Changes between Version 5 and Version 6 of Commentary/Compiler/Backends/LLVM/WIP


Ignore:
Timestamp:
Mar 24, 2010 4:20:24 AM (4 years ago)
Author:
dterei
Comment:

added a whole bunch of notes about todo's

Legend:

Unmodified
Added
Removed
Modified
  • Commentary/Compiler/Backends/LLVM/WIP

    v5 v6  
    44This page is meant to collect together information about people working on (or interested in working on) LLVM in GHC, and the projects they are looking at.  See also the [wiki:Commentary/Compiler/NewCodeGen state of play of the whole back end]. 
    55 
    6  * David Terei is going to be at Microsoft Research for an internship to work on LLVM at the end of April 
     6 * David Terei is going to be at Microsoft Research for an internship to work on LLVM from the middle of April 
    77 * There is interest on working on LLVM as part of the Google Summer of Code 
    88   * The GSoC timeline is May 24th - August 9th (http://socghop.appspot.com/document/show/gsoc_program/google/gsoc2010/faqs#timeline), so this overlaps with David's internship 
     
    1010   * Alp Mestanogullari (http://alpmestan.wordpress.com/, http://twitter.com/alpmestan) is interested in working on a SoC project on LLVM 
    1111 
    12 For possible TODO items, see [wiki:Commentary/Compiler/Backends/LLVM/DevelopmentNotes Development Notes & Bugs] 
     12== Big Ticket Items == 
     13 
     14=== LLVM IR Representation === 
     15 
     16The 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.  
     17 
     18The current design is overly complicated and could be faster. It uses String + show operations for printing for example when it should be using FastString + Outputable. Before simplifying this design though it would be good to investigate using the LLVM API instead of the assembly language for interacting with LLVM. This would be done most likely by using the pre-existing Haskell LLVM API bindings found [http://hackage.haskell.org/package/llvm here]. This should hopefully provide a speed up in compilation speeds which is greatly needed since the LLVM back-end is ~2x slower at the moment. 
     19 
     20=== TABLES_NEXT_TO_CODE === 
     21 
     22Fix the [wiki:Commentary/Compiler/Backends/LLVM/Issues#TABLES_NEXT_TO_CODE TNTC] issue. 
     23 
     24There are two possible approaches: 
     25  * Add the needed features to LLVM so that it could natively support this feature. This could be quite a difficult and ambitious project as getting the feature into LLVM would mean fixing all the back-ends so that they support the feature. And you would need to convince the LLVM developers that this feature is worth them supporting which might be a hard sell if they think only GHC would use it. 
     26  * A new ''evil mangler'' to process the assembly produced by LLVM and patch it to support the feature. This wouldn't be so bad if the new ''evil mangler'' was written in Haskell and only did this one fairly simple thing (current ''evil mangler'' is in perl and does a fair few things in addition to TNTC). 
     27 
     28For either approach, looking at how LLVM's new metadata feature may assist would be very useful: 
     29 
     30 * [http://llvm.org/docs/LangRef.html#metadata] 
     31 * [http://nondot.org/sabre/LLVMNotes/ExtensibleMetadata.txt] 
     32 * [http://nondot.org/sabre/LLVMNotes/EmbeddedMetadata.txt] 
     33 
     34=== Optimise the output of the LLVM Back-end === 
     35 
     36The LLVM back-end at the moment generally takes the most straight-forward approach to compiling Haskell (Cmm really) to LLVM. LLVM is designed in such a way that this is how things should be by and large done. Its instruction set is designed to be simple and generally have one way to approach a problem (especially when coming from fairly similar Cmm), you are encouraged to rely on the optimisation passes of LLVM to handle fixing things up. However, this doesn't mean there isn't potentially some room for improvement, especially since we simply don't know if there is or isn't. The LLVM back-end is new and experiments and benchmarks need to be done to figure out its limits and places it can be improved. Some quick ideas: 
     37 
     38 * Update the back-end to use some of the new features of LLVM 2.6 and 2.7. Currently it only uses features of 2.5. (e.g could maybe use the new LLVM integer specific add operation to detect overflow rather then the current custom code to do it). One quick improvement is that as of 2.7 the LLVM assembler ('llvm-as') stage in the LLVM back-end pipeline isn't needed now at the LLVM optimiser tool ('opt') can be directly given LLVM assembly now as well as LLVM bitcode. 
     39 * Look into the various [http://llvm.org/docs/LangRef.html#paramattrs parameter attributes] and [http://llvm.org/docs/LangRef.html#fnattrs function attributes] that LLVM supports and how they should be used by the LLVM back-end. (e.g the noalias parameter attribute should probably be used). 
     40 * Look at the various [http://llvm.org/docs/LangRef.html#intrinsics intrinsic functions] supported by LLVM. Some of them could maybe be used to replace existing code in LLVM or calls to the rts. (e.g Cmm expects support of a fair number of basic math operations [e.g sin], for which LLVM intrinsic functions exists. However the back-end currently calls the C library for them). 
     41 
     42=== Update the Back-end to use the new Cmm data types / New Code Generator === 
     43 
     44There is ongoing work to produce a new, nicer, more modular code generator for GHC (the slightly confusingly name code generator in GHC refers to the pipeline stage where the Core IR is compiled to the Cmm IR). The LLVM back-end could be updated to make sure it works with the new code generator and does so in an efficient manner. 
     45 
     46=== LLVM's Link Time Optimisations === 
     47 
     48One of LLVM's big marketing features is its support for link time optimisation. This does thinks such as in-lining across module boundaries, more aggressive dead code elimination... ect). The LLVM back-end could be updated to make use of this. 
     49 
     50 * [http://llvm.org/releases/2.6/docs/LinkTimeOptimization.html] 
     51 * [http://llvm.org/docs/GoldPlugin.html] 
     52 
     53=== Optimise LLVM for the type of Code GHC produces === 
     54 
     55At the moment only a some fairly basic benchmarking has been done of the LLVM back-end. Enough to give an indication of how it performs on the whole (well as far as you trust benchmarks anyway) and of what it can sometimes achieve. However this is by no means exauhstive or probably even close to it and doesn't give us enough information about the areas where LLVM performs badly. The LLVM optimisation pass also at the moment just uses the standard '-O[123]' levels, which like GCC entail a whole bunch of optimisation passes. These groups are designed for C programs mostly. 
     56 
     57So: 
     58 * More benchmarking, particularly finding some bad spots for the LLVM back-end and generating a good picture of the characteristics of the back-end. 
     59 * Look into the LLVM optimiser, e.g perhaps some more work in the style of [http://donsbot.wordpress.com/2010/03/01/evolving-faster-haskell-programs-now-with-llvm/ Don's work] 
     60 * Look at any new optimisation passes that could be written for LLVM which would help to improve the code it generates for GHC. 
     61 * Look at general fixes/improvement to LLVM to improve the code it generates for LLVM (e.g at the moment LLVM performs a lot of redundant stack manipulation in the code in generates for GHC, would be good to fix this up). 
     62 
     63=== LLVM Cross Compiler / Port === 
     64 
     65This is more of an experimental idea but the LLVM back-end looks like it would make a great choice for Porting LLVM. That is, instead of porting LLVM through the usual route of via-C and then fixing up the NCG, just try to do it all through the LLVM back-end. As LLVM is quite portable and supported on more platforms then GHC, it would be an interesting and valuable experiment to try to port GHC to a new platform by simply getting the LLVM back-end working on it. (The LLVM back-end works in both unregistered and registered mode, another advantage for porting compared to the C and NCG back-ends). 
     66 
     67It would also be interesting to looking into improving GHC to support cross compiling and doing this through the LLVM back-end as it should be easier to fix up to support this feature than the C or NCG back-ends. 
     68 
     69=== Stabilise / Bug Fixing === 
     70 
     71The back-end needs a fair amount of love and care just to get it into a state where it could be used as the default back-end by GHC if desired. 
     72 
     73 * '''Platform support''': Only supports x86 Linux. There are a number of serious bugs on Mac OS X. Windows hasn't been tested. SPARC also hasn't been tested and would need to have changes made in LLVM so that the SPARC LLVM back-end supported the GHC calling convention. 
     74 * Has been a report the back-end interacts badly with the '-dynamic' GHC flag. 
     75 * Back-end hasn't been thoroughly tested across the full range of GHC configurations (e.g threaded...) 
     76 * LLVM back-end is out of tree currently.