Fix typeOf const struct whose type is named. Previously, the explicit struct type was emitted instead of the associated type name, making the AST uneccesarily verbose.

`extractValue` is modified so that it can reconstruct its return type even in
the presence of `NamedTypeReference`. This implies a MonadModuleBuilder dependency as in `gep` because named type references need to be looked up in the module builder state.

Co-authored-by: Guerric Chupin <guerric@chnik.fr>
Co-authored-by: Andrew Anderson <aanderso@tcd.ie>

* Handle NamedTypeReference in extractValue

extractValue would error if given an alias to a struct/array,
since it relied on extractValueType that doesn't know about
aliases.

Now it calls out the same function gep uses for type resolution.
Because gep also accepts pointers and vector types, we check
beforehand that only the the expected types were given

Fixes llvm-hs#295

* Add missing HasCallStack for gep / extractValue

Errors in `gep` would not show the call stack due to this

Co-authored-by: Andrew Anderson <aanderso@tcd.ie>

- Fix all outstanding warnings when building against LLVM 9.0.1
- Remove deprecated `ListT` instances
- Remove redundant imports
- Correct hard-coded assumption that size_t is 32 bits in `LLVM.AST.Constant`
- Add a matching non-constant sizeof in `LLVM.IRBuilder.Instruction`
- Update test for extractValue in `LLVM.Test.Constants` and add a link to the upstream bug filed on the LLVM bugzilla

There is still a lot of stuff missing but given that the existing API
is going to go away in the next major LLVM release, I’m going to merge
this in the current state so we have at least something after that.

This reverts commit 5eac408.

This functionality is not in LLVM 9 and so shouldn't be on this branch

* Add support for atomicrmw on floating point types

* Generate rmw operation enum instead of defining separately

Update llvm-hs from LLVM 9 to LLVM 12.

List of changes:

- Attributes
  - Various minor changes, not listed in detail.
- Debug information
  - DIFlag: BlockByrefStruct and ArgumentNotModified removed.
  - DISubrange
    - New arguments: upperBound and stride.
    - Arguments {lowerBound,upperBound,stride} are generalized and may be
      DIVariable or DIExpression in addition to Constant Int64.
  - DIModule
    - New arguments: apiNotesFile and lineNo.
    - Removed arguments: isysroot.
- Instructions
  - New instructions: freeze.
  - Modified instructions
    - shufflevector: now takes mask as [Int32] instead of Constant.
- JIT
  - Some deprecated OrcJIT v1 APIs are now removed: VModuleKey, SymbolResolver.
  - Start updating APIs and tests using OrcJIT v2 APIs. Work-in-progress.
- Target
  - Target option flag changes, not listed in detail.
- Transforms
  - Some existing transforms were removed or merged into others.
    Some had options modified.

- Newly added APIs
  - Added Triple (target triple) API to llvm-hs-pure, along with string
    conversion and parsing.
  - Added debugging utilities: dumpMetadata.
- Tests
  - Textual LLVM IR changes
    - Function parameters are now printed with local identifiers (e.g. `%0`).
    - Instruction alignment arguments must now be explicit: 0 no longer seems
      like a valid default value.
  - Instrumentation tests
    - Update MemorySanitizer test to run only on supported platforms: matching
      the implementation at https://github.com/llvm/llvm-project/blob/release/12.x/llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp.
  - Optimization tests
    - Add a more robust "is module vectorized?" check for the SLPVectorization
      test.

Current status: all tests pass via `stack test`.

Todos:
- Relevant comments to resolve: `{FIXME,TODO,NOTE}(llvm-12)`
- Lots of code cleanup and API design finalization for OrcJIT v2.
- Update installation instructions for llvm-hs@llvm-12.
- Fix CI for llvm-hs@llvm-12.
  - Vote for using GitHub Actions instead of Travis CI.

Co-authored-by: Adam Paszke <apaszke@google.com>

llvm/llvm-project@65fd034

The symbols they call in the LLVM C bindings are not resolved (as of 12 Apr 2021 and possibly before). (llvm-hs#322)

It seems reasonable to expect users of llvm-hs to be able to compose `writeFile stderr <$> ppLlvm <$> moduleAST`

Since LLVM 12 changed the mask argument of `shufflevector` to a constant vector of `Int32`, `ppLlvm` requires a `Typed` instance for `[Int32]` for the above snippet to work properly.

…y LLVM components, remove hardcoded parallelism limits and let cmake figure things out automatically

…e getElementPtrType or getElementType

…nings generated by code in the LLVM release/9.x branch

…ileUnit -- change it in llvm-hs also to avoid divergence

* Apply fix for llvm-hs#315 and update C++ standard version (LLVM 12 uses C++14)

* Update build options and README to address llvm-hs#315

* Handle NamedTypeReference in typeOf

* Fix warning due to unused captures

* Add a test for the resolution of typedefs (i.e. NamedTypeReference) in constant expressions

* Add support for more DWARF operators (partially addresses llvm-hs#306)

* Print an explicit error message for unhandled top-level DINodes, instead of the inscrutable "Unknown sublass id for DINode: <integer>" error message.

* Rather than crash when we encounter a DINode for which lifting to Haskell is not implemented, simply ignore it. This way, the input program at least loads, with some debug information missing. Otherwise, llvm-hs simply cannot be used with IR produced from the LLVM tools with debug info emission turned on (for example, 'clang -g' will produce IR that crashes llvm-hs).

* Fix two type signatures that got lost in the cherry-picking from the llvm-9 branch

* Build fix for GHC-9

Fixes llvm-hs#338

* Build fix for GHC-8.6

* Build fix for GHC>=8.10

Fixes llvm-hs#315 (properly)

Co-authored-by: Andrew Anderson <aanderso@tcd.ie>

…ith or without RTTI. Reinstate the unfortunately necessary stanza in . Fixes llvm-hs#354.

Fixes llvm-hs#354

…type of a getelementptr instruction, show the invalid index so the user has some chance of fixing the problem.

…strip off an index while processing 'NamedTypeReference's in constant expressions

…diagnose what is causing the failures. Audit LLVM.Internal.FFI.Type for duplication of the LLVM C API and remove it (addresses llvm-hs#358).

…o remove old, stale text, and encourage creation of issues for incomplete LLVM API coverage.

…I layer to make it easier to debug when things go wrong

…resses llvm-hs#363)

Add int16 helper function

* IR-builder: use first terminator during codegen

Previously, the latest terminator instruction was always used (`br`,
`condBr`, `ret`, ...). This limited composability when creating
combinators based on the IR builder.

Now, we keep track of the first terminator which is more in line with
how LLVM semantics work (terminators that appear later in a basic block
are dead code).

With this change, we can now create control flow combinators that work
like we expect them to:

```haskell
-- A combinator for a "one sided if statement"
if' condition asm = mdo
  condBr condition begin end
  begin <- block `named` "if.begin"
  asm
  br end
  end <- block `named` "if.end"
  return ()

example = do
  function "f" [(AST.i1, "a"), (AST.i1, "b")] AST.i1 $ \[a, b] -> mdo

    entry <- block `named` "entry"; do
      if' a $ do
        if' b $ do
          -- Previously, this ret instruction would not be emitted!
          ret (bit 0)

      ret (bit 1)
```

* Add tests for emitting only first terminator

* Update CHANGELOG

* Add nested control flow test

`cxx-flags` aren't applied to `c-sources`, causing build errors on macOS.

llvm-12 has gotten quite a few updates that weren't included here, so
this is an attempt to sync them up.