Difference between revisions of "Toolchain"
(→The nu+ LLVM Structure) |
|||
Line 57: | Line 57: | ||
''elf2hex'' is a tool required for memory image generation. | ''elf2hex'' is a tool required for memory image generation. | ||
− | + | * [[Frontend | A custom version of the Clang frontend]] | |
− | + | * [[Backend | A native nu+ backend]] | |
− | + | * [[Tools| Some external tools that are used by llvm, i.e. linker, disassembler, etc..]] | |
− | + | * [[How to add a new backend to llvm | registration phase]] was required. | |
− | |||
− | |||
− | |||
== Testing == | == Testing == |
Revision as of 14:48, 29 March 2019
The nu+ toolchain is a collection of tools required to compile a nu+ executable application. It is based on the LLVM project leveraging on a custom version of the Clang Frontend and on a from-scratch implementation of the Backend. The toolchain comes with a modified version of LLD, an elf2hex tool to generate nu+ compatible memory images and lastly a custom version of the objdump tool for debug purposes. A custom implementation of the libc libraries is provided.
The toolchain supports the compilation of C-based and OpenCL C kernels.
Contents
Building Toolchain on Ubuntu Linux environment
This section shows how to build the nu+ toolchain in an Ubuntu Linux environment. The following steps are still valid in any other Unix-based system despite of package manager related differences.
Required Software
The nu+ toolchain installation relies on the following dependencies:
- Git
- GCC
- CMake
- Python 2.7
- libxml
- Zlib
- Bison
- Flex
- Libedit
- Swig
- ncurses libraries
- Ninja
The following terminal command may be used to install all required software:
$ sudo apt install libxml2-dev git cmake gcc g++ python bison flex zlib1g-dev swig python-dev libedit-dev libncurses5-dev ninja
Building Process
First, you have to obtain the source code from the official repository by typing the following command:
$ git clone https://gitlab.com/catecio/NuPlusLLVM.git <clone_directory>
The repository contains a helper script, setup.sh to make easier the installation process. To build a new implementation of the nu+ toolchain just type:
$ ./setup.sh -n
This command starts the compilation process, installing the toolchain in Release Mode in /usr/local/llvm-nuplus
.
If a debug version is required, add the -d
flag to setup.sh. You can also choose the number of threads for the compilation process, by using the -t=<number_of_threads>
parameter.
At the end of the compilation process, libraries are required to be linked to the installation folder:
$ ./setup.sh -l
Now, you can use the toolchain to build your own application.
The nu+ LLVM Structure
The nu+ toolchain relies on the LLVM project, version 7.0, providing customized implementation of its libraries to make possible generating nu+ executable kernels.
Clang is the compiler frontend for nu+, extended to handle the token-recognition of custom intrinsic functions.
LLVM core libraries are used to implement a custom backend for nu+, to manage the code lowering on the target-device.
LLD is adapted to match nu+ architecture requirements for linking.
objdump is provided to disassemble and analyze the generated code.
elf2hex is a tool required for memory image generation.
- A custom version of the Clang frontend
- A native nu+ backend
- Some external tools that are used by llvm, i.e. linker, disassembler, etc..
- registration phase was required.
Testing
To test the code generation process, the LLVM testing infrastructure has been used.
The script "run_tests" can be used to run the nu+ testing suite. The main purpose is to perform regression tests, so that it is possible to determine if a change in the back-end has any negative consequence.
The tool used is the "llvm-lit", a python script that parses the test files and executes the testing commands.
The code generation tests are contained in "compiler/test/CodeGen/NuPlus" and cover the main LLVM IR operations. A code generation test file is a collection of functions written in LLVM IR with commands directed to the llvm-lit and FileCheck tools. These commands are written as comments. The main commands used are RUN, CHECK, and CHECK-LABEL.
The RUN line tells lit how to run the test. If there are no RUN lines, lit will issue an error while running a test. The line syntax is similar to a shell’s syntax for pipelines. To examine the output and check if the test has been passed, the FileCheck tool must be used.
The RUN line used for all the tests is:
; RUN: llc -march=nuplus < %s | FileCheck %s
This tells lit to run llc with nu+ architecture as target, to give the output file to the FileCheck tool.
The CHECK_LABEL and the CHECK lines are interpreted by the FileCheck tool. It compares the llc output file with the CHECK_LABEL and CHECK lines. The comparison is done in sequence, although there are directives that let the checking not be performed in sequential order, see [1] for further details.
Libraries
nu+ is provided with different kind of libraries that are contained in the libs/ folder. For more information, check The nu+ libraries.
Makefiles and linker scripts
TODO: completare
Debug and typical errors
For more information about the debug support and some typical errors, check How to debug LLVM.