|Version 9 (modified by dterei, 5 years ago) (diff)|
The LLVM backend
David Terei wrote a new code generator for GHC which targets the LLVM compiler infrastructure. Most of the work was done as part of an honours thesis at the University of New South Wales under the supervision of Manuel Chakravarty. Its now at a stage where it is under consideration to be merged into GHC mainline.
The thesis paper which offers a detailed performance evaluation, as well as the motivation and design of the back-end can be found at:
As the back-end isn't currently in GHC head, you need to follow the steps below to get it up and running.
- GHC Patch (applies to GHC): http://www.cse.unsw.edu.au/~davidt/downloads/ghc-llvmbackend-full.gz
- LLVM Patch (applies to LLVM): http://www.cse.unsw.edu.au/~davidt/downloads/llvm-ghc.patch
These are the patches that you should be working with, they are 'stable'. The back-end also lives in a Git repository where the actual development work is done, this can be found at https://cgi.cse.unsw.edu.au/~davidt/cgit/cgit.cgi/Thesis%20GHC%20Dev/
Apply the darcs patch linked above to GHC head. This will make some changes across GHC, with the bulk of the new code ending up in 'compiler/llvmGen'.
To build GHC you need to add two flags to build.mk, they are:
GhcWithLlvmCodeGen = YES GhcEnableTablesNextToCode = NO
The LLVM code generator doesn't support at this time the TABLES_NEXT_TO_CODE optimisation due to limitations with LLVM.
You will also need LLVM installed on your computer to use the back-end. Version 2.6 or SVN trunk is supported. If you want to use the back-end in an unregistered ghc build, then you can use a vanilla build of LLVM. However if you want to use a registered GHC build (very likely) then you need to patch LLVM for this to work using the patch provided above.
LLVM is very easy to build and install. It can be done as follows:
$ svn co http://llvm.org/svn/llvm-project/llvm/trunk llvm $ cd llvm $ patch -p0 -i ~/llvm-ghc.patch $ ./configure --enable-optimized # probably also want to set --prefix $ make $ make install
Just make sure this modified version of LLVM is on your path and takes precedence over any other builds.
Once GHC is built, you can trigger GHC to use the LLVM back-end with the -fllvm flag. There is also a new -ddump-llvm which will dump out the LLVM IR code generated (must be used in combination with the -fllvm flag. (or use the -keep-tmp-files flag).
ghc --info should also now report that it includes the llvm code generator.
The ghc-core tool also supports the llvm backend, and will display the generated assembly code for your platform.
Supported Platforms & Correctness
- Linux x86-32/x86-64 are currently well supported. The back-end can pass the test suite and build a working version of GHC (bootstrap test).
- Mac OS X 10.5 currently has a rather nasty bug with any dynamic lib calls (all libffi stuff) [due to the stack not being 16byte aligned when the calls are made as required by OSX ABI for the curious]. Test suite passes except for most the ffi tests.
- Other platforms haven't been tested at all. As using the back-end with a registered build of GHC requires a modified version of LLVM, people wanting to try it out on those platforms will need to either make the needed changes to LLVM themselves, or use an unregistered build of GHC which will work with a vanilla install of LLVM. (A patch for LLVM for x86 is linked to below.)
(All done on linux/x86-32)
A quick summary of the results are that for the 'nofib' benchmark suite, the LLVM code generator was 3.8% slower than the NCG (the C code generator was 6.9% slower than the NCG). The DPH project includes a benchmark suite which I (David Terei) also ran and for this type of code using the LLVM back-end shortened the runtime by an average of 25% compared to the NCG. Also, while not included in my thesis paper as I ran out of time, I did do some benchmarking with the 'nobench' benchmark suite. It gave performance ratios for the back-ends of around:
A nice demonstration of the improvements the LLVM back-end can bring to some code though can be see at http://donsbot.wordpress.com/2010/02/21/smoking-fast-haskell-code-using-ghcs-new-llvm-codegen/
LLVM Back-end Design
The initial design tries to fit into GHC's current pipeline stages as seamlessly as possible. This allows for quicker development and focus on the core task of LLVM code generation.
The LLVM pipeline works as follows:
- New path for LLVM generation, separate from C and NCG. (path forks at compiler/main/CodeOutput.lhs, same place where C and NCG fork).
- LLVM code generation will output LLVM assembly code.
- The LLVM assembly code is translated to an object file as follows
- First, there is an 'LlvmAs' phase which generates LLVM bitcode from LLVM assembly code (using the llvm-as tool).
- The LLVM optimizer is run which is a series of bitcode to bitcode optimization passes (using the llc tool).
- Finally an object file is created from the LLVM bitcode (using the llc tool)
- This brings the LLVM path back to the other back-ends.
- The final state is the Link stage, which uses the system linker as with the other back-ends.
Here is a diagram of the pipeline:
Cmm -> (codeOutput) --->(ncg) Assembler -->(mangler, splitter) --> ('As' phase) -----> Object Code --> (link) --> executable \---> LLVM Assembler --> LLVM Optimizer --> ('llc' phase) -----/
This approach was the easiest and thus quickest way to initially implement the LLVM back-end. Now that it is working, there is some room for additional optimisations. A potential optimisation would be to add a new linker phase for LLVM. Instead of each module just being compiled to native object code ASAP, it would be better to keep them in the LLVM bitcode format and link all the modules together using the LLVM linker. This enable all of LLVM's link time optimisations. All the user program LLVM bitcode will then be compiled to a native object file and linked with the runtime using the native system linker.
The biggest problem is that LLVM doesn't provide all the features we need. The two issues below, 'Register Pinning' and 'TNTC' are the primary examples of this. While there is a patch for LLVM to partially correct fix this, this is a problem in itself as we now must include in GHC our own version of LLVM. Eventually we need to either get the changes we need included in LLVM or improve LLVM so that the features we require could be included dynamically.
The new back-end supports a custom calling convention to place the STG virtual registers into specific hardware registers. The current approach taken by the C back-end and NCG of having a fixed assignment of STG virtual registers to hardware registers for performance gains is not implemented in the LLVM back-end. Instead, it uses a custom calling convention to support something semantically equivalent to register pinning. The custom calling convention passes the first N variables in specific hardware registers, thus guaranteeing on all function entries that the STG virtual registers can be found in the expected hardware registers. This approach is believed to provide better performance than the register pinning used by NCG/C back-ends as it keeps the STG virtual registers mostly in hardware registers but allows the register allocator more flexibility and access to all machine registers.
GHC for heap objects places the info table (meta data) and the code adjacent to each other. That is, in memory, the object firstly has a head structure, which consists of a pointer to an info table and a payload structure. The pointer points to the bottom of the info table and the closures code is placed to be straight after the info table, so to jump to the code we can just jump one past the info table pointer. The other way to do this would be to have the info table contain a pointer to the closure code. However this would then require two jumps to get to the code instead of just one jump in the optimised layout. Achieving this layout can create some difficulty, the current back-ends handle it as follows:
- The NCG can create this layout itself
- The C code generator can't. So the Evil Mangler rearranges the GCC assembly code to achieve the layout.
There is a build option in GHC to use the unoptimised layout and instead use a pointer to the code in the info table. This layout can be enabled/disabled by using the compiler #def TABLES_NEXT_TO_CODE. As LLVM has no means to achieve the optimised layout and we don't wish to write an LLVM sister for the Evil Mangler, the LLVM back-end currently uses the unoptimised layout. This apparently incurs a performance penalty of 5% (source, Making a Fast Curry: Push/Enter vs. Eval/Apply for Higher-order Languages, Simon Marlow and Simon Peyton Jones, 2004).
Shared Code with NCG
It is probable that some of the code needed by the LLVM back-end is already implemented for the NCG back-end. Some examples of this code would be the following two functions in compiler/main/AsmCodeGen.lhs:
- Changes assignments to global registers to instead assign to the RegTable, used for non-pinned virtual registers.
- Optimises the cmm code, in particular it changes loads from global registers to instead load from the RegTable.
LLVM IR Representation
The LLVM IR is modeled in GHC using an algebraic data type to represent the first order abstract syntax of the LLVM assembly code. The LLVM representation lives in the 'Llvm' subdirectory and also contains code for pretty printing. This is the same approach taken by EHC's LLVM Back-end, and we adapted the module developed by them for this purpose.
It is an open question as to if this binding should be split out into its own cabal package. Please contact the GHC mailing list if you think you might be a user of such a package.
The GHC patch has been validated to make sure it won't break anything. This is just compiling and running GHC normally but with the LLVM back-end code included. It doesn't actually test the LLVM code generator, just makes sure it hasn't broken the NCG or C code generator.
OVERALL SUMMARY for test run started at Do 18. Feb 11:21:48 EST 2010 2457 total tests, which gave rise to 9738 test cases, of which 0 caused framework failures 7573 were skipped 2088 expected passes 76 expected failures 0 unexpected passes 1 unexpected failures Unexpected failures: user001(normal)
OVERALL SUMMARY for test run started at Thu 18 Feb 15:28:32 EST 2010 2458 total tests, which gave rise to 9739 test cases, of which 0 caused framework failures 7574 were skipped 2087 expected passes 77 expected failures 0 unexpected passes 1 unexpected failures Unexpected failures: T1969(normal)
Mac OS X 10.5/x86-32:
OVERALL SUMMARY for test run started at Thu Feb 18 12:35:49 EST 2010 2458 total tests, which gave rise to 9122 test cases, of which 0 caused framework failures 6959 were skipped 2085 expected passes 76 expected failures 0 unexpected passes 2 unexpected failures Unexpected failures: T1969(normal) ffi005(optc)
All of the test failures fail for me with a unmodified GHC head build as well as when the LLVM patch is included, so the llvm patch isn't introducing any new failures.