not). Code generation uses it.
- Added a policy that uses it.

Tested on pipeline benchmark.

Compiles correcctly for all targets and target combinations,
pending fix in python script for llvm 4.0 -> llvm 3.4

for allocation nodes.

(benchmark not updated)

Bug fix in nvptx and spir backends related to adding arguments to
node functions.

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

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.

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)

Added singleNodeStream.ll and twoNodeStream.ll test cases

(1) Changes to DFGraph.h for identifying streaming graphs and edges
(2) Added/modified launch, wait, push, pop intrinsics for streaming graphs
(3) Changes to DFG2LLVM_X86.cpp and DFG2LLVM.h pass to support code gen for streaming
(4) BuildDFG.cpp modified to create streaming edges
(5) Modified visc-rt.cpp and visc-rt.h to add runtime support for streaming

Created Abstract class DFG2LLVM. The two passes DFG2LLVM_NVPTX and DFG2LLVM_X86 inherit from this class. Common features/functions have been put in the abstract class DFG2LLVM

code enquires about its ancestor. Enables to run visc parboil benchmarks on x86
alone

(2) Modified unittests to have llvm.visc.init and cleanup intrinsics. Now they
all pass

(3) Modified visc.mk to easily compile visc version for just x86. Pass
TARGET=x86 as option to make command

(1) Working on implementation of further lowereing the visc api by breaking the
DFG launch call into init, launch, wait, cleanup. There are issues pending when
these require some kind of inter procedural analysis to link init, launch, wait
and cleanup. This is necessary during code generation.
(2) Also contains changes which simplify GenVISC pass by using macro for
identifying various source __visc__ calls

(2) Added dummy function calls for visc initialization and visc clean up to
benchmarks and added support for it to all the passes.
(3) Timers are now only printed once
(4) Added extra timers to parboil header file

pthread creation overhead added

global variable s for timers.

    This would help to use timers in any pass
(2) Added a command line argument -visc-timers to DFG2LLVM X86/NVPTX passes to
    turn on/off timers based on the command line argument
(3) Added stencil makefile which was missing till now from repo
(4) DFG2LLVM_NVPTX uses timers now

(2) Changes to Makefile of passes to ignore C++11 warnings

    set. This is important so as not to fail in case of 2-level PTX kernel,
    where the intermediate node does not have a gen function, but would pass
    because its parent has one.
(2) Modified PTX pass to
    i.  Correct mapping of arguments between intermediate and kernel launch node
    ii. Allow both constants and arguments as node limit values
    iii.Generate get_group_id and get_global_id correctly

M    Transforms/DFG2LLVM_X86/DFG2LLVM_X86.cpp
M    Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp

is incompatible with visc-rt
(2) Set GenFunc pointer to NULL. Completely removed FMap from both backend
passes. Now generated func info is contained in DFNode itself.
(3) Added Makefile to visc-rt for easy compilation to generate visc-rt.ll

pointer, size ... format for PTX nodes
(2) refactoring DFG2LLVM passes. Moving all the ptx code to ptx pass.
M    llvm/projects/visc-rt/visc-rt.cpp
M    llvm/lib/Transforms/DFG2LLVM_X86/DFG2LLVM_X86.cpp
M    llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp