[LLVMdev] Proposal for GSoc project - Compile programs with LLVM compiler

[Török Edwin]

Kumaripaba Miyurusara Atukorala wrote:
>
> hello everybody,
>
> I'm planning to compile a large codebase (Ubuntu) with LLVM compiler
> as my GSoc project 2008. I've given the my proposal below. I'll be
> really great full  if you can provide  your feedback on improving this
> proposal.
>

As Anton has said, Gentoo might be a better choice for this. But Ubuntu
should be ok too, if you like it better ...

>
> 1. Project title
>
> Compile programs with LLVM compiler
>
> 2. Abstract
>
> LLVM[1] is a Low Level Virtual Machine. It basically consists of two
> modules. They are the LLVM suite which consists of all the tools,
> libraries and header files needed to use the low level virtual machine
> and the GCC front end. The GCC front end contains a version of GCC
> that compiles C and C++ code in to LLVM bitcode. There is a third,
> optional piece called llvm-test. It is a suite of programs with a
> testing harness that can be used to further test LLVM's functionality
> and performance.
>
> The main objective of this project is to improve the LLVM
> testsuite[2].  The basic approach followed in achieving this objective
> is compiling a large code base that has not yet been compiled with
> LLVM and converts the build system to be compatible with build system
> of the LLVM testsuite. 
>
> In this project Ubuntu is used as the code base and initially it will
> be built using LLVM. Then the build system is tested and files bug
> reports for any issues that might hit and help track down problems.
>
> The development progress can be tracked from here [3].
>
> The complete project proposal is available here [4].
>
>  
>
> 3. Deliverables
>
> ·         Updated build system
>
> ·         Bugs reported to the bug tracker
>
> ·         Document (associated work)
>
> ·         Improved test programs
>
> ·         A guide for future continuation of this project or similar
> projects
>
>  
>
> 4. Benefits to the LLVM community
>
> The following will be available to the LLVM community on successful
> completion of the project.
>
> ·         An improved testsuite
>
> ·         A guide (document) to follow similar projects in future
>
>  5. Overview
>
> This section contains my work in this project.
>
> The main task is to build the entire Ubuntu distribution in LLVM. This
> will add new testcases and benchmarks for the use of LLVM.  A large
> testsuite is very important since it provides a lot of coverage of
> programs and enables us to spot and improve any problematic areas in
> the compiler.
>

Ubuntu has 13000+ packages, even if you count only those in main, there
are 3084 source packages.
Handling all of them is beyond what a single person can achieve during
the summer.

I would say handle 'Essential: yes'/'Build-Essential: yes' packages as a
first step (25, resp. 22 packages). Then if time permits try to handle
as much from the remaining packages as possible.

> The procedure followed in this project is discussed below.
>
> Initially the entire Ubuntu distribution will be compiled with the
> LLVM. If the compilation fails then the set of projects that
> successfully build with LLVM will be selected.  Bugs are filed for the
> projects that fail with reduced testcases.  If the compilation
> succeeds the build system will be converted to be compatible with the
> LLVM Programs testsuite. This can be checked in to SVN and the
> automated tester can use it to track progress of the compiler.
>

Is the goal only building packages? Or does it include testing built
packages as well?
I would suggest at least running testsuites for packages that have them.
Also what architecture are you building on? x86 32-bit?
>
> The next step is to build an image from projects that are successfully
> built with LLVM and merge projects built with gcc for those who failed
> in LLVM.  Then this image is booted and the problems that encounter
> during this process will be tracked down.  This process will be
> repeated until the entire system is successfully built with LLVM and
> booted as expected.
>
> When testing the code, it can be run with a variety of optimizations,
> and with all the back-ends: CBE, llc, and lli.
>
> 6. Project plan
>
> The basic steps in this project are listed below.
>
> 1.       Compile the Ubuntu distribution in gcc and LLVM
>
> 2.       If the compilation succeeds convert the build system to LLVM
> Program testsuite compatible mode
>
> 3.       If fails,
>
> a.       File bugs
>
> b.       Compile project by project in LLVM and select the projects
> that successfully compiles in LLVM.
>
> c.       For the projects that fail in LLVM compile them in gcc
>
> 4.       Build an image from projects successfully built with LLVM and
> merge in projects built with gcc in 3.c
>
> 5.       Try to boot the image from step 4 and track the problems.
>
> 6.       Continue the process until step 2 works.
>
> According to these steps I can break down this project in to three
> major phases
>
>  
>
>  
>
> 1^st Phase
>
> Compile Ubuntu distribution in gcc and LLVM. This is the first most
> steps and it will require about 3 weeks.
>

Well, Ubuntu is already built by gcc, I think you could focus on
building with LLVM. Building with gcc is useful when you want to track
down problems.

Best regards,
--Edwin