OpaqueClosures cause the LLVM verifier to error or segfault

[vtjnash]

For example:

julia> code_llvm(()) do; Base.Experimental.@opaque(@noinline x::Int->println(x))(1) end                                                                                                                                                                                                          
Global is referenced in a different module!                                                                                                                                                                                                                                                      
{}* @"*Main.Base.println#491.jit"                           
; ModuleID = '#9'                                                                                                                               
  %"*Main.Base.println#491" = load {}*, {}** @"*Main.Base.println#491", align 8, !dbg !DILocation(line: 1, scope: <0x556b30a08d70>), !tbaa !27, !invariant.load !0, !alias.scope !29, !noalias !30                                                                                               
void ({} addrspace(11)*, i64)* @"julia_opaque closure_489"                                                                                      
; ModuleID = 'opaque closure'                                                                                                                   
Failed to verify module '#9', dumping entire module!                                                                                            

[gbaraldi]

Is this a new error?

[maleadt]

Looks like a new error (requires LLVM assertions):

441fcb1e06513ceff4d5ca4e111863c2ebba957b is the first bad commit
commit 441fcb1e06513ceff4d5ca4e111863c2ebba957b
Author: pchintalapudi <34727397+pchintalapudi@users.noreply.github.com>
Date:   Mon Jul 31 17:58:04 2023 +0000

    Merge opaque closure modules with the rest of the workqueue (#50724)

    This sticks the compiled opaque closure module into the
    `compiled_functions` list of modules that we have compiled for the
    particular `jl_codegen_params_t`. We probably should manage that vector
    in codegen_params, since it lets us see if a particular codeinst has
    already been compiled but not yet emitted.

 src/aotcompile.cpp |  11 +++---
 src/codegen.cpp    | 112 +++++++++++++++++++++++------------------------------
 src/jitlayers.cpp  |  19 +++++----
 src/jitlayers.h    |  25 +++++++-----
 4 files changed, 78 insertions(+), 89 deletions(-)

ref #50724, which was backported, so putting this on the milestone...

[gbaraldi]

I looked a bit into this, and it's the code that we execute to make the printing better looking that causes the module to become invalid