![Example
collat :: Int -> Word32 -> Int Run-time (ms)
collat c 1 = c 2000
collat c n | even n = 1800
collat (c+1) $ n `div` 2 1600
| otherwise = 1400
collat (c+1) $ 3 * n + 1
1200
1000
pmax x n = x `max` (collat 1 n, n)
800
600
main = print $ foldl pmax (1,1)
400
[2..1000000]
200
0
LLVM C NCG
Different, updated run-times compared to paper](https://image.slidesharecdn.com/ghc-llvm-hs10-slides-111018142902-phpapp02/85/Haskell-Symposium-2010-An-LLVM-backend-for-GHC-4-320.jpg)
![Handling Tables-Next-To-Code
Use GNU Assembler sub-section feature.
• Allows code/data to be put into numbered sub-section
• Sub-sections are appended together in order
• Table in <n>, entry code in <n+1>
.text 12
sJ8_info:
movl ...
movl ...
jmp ...
[...]
.text 11
sJ8_info_itable:
.long ...
.long 0
.long 327712](https://image.slidesharecdn.com/ghc-llvm-hs10-slides-111018142902-phpapp02/85/Haskell-Symposium-2010-An-LLVM-backend-for-GHC-13-320.jpg)
Haskell Symposium 2010: An LLVM backend for GHC
- 1. AN LLVM BACKEND FOR GHC David Terei & Manuel Chakravarty University of New South Wales Haskell Symposium 2010 Baltimore, USA, 30th Sep
- 2. What is ‘An LLVM Backend’? • Low Level Virtual Machine • Backend Compiler framework • Designed to be used by compiler developers, not by software developers • Open Source • Heavily sponsored by Apple
- 3. Motivation • Simplify • Reduce ongoing work • Outsource! • Performance • Improve run-time
- 4. Example collat :: Int -> Word32 -> Int Run-time (ms) collat c 1 = c 2000 collat c n | even n = 1800 collat (c+1) $ n `div` 2 1600 | otherwise = 1400 collat (c+1) $ 3 * n + 1 1200 1000 pmax x n = x `max` (collat 1 n, n) 800 600 main = print $ foldl pmax (1,1) 400 [2..1000000] 200 0 LLVM C NCG Different, updated run-times compared to paper
- 5. Competitors C Backend (C) Native Code Generator (NCG) • GNU C Dependency • Huge amount of work • Badly supported on • Very limited portability platforms such as Windows • Does very little • Use a Mangler on optimisation work assembly code • Slow compilation speed • Takes twice as long as the NCG
- 6. GHC’s Compilation Pipeline C Core STG Cmm NCG LLVM Language, Abstract Machine, Machine Configuration Heap, Stack, Registers
- 7. Compiling to LLVM Won’t be covering, see paper for full details: • Why from Cmm and not from STG/Core • LLVM, C-- & Cmm languages • Dealing with LLVM’s SSA form • LLVM type system Will be covering: • Handling the STG Registers • Handling GHC’s Table-Next-To-Code optimisation
- 8. Handling the STG Registers STG Register X86 Register Implement either by: Base ebx • In memory Heap Pointer edi • Pin to hardware registers Stack Pointer ebp R1 esi NCG? • Register allocator permanently stores STG registers in hardware C Backend? • Uses GNU C extension (global register variables) to also permanently store STG registers in hardware
- 9. Handling the STG Registers LLVM handles by implementing a new calling convention: STG Register X86 Register Base ebx Heap Pointer edi Stack Pointer ebp R1 esi define f ghc_cc (Base, Hp, Sp, R1) { … tail call g ghc_cc (Base, Hp’, Sp’, R1’); return void; }
- 10. Handling the STG Registers • Issue: If implemented naively then all the STG registers have a live range of the entire function. • Some of the STG registers can never be scratched (e.g Sp, Hp…) but many can (e.g R2, R3…). • We need to somehow tell LLVM when we no longer care about an STG register, otherwise it will spill and reload the register across calls to C land for example.
- 11. Handling the STG Registers • We handle this by storing undef into the STG register when it is no longer needed. We manually scratch them. define f ghc_cc (Base, Hp, Sp, R1, R2, R3, R4) { … store undef %R2 store undef %R3 store undef %R4 call c_cc sin(double %f22); … tail call ghc_cc g(Base,Hp’,Sp’,R1’,R2’,R3’,R4’); return void; }
- 12. Handling Tables-Next-To-Code Un-optimised Layout Optimised Layout How to implement in LLVM?
- 13. Handling Tables-Next-To-Code Use GNU Assembler sub-section feature. • Allows code/data to be put into numbered sub-section • Sub-sections are appended together in order • Table in <n>, entry code in <n+1> .text 12 sJ8_info: movl ... movl ... jmp ... [...] .text 11 sJ8_info_itable: .long ... .long 0 .long 327712
- 14. Handling Tables-Next-To-Code LLVM Mangler C Mangler • 180 lines of Haskell (half • 2,000 lines of Perl is documentation) • Needed for every platform • Needed only for OS X
- 15. Evaluation: Simplicity Code Size LLVM 25000 • Half of code is representation of LLVM language 20000 C 15000 • Compiler: 1,100 lines • C Headers: 2,000 lines 10000 • Perl Mangler: 2,000 lines 5000 NCG • Shared component: 8,000 lines 0 • Platform specific: 4,000 – 5,000 LLVM C NCG for X86, SPARC, PowerPC
- 16. Evaluation: Performance Run-time against LLVM (%) Nofib: 3 • Egalitarian benchmark suite, everything is equal 2.5 • Memory bound, little 2 room for optimisation once at Cmm stage 1.5 1 0.5 0 NCG C -0.5
- 17. Repa Performance runtimes (s) 16 14 12 10 8 LLVM NCG 6 4 2 0 Matrix Mult Laplace FFT
- 18. Compile Times, Object Sizes Compile Times Vs Object Sizes Vs LLVM LLVM 0 60 NCG C -2 40 -4 20 -6 0 NCG C -8 -20 -10 -40 -12 -60 -80 -14
- 19. Result • LLVM Backend is simpler. • LLVM Backend is as fast or faster. • LLVM developers now work for GHC!
- 20. Get It • LLVM • Our calling convention has been accepted upstream! • Included in LLVM since version 2.7 http://llvm.org • GHC • In HEAD • Should be released in GHC 7.0 • Send me any programs that are slower!
- 21. Questions?
- 22. Why from Cmm? A lot less work then from STG/Core But… Couldn’t you do a better job from STG/Core? Doubtful… Easier to fix any deficiencies in Cmm representation and code generator
- 23. Dealing with SSA LLVM language is SSA form: • Each variable can only be assigned to once • Immutable How do we handle converting mutable Cmm variables? • Allocate a stack slot for each Cmm variable • Use load and stores for reads and writes • Use ‘mem2reg’ llvm optimisation pass • This converts our stack allocation to LLVM variables instead that properly obeys the SSA requirement
- 24. Type Systems? LLVM language has a fairly high level type system • Strings, Arrays, Pointers… When combined with SSA form, great for development • 15 bug fixes required after backend finished to get test suite to pass • 10 of those were motivated by type system errors • Some could have been fairly difficult (e.g returning pointer instead of value)