Remove literal_static_pointer_val from literal_pointer_val implementation

[pchintalapudi]

This eliminates most of the literal pointers outside of ccalls and directly JIT-ted code, except instead of adding names we instead just use imaging mode and fix up the visualization in jl_get_llvmf_defn.

Depends on #50631

[vchuravy]

cc: @maleadt

Have you tested GPUCompiler against this?

[pchintalapudi]

Have not tested GPUCompiler, though whatever method GPUCompiler is currently using to handle global_targets should still apply here.

[gbaraldi]

It's failing the literal string comparison tests but it seems to work .

[maleadt]

This broke RuntimeGeneratedFunctions.jl: https://s3.amazonaws.com/julialang-reports/nanosoldier/pkgeval/by_date/2023-07/26/RuntimeGeneratedFunctions.primary.log

Program aborted due to an unhandled Error:
Symbols not found: [ *Core.getfield#1 ]

[9] signal (6.-6): Aborted
in expression starting at /home/pkgeval/.julia/packages/RuntimeGeneratedFunctions/Avj6S/test/runtests.jl:152
gsignal at /lib/x86_64-linux-gnu/libc.so.6 (unknown line)
abort at /lib/x86_64-linux-gnu/libc.so.6 (unknown line)
_ZNK4llvm5Error19fatalUncheckedErrorEv at /opt/julia/bin/../lib/julia/libLLVM-15jl.so (unknown line)
assertIsChecked at /source/usr/include/llvm/Support/Error.h:264 [inlined]
~Error at /source/usr/include/llvm/Support/Error.h:226
~JITSymbol at /source/usr/include/llvm/ExecutionEngine/JITSymbol.h:322 [inlined]
operator() at /source/src/jitlayers.cpp:1510 [inlined]
withModuleDo<JuliaOJIT::addModule(llvm::orc::ThreadSafeModule)::<lambda(llvm::Module&)> > at /source/usr/include/llvm/ExecutionEngine/Orc/ThreadSafeModule.h:136 [inlined]
addModule at /source/src/jitlayers.cpp:1492
jl_add_to_ee at /source/src/jitlayers.cpp:1964

Maybe related to opaque closures? I'll see if I can come up with a minimal reproducer.

EDIT: Diffractor.jl has the same issue (cc @oxinabox).

[maleadt]

MWE:

@generated function foo()
    quote
        $(Symbol("wat"))
        $(Expr(:opaque_closure, :(() -> bar)))
    end
end
foo()

Requires Julia#master with LLVM assertions.

[maleadt]

Doesn't reproduce on very latest master though, I guess #50650 may have fixed this?
EDIT: RGF works now, but Diffractor still doesn't.

[pchintalapudi]

This is unfortunately not a convenient missing symbol, but is a memory corruption somewhere. We seem to be allocating two llvm::GlobalVariable objects at the same address (testing on master before #50650 with the MWE), and I can't debug further because GlobalVariable overrides operator new and has its own custom allocator, which removes the normal forms of memory debugging.

[gbaraldi]

Are we freeing the GlobalVariable objects somewhere?

[pchintalapudi]

No, hardware watch doesn't detect an intervening destructor call.

[Keno]

Possibly the same issue as #48742?

[pchintalapudi]

That does seem likely, since we'd be freeing the associated module and potentially the globalvariable there too. @vtjnash the WeakVH is too expensive, but can we just make global_targets a map of void* -> std::string? Alternatively we could generate a new private global every time we call julia_pgv, and then mark it up with information in the GV attributes that allows us to merge duplicate globals post-codegen.

[vtjnash]

That code really just should not call jl_merge_module, since we don't support that. The output of codegen is supposed to be the result of doing emission of one thing into a single module, so that we can handle errors and incremental work.

[maleadt]

I'm also getting the following error on master with Metal.jl:

error: Linking globals named '*Core.Intrinsics.bitcast#1': symbol multiply defined!

Reverting to before this PR fixes that, although another PR may be the actual cause of course.

[pchintalapudi]

Does it still fail on #50724 ?

[maleadt]

Does it still fail on #50724 ?

Yeah.

EDIT: this is when GPUCompiler invokes the linker; I'll have a closer look myself first.