August 2015 - ghc-tickets

[GHC] #8971: Native Code Generator 7.8.1 RC2 is not as optimized as 7.6.3...
by GHC 09 Dec '16

09 Dec '16

#8971: Native Code Generator 7.8.1 RC2 is not as optimized as 7.6.3... ------------------------------+-------------------------------------------- Reporter: | Owner: GordonBGood | Status: new Type: bug | Milestone: Priority: normal | Version: 7.8.1-rc2 Component: Compiler | Operating System: Unknown/Multiple Keywords: | Type of failure: Runtime performance bug Architecture: | Test Case: Unknown/Multiple | Blocking: Difficulty: Unknown | Blocked By: | Related Tickets: | ------------------------------+-------------------------------------------- The output assembly code is not as optimized for the Windows 32-bit version 7.8.1 RC2 compiler as the Windows 7.6.3 compiler (32-bit) when the option switches are exactly the same although it may not be limited to only the Windows platform; this has a negative impact on execution time for tight loops of about a factor of two times slower. The following code will reproduce the problem: {{{#!haskell -- GHC_NCG_OptimizationBug.hs -- it seems the Haskell GHC 7.8.1 NCG Native Code Generator (NCG) doesn't -- optimize as well for (at least) the x86 target as version 7.6.3 {-# OPTIONS_GHC -O3 -rtsopts -v -dcore-lint -ddump-asm -ddump-to-file -dumpdir . #-} -- or O2 import Data.Bits import Control.Monad.ST (runST,ST(..)) import Data.Array.Base -- Uses a very simple Sieve of Eratosthenes to 2 ^ 18 to prove it. accNumPrimes :: Int -> Int accNumPrimes acc = acc `seq` runST $ do let bfSz = (256 * 1024 - 3) `div` 2 bfLmtWrds = (bfSz + 1) `div` 32 bufw <- newArray (0, bfLmtWrds) (-1) :: ST s (STUArray s Int Int) -- to clear the last "uneven" bit(s) unsafeWrite bufw bfLmtWrds (complement ((-2) `shiftL` (bfSz .&. 31))) bufb <- (castSTUArray :: STUArray s Int Int -> ST s (STUArray s Int Bool)) bufw let cullp i = let p = i + i + 3 in let s = (p * p - 3) `div` 2 in if s > bfSz then let count i sm = do sm `seq` if i > bfLmtWrds then return (acc + sm) else do wd <- unsafeRead bufw i count (i + 1) (sm + (popCount wd)) in count 0 1 -- use '1' for the '2' prime not in the array else do v <- unsafeRead bufb i if v then let cull j = do -- very tight inner loop if j > bfSz then cullp (i + 1) else do unsafeWrite bufb j False cull (j + p) in cull s else cullp (i + 1) cullp 0 main = -- run the program a number of times to get a reasonable time... let numloops = 2000 in let loop n acc = acc `seq` if n <= 0 then acc else loop (n - 1) (accNumPrimes acc) in print $ loop numloops 0 }}} The above code takes almost twice as long to run when compiled under 7.8.1 RC2 for Windows (32-bit) as it does for the version 7.6.3 compiler (both 32-bit compilers). The -ddump-simpl Core dump is almost identical between the two, which is also evidenced by that using the -fllvm LLVM compiler back end switch for each results in code that runs at about the same speed for each compiler run (which would use the same Core output as used for NCG, right?). Under Windows, the compilation and run for 7.8.1 RC2 goes like this: {{{ *Main> :! E:\ghc-7.8.0.20140228_32\bin\ghc --make -pgmlo "E:\llvm32\build\Release\bin\opt" -pgmlc "E:\llvm32\build\Release\bin\llc" "GHC_NCG_OptimizationBug.hs" compile: input file WindowsVsLinuxNCG.hs Created temporary directory: C:\Users\Gordon\AppData\Local\Temp\ghc15460_0 *** Checking old interface for main:Main: *** Parser: *** Renamer/typechecker: [1 of 1] Compiling Main ( GHC_NCG_OptimizationBug.hs, GHC_NCG_OptimizationBug.o ) *** Desugar: Result size of Desugar (after optimization) = {terms: 260, types: 212, coercions: 0} *** Core Linted result of Desugar (after optimization): *** Simplifier: Result size of Simplifier iteration=1 = {terms: 213, types: 136, coercions: 52} *** Core Linted result of Simplifier: Result size of Simplifier iteration=2 = {terms: 215, types: 148, coercions: 67} *** Core Linted result of Simplifier: Result size of Simplifier iteration=3 = {terms: 209, types: 135, coercions: 51} *** Core Linted result of Simplifier: Result size of Simplifier = {terms: 209, types: 135, coercions: 42} *** Core Linted result of Simplifier: *** Specialise: Result size of Specialise = {terms: 209, types: 135, coercions: 42} *** Core Linted result of Specialise: *** Float out(FOS {Lam = Just 0, Consts = True, PAPs = False}): Result size of Float out(FOS {Lam = Just 0, Consts = True, PAPs = False}) = {terms: 286, types: 185, coercions: 42} *** Core Linted result of Float out(FOS {Lam = Just 0, Consts = True, PAPs = False}): *** Float inwards: Result size of Float inwards = {terms: 286, types: 185, coercions: 42} *** Core Linted result of Float inwards: *** Simplifier: Result size of Simplifier iteration=1 = {terms: 502, types: 393, coercions: 103} *** Core Linted result of Simplifier: Result size of Simplifier iteration=2 = {terms: 428, types: 326, coercions: 29} *** Core Linted result of Simplifier: Result size of Simplifier iteration=3 = {terms: 420, types: 321, coercions: 29} *** Core Linted result of Simplifier: Result size of Simplifier = {terms: 420, types: 321, coercions: 29} *** Core Linted result of Simplifier: *** Simplifier: Result size of Simplifier iteration=1 = {terms: 418, types: 318, coercions: 29} *** Core Linted result of Simplifier: Result size of Simplifier = {terms: 418, types: 318, coercions: 29} *** Core Linted result of Simplifier: *** Simplifier: Result size of Simplifier iteration=1 = {terms: 475, types: 383, coercions: 32} *** Core Linted result of Simplifier: Result size of Simplifier iteration=2 = {terms: 444, types: 336, coercions: 9} *** Core Linted result of Simplifier: Result size of Simplifier = {terms: 444, types: 336, coercions: 9} *** Core Linted result of Simplifier: *** Demand analysis: Result size of Demand analysis = {terms: 444, types: 336, coercions: 9} *** Core Linted result of Demand analysis: *** Worker Wrapper binds: Result size of Worker Wrapper binds = {terms: 579, types: 457, coercions: 9} *** Core Linted result of Worker Wrapper binds: *** Simplifier: Result size of Simplifier iteration=1 = {terms: 510, types: 415, coercions: 9} *** Core Linted result of Simplifier: Result size of Simplifier = {terms: 420, types: 322, coercions: 9} *** Core Linted result of Simplifier: *** Float out(FOS {Lam = Just 0, Consts = True, PAPs = True}): Result size of Float out(FOS {Lam = Just 0, Consts = True, PAPs = True}) = {terms: 426, types: 326, coercions: 9} *** Core Linted result of Float out(FOS {Lam = Just 0, Consts = True, PAPs = True}): *** Common sub-expression: Result size of Common sub-expression = {terms: 424, types: 326, coercions: 9} *** Core Linted result of Common sub-expression: *** Float inwards: Result size of Float inwards = {terms: 424, types: 326, coercions: 9} *** Core Linted result of Float inwards: *** Liberate case: Result size of Liberate case = {terms: 1,824, types: 1,259, coercions: 9} *** Core Linted result of Liberate case: *** Simplifier: Result size of Simplifier iteration=1 = {terms: 608, types: 422, coercions: 9} *** Core Linted result of Simplifier: Result size of Simplifier iteration=2 = {terms: 604, types: 413, coercions: 9} *** Core Linted result of Simplifier: Result size of Simplifier iteration=3 = {terms: 604, types: 413, coercions: 9} *** Core Linted result of Simplifier: Result size of Simplifier = {terms: 604, types: 413, coercions: 9} *** Core Linted result of Simplifier: *** SpecConstr: Result size of SpecConstr = {terms: 708, types: 505, coercions: 9} *** Core Linted result of SpecConstr: *** Simplifier: Result size of Simplifier iteration=1 = {terms: 702, types: 499, coercions: 9} *** Core Linted result of Simplifier: Result size of Simplifier = {terms: 608, types: 405, coercions: 9} *** Core Linted result of Simplifier: *** Tidy Core: Result size of Tidy Core = {terms: 608, types: 405, coercions: 9} *** Core Linted result of Tidy Core: *** CorePrep: Result size of CorePrep = {terms: 825, types: 489, coercions: 9} *** Core Linted result of CorePrep: *** Stg2Stg: *** CodeOutput: *** New CodeGen: *** CPSZ: *** CPSZ: *** CPSZ: *** CPSZ: *** CPSZ: *** CPSZ: *** CPSZ: *** CPSZ: *** CPSZ: *** CPSZ: *** CPSZ: *** CPSZ: *** CPSZ: *** CPSZ: *** CPSZ: *** Assembler: "E:\ghc-7.8.0.20140228_32\lib/../mingw/bin/gcc.exe" "-U__i686" "-fno- stack-protector" "-DTABLES_NEXT_TO_CODE" "-I." "-x" "assembler-with-cpp" "-c" "C:\Users\Gordon\AppData\Local\Temp\ghc15460_0\ghc15460_2.s" "-o" "GHC_NCG_OptimizationBug.o" Linking GHC_NCG_OptimizationBug.exe ... *Main> :! GHC_NCG_OptimizationBug +RTS -s 46000000 32,965,096 bytes allocated in the heap 7,032 bytes copied during GC 41,756 bytes maximum residency (2 sample(s)) 19,684 bytes maximum slop 2 MB total memory in use (0 MB lost due to fragmentation) Tot time (elapsed) Avg pause Max pause Gen 0 61 colls, 0 par 0.00s 0.00s 0.0000s 0.0000s Gen 1 2 colls, 0 par 0.00s 0.00s 0.0001s 0.0001s INIT time 0.00s ( 0.00s elapsed) MUT time 1.73s ( 1.73s elapsed) GC time 0.00s ( 0.00s elapsed) EXIT time 0.00s ( 0.00s elapsed) Total time 1.73s ( 1.73s elapsed) %GC time 0.0% (0.0% elapsed) Alloc rate 19,006,902 bytes per MUT second Productivity 100.0% of total user, 100.2% of total elapsed }}} whereas under version 7.6.3 goes like this: {{{ *Main> :! E:\ghc-7.6.3_32\bin\ghc --make -pgmlo "E:\llvm32\build\Release\bin\opt" -pgmlc "E:\llvm32\build\Release\bin\llc" "GHC_NCG_OptimizationBug.hs" compile: input file GHC_NCG_OptimizationBug.hs Created temporary directory: C:\Users\Gordon\AppData\Local\Temp\ghc28200_0 *** Checking old interface for main:Main: *** Parser: *** Renamer/typechecker: [1 of 1] Compiling Main ( GHC_NCG_OptimizationBug.hs, GHC_NCG_OptimizationBug.o ) *** Desugar: Result size of Desugar (after optimization) = {terms: 247, types: 212, coercions: 0} *** Core Linted result of Desugar (after optimization): *** Simplifier: Result size of Simplifier iteration=1 = {terms: 198, types: 132, coercions: 35} *** Core Linted result of Simplifier: Result size of Simplifier iteration=2 = {terms: 200, types: 144, coercions: 43} *** Core Linted result of Simplifier: Result size of Simplifier iteration=3 = {terms: 194, types: 131, coercions: 57} *** Core Linted result of Simplifier: Result size of Simplifier = {terms: 194, types: 131, coercions: 39} *** Core Linted result of Simplifier: *** Specialise: Result size of Specialise = {terms: 194, types: 131, coercions: 39} *** Core Linted result of Specialise: *** Float out(FOS {Lam = Just 0, Consts = True, PAPs = False}): Result size of Float out(FOS {Lam = Just 0, Consts = True, PAPs = False}) = {terms: 277, types: 191, coercions: 39} *** Core Linted result of Float out(FOS {Lam = Just 0, Consts = True, PAPs = False}): *** Float inwards: Result size of Float inwards = {terms: 277, types: 191, coercions: 39} *** Core Linted result of Float inwards: *** Simplifier: Result size of Simplifier iteration=1 = {terms: 514, types: 403, coercions: 103} *** Core Linted result of Simplifier: Result size of Simplifier iteration=2 = {terms: 420, types: 317, coercions: 29} *** Core Linted result of Simplifier: Result size of Simplifier iteration=3 = {terms: 412, types: 312, coercions: 29} *** Core Linted result of Simplifier: Result size of Simplifier = {terms: 412, types: 312, coercions: 29} *** Core Linted result of Simplifier: *** Simplifier: Result size of Simplifier iteration=1 = {terms: 410, types: 309, coercions: 29} *** Core Linted result of Simplifier: Result size of Simplifier = {terms: 410, types: 309, coercions: 29} *** Core Linted result of Simplifier: *** Simplifier: Result size of Simplifier iteration=1 = {terms: 455, types: 364, coercions: 32} *** Core Linted result of Simplifier: Result size of Simplifier iteration=2 = {terms: 422, types: 317, coercions: 9} *** Core Linted result of Simplifier: Result size of Simplifier = {terms: 422, types: 317, coercions: 9} *** Core Linted result of Simplifier: *** Demand analysis: Result size of Demand analysis = {terms: 422, types: 317, coercions: 9} *** Core Linted result of Demand analysis: *** Worker Wrapper binds: Result size of Worker Wrapper binds = {terms: 536, types: 427, coercions: 9} *** Core Linted result of Worker Wrapper binds: *** Simplifier: Result size of Simplifier iteration=1 = {terms: 480, types: 391, coercions: 9} *** Core Linted result of Simplifier: Result size of Simplifier = {terms: 400, types: 306, coercions: 9} *** Core Linted result of Simplifier: *** Float out(FOS {Lam = Just 0, Consts = True, PAPs = True}): Result size of Float out(FOS {Lam = Just 0, Consts = True, PAPs = True}) = {terms: 408, types: 311, coercions: 9} *** Core Linted result of Float out(FOS {Lam = Just 0, Consts = True, PAPs = True}): *** Common sub-expression: Result size of Common sub-expression = {terms: 406, types: 311, coercions: 9} *** Core Linted result of Common sub-expression: *** Float inwards: Result size of Float inwards = {terms: 406, types: 311, coercions: 9} *** Core Linted result of Float inwards: *** Liberate case: Result size of Liberate case = {terms: 1,186, types: 824, coercions: 9} *** Core Linted result of Liberate case: *** Simplifier: Result size of Simplifier iteration=1 = {terms: 585, types: 411, coercions: 9} *** Core Linted result of Simplifier: Result size of Simplifier iteration=2 = {terms: 569, types: 392, coercions: 9} *** Core Linted result of Simplifier: Result size of Simplifier iteration=3 = {terms: 569, types: 392, coercions: 9} *** Core Linted result of Simplifier: Result size of Simplifier = {terms: 569, types: 392, coercions: 9} *** Core Linted result of Simplifier: *** SpecConstr: Result size of SpecConstr = {terms: 746, types: 566, coercions: 9} *** Core Linted result of SpecConstr: *** Simplifier: Result size of Simplifier iteration=1 = {terms: 739, types: 560, coercions: 9} *** Core Linted result of Simplifier: Result size of Simplifier iteration=2 = {terms: 762, types: 546, coercions: 9} *** Core Linted result of Simplifier: Result size of Simplifier = {terms: 642, types: 402, coercions: 9} *** Core Linted result of Simplifier: *** Tidy Core: Result size of Tidy Core = {terms: 642, types: 402, coercions: 9} *** Core Linted result of Tidy Core: writeBinIface: 10 Names writeBinIface: 34 dict entries *** CorePrep: Result size of CorePrep = {terms: 779, types: 483, coercions: 9} *** Core Linted result of CorePrep: *** Stg2Stg: *** CodeOutput: *** CodeGen: *** Assembler: "E:\ghc-7.6.3_32\lib/../mingw/bin/gcc.exe" "-fno-stack-protector" "-Wl ,--hash-size=31" "-Wl,--reduce-memory-overheads" "-I." "-c" "C:\Users\Gordon\AppData\Local\Temp\ghc28200_0\ghc28200_0.s" "-o" "GHC_NCG_OptimizationBug.o" Linking GHC_NCG_OptimizationBug.exe ... *Main> :! GHC_NCG_OptimizationBug +RTS -s 46000000 32,989,396 bytes allocated in the heap 4,976 bytes copied during GC 41,860 bytes maximum residency (2 sample(s)) 19,580 bytes maximum slop 2 MB total memory in use (0 MB lost due to fragmentation) Tot time (elapsed) Avg pause Max pause Gen 0 61 colls, 0 par 0.00s 0.00s 0.0000s 0.0000s Gen 1 2 colls, 0 par 0.00s 0.00s 0.0001s 0.0001s INIT time 0.00s ( 0.00s elapsed) MUT time 0.64s ( 0.64s elapsed) GC time 0.00s ( 0.00s elapsed) EXIT time 0.00s ( 0.00s elapsed) Total time 0.66s ( 0.64s elapsed) %GC time 0.0% (0.1% elapsed) Alloc rate 51,495,642 bytes per MUT second Productivity 100.0% of total user, 102.3% of total elapsed }}} Looking at the ASM dump for the innermost tight culling loop reveals the problem, with 7.8.1 RC2 outputting as follow: {{{ _n3nx: movl 76(%esp),%ecx _c3gf: cmpl %ecx,%eax jg _c3jB _c3jC: movl %eax,%edx sarl $5,%edx movl %ecx,76(%esp) movl $1,%ecx movl %ecx,280(%esp) movl %eax,%ecx andl $31,%ecx movl %eax,292(%esp) movl 280(%esp),%eax shll %cl,%eax xorl $-1,%eax movl 64(%esp),%ecx addl $8,%ecx movl (%ecx,%edx,4),%ecx andl %eax,%ecx movl 64(%esp),%eax addl $8,%eax movl %ecx,(%eax,%edx,4) movl 292(%esp),%eax addl $3,%eax jmp _n3nx }}} and 7.6.3 outputting as follows: {{{ .text .align 4,0x90 .long 1894 .long 32 s1GZ_info: _c1YB: cmpl 16(%ebp),%esi jg _c1YE movl %esi,%edx sarl $5,%edx movl $1,%eax movl %esi,%ecx andl $31,%ecx shll %cl,%eax xorl $-1,%eax movl 12(%ebp),%ecx movl 8(%ecx,%edx,4),%ecx andl %eax,%ecx movl 12(%ebp),%eax movl %ecx,8(%eax,%edx,4) addl 4(%ebp),%esi jmp s1GZ_info _c1YE: movl 8(%ebp),%esi addl $8,%ebp jmp s1GB_info }}} The second code is clearly much more efficient, with the only memory access reading/writing the sieve buffer array and one register reload of the prime value to add to the current position index, whereas the first (7.8.1 RC2) code has three register spills and five register re-loads, almost as if debugging were still turned on. This bug was tested under Windows, but likely applies to other platforms, at least for 32-bit versions but also possibly to others. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/8971> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 30

[GHC] #10371: GHC fails to inline and specialize a function
by GHC 07 Dec '16

07 Dec '16

#10371: GHC fails to inline and specialize a function -------------------------------------+------------------------------------- Reporter: | Owner: MikeIzbicki | Status: new Type: bug | Milestone: Priority: normal | Version: 7.10.1 Component: Compiler | Operating System: Unknown/Multiple Keywords: | Type of failure: None/Unknown Architecture: | Blocked By: Unknown/Multiple | Related Tickets: Test Case: | Blocking: | Differential Revisions: | -------------------------------------+------------------------------------- I have an alternative Prelude library called [subhask](https://github.com/mikeizbicki/subhask) that redefines the numeric type class hierarchy. I'm trying to update it to work with GHC 7.10, but there is a major inlining bug that is killing performance. The code below demonstrates the issue. It first defines a distance function over 2 vectors, then measures the performance using criterion. (It requires the subhask to compile.) {{{ {-# LANGUAGE BangPatterns #-} import Control.DeepSeq import Criterion.Main import qualified Data.Vector.Unboxed as VU import qualified Data.Vector.Generic as VG import qualified Prelude import SubHask -- distance_standalone :: VU.Vector Float -> VU.Vector Float -> Float distance_standalone v1 v2 = sqrt $ go 0 0 where go !tot !i = if i>VG.length v1-4 then goEach tot i else go tot' (i+4) where tot' = tot +(v1 `VG.unsafeIndex` i-v2 `VG.unsafeIndex` i) *(v1 `VG.unsafeIndex` i-v2 `VG.unsafeIndex` i) +(v1 `VG.unsafeIndex` (i+1)-v2 `VG.unsafeIndex` (i+1)) *(v1 `VG.unsafeIndex` (i+1)-v2 `VG.unsafeIndex` (i+1)) +(v1 `VG.unsafeIndex` (i+2)-v2 `VG.unsafeIndex` (i+2)) *(v1 `VG.unsafeIndex` (i+2)-v2 `VG.unsafeIndex` (i+2)) +(v1 `VG.unsafeIndex` (i+3)-v2 `VG.unsafeIndex` (i+3)) *(v1 `VG.unsafeIndex` (i+3)-v2 `VG.unsafeIndex` (i+3)) goEach !tot !i = if i>= VG.length v1 then tot else goEach tot' (i+1) where tot' = tot+(v1 `VG.unsafeIndex` i-v2 `VG.unsafeIndex` i) *(v1 `VG.unsafeIndex` i-v2 `VG.unsafeIndex` i) main = do let v1 = VU.fromList [1..200] :: VU.Vector Float v2 = VU.fromList [1..200] :: VU.Vector Float deepseq v1 $ deepseq v2 $ return () defaultMain [ bench "distance_standalone" $ nf (distance_standalone v1) v2 ] }}} Here are the results of compiling and running using GHC 7.10 and 7.8: {{{ $ ghc-7.10.1 Main.hs -O2 -fforce-recomp -ddump-to-file -ddump-simpl && ./Main [1 of 1] Compiling Main ( Main.hs, Main.o ) Linking Main ... benchmarking distance_standalone time 8.135 μs (8.121 μs .. 8.154 μs) 1.000 R² (1.000 R² .. 1.000 R²) mean 8.188 μs (8.158 μs .. 8.250 μs) std dev 139.3 ns (66.05 ns .. 250.4 ns) variance introduced by outliers: 15% (moderately inflated) }}} {{{ $ ghc-7.8.2 Main.hs -O2 -fforce-recomp -ddump-to-file -ddump-simpl && ./Main [1 of 1] Compiling Main ( Main.hs, Main.o ) Linking Main ... benchmarking distance_standalone time 733.2 ns (732.9 ns .. 733.6 ns) 1.000 R² (1.000 R² .. 1.000 R²) mean 734.1 ns (733.7 ns .. 734.5 ns) std dev 1.458 ns (1.262 ns .. 1.754 ns) }}} As you can see, GHC 7.10 is 10x slower. Looking through the core output shows that the cause of this is that GHC 7.8 is properly specializing the code whereas GHC 7.10 is not. If you uncomment the type signature before the `distance_standalone` function then both compilers perform at the faster speed. I believe the cause of this may be related to the complicated class numeric class hierarchy in SubHask. If you comment out the lines: {{{ import qualified Prelude import SubHask }}} then GHC uses the Prelude hierarchy instead of SubHask's hierarchy, and both compilers generate the faster program. There's one last wrinkle. If you define the `distance_standalone` function in a different file. Then in GHC 7.10, the `INLINE` and `INLINABLE` pragmas do absolutely nothing. Not only does the resulting code not get inlined, but if I add the specialization: {{{ {-# SPECIALIZE distance_standalone :: VU.Vector Float -> VU.Vector Float -> Float #-} }}} to the Main file, I get an error message saying something like: {{{ bench/Vector.hs:18:1: Warning: You cannot SPECIALISE ‘distance_standalone{v ru1}’ because its definition has no INLINE/INLINABLE pragma (or its defining module ‘subhask-0.1.0.0@subha_LNZiQvSbo8Z0VdLTwuvkrN:SubHask.Algebra’ was compiled without -O) }}} I get this message despite the fact that the defining module was compiled with `-O2` and the function had an `INLINABLE` pragma. If I add the specialization pragma to the defining module, then I don't get the warning, but the code still doesn't specialize properly. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/10371> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 4

[GHC] #9320: Inlining regression/strangeness in 7.8
by GHC 07 Dec '16

07 Dec '16

#9320: Inlining regression/strangeness in 7.8 -------------------------------------+------------------------------------- Reporter: dolio | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 7.8.2 Keywords: | Differential Revisions: Operating System: Unknown/Multiple | Architecture: Type of failure: Runtime | Unknown/Multiple performance bug | Difficulty: Unknown Test Case: | Blocked By: Blocking: | Related Tickets: -------------------------------------+------------------------------------- A couple days ago, it was reported to me that vector-algorithms had a significant performance regression (~20x) on GHC 7.8.2. The problem stems from a lack of inlining and specialization of some of the functions that were previously handled in 7.6 and earlier. The following is a reduced test case (the vector and primitive packages are required): {{{ module A (test) where import Control.Monad.ST import Control.Monad import Control.Monad.Primitive import Data.Vector.Generic.Mutable as U test :: (PrimMonad m, MVector v a, Num a) => Int -> v (PrimState m) a -> m a -- test :: (MVector v a, Num a) => Int -> v s a -> ST s a test 0 v = liftM (+1) $ unsafeRead v 0 test n v = do long1 v test (n-1) v {-# INLINABLE test #-} long1, long2, long3, long4 :: (PrimMonad m, MVector v a) => v (PrimState m) a -> m () long1 v = long2 v >> long2 v >> long2 v >> long2 v long2 v = long3 v >> long3 v >> long3 v >> long3 v long3 v = long4 v >> long4 v >> long4 v >> long4 v long4 v = unsafeRead v 0 >>= unsafeWrite v 0 {-# INLINE long1 #-} {-# INLINE long2 #-} {-# INLINE long3 #-} {-# INLINE long4 #-} }}} {{{ module Main (main) where import Control.Monad.ST import Data.Vector.Unboxed.Mutable as U hiding (read) import System.Environment import Unsafe.Coerce import GHC.Prim import A test0 :: Int -> MVector s Int -> ST s Int test0 n v = test n v {-# NOINLINE test0 #-} test1' :: Int -> MVector Any Int -> ST Any Int test1' n v = test n v {-# NOINLINE test1 #-} test1 :: Int -> MVector a Int -> ST a Int test1 = unsafeCoerce test1' main = getArgs >>= \(n:b:_) -> print $ runST $ do v <- new 1 write v 0 0 (if read b then test0 else test1) (read n) v }}} Module `A` exports a single function, `test`. This function is engineered to be quite large, by inlining several other functions into it, and it is itself marked INLINABLE. Then the `Main` module uses this function in two different ways: * `test0` uses `test` at a type that is compatible with `runST` * `test1'` uses `test` at a completely monomorphic type, which is then coerced to a `runST` compatible type in `test1` On 7.6 I believe (though have not checked) that there will be little or no performance difference between `test0` and `test1`. However, on 7.8.2 (and, I have been assured, 7.8.3), there is a massive speed pentalty for `test0`; about 70x on my machine. This seems to be due to no inining or specialization for its use of `test`, which can be seen from `-ddump- simpl`. However, if one changes the type of `test` in `A` to be specific to `ST s` rather than using `PrimMonad`, there is no performance difference, even on 7.8.2. So, the choice to inline and specialize seems to hinge on the instantiation of all the class constraints to monomorphic types containing no variables, rather than just types that resolve all overloading. I myself did not notice this problem, because my benchmark suite uses `IO`, which is a concrete instantiation of the type, and doesn't exhibit this problem. I have temporarily 'fixed' vector-algorithms by moving back to `INLINE` pragmas, but `INLINABLE` is actually preferable in that case, because it generates faster code than `INLINE` when the optimizations actually fire. My test case here does not illustrate that well, though. Is it safe to assume that this was not an intentional change? It's a rather weird rule (to me) if it was. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/9320> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 5

[GHC] #8668: SPECIALIZE silently fails to apply
by GHC 07 Dec '16

07 Dec '16

#8668: SPECIALIZE silently fails to apply ----------------------------------+------------------------------------- Reporter: crockeea | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 7.6.2 Keywords: | Operating System: Unknown/Multiple Architecture: x86_64 (amd64) | Type of failure: None/Unknown Difficulty: Unknown | Test Case: Blocked By: | Blocking: Related Tickets: | ----------------------------------+------------------------------------- I have a small example where GHC refuses to specialize a call to `(+)`, compiling with -O2. The two files are Foo.hs (http://lpaste.net/98464) and Main.hs (http://lpaste.net/98342) There seem to be two problems: 1. The active `SPECIALIZE` pragma should be applied, but isn't. This can be seen by comparing the core and runtimes of `fcTest` (slow) vs `vtTest` (fast). I need this version of the pragma in my real code as the phantom type `m` is reified, so I need to specialize the vector code without specifying the phantom type. 2. I can get `fcTest` to run fast if I use the commented-out `SPECIALIZE` pragma instead. However, that pragma seems very straightforward to me (all types are concrete). The docs indicate that GHC should automatically specialize in most cases, why does it not specialize to the commented-out signature automatically? This problem is also posted here: http://stackoverflow.com/questions/21071706/specialization-with- constraints -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/8668> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 26

[GHC] #9314: Huge space leak of GHC API 7.8.x
by GHC 07 Dec '16

07 Dec '16

#9314: Huge space leak of GHC API 7.8.x ------------------------------------+------------------------------------- Reporter: kazu-yamamoto | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: GHC API | Version: 7.8.3 Keywords: | Operating System: Unknown/Multiple Architecture: Unknown/Multiple | Type of failure: Other Difficulty: Unknown | Test Case: Blocked By: | Blocking: Related Tickets: | ------------------------------------+------------------------------------- GHC API 7.8.x uses much more memory than GHC API 7.6.x. Attached two files demonstrate this: - A.hs -- Simple program using GHC API (copied from Wiki) - B.hs -- A target file, just hello world You can compile A.hs as follows: {{{ % ghc A.hs -package ghc -package ghc-paths }}} A.hs stays in 10 seconds. So, we can investigate its memory usage with the "top" command.The following is the result: {{{ Mac (64bit) Linux (64bit) GHC 7.6.3: 20MB 4MB GHC 7.8.3: 106MB 13MB }}} -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/9314> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 25

[GHC] #8460: Annotation reification with types in TH
by GHC 07 Dec '16

07 Dec '16

#8460: Annotation reification with types in TH -------------------------------------------+------------------------------- Reporter: errge | Owner: Type: feature request | Status: new Priority: normal | Milestone: 7.8.1 Component: Compiler | Version: 7.7 Keywords: | Operating System: Architecture: Unknown/Multiple | Unknown/Multiple Difficulty: Easy (less than 1 hour) | Type of failure: Blocked By: | None/Unknown Related Tickets: #8397 | Test Case: | Blocking: -------------------------------------------+------------------------------- The attached patch adds two simple helper functions to support typed reification of annotations. This was not in the original #8397 patch only because the typed branch was not merged at that time. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/8460> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 7

[GHC] #9813: Error when reifying type constructor
by GHC 07 Dec '16

07 Dec '16

#9813: Error when reifying type constructor -------------------------------------+------------------------------------- Reporter: owst | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 7.8.3 Keywords: | Operating System: Architecture: Unknown/Multiple | Unknown/Multiple Difficulty: Unknown | Type of failure: Blocked By: | None/Unknown Related Tickets: | Test Case: | Blocking: | Differential Revisions: -------------------------------------+------------------------------------- The following code works under ghc-7.6.3 and template-haskell-2.8.0.0: {{{#!hs {-# LANGUAGE TemplateHaskell #-} data Huh = ThisDefinitely | UsedToWork constructorNames :: String constructorNames = $(do ty <- reify ''Huh let showCon (NormalC n _) = nameBase n strs = case ty of (TyConI (DataD _ _ _ cons _)) -> map showCon cons return . LitE . StringL $ concat strs) main = putStrLn constructorNames }}} Printing the following at compile time: {{{ [1 of 1] Compiling Main ( Main.hs, Main.o ) Loading package ghc-prim ... linking ... done. Loading package integer-gmp ... linking ... done. Loading package base ... linking ... done. Loading package array-0.4.0.1 ... linking ... done. Loading package deepseq-1.3.0.1 ... linking ... done. Loading package containers-0.5.0.0 ... linking ... done. Loading package pretty-1.1.1.0 ... linking ... done. Loading package template-haskell ... linking ... done. Linking Main ... }}} and then successfully executing: {{{ $ ./Main ThisDefinitelyUsedToWork }}} However, using ghc 7.8.3/template-haskell-2.9.0.0 I get the following compile error: {{{ [1 of 1] Compiling Main ( Main.hs, Main.o ) Loading package ghc-prim ... linking ... done. Loading package integer-gmp ... linking ... done. Loading package base ... linking ... done. Loading package array-0.5.0.0 ... linking ... done. Loading package deepseq-1.3.0.2 ... linking ... done. Loading package containers-0.5.5.1 ... linking ... done. Loading package pretty-1.1.1.1 ... linking ... done. Loading package template-haskell ... linking ... done. Main.hs:8:22: ‘Huh’ is not in the type environment at a reify In the splice: $(do { ty <- reify ''Huh; let showCon (NormalC n _) = ... ....; return . LitE . StringL $ concat strs }) }}} Is this expected? I couldn't see anything in the GHC release notes to suggest this should no longer work. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/9813> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 46

[GHC] #10383: AArch64: get GHC Calling convention working
by GHC 07 Dec '16

07 Dec '16

#10383: AArch64: get GHC Calling convention working -------------------------------------+------------------------------------- Reporter: erikd | Owner: erikd Type: feature | Status: new request | Milestone: 7.12.1 Priority: normal | Version: 7.11 Component: Compiler | Operating System: Unknown/Multiple Keywords: | Type of failure: Building GHC Architecture: aarch64 | failed Test Case: | Blocked By: Blocking: | Related Tickets: Differential Revisions: | -------------------------------------+------------------------------------- Building on AArch64 via the LLVM abckend has been working for some time. Its now time to get `("Unregisterised", "NO")` working as well. Enabling GHC calling convention can be done with this patch to `configure.ac`: {{{ diff --git a/configure.ac b/configure.ac index d5d9ab3..a11e5af 100644 --- a/configure.ac +++ b/configure.ac @@ -241,7 +241,7 @@ AC_SUBST(SOLARIS_BROKEN_SHLD) dnl ** Do an unregisterised build? dnl -------------------------------------------------------------- case "$HostArch" in - i386|x86_64|powerpc|arm) + i386|x86_64|powerpc|arm|aarch64) UnregisterisedDefault=NO ;; *) }}} but when building it this way, the stage2 compiler dies will `Illegal instructon`. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/10383> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 41

[GHC] #10332: AArch64 : divbyzero test fails
by GHC 07 Dec '16

07 Dec '16

#10332: AArch64 : divbyzero test fails -------------------------------------+------------------------------------- Reporter: erikd | Owner: Type: bug | Status: new Priority: normal | Milestone: 7.12.1 Component: Compiler | Version: 7.11 Keywords: | Operating System: Unknown/Multiple Architecture: aarch64 | Type of failure: Incorrect result Test Case: | at runtime testsuite/tests/rts/divbyzero.hs | Blocked By: Blocking: | Related Tickets: Differential Revisions: | -------------------------------------+------------------------------------- On AArch64/Linux the test in `testsuite/tests/rts/divbyzero.hs` which is: {{{ main :: IO () main = print (5 `divInt` 0) }}} just prints "0" and exits with a zero status code. From the "ARMv8 Instriction Set Overview" document I found (wasn't able to find anything more recent): https://www.element14.com/community/servlet/JiveServlet/downloadBody/41836-… Section 3.6 says "There is no hardware check for “divide by zero”, but this check can be performed in the shadow of a long latency division. A divide by zero writes zero to the destination register." Looks like we need extra code to check for this, but not sure how to report it. Should probably look at what GCC and Clang do. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/10332> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 2

[GHC] #10174: AArch64 : ghc-stage2 segfaults compiling libraries/parallel
by GHC 07 Dec '16

07 Dec '16

#10174: AArch64 : ghc-stage2 segfaults compiling libraries/parallel -------------------------------------+------------------------------------- Reporter: erikd | Owner: Type: bug | Status: new Priority: normal | Milestone: 7.12.1 Component: Compiler | Version: 7.11 Keywords: | Operating System: Unknown/Multiple Architecture: aarch64 | Type of failure: Building GHC Test Case: | failed Blocking: | Blocked By: Differential Revisions: | Related Tickets: -------------------------------------+------------------------------------- Compiling git HEAD (e02ef0e6d4eefa5f0) on AArch64 hardware using ghc-7.6.3 from Debian as the bootstrap compiler. LLVM is version 3.6. It fails with: {{{ "inplace/bin/ghc-stage2" -hisuf hi -osuf o -hcsuf hc -static -H64m -O0 -fllvm \ -this-package-key paral_2NTMt5X9WUG6LNupBFIZti -hide-all-packages -i \ -ilibraries/parallel/. -ilibraries/parallel/dist-install/build \ -ilibraries/parallel/dist-install/build/autogen \ -Ilibraries/parallel/dist-install/build \ -Ilibraries/parallel/dist-install/build/autogen -Ilibraries/parallel/. \ -optP-include -optPlibraries/parallel/dist- install/build/autogen/cabal_macros.h \ -package-key array_BLJREAlFJ4zJ2kSDphVieY -package-key base_8gvrDSBdaidLA14EDtK6ja \ -package-key conta_1uqbEcrmZiO1C91Z8ePoyI -package-key deeps_Bw45clVBNVT4S1LLyUOPfh \ -Wall -feager-blackholing -XHaskell2010 -O -fllvm -no-user-package- db -rtsopts \ -odir libraries/parallel/dist-install/build \ -hidir libraries/parallel/dist-install/build \ -stubdir libraries/parallel/dist-install/build \ -dynamic-too -c libraries/parallel/./Control/Parallel/Strategies.hs \ -o libraries/parallel/dist- install/build/Control/Parallel/Strategies.o \ -dyno libraries/parallel/dist- install/build/Control/Parallel/Strategies.dyn_o libraries/parallel/ghc.mk:5: recipe for target 'libraries/parallel/dist-install/build/Control/Parallel/Strategies.o' failed make[1]: *** [libraries/parallel/dist- install/build/Control/Parallel/Strategies.o] Segmentation fault }}} It seems that `inplace/lib/bin/ghc-stage1` (which is working fine) is statically linked to all the Haskell libraries where as `inplace/lib/bin /ghc-stage2` is dynamically linked which would explain why stage1 works fine and stage2 segfaults. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/10174> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 8