Stack overflow / segv in attributor on minimal subset of libc

[JonChesterfield]

Discovered while using parts of llvm libc as a debugging tool. Symptom was clang crashing at O1 or higher.

target datalayout = "e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32-p7:160:256:256:32-p8:128:128-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:\
64-S32-A5-G1-ni:7:8"
target triple = "amdgcn-amd-amdhsa"

%struct.foo = type { %struct.pluto, ptr, i64 }
%struct.pluto = type { [512 x i8], ptr }

@global.2 = internal addrspace(1) global %struct.foo { %struct.pluto zeroinitializer, ptr addrspacecast (ptr addrspace(1) @global.2 to ptr), i64 0 }

; Function Attrs: nounwind                                                                                                                                                                                         
define internal void @hoge(ptr noundef %arg)  {
bb:
  %load = load i64, ptr addrspace(1) getelementptr inbounds (%struct.foo, ptr addrspace(1) @global.2, i64 0, i32 2), align 8

  ret void
}

!0 = !{!"clang version 17.0.0 (git@github.com:llvm/llvm-project.git f717c2d4f2b7663317e46af069aaf88edca760de)"}

Reproduce

opt -mtriple=amdgcn-amd-amdhsa -amdgpu-attributor bugpoint-reduced-simplified.ll

Godbolt crashing (had to delete the target layout string) https://godbolt.org/z/MT5vbj5cx

[arsenm]

This is just directly recursing on a global used to initialize itself. Do we still not have a generic utility to walk through ConstantExprs correctly?

[jhuber6]

I think I actually remember something like this coming up, specifically in NVPTX and maybe AMDGPU I think we cannot use a self-referential initializer. This is something the libc implementation of the File ended up doing but I thought I statically trimmed that.

In any case, we should probably have a nicer failure mode if it is indeed illegal to initialize in this way.

[JonChesterfield]

This is just directly recursing on a global used to initialize itself. Do we still not have a generic utility to walk through ConstantExprs correctly?

There was a generic utility to expand them into instructions, but it didn't work. I think I've deleted it. Wouldn't apply to the global anyway. I don't know of a traversal API, the on the fly deduplication makes the things quite awkward to deal with.

A global referring to itself sounds fine to me, if the backed blows up on that we should fix that too.

[llvmbot]

@llvm/issue-subscribers-backend-amdgpu