lC9g97TdMe6If0k28_jdKr6GbEIaeX9krQnqXnUxumxNfb_uTEYMqjl_Ku2NCnYm4SDtsmCI-TNfLtSdYLjIdG-gxjl4i0cl-MLrb0HQn25_AEUSGfIvnPw2lJwZOyZ3DBULiUSzxj-RO34m_ycJDyuN-w-FrT_L9zHMvc-gUzVn-HzOaHTAATU7DcvDnS_Rb_gjFIPpUWP4Kmy49pdlB8--IlZ-x6MSdG1VVLU2L1pQPr635LhNjzvop04Jfkh-w7IvyTSgbdmQ93_Q_XS-Ynp-s8E6Ei-iUmjxcGztczzz1XKdUhvr7Mdzv39c8vzKTQ0CX_xOZon_Da94xxXWQv5W85zd6b8B_I-Wz6jcJGb1KyOULfHyLQtEphWa_DoW4ersTjriyG2o6HVK-sOYBdSiX9325RyKUxhcHCirGgtZKO4Gt_QIOWl9ESxRAgzhmEN96B3mq5xx96xRyjcv-Pp4VOa-05Mp2EHeM2Mu6uAZ_PYnn0VTcG1Ynt1uSNtzCNlgr-WYukHlizwcJ6tcAibnRSAdVbL7Rztsu9DqH-3y4uA_uDkb6Q93MfQPpkk3dB0UsMtQCfZTCtiC_pZoPLXROjvejfu7a6qxuLksueOtGNlpMxIcnse5Qkz8HKoCVHkfUoBtjoZwXReeDCj4Z4Vxu04puzQYM2A_Nguw7iBuca-LJdGxu2dlR-l58YTSMq4AS45yCo7gVGA5bLLaD8k571UJDrR4RXW4BcHii-mMpf6KVe3ChEax2Upc02DemVTyfp8tFFkfZuFqlVZ7mcuyVb2l13WfeGdTebRuU_exZRPsYYgaNPgI2c85AFQI47my7arzfOj6Qk_f4rBHjrpgiPngJx3H_M0ES_AUdeA1NBbqD96GvGDercr6IFllaxTM5X8RxmUVJXlXpPFrW-TJdJOPARrfCcYjWoUC_SWz31dXVcyMT-JgAOBTDoMHhh_GLOByG7z-_45fsdqxW3GZGeZLEeRrPsikt0yzL6zRN8mVMsxwJQkivdhoob-x6bFfBLATcYbJVzrvjJBpIHHlEATLkyw7NoF1tcQTbvWxbshNpN-PgiVVwwz-1-uEj">