Upgrade to support LLVM 4.0 (development)

[dylanmckay]

I get this in the tip of avr-support-4.0

configure --enable-debug --disable-docs --enable-llvm-assertions --enable-debug-assertions --enable-rustbuild
make rustc-stage1

Building stage1 std artifacts (x86_64-apple-darwin -> x86_64-apple-darwin)
   Compiling core v0.0.0 (file:///Users/dylanmckay/projects/avr-rust/rust/src/libcore)

Assertion failed: (isa<X>(Val) && "cast<Ty>() argument of incompatible type!"), function cast, file /Users/dylanmckay/projects/builds/avr-rust/rust/build/x86_64-apple-darwin/llvm/include/llvm/Support/Casting.h, line 237.
error: Could not compile `core`.

This triggers it

 build/x86_64-apple-darwin/stage1/bin/rustc /Users/dylanmckay/projects/avr-rust/rust/src/libcore/lib.rs --crate-name core --crate-type lib -C debug-assertions=off -C metadata=877272a66b7a4e4d -C extra-filename=-877272a66b7a4e4d --out-dir ./libcore --target x86_64-apple-darwin -L dependency=/Users/dylanmckay/projects/builds/avr-rust/rust/build/x86_64-apple-darwin/stage1-std/x86_64-apple-darwin/debug/deps --cfg stage1 --sysroot /Users/dylanmckay/projects/builds/avr-rust/rust/build/x86_64-apple-darwin/stage1 -g -C debug-assertions=y -C codegen-units=1 --emit llvm-ir -o libcore.ll

[shepmaster]

(lldb) p *Val
(llvm::Metadata) $0 = (SubclassID = '\x90', Storage = '�', SubclassData16 = 4046, SubclassData32 = 1)

Coming from RustWrapper.cpp:

-> 319  DEFINE_ISA_CONVERSION_FUNCTIONS(Metadata, LLVMRustMetadataRef)

   327  DIT* unwrapDIptr(LLVMRustMetadataRef ref) {
-> 328      return (DIT*) (ref ? unwrap<MDNode>(ref) : NULL);
   329  }

   566  #if LLVM_VERSION_GT_OR_EQ(4,0)
-> 567          unwrapDI<DIExpression>((LLVMRustMetadataRef) Val),
   568  #else

From rustc_trans::debuginfo::metadata::create_global_var_metadata in metadata.rs:

-> 1939         llvm::LLVMRustDIBuilderCreateStaticVariable(DIB(cx),

[shepmaster]

Hard coding that to nullptr gets to the next error:

Program aborted due to an unhandled Error:
Error value was Success. (Note: Success values must still be checked prior to being destroyed).

appears to be coming from:

   160  LLVMRustArchiveIteratorFree(LLVMRustArchiveIteratorRef rai) {
-> 161      delete rai;
   162  }

[shepmaster]

temporarily removing the error check and I've gotten back to compiling stage2. That seems to indicate that the end of this set of upgrades is in sight!

[dylanmckay]

This is way more work that I remember it being last time lol

[shepmaster]

@dylanmckay heads up - I'm going to do a rebase on Rust master and squash some things together. I want to post something to upstreams issue tracker about upgrading to 4.0 because it seems so difficult. Do you have any outstanding work?

[nateozem]

shepmaster: "temporarily removing the error check and I've gotten back to compiling stage2"

I'm looking into this issue, but I'm not sure if anybody settled a solution for this.
Will the error check be removed or is there a way to satisfy the test?

[shepmaster]

@nateozem it's tricky 😈

There are two competing projects to keep up with: Rust and (AVR-)LLVM. Working against any given "older" version is potentially a waste of time; things might have changed between then and the present and there would be wasted work.

On my laptop (and not pushed anywhere yet), I am attempting to integrate avr-llvm/llvm@3fa7cf1 and rust-lang/rust@cae6ab1, which are the latest commits in each repository at this point in time. This also means that I am getting a new error, different from the others above. 😸

Generally, my approach has been "quickly hack up Rust to get it to compile, then go back and tease out each fix to prepare a proper PR". I'm still in the "quickly hack up" part, so saying whether any given hack will be preserved is difficult at best.

[shepmaster]

I've also created rust-lang#37609 to potentially get suggestions and guidance from people more familiar with the marriage of Rust + LLVM, as well as a way to ease getting the changes upstreamed.

[dylanmckay]

No outstanding work @shepmaster.

Your description on rust-lang#37609 sounds 💯

[shepmaster]

Ok, I've rebased, cleaned up, and force-pushed to avr-support-4.0 (de6eefe581ae70452324110cea534a970639b9f0). I was able to completely build the compiler with this branch and an out-of-tree build of AVR-LLVM with Rust support (avr-llvm/llvm@d93e840). I have not attempted to actually compile any AVR code with it 😇

[dsvensson]

Sorry to add noise here, but I'm super excited about the AVR support and just tried to build with ../configure --target=avr-atmel-none --disable-jemalloc, and in the Entering 'src/rust-installer I get an error that AVR target doesn't exist. Is that just reality, or is the README not updated to the latest state of things?

[shepmaster]

he README not updated to the latest state of things?

Not updated, and the code isn't really ready yet. I'm working on it, I swear!

[dylanmckay]

By the way @shepmaster, there is technically one difference between the downstream and upstream versions of the AVR backend.

In downstream, the backend is setup in CMake to be a first-class backend (enabled by default), but upstream it is set up as an "experimental" backend, which must be enabled by cmake -DLLVM_EXPERIMENTAL_TARGETS_TO_BUILD=AVR.

We will need to add this flag when we add AVR support to Rust properly.