- while working on NMODL + LLVM, we don't worry that much
    about Python bindings by default
  - so lets disable them by default

  * added NMODL_ENABLE_LLVM option to enable/disable
    llvm support in nmodl
  * LLVMHelper.cmake added to help with linking LLVM libraries
     - clang might need to use libstdc++ or libc++ linking
     - on BB5, using GCC with LLVM libraries is fine. But using
       clang results into lots of link error. Adding -stdlib=libstd++
       solves the issue
     - use check_cxx_source_compiles to find out which cxx flag is needed

 - added llvm dir under codegen where LLVM code generation
   work will live
 - llvm codegen visitor created that can be used as template
   for initial work
 - cmake adapted to enable llvm codegen based on CMake option
 - simple procedure.mod added that can be initial target for
   testing
 - new CLI option --llvm that runs LLVM codegen visitor
 - Enable CXX 14 because new LLVM versions require it

 - install llvm via brew
 - set LLV_DIR variable so that CMake can find llvm-config

  - print table with different build options, flags and paths
    used that can be helpful for debugging
  - fix git revision date for older git version
  - update INSTALL.md with correct brew paths for flex and bison

  - test/unit/codegen/llvm.cpp added for unit testing
    LLVM code generation visitor
  - ./bin/testcodegen binary can be used to launch
    LLVM codegen specific tests
  - multiple llvm_map_components_to_libnames removed
  - update procedure.mod with simple examples for IR generation

* Added LLVM code generation for `ProcedureBlock`.
* Added code generation routines for double, integer and
   boolean variable types.
* Added binary and unary operator code generation:
     - Supported binary operators: +, -, *, /.
     - Supported unary operators: -.
     - Assignment (=) is also supported.
* Added regex matching unit tests for LLVM code generation.
* Fixed Travis CI/builds.

fixes #451, fixes #452, fixes #456

Co-authored-by: Pramod Kumbhar <[email protected]>

* LLVM code generation for `FunctionBlock` is now supported.
* Terminators in function or procedure blocks are enforced:
      - Every procedure must have `ret void` instruction.
      - Every function returns a double, specified by `ret_<function_name>`.
* For local symbol table, code generation now uses LLVM's builtin
`llvm::ValueSymbolTable`.


fixes #454, fixes #469

* Add option to run LLVM optimisation passes
  - update CLI argument from --llvm to llvm --ir --opt
  - --ir runs CodegenLLVMVicitor and emits LLVM IR
  - if --opt is passed, we run basic LLVM optimisation passes
  - update simple test to check optimisation passes
* Add function example in procedure.mod
* Add test for LLVM optimisation passes and dead code removal

This patch adds support for function call code generation, particularly:

- User-defined procedures and functions can now lowered to LLVM IR.
- A framework for external method calls (e.g. sin, exp, etc.) has been created, currently `exp` and `pow` are supported.
- Corresponding tests added.

fixes #472

LLVM code generation for `IndexedName`s.

- Added code generation for initialising arrays in LOCAL blocks (with both integer constants and macros).
- Added support for indexing arrays.

fixes #467

NMODL AST needs various transformation to generate C++
code or LLVM IR. This PR is begining of AST transformations
to simplify code generation backends.

* New CodegenLLVMHelperVisitor to perform various AST
  transformations to simplify code generation for various
  backends and simulators.
* CodegenLLVMHelperVisitor is currently limited to LLVM
  backend to simplify initial implementation and keep
  C++ based backends working.
* CodegenLLVMHelperVisitor now handles FUNCTIONS and
  PROCEDURES blocks
  -  Replace LocalListStatement with CodegenVarStatement
  - Added new AST types for code generation
    - CodegenVar to represent variable used for code generation
    - CodegenVarType to represent codegen variable
    - CodegenVarListStatement to represent list of CodegenVar
    - CodegenStruct will be used in future to represent struct
      like NrnThread or Mechanism class

See #474

* Added new BinaryOp for += and -=
  * Added string_to_binaryop function
  * Added Void node type to represent void return type
  * Added CodegenAtomicStatement for ion write statements
  * llvm helper started handling visit_nrn_state_block
    - NrnStateBlock is being converted into CodegenFunction
    - for loop body with solution blocks created
    - voltage and node index initialization code added
    - read and write ion statements are handled
  * Some of the functions are now moved into CodegenInfo

Co-authored-by: Ioannis Magkanaris <[email protected]>

This commit introduces a functionality to execute functions from MOD file via LLVM jit.

For that, there is now:
- `JITDriver` class that, given a LLVM IR module, set ups the JIT compiler and is able to look up a function and execute it.
- `Runner` class that wraps around JIT driver. It helps to initialise JIT with LLVM IR module only once, and then run multiple functions from it. 

To execute functions, `nmodl_llvm_runner` executable is used. It takes a single mod file and a specified entry-point function, and runs it via LLVM code generation pipeline and JIT driver. Only functions with double result types are supported at the moment.

For example, for MOD file `foo.mod`:
```
FUNCTION one() {
    one = 1
}

FUNCTION bar() {
    bar = one() + exp(1)
}
```
running `nmodl_llvm_runner -f foo.mod -e bar` gives
```
Result: 3.718282
```

Tests for execution of generated  IR have been added as well.

fixes #482

Co-authored-by: Pramod Kumbhar <[email protected]>

* Added more bin ops and refactored code
   - Now, there are code generation functions for all comparison
      and logical operators.
   - Code generation functions are now split based on the expression "type"
      (assignment, arithmetic, comparison, logical). Moreover, the lhs and rhs
      expression results can be both double and integer. This is important for
      control flow code generation and for the new AST node CodegenVarType.
* Added support for NOT op
* Added default type flag to switch between float and double
* Added tests for single precision
* Renames LLVM test file to codegen_llvm_ir.cpp to follow convention.
* NOTE : Tests for new operators will be added when the first control
                flow node (most likely FOR node) will land.

fixes #453

* visit_statement_block of all FUNCTION and PROCEDURE
    blocks was called resulting in changing LOCAL
    statement to DOUBLE statement
  * As statement block doesn't need to be visited for this
    purpose, rename function to convert_local_statement
  * Call convert_local_statement only when required i.e.
    only when codegen function creation time.

fixes #491

* Handle CodegenVarType type in JSON printer
  - As AstNodeType is enum type and node itself,
    we need to print that explicitly
* Indent json visitor jinja template
 - initially template was not indented as code generated
   was not looking good
 - now all generated code is automatically clang-formatted
   so it's less of a concern. Readability is important.

fixes #493

* LLVM Helper visitor now can return a vector of `CodegenFunction`s.
* LLVM Helper visitor has been integrated into LLVM visitor:
   - The type of variables is still double by default, but can also be inferred from `CodegenVarType` node.
   - Procedure's return type changed to int (so that error codes can be returned in the future). 
   - New visitor functions added: for `CodegenReturn`, `CodegenFunction`, `CodegenVarList` and `CodegenVarType`.

* Added a new code generation function for conditional statements (`if`, `else if`, `else` and their nested variations).
* Added tests for the new code generation:
   - IR unit tests.
   - Execution tests.
* Fixed FP and integer comparison ordering in macros.

fixes #468

Added error handling when a non-scope value is looked up. Before, such a lookup would yield a nullptr, therefore leading to a segmentation fault. This PR adds a lookup function that wraps around value symbol lookup, and throws an error with a message if nullptr is returned.

Added support for WHILE statement code generation. Corresponding tests for IR generation and execution were also added.

Additional visitor for StatementBlock was added to reduce code duplication.

fixes #500

* Moved info related function to codegen_info
  - Moved get_float_variables, codegen_int_variables,
     codegen_global_variables, codegen_shadow_variables
     into CodegenHelper
  - Move small utility functions from CodegenCVisitor to codeged_utils
* Add proper variables to the mech_Instance
* Adding LLVMStructBlock
* Added test and visitor
* Fix llvm codegen tests with x[0-9].*

- Added support for string function arguments. These are
   converted into global `i8` array values.
- Added support for `printf` function call with variable number 
   of arguments.
- Refactored function/procedure call argument processing into
   a separate function.

fixes #510

* Move code gen specific InstanceStruct node to codegen.yaml
  - nmodl.yaml file is more for language constructs
  - InstanceStruct is specific for code generation and hence
    move it to codegen.yaml
* Update CI scripts
* fix cmake-format with v==0.6.13

 - instance structure now contains all global variables
 - instance structure now contains index variables for ions
 - nrn_state kernel now has all variables converted to instance
 - InstanceVarHelper added to query variable and it's location
* Support for codegen variable with type
* Add nmodl_to_json helper added in main.cpp
* Added --vector-width CLI option
* Add instance struct argument to nrn_state_hh
* Add comments as TODOs to support LLVM IR generation

Note that this commit and next commit (Part II) are required to
make LLVM IR code generation working. Vector IR generation is
working except indirect indexes. See comment in #531.

  - remove undefined visit_codegen_instance_var
  - Improved member creation for instance struct
  - Instance struct type generation for kernel arguments
  - Proper integration of instance struct
  - Added scalar code generation for the kernel
  - Removed instance test since it is not created explicitly anymore
  - Fixed ordering for precision and width in LLVM Visitor
  - Added vector induction variable
  - Vectorised code for compute with direct loads fully functional
  - Instance naming fixed
  - (LLVM IR) Fixed compute vector code generation types
  -  refactoring : improve coversion of double to int for
    the loop expressions

This PR adds a unit test to check LLVM instructions generated for
the scalar kernel, particularly:

- FOR loop blocks.

- Induction variable increments and comparisons.

- Correct loads through GEPs from the struct.

Test for vectorised code generation would be added in a separate
PR or when full vectorisation support (indirect indexing) would
land.

Improved index code generation within the LLVM pipeline.
The following issues were addressed:

Array indices are i64 per LLVM's addressing convention.
This means that if the value is not a constant, an additional
sext instruction must be created.

Bounds check is removed since it requires a certain analysis
on the index value. This can be addressed in a separate PR.

`IndexedName` code generation is separated into 2 functions
The first, `get_array_length()` is responsible for array initialisation,
the second, `get_array_index()`, for indexing. In latter case, we
support the following cases:
```
...
// Indexing with an integer constant
k[0] = ...

// Indexing with an integer expression
k[10 - 10] 

// Indexing with a `Name` AST node that is an integer
// (in our case a FOR loop induction variable or a variable
// with `CodegenVarType` == `Integer`
k[id] = ...      
k[ena_id] = ...
```
Note that the case:
```
// id := loop integer induction variable
k[id + 1] = ... 
```
is not supported for 2 reasons:

On the AST level, as per #545 the expression would
contain a Name and not VarName node that fails the
code generation.

The case only arises in the kernel functions like state_update,
where indexing is "artificially" created with indexing by a Name
only.

fixes #541

* CodegenLLVMHelperVisitor improved without hardcoded parameters
* Added get_instance_struct_ptr to get instance structure for variable information
* test/unit/codegen/codegen_data_helper.cpp : first draft implementation
   of codegen data helper
* Added test for typecasting to the proper struct type

Co-authored-by: Pramod Kumbhar <[email protected]>

* Implement remainder loop along with main vector loop
* Add unit test for the same

fixes #532

* return value in PROCEDURE block was not initialised
* do the initialisation as part of ASTR transformation
* remove initialisation specific code from LLVM visitor

fixes #530

* Added support for arguments in the JIT llvm runner
* Adjusted tests and added a simple kernel test
* Removed printfs from the kernel
* Fixed kernel number of arguments check
* Initial integration of dataHelper for kernel tests
* Implemented a test to check the scalar kernel execution

* Added renaming for loop local variables in CodegenForStatement
* Fixed trip count in main loop and removed epilogue loop for scalar case
* Refactored loop remainder tests and added a scalar case
* Change `reminder` to `epilogue` in the test

* Add gather support
* Fixed vectorisation patterns and added simple JIT tests
* Added IR regex test for gather

Added several minor improvement to the current pipeline
infrastructure. Particularly, the following was addressed:

- The generated IR module is now verified after running the
visitor
- The kernel is checked if it can be vectorised or not
- The generated IR can be dumped to `.ll` file with
`-o <filename>`
- Printing LLVM IR is moved to debug mode

* avoid local variables inside loop to not have allocas
* this was causing stack overview for large instance count

Introduced the benchmarking for LLVM code generation pipeline.
For that, new options have been added:

```
benchmark
  LLVM benchmark option
  Options:
    --run                                                               Run LLVM benchmark (false)
    --instance-size INT                                       Instance struct size (10000)
    --repeat INT                                                    Number of experiments for benchmarking (100)
    --backend TEXT:{avx2, default, sse2}         Target's backend (default)
```

The JIT runner has also been modified to extract the target
information correctly, and disable available CPU features for
benchmarking a specific backend.

Example:
```
$ nmodl hh.mod llvm --ir --vector-width 1 benchmark --run --instance-size 100 --repeat 2 --backend default

Created LLVM IR module from NMODL AST in 0.006765817

Benchmarking kernel 'nrn_state_hh'
Experiment 0: compute time = 0.013977749
Experiment 1: compute time = 0.004847989
Average compute time = 0.0058550929
```

Co-authored-by: Pramod Kumbhar <[email protected]>

- allocate instance data only once
- store memory size with instance data
- print memory size while running benchmarking kernel

- LOCAL statement was not deleted correctly
- Instead of getting first element from statement vector,
   use local statement pointer to erase it from the node.
  
Related to #594

* Made compatible with LLVM 13 and replaced void* with i8*

  - if LOCAL variable was declared inside DERIVATIVE block then
    we were getting error:
      "Stored value type does not match pointer operand type!"
  - the error was happening because scalar variable from epilogue
    loop was conflicting with the vector type variable in main loop
  - to avoid conflict between main and epilogue loop, rename all
    local variables in epilogue.
  - bug fix for recursive handling of LocalList statement

fixes #594

 * In order to use VecLibReplace pass, we need LLVM 13 / trunk
 * Change ubuntu image on azure from 16.04 to 18.04
 * Install llvm-13 nightly snapshot
 * Enable LLVM build on Ubuntu
 * For Mac OS use pre-built binary package from https://github.com/pramodk/llvm-nightly
 * We will see if we get OS X bottle from BlueBrain/homebrew-tap/pull/7

Added support for replacing LLVM IR maths intrinsics with vector
maths functions from Accelerate, libmvec, MASSV, and SVML. To
trigger the replacement, a new `--veclib` option should be used.
This is only supported on LLVM 13+.

Example:
```
$ bin/nmodl hh.mod llvm --ir --vector-width 4 --veclib SVML
```

fixes #589

Co-authored-by: Pramod Kumbhar <[email protected]>

* Integrated veclibs in benchmark and added shared libs support for JIT
* Tested on BBP Ubuntu Linux box
* Make sure to set LD_LIBRARY_PATH for Intel library dir

- std::ofstream().rdbuf() was used but as it was a local object,
   it becomes invalid at the end of function scope
 - make std::ofstream as member variable

The following is added:

1. Dumping object files in JIT.

A functionality to dump (enabled by default) the generated from LLVM IR file binary to `.o` has been added to benchmarking. Now, in addition to logs, a `v<vector_width>_<mod_filename>.o` is generated. The reasons it is an object file and not an assembly (hence not included in logs) are the following:

- LLVM does not have library functions that take the object and turn back into assembly, but rather `object -> file -> assembly` path. It also has a `llvm-objdump` tool, but it is intended as a command-line utility and does not have a well-defined API.

- Writing custom functions to produce a readable assembly is not a priority. Also, mimicking `objdump` functionality would be difficult.

- Both `objdump` and `llvm-objdump` can be used to isnpect the `.o` file manually.

2. Refactoring of `Runner` class.

In addition to the support of dumping the binary, `Runner`and `JITDriver` classes were refactored to have a nicer OOP-style.

fixes #611

Co-authored-by: Pramod S Kumbhar <[email protected]>

This PR adds two flags to the benchmarking pipeline:

* `--opt-level-ir`: This flag is used to run `-On` passes on the
   generated LLVM IR module.
* `--opt-level-codegen`: This flag is used for setting optimisation level
    for machine code generation inside the JIT target machine.
* As an example:

```bash
$ ./nmodl file.mod \
     llvm --ir --vector-width 1 \
     benchmark --run --instance-size 10000000 --repeat 20 --opt-level-ir 2 --opt-level-codegen 2
```

fixes #616

Added debug support to LLVM code generation pipeline. Currently,
only basic support was added:

1. Debug information about functions (name)
2. Debug information about module

**What has been changed and added**

1. A new class `DebugBuilder` was created. It is used as a wrapper
around LLVM's `DIBuilder` and holds important information such as
`LLVMContext`, debug file and compile unit. It also wraps `DIBuilder`'s
functionality into a more suitable API.

2. A temporary `Location` struct has been added. It encapsulates the
location of the source AST construct and reflects `ModToken` on LLVM
code generation level. It is only used if the location of the source NMODL
function is known.

3. LLVM visitor know takes an extra `add_debug_information` flag and
handles debug information creation. Fore readability, `IRBuilder` was
renamed to `ir_builder`.

4. JIT runner is now able to listen for GDB, perf (build LLVM with
`-DLLVM_USE_PERF=ON`) and VTune (build LLVM with
`-DLLVM_USE_INTEL_JITEVENTS=ON`) events.

5. Necessary cmake changes were added to optionally support JIT event
listeners (`-DNMODL_HAVE_JIT_EVENT_LISTENERS`).

**How to generate debug information**

Debug information is attached to every function, procedure or artificially
created kernel (and corresponding wrappers). Debug information is enable
by default, so to turn it off use ` --disable-debug-info` flag. For example,
the given NMODL
```nmodl
1   FUNCTION func(x) {
2     func = x
3   }
4
5   PROCEDURE proc() {}
```
is transformed (running `./bin/nmodl <filename>.mod llvm --ir`) into
```llvm
define double @func(double %x1) !dbg !4 {
  ; ...
}

define i32 @proc() !dbg !6 {
  ; ...
}

!llvm.dbg.cu = !{!0}
!llvm.module.flags = !{!3}

!0 = distinct !DICompileUnit(language: DW_LANG_C, file: !1, producer: "NMODL-LLVM", isOptimized: false, runtimeVersion: 0, emissionKind: FullDebug, enums: !2)
!1 = !DIFile(filename: "foo", directory: ".")
!2 = !{}
!3 = !{i32 2, !"Debug Version", i32 3}
!4 = distinct !DISubprogram(name: "func", linkageName: "func", scope: null, file: !1, line: 1, type: !5, scopeLine: 1, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !0, retainedNodes: !2)
!5 = !DISubroutineType(types: !2)
!6 = distinct !DISubprogram(name: "proc", linkageName: "proc", scope: null, file: !1, line: 5, type: !5, scopeLine: 5, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !0, retainedNodes: !2)
```

fixes #592 #612 

Co-authored-by: Pramod Kumbhar <[email protected]>

  * in the new deployment, pgi module is changed to nvhpc
  * fix gitlab CI script accordingly

* Fix `append_statements_from_block` function in LLVM helper visitor.
* Before, if nonspecific current was not specified, the whole `BREAKPOINT`
   block would be added to the kernel body.
* This led to cases when `SOLVE` block was together with the actual
    solution to `DERIVATIVE`

Improves the code structure for the LLVM code generation pipeline

The following changes were added:

1. New IR builder class.

Before, LLVM visitor just simply used `llvm::IRBuilder<>` class to
generate instructions. Recently, this (as well as adding the functionality
to the visitor on the go) had led to code duplication and it became hard
to introduce new features nicely.

Hence, a special `IRBuilder` class is now used. This class is a wrapper
around `llvm::IRBuilder<>` that keeps track of certain IR generation
specific fields (that are unrelated to the visitor), defines an API that the
visitor can use to generate LLVM IR.

Also, this IR builder has been designed to be nearly fully-independent from
NMODL AST nodes. this allows it to be more generic and to be more
extensible.

2. Visitor clean-up

 LLVM visitor has been refactored to take the new IR builder class into
account. Also, the functions were reordered, refactored and renamed to
better reflect the intended use and provide encapsulation.

3. Scatter preparation

The functionality of the generating code for `CodegenInstanceVar` node
has been extended with `read_from_or_write_to_instance(...)` function.
Now, an optional `value_to_store` is passed to indicate whether the code
needs to be generated for reading the instance variable or writing to it.


fixes #538

  * Fixed CodegenAtomicStatement initialisation
  * Removed unused variable and changed comment

* Instance data structure initialization had following bug
   - instance struct has int member variables which act as
     offsets to other vectors (e.g. node_index, na_ion_index)
   - these variables were initialized from 1 to N where N was
     incremented always without considering the upper bound on
     for offset.
* With this fix
   - index / integer variables are always initialized from
     0 to N-1.
   - Variables are initialised 1e-5 prevision so that we have
     reaosanbly bigger values
   - Update tests to check offset from 0 to N-1

Added basic support to transform indirect writes into
`llvm.masked.scatter` intrinsic. Currently, the scatter functionality is
limited to non-atomic writes and assignment (e.g. `+=` operator is
not yet supported). Hence, a warning is logged to the console
indicating all limitations.

Corresponding IR and execution tests were also added.

fixes #539

* Move benchmark + JIT related code from src/codegen/llvm
  to test/benchmark
* Common execution of CodegenLLVMVisitor for llvm --ir and
  benchmark option. With this, ast transformed for LLVM
  code generation is dumped to file.
* Previous object file is removed (if exist) so that output
  file name is same / deterministic
* Benchmark output is always printed to stdout via common
  logger object
* Remove unnecessary LLVMBuildInfo struct

Previously, there was no metadata and attributes associated with the
instance struct pointer, compute kernels or loops. This commit fixes
this.

- New instance struct attributes

Since all pointers contained in the instance struct do not alias, we add
a `noalias` (LLVM's `__restrict` alternative) attribute to it. In addition,
we add `nocapture` (No capturing occurs in the function) and
`readonly` (Struct pointer is not written to) attributes.

This means that some load instructions can be moved out from the loop
body. Example:
```llvm
; BEFORE
for.body.lr.ph:                                   ; preds = %0
  %5 = getelementptr inbounds %avx__instance_var__type, %avx__instance_var__type* %mech1, i64 0, i32 1
  br label %for.body

for.body:                                         ; preds = %for.body.lr.ph, %for.body
  %15 = load double*, double** %5, align 8
  ; ...


; AFTER
for.body.lr.ph:                                   ; preds = %0
  %5 = getelementptr inbounds %avx__instance_var__type, %avx__instance_var__type* %mech1, i64 0, i32 1
  %6 = load double*, double** %5, align 8
  br label %for.body
```

- New function attributes

Now, compute kernels are marked with `nofree` and `nounwind`
attributes.

- Loop metadata

Also, loop metadata is added to scalar kernels, specifying that no
vectorization is needed. The reason for this is because we want to
benchmark truly scalar kernels, and disable LLVM's vectorization if
necessary.

Note that for vector loop epilogue there is no metadata that disables
vectorization.

fixes #607

- fixes the case, where loaded value was taken from the stack, but was never actually put there

Added support for vector predication. Currently, we support a very basic
predication pattern (that will be extended in the future):
```c++
IF (/*condition*/) {
  // code here, no nested conditionals
} ELSE {
  // code here, no nested conditionals
} 
```
**What has been changed and added**

1. Removed vectorization check

Before, in the `FOR` statement visitor we were checking whether the code
can be vectorized. After refactoring `llvm::IRBuilder<>` into a separate class,
there is no interface to reset the builder's vector width. Hence, this check leads
to visitor having scalar vector width of 1, and builder having the same vector width.
```c++
if (!can_vectorize(node, sym_tab)) {
    vector_width = 1;
     ir_builder.generate_scalar_code();
}
```
In order to avoid any issues, this check is simply removed and will be added in
the separate PR.

2. Predication support

- `can_vectorize` has been changed to support a single `IF` or `IF/ELSE` pair.
- A special vectorized `IF` AST node visitor has been added.
- If generating code within `IF` AST node, instructions are masked.

3. Added execution and IR tests

fixes #539

* Improved cmake versioning of LLVM
* Added ^ support
* Added more math functions intrinsics with tests
* Added compute time variance and min/max times in benchmarking output

* With this PR alloca instructions are always inserted in the beginning
   of the function entry block. This is done to avoid them in the while or
   for loops, where allocations per iteration cause stack overflow
   (if the IR is not optimized).
* Insertion point for allocas is the enetry block now

See #653

Added support for fast math flags in LLVM backend. Currently,
the user can specify them via command-line (this approach was
chosen for easier benchmarking). The specified flags are named
exactly  the same as in LLVM. This feature is useful to enable
previously unsafe FP-math optimizations. For example, fused-multiply-add
instructions can now be generated when lowering LLVM IR to assembly
or executing via JIT.

Example:
```c++
// fma.mod
FUNCTION fma(a, b, c) {
    fma = (a * b) + c
}
```
```bash
$ ./nmodl fma.mod --verbose debug llvm --ir --fmf nnan contract afn --opt
```
```llvm
define double @fma(double %a, double %b, double %c) {
  %1 = fmul nnan contract afn double %a, %b
  %2 = fadd nnan contract afn double %1, %c
  ret double %2
}
```

- NMODL parser uses VarName on the LHS of assignment expression
- Inline visitor was using Name on the LHS of assignment expression

Related to #667

* Added support for `libsystem_m` and `SLEEF` vector libraries. The
first one is supported by LLVM internally, so it comes for free with
LLVM 13. For `SLEEF`, basic support was added for AArch64 and
x86 architectures. Currently, we support
- `exp`
- `pow`

* Added corresponding IR checks for `libsystem_m` and
`SLEEF` (both AArch64 and x86).

* Updated LLVM binaries for MAC OS CI, as well as for latest LLVM 13
(trunk) to fix link errors for Darwin vector library.

Co-authored-by: Pramod Kumbhar <[email protected]>