[LLVMdev] running a module pass via opt on multiple bitcode files

[John Criswell]

On 8/19/11 4:36 PM, Jinwook Shin (HOTWING) wrote:
>
> Thanks John. Your approach worked fine for my small toy program. Now, 
> I would like to run my module pass on a huge project (still a single 
> executable) consisting of a few thousand bit code files scattered in 
> many different directories. And some of them are static libraries. Do 
> you still think it's a good idea to manually gather and link them with 
> llvm-ld? I've seen other module pass implementations where they 
> directly patch a specific version of LLVM or clang, instead of running 
> a shared object implementation with opt. Which would be the ideal 
> approach to analyze a program with a module pass? Thanks.
>

The question is whether your pass requires having the whole program to 
work (or to be effective) or whether it can be done on each compilation 
unit separately.

If you can do it on separate compilation units, integrate it into 
Clang.  We did this for some of the SAFECode passes in mainline 
SAFECode; you can use it as an example.

If your pass must do whole program analysis (or simply works better if 
it does), then you want to add your pass to libLTO and enable the use of 
link-time optimization using the -emit-llvm option on the llmv-gcc/clang 
command line.

Note that all libLTO passes are enabled when compiling using link-time 
optimization.  If you replace libLTO on your system, your pass is going 
to be run for every program compiled using link-time optimization 
(unless you can make your pass conditional on some optimization level 
like -O4 or -O5 that gets communicated from the compiler front-end to 
the linker to the plugin).  If you're doing instrumentation, you might 
want to do what we did in SAFECode: perform instrumentation in the 
front-end and then use whole-program analysis to change/optimize the 
instrumentation in libLTO.  That way, our libLTO transforms can always 
be turned on; for uninstrumented code, they become a no-op.

-- John T.


> *From: *John Criswell <criswell at illinois.edu 
> <mailto:criswell at illinois.edu>>
> *Organization: *University of Illinois
> *Date: *Tue, 12 Jul 2011 16:19:06 -0500
> *To: *Jinwook Shin <Jinwook.Shin at microsoft.com 
> <mailto:Jinwook.Shin at microsoft.com>>
> *Cc: *"llvmdev at cs.uiuc.edu <mailto:llvmdev at cs.uiuc.edu>" 
> <llvmdev at cs.uiuc.edu <mailto:llvmdev at cs.uiuc.edu>>
> *Subject: *Re: [LLVMdev] running a module pass via opt on multiple 
> bitcode files
>
> On 7/12/11 4:11 PM, Jinwook Shin (HOTWING) wrote:
>
> Hi,
>
> I'm trying to implement a module pass as a share object to process an 
> entire program as a unit. The target program is built from multiple 
> object files, hence multiple bitcode files. However, it seems that opt 
> does not take the whole program, but just one bitcode file, "test.bc" 
> in the example run shown below.
>
> $ opt -load mypass.dylib -mymodulepass < test.bc > /dev/null
>
> How do I run my module pass in a library on a whole program?
>
>
> Link the bitcode files together into a single bitcode file with 
> llvm-ld.  You can then run the single bitcode file through opt.
>
> -- John T.
>
>
> Thanks.
>
>
>
>
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu  <mailto:LLVMdev at cs.uiuc.edu>          http://llvm.cs.uiuc.eduhttp://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20110819/5ccb4b2a/attachment.html>