Changelog for llvm-hs-8.0.0

8.0.0 (2019-03-10)

Upgrade to LLVM 8
It is now possible to define multiple variables with an empty name without causing a name collision. Variables with an empty name cannot be referenced. This fixes an issue caused by using IRBuilder’s extern helper.

Support llvm-config executables named llvm-config-7 in addition to llvm-config-7.0 and llvm-config.

Throw an EncodeException if a local variable is defined multiple times. This is consistent with llc which also forbids reusing variable names.
Update for LLVM 7.0:
- The ORC JIT API has been changed to reflect the changes in LLVM itself. The test suite should give you a good idea of how to adapt your code but here is a summary:
  - There is now a separate ExecutionSession that tracks module handles.
  - SymbolResolver only has a single resolver function.
  - Creating an ObjectLinkingLayer requires a way to get the symbol resolver for a module.
  - Creating a CompileOnDemandLayer requires a way to set and get the symbol resolver for a module.

findSymbol and findSymbolIn now return Either JITSymbolError JITSymbol if the JIT symbol has the error flag set or the address is 0. (This is consistent with how LLVM treats JIT symbols).
The return type of SymbolResolverFn has been changed to Either JITSymbolError JITSymbol. It is fine to return a 0 address in Right so existing resolvers can be adapted by wrapping the result in Right.
Fixed a bug where instructions were constant-folded during encoding. This caused problems since the API available on a Constant is not the same as the one on an Instruction (e.g., we cannot set metadata on a Constant).
Fix use-after-free in createObjectFile.
Add withObjectFile wrapper for createObjectFile and disposeObjectFile.
Add API for looking up symbols in the LinkingLayer.
Add LLVM.OrcJIT.CompileLayer and LLVM.OrcJIT.LinkingLayer modules that reexport the internal modules. The findSymbol/findSymbolIn methods that have previously been exported from LLVM.OrcJIT can now be found in LLVM.OrcJIT.CompileLayer.

Remove field prefixes from DIDerivedType, DIBasicType and DISubroutineType to make the API consistent with the other debug metadata types.
Change the type of the scope fields in DIModule and DINamespace to Maybe (MDRef DIScope) to reflect that they can be optional.

Remove the MetadataNodeReference constructor. References to metadata nodes are now encoded using the polymorphic MDRef type.
Add support for encoding and decoding debug metadata. Thanks to @xldenis who started that effort!
Drop support for GHC 7.10.
Support decoding/encoding of metadata in GlobalVariable and Function.
Fix check that the type of GlobalReference is correct in the presence of automatic renamings due to name collisions.
Extract LinkingLayer into a separate module.

Support for LLVM 6.0, take a look at the changelog of llvm-hs-pure for details.
Add AggregateZero for zero-initializing structs, arrays and vectors. Previously Null was used for null pointers as well as zero-inializing aggregates. The new behavior reflects LLVM’s internal representation and the C++-API.
Enforce that Null is only used on pointer types. Existing uses of Null on arrays, structs and vector must be changed to the newly introduced AggregateZero.

Fix argument order in LLVM_Hs_CreateTargetMachine. This affects withTargetMachine and withHostTargetMachine.
Add support for MCTargetOptions.

Set target options in withTargetMachine. Previously the options passed there were simply ignored.
Fix decoding of constant vectors.
Fix decoding of function attributes in calls.

Support for LLVM 5.0

We only give a summary of the changes affecting the public API of llvm-hs here. Please refer to the official release notes for LLVM 5.0 for an overview of all changes in LLVM 5.0.
- The X86_64_Win64 calling convention is now called Win64.
- There is a new Speculatable function attribute.
- The CrossThread synchronization scope has been removed. There is now a new System synchronization scope.
- The OrcJIT-API now operates on individual modules instead of sets of modules.
- The lessPreciseFloatingPointMultiplyAddOption field has been removed from the target options.
- The compressDebugSections option field is now of type DebugCompressionType instead of Bool.
- The BasicBlockVectorize pass has been removed. You should use SuperwordLevelParallelismVectorize instead.
Throw 'EncodeException' when the type supplied in a 'GlobalReference' does not match the type of the expression.
Throw 'EncodeException' when the result of instructions returning void is named using ':='.

Revamp OrcJIT API
- The user facing API is now exposed using LLVM.OrcJIT.
- All user facing functions have been documented.
- In addition the bracket-style API, there are now new* and dispose* functions making it easier to ingegrate OrcJIT in custom monad transformer stacks.
- There is a new CompileLayer typeclass which abstracts over the various compile layers in OrcJIT.
Support QuickCheck 2.10

Switch most of the API from String to ByteString.
Switch from ExceptT to using exceptions. See LLVM.Exception for an overview of the exceptions potentially thrown.