- Changed all benchmarks accordingly
- GenVISC implementation updated accordingly

(benchmark not updated)

(CPU/GPU/SPIR) based on policy.
Support for scheduling per iteration included.
Further testing required.

Fixed spmv. LLVM 4.0 allows argument type promotion which was breaking the type check between __visc__node call and kernel function

Fixed visc_sh sgemm benchmark. The bug was that mtc bind when fixing function type was not handled correctly

Added parsing support for visc in,out,inout attributes. Commit ef703216 in previous llvm 3.4 based repo

Added the function to clone any function into a different typesHandiling two cases: (1) The caller is attempting to add data size argument after every pointer(2) The caller wants to add some extra arguments at the end

The SymVT stuff has not been moved yet, I removed it for now.
The test/VISC tests also have not been moved yet, so they may not
work.

Added argument to __visc__edge to allow user to specify 1-to-1 or all-to-all edge. all-to-all = true, one-to-one = false

Changed GenVISC pass to automatically infer return type of internal nodes from bind output intrinsics

So no need to generate backend specific intrinsics for PTX backend.
Now only using nvvm specifc intrinsic for rsqrt

get_global_size
(2) Code gen for nvvm intrinsics in GenVISC pass
(3) Commented out code added to GenVISC and DFG2LLVM_X86 to allow wait
instruction in a different BB than launch (with phi nodes as uses of launch)