January 2017 - ghc-tickets

[GHC] #12413: Compact regions support needs some discussion in the release notes
by GHC 23 Feb '17

23 Feb '17

#12413: Compact regions support needs some discussion in the release notes -------------------------------------+------------------------------------- Reporter: bgamari | Owner: Type: task | Status: new Priority: highest | Milestone: 8.2.1 Component: Documentation | Version: 8.0.1 Keywords: | Operating System: Unknown/Multiple Architecture: | Type of failure: None/Unknown Unknown/Multiple | Test Case: | Blocked By: Blocking: | Related Tickets: Differential Rev(s): | Wiki Page: -------------------------------------+------------------------------------- And perhaps elsewhere in the users guide. It was merged as cf989ffe490c146be4ed0fd7e0c00d3ff8fe1453. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/12413> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 5

[GHC] #12798: LLVM seeming to over optimize, producing inefficient assembly code...
by GHC 23 Feb '17

23 Feb '17

#12798: LLVM seeming to over optimize, producing inefficient assembly code... -------------------------------------+------------------------------------- Reporter: GordonBGood | Owner: Type: bug | Status: new Priority: normal | Milestone: 8.2.1 Component: Compiler | Version: 8.0.1 (LLVM) | Keywords: | Operating System: Unknown/Multiple Architecture: | Type of failure: Runtime Unknown/Multiple | performance bug Test Case: | Blocked By: Blocking: | Related Tickets: Differential Rev(s): | Wiki Page: -------------------------------------+------------------------------------- Since in many cases, the use of the LLVM backend is the only way to avoid the NCG's poor register allocation (ticket #8971), this is a concern that using "-fllvm" is producing overly complex code through a (seeming) failed attempt to optimize. The following code uses a very simple "odds-only" implementation of the Sieve of Eratosthenes with a very tight inner culling loop limited to using a 16 Kilobyte buffer (<= the size of most modern CPU L1 data cache size) to reproduce the problem; it uses a "twos" Look Up Table (LUT) for better speed than using a variable shift left operation for setting the composite bits in the buffer array as it (should) take the same number of registers and the array look-up instruction is easier for the CPU to fuse than a variable shift left: {{{#!hs -- GHC_EfficiencyBug.hs {-# LANGUAGE FlexibleContexts #-} {-# OPTIONS_GHC -O3 -fllvm -rtsopts -keep-s-files #-} -- or -O2 import Data.Word import Data.Bits import Data.Array.ST (runSTUArray) import Data.Array.Base numLOOPS = 10000 :: Int twos :: UArray Int Word32 twos = listArray (0, 31) [1 `shiftL` i | i <- [0 .. 31]] soe :: () -> [Word32] soe() = 2 : [fromIntegral i * 2 + 3 | (i, False) <- assocs bufb] where bufb = runSTUArray $ do let bfLmt = (256 * 1024) `div` 2 - 1 -- to 2^18 + 2 is 128 KBits - 1 = 16 KBytes cmpstsb <- newArray (0, bfLmt) False :: ST s (STUArray s Int Bool) cmpstsw <- (castSTUArray :: STUArray s Int Bool -> ST s (STUArray s Int Word32)) cmpstsb let loop n = -- cull a number of times to test timing if n <= 0 then return cmpstsb else let cullp i = let p = i + i + 3 in let s = (p * p - 3) `div` 2 in if s > bfLmt then loop (n - 1) else do isCmpst <- unsafeRead cmpstsb i if isCmpst then cullp (i + 1) else -- is Prime let cull j = -- very tight inner loop where all the time is spent if j > bfLmt then cullp (i + 1) else do let sh = unsafeAt twos (j .&. 31) -- (1 `shiftL` (j .&. 31))) let w = j `shiftR` 5 ov <- unsafeRead cmpstsw w unsafeWrite cmpstsw w (ov .|. sh) cull (j + p) in cull s in cullp 0 loop numLOOPS main = print $ length $ soe() }}} The main culling is repeated "numLOOPS" times to get a reasonable execution time for accurate timing and to make the time required to use the List comprehension to determine the number of found primes (the answer) a negligible part of the execution time. Timing results can be produced by running "./GHC_EfficiencyBug +RTS -s". The desired assembly code result for the tight inner loop is as for the Rust/LLVM compiler, in this case for x86_64 64-bit code: {{{ .p2align 4, 0x90 .LBB10_27: movq %rcx, %rdx shrq $5, %rdx movl %ecx, %esi andl $31, %esi movl (%rbp,%rsi,4), %esi orl %esi, (%r14,%rdx,4) addq %rax, %rcx .LBB10_26: cmpq %r13, %rcx jb .LBB10_27 }}} The above code is extremely efficient on a CPU that is not cache bottle necked (such as the AMD Bulldozer series are) and takes just about three clock cycles per inner composite culling loop on Intel Sky Lake; it is just as efficient for x86 code since there are only seven registers used in this inner loop. Due to this attempt at "over-optimization", the GHC/LLVM backend produces the following x86_64 64-bit code: {{{ .align 16, 0x90 .LBB34_2: # %c8T2 # =>This Inner Loop Header: Depth=1 movq %rcx, %rsi sarq $5, %rsi movl %r8d, %edi andl $124, %edi movl 16(%rax,%rdi), %edi orl %edi, 16(%r11,%rsi,4) addq %r14, %rcx addq %rdx, %r8 cmpq %r10, %rcx jle .LBB34_2 }}} As can be seen, instead of just masking the "twos" index register by 31 (0x1F), the code is using two extra separate registers to contain "(j * 4)" increment and the accumulated index, which increment is added to the "twos" index register per loop and masked by 124 (0x7C or 0x1F times 4), requiring an extra two registers and an extra instruction for the extra addition. This isn't a problem as to the number of registers for x86_64 code which has more than enough, but it adds the extra instruction execution time of one third of a CPU clock cycle (I know, only one ninth extra time). However, for 32-bit x86 code with barely enough registers previously, the use of the extra registers triggers a chain of three register reloads as can be seen in the following assembly code: {{{ .align 16, 0x90 LBB33_2: # %c8Wb # =>This Inner Loop Header: Depth=1 movl %ebx, %ebp sarl $5, %ebp movl %edi, %ecx andl $124, %ecx movl %esi, %edx movl %eax, %esi movl 36(%esp), %eax # 4-byte Reload movl 8(%eax,%ecx), %ecx movl %esi, %eax movl %edx, %esi orl %ecx, 8(%esi,%ebp,4) addl 32(%esp), %ebx # 4-byte Folded Reload addl 28(%esp), %edi # 4-byte Folded Reload cmpl %eax, %ebx jle LBB33_2 }}} '''The above code runs about 25% slower than it should on Intel Sky Lake for this 32-bit code.''' This was tested for GHC version 8.0.1 under both Windows and Linux for both 32-bit and 64-bit code with identical results for each native code width. The code was also tested for 32 and 64 bit code produced by the NCG; for this specific problem, NCG takes the simple approach and does not waste the extra register. However, due to the inefficient allocation of registers as per ticket #8971, not moving the loop completion check to the end of the loop and thus requiring an extra jump instruction, and not combining the read/modify/write into a single instruction, it is still slower (much slower for 32-bit code) than the LLVM produced code. As its problems are known, I have not documented the NCG code. Conclusion: This may seem like a nit picky type of bug as in some use cases the execution time cost is very small, but it may be an indication of problems in other use cases that cause more serious effects on execution speed. It is my feeling that for such low level somewhat imperative types of code, GHC should really produce code that is as fast as C/C++/Rust. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/12798> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 7

[GHC] #12863: Associated data families don't use superclasses when deriving instances
by GHC 22 Feb '17

22 Feb '17

#12863: Associated data families don't use superclasses when deriving instances -------------------------------------+------------------------------------- Reporter: Iceland_jack | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 8.0.1 Keywords: | Operating System: Unknown/Multiple Architecture: | Type of failure: None/Unknown Unknown/Multiple | Test Case: | Blocked By: Blocking: | Related Tickets: Differential Rev(s): | Wiki Page: -------------------------------------+------------------------------------- {{{#!hs class Functor (TotalMap key) => Key key where data TotalMap key :: Type -> Type (¡) :: TotalMap key val -> key -> val }}} a constraint `Key key` entails a `Functor (TotalMap key)` constraint, using the notation of [https://hackage.haskell.org/package/constraints constraints] {{{#!hs prf1 :: Key key :- Functor (TotalMap key) prf1 = Sub Dict prf2 :: (Key key1, Key key2) :- (Functor (TotalMap key1), Functor (TotalMap key2)) prf2 = prf1 *** prf1 }}} Yet these proofs are not used in {{{#!hs -- tzLp.hs:28:89-95: error: … -- • No instance for (Functor (TotalMap key1)) -- arising from the first field of ‘PairMap’ -- (type ‘TotalMap key1 (TotalMap key2 val)’) -- Possible fix: -- use a standalone 'deriving instance' declaration, -- so you can specify the instance context yourself -- • When deriving the instance for (Functor (TotalMap (key1, key2))) -- Compilation failed. instance (Key key1, Key key2) => Key (key1, key2) where data TotalMap (key1, key2) val = PairMap (TotalMap key1 (TotalMap key2 val)) deriving Functor (¡) :: TotalMap (key1, key2) val -> ((key1, key2) -> val) PairMap keys ¡ (k1, k2) = keys ¡ k1 ¡ k2 }}} I would expect it to work like {{{#!hs deriving instance (Key key1, Key key2) => Functor (TotalMap (key1, key2)) }}} Same problem occurs with only a single constraint {{{#!hs -- tzLp.hs:33:14-20: error: … -- • No instance for (Functor (TotalMap key)) -- arising from the first field of ‘Id’ (type ‘TotalMap key val’) -- Possible fix: -- use a standalone 'deriving instance' declaration, -- so you can specify the instance context yourself -- • When deriving the instance for (Functor -- (TotalMap (Identity key))) -- Compilation failed. instance Key key => Key (Identity key) where data TotalMap (Identity key) val = Id (TotalMap key val) deriving Functor }}} which should work like {{{#!hs deriving instance Key key => Functor (TotalMap (Identity key)) -- or deriving instance Functor (TotalMap key) => Functor (TotalMap (Identity key)) }}} -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/12863> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 3

[GHC] #10343: Make Typeable track kind information better
by GHC 21 Feb '17

21 Feb '17

#10343: Make Typeable track kind information better -------------------------------------+------------------------------------- Reporter: oerjan | Owner: Type: feature | Status: new request | 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: 9858 Test Case: typeOf :: | Typeable (a::k) => Proxy a -> | TypeRep | Blocking: | Differential Revisions: | -------------------------------------+------------------------------------- With all the apparent exploitable bugs now fixed, #9858 seems to be settling down. Simon P.J. asked me to make a new ticket for some quibbles I still have with limitations of the new design. These started with some examples that I discussed there pre-implementation with Richard Eisenberg, and which he concluded could be supported by a "sufficently smart solver". In the new (current) design, a `TypeRep` ''does'' contain kind information, which as I understand it is a list of `KindRep`s for the explicit kind parameters applied to the type constructor after type inference. This is sufficient to prevent confusing the same type constructor used at different kinds, and also easily supports type application ''provided'' the `TypeRep`s of the parts are known. For this, the internal `Typeable` constraint solver only ever needs to combine `TypeRep`s, never decompose them. Since no one else complained in #9858, I assume this is enough for what people are currently using `Typeable` for. However: I think the current `TypeRep` representation for `Typeable (a :: k)` doesn't contain enough information to generally extract the `KindRep` for `k` itself, much less to decompose it, even if the internal `Typeable` solver were to try. This prevents some polykinded code that is intuitively sound, and which "worked" in the old system, although if only because it didn't try to track kinds at all. For example, the following expression, with `PolyKinds` enabled: {{{#!hs typeOf :: Typeable a => Proxy a -> TypeRep }}} In the old system, this worked simply by combining the `TypeRep` for `Proxy` with the `TypeRep` for `a`. But in the new system, it fails, because the `TypeRep` for `Proxy` needs to know the kind of `a`, which is neither statically known nor available from the `TypeRep` of `a`. You would need to have compile-time information about the type constructor of `a` to even know how to build the `KindRep` from the kind arguments. This problem with constructing `TypeRep`s would arise whenever a polykinded type constructor is applied to a type whose kind is not a known monomorphic one. The second similar one I can think of is {{{#!hs typeRep :: Typeable a => Proxy (Typeable a) -> TypeRep }}} In an ideal system, we can also find use cases that require not just extracting the `KindRep` for a type, but also decomposing it. Here is my very hypothetical kitchen sink example from [comment:46:ticket:9858] (this example was made up as a test case before I knew what the new solver would support): {{{#!hs {-# LANGUAGE PolyKinds #-} {-# LANGUAGE DataKinds #-} {-# LANGUAGE KindSignatures #-} {-# LANGUAGE AutoDeriveTypeable #-} import Data.Typeable newtype C a b = C () step :: Proxy (a :: [k] -> *) -> Proxy (C a :: [k -> *] -> *) step Proxy = Proxy nest :: Typeable (a :: [k] -> *) => Int -> Proxy a -> TypeRep nest 0 x = typeRep x nest n x = nest (n-1) (step x) main = print $ nest 10 (Proxy :: Proxy (C () :: [()] -> *)) }}} Here, to make this compile, the solver would need to derive `Typeable (C a :: [k -> *] -> *]` from `Typeable (a :: [k] -> *)`, which at runtime requires first extracting the `KindRep` of `k` from the `TypeRep` of `a`, and then building the `TypeRep` for the correctly kinded `C`. At least the former part seems essentially impossible with the current `TypeRep` representation, even with a more clever solver. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/10343> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 38

[GHC] #13163: Make type import/export API Annotation friendly
by GHC 21 Feb '17

21 Feb '17

#13163: Make type import/export API Annotation friendly -------------------------------------+------------------------------------- Reporter: alanz | Owner: alanz Type: bug | Status: new Priority: normal | Milestone: 8.2.1 Component: Compiler | Version: 8.0.1 (Parser) | Keywords: | Operating System: Unknown/Multiple Architecture: | Type of failure: Incorrect API Unknown/Multiple | annotation Test Case: | Blocked By: Blocking: | Related Tickets: Differential Rev(s): | Wiki Page: -------------------------------------+------------------------------------- At the moment an export of the form {{{#!hs type C(..) }}} is parsed by the rule {{{ | 'type' oqtycon {% amms (mkTypeImpExp (sLL $1 $> (unLoc $2))) [mj AnnType $1,mj AnnVal $2] } }}} This means that the origiinal `oqtycon` loses its location which is then retained in the `AnnVal` annotation. The problem is if the `oqtycon` has its own annotations, these get lost. e.g. in {{{#!hs type (?)(..) }}} the parens annotations for `(?)` get lost. This can be solved by introducing a new `IE` constructor `IEVarType`, treated exactly the same as `IEVar` in the rest of the processing. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/13163> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 5

[GHC] #11526: unsafeLookupStaticPtr should not live in IO
by GHC 21 Feb '17

21 Feb '17

#11526: unsafeLookupStaticPtr should not live in IO -------------------------------------+------------------------------------- Reporter: edsko | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 8.0.1-rc1 Keywords: | Operating System: Unknown/Multiple Architecture: | Type of failure: None/Unknown Unknown/Multiple | Test Case: | Blocked By: Blocking: | Related Tickets: Differential Rev(s): | Wiki Page: -------------------------------------+------------------------------------- `GHC.StaticPtr` provides {{{#!hs unsafeLookupStaticPtr :: StaticKey -> IO (Maybe (StaticPtr a)) }}} I don't think this function should live in IO. It is morally pure (the SPT that it needs to access is static for the duration of the program after all) and when this lives in IO this function cannot be used in `Binary` instances (without using `unsafePerformIO`, obviously). -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/11526> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 14

[GHC] #11817: Add proper support for weak symbols to the runtime linker
by GHC 21 Feb '17

21 Feb '17

#11817: Add proper support for weak symbols to the runtime linker -------------------------------------+------------------------------------- Reporter: Phyx- | Owner: Type: feature | Status: new request | Priority: normal | Milestone: 8.2.1 Component: Runtime | Version: 8.1 System (Linker) | Keywords: | Operating System: Unknown/Multiple Architecture: | Type of failure: None/Unknown Unknown/Multiple | Test Case: | Blocked By: Blocking: | Related Tickets: #11223 Differential Rev(s): | Wiki Page: -------------------------------------+------------------------------------- Currently weak symbols support is only partially supported by the runtime linker. The current support is mostly there to provide `COMDAT` support. In #11223 this support has been extended but there is still some work needed to finish weak symbols support. Also there needs some investigation into how this support should look like. E.g. are we only supporting linking weak symbols in C or do we want to support it in Haskell as well. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/11817> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 6

[GHC] #11011: Type-indexed TypeReps, Static Pointers and Distributed Closures
by GHC 21 Feb '17

21 Feb '17

#11011: Type-indexed TypeReps, Static Pointers and Distributed Closures -------------------------------------+------------------------------------- Reporter: bjmprice | Owner: Type: feature | Status: new request | Priority: normal | Milestone: Component: Compiler | Version: Keywords: | Operating System: Unknown/Multiple Architecture: | Type of failure: None/Unknown Unknown/Multiple | Test Case: | Blocked By: Blocking: | Related Tickets: Differential Rev(s): | Wiki Page: -------------------------------------+------------------------------------- There is a mock-up of a new API for type indexed type representations, basic Dynamic functionality, static pointers and distributed closures (https://github.com/brprice/typeableT) based on https://ghc.haskell.org/trac/ghc/wiki/TypeableT. We would like to invite comments and discussion from the community using this ticket as a more permenant home that email. One particular point that would benefit from more eyes is the polymorphic static pointers support. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/11011> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 52

[GHC] #12636: ProfHeap's printf modifiers are incorrect
by GHC 21 Feb '17

21 Feb '17

#12636: ProfHeap's printf modifiers are incorrect ----------------------------------------+--------------------------------- Reporter: Phyx- | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: Runtime System | Version: 8.0.1 Keywords: newcomer | Operating System: Windows Architecture: Unknown/Multiple | Type of failure: None/Unknown Test Case: | Blocked By: Blocking: | Related Tickets: Differential Rev(s): | Wiki Page: ----------------------------------------+--------------------------------- during compile I noticed {{{ rts\ProfHeap.c: In function 'dumpCensus': rts\ProfHeap.c:768:26: error: warning: format '%lu' expects argument of type 'long unsigned int', but argument 3 has type 'long long unsigned int' [-Wformat=] fprintf(hp_file, "VOID\t%lu\n", ^ rts\ProfHeap.c:770:26: error: warning: format '%lu' expects argument of type 'long unsigned int', but argument 3 has type 'long long unsigned int' [-Wformat=] fprintf(hp_file, "LAG\t%lu\n", ^ rts\ProfHeap.c:772:26: error: warning: format '%lu' expects argument of type 'long unsigned int', but argument 3 has type 'long long unsigned int' [-Wformat=] fprintf(hp_file, "USE\t%lu\n", ^ rts\ProfHeap.c:774:26: error: warning: format '%lu' expects argument of type 'long unsigned int', but argument 3 has type 'long long unsigned int' [-Wformat=] fprintf(hp_file, "INHERENT_USE\t%lu\n", ^ rts\ProfHeap.c:776:26: error: warning: format '%lu' expects argument of type 'long unsigned int', but argument 3 has type 'long long unsigned int' [-Wformat=] fprintf(hp_file, "DRAG\t%lu\n", ^ }}} -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/12636> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 11

[GHC] #13152: Provide a mechanism to notify build system when .hi file is ready
by GHC 20 Feb '17

20 Feb '17

#13152: Provide a mechanism to notify build system when .hi file is ready -------------------------------------+------------------------------------- Reporter: rwbarton | Owner: Type: feature | Status: new request | Priority: normal | Milestone: 8.2.1 Component: Driver | Version: 8.1 Keywords: | Operating System: Unknown/Multiple Architecture: | Type of failure: None/Unknown Unknown/Multiple | Test Case: | Blocked By: Blocking: | Related Tickets: Differential Rev(s): | Wiki Page: -------------------------------------+------------------------------------- In one-shot mode GHC typically finishes writing the interface file around halfway through compilation. A dependent module (if it doesn't use Template Haskell) only needs the interface file to start building. If we could start building a module as soon as all its dependencies' interface files are ready, we would cut the critical path by about a factor of 2; and parallelism in the GHC build tree is currently low enough that this should give significantly lower build times even on only modestly parallel systems. The first obstruction to doing this is that there isn't a good way to know when the interface file becomes ready. I propose the following simple and flexible mechanism: * Add a GHC command-line argument `-finterface-file-finished=N,F,str`. When GHC has finished writing the interface file it uses the `send` system call to send the string `str` to file descriptor `N` using flags `F`. The build system might then invoke GHC with file descriptor `N` open to a UNIX datagram socket, for example, and generate a unique `str` for each interface file dependency. Alternative suggestions are more than welcome. I tentatively milestoned this for 8.2 since it would be nice to have this available in the bootstrapping compiler when we switch to Hadrian, and the GHC-side implementation should be simple. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/13152> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 9