October 2005 - Glasgow-haskell-users

RE: GHC-6.4.1 on FreeBSD-amd64 port progress
by Simon Marlow 30 May '06

30 May '06

On 25 October 2005 10:01, Wilhelm B. Kloke wrote: > Simon Marlow <simonmar(a)microsoft.com> schrieb: >> On 24 October 2005 19:04, Wilhelm B. Kloke wrote: >> >> Is it that it does not find the libraries, or does it find the >> libraries but still gets a bunch of undefined references? If the >> latter, then there could be an issue with the mangler. Try an 'nm' >> on libHSbase.a to see if the missing symbols are actually defined. > > Nm shows that the problem may come from splitting. The missing symbols > are marked as locals (small d); being in different .o modules they > are not resolved. Try with splitting off: set SplitObjs=NO in your mk/build.mk. Cheers, Simon

3 2

Re: C minus minus? No, C monad monad.
by Simon Peyton-Jones 10 Nov '05

10 Nov '05

John Interesting message. http://repetae.net/john/repos/jhc/docs/c-minus-monad.txt But I'm not completely convinced. Firstly, C-- *continuations* are quasi-first class values: you can pass them as arguments to functions, and store them in data structures. In contrast, your continuations are more akin to C-- *labels*. These are definitely not first class. They just give a textual way of describing a control flow graph. That's much closer to what you mean to do (with join points etc). Of course, labels allow you to build a cyclic graph. Your hack for building cycles feels to me like using the monadic story for something it wasn't really meant for. Somehow, a set of labelled blocks seems more more direct. I'm also not convinced that you are getting anything much from your monadic presentation that you would not also get, sometimes more directly, from a C-- representation. What, do you think you gain? There is a long and distinguished history to regarding lambda code as an intermediate language that can go all the way to machine code (e.g. the lambda-the-ultimate papers; the T compiler; Appel's book etc). Using a monadic, rather than CPS, framework may be better -- though it's a tricky case to make. But in fact, few compilers do take this idea right through. It's hard to say why, except that it can be a bit clumsy to make one hammer that is truly comfortable on many different nails. At the moment, I'd still incline to translating Grin into C-- (or something like it), and optimising there. But maybe you have good reasons not to do that. Simon

2 1

RE: jhc vs ghc and the surprising result involving ghc generatedassembly.
by Simon Peyton-Jones 03 Nov '05

03 Nov '05

All this interesting talk of back ends is somewhat topical at GHC HQ....It may well be that we can share some technology here. GHC uses C-- as its intermediate representation; Core ==> C-- either==> assembly code or==> C Currently the Core->C-- step is done in one big go, by the code generator. The C-- code uses NO native C-- calls; instead, it uses tail calls only, and builds a totally explicit stack as a data structure in the heap. But Chris Thompson-Walsh, as his undergrad project, is writing a CPS converter for C--, so we can do this Core ==> C-- ==> C-- with tail calls only The first step is code generation, converting lambda code to C--, but now this is C-- with native C-- calls. The stack is implicit, entirely under C--'s control. The second step is the CPS conversion. After the CPS conversion the stack is totally explicit. One reason for doing all this is that it factors the rather-complicated code generator into two pieces. The second reason is that if we had a full-blown C-- compiler available, supporting the full runtime interface, we could print out the C-- before the CPS conversion, and use the C-- compiler for the rest. Anyway, apart from the first code-gen step, not much of this process is particularly GHC specific. At the moment it's entirely implemented in Haskell. One of the things Chris is going to look at (if he has time) is packaging the Cmm data type, its various useful functions (e.g. data flow analysis), and his CPS converter, as a Cabal package that GHC just uses. Then other people could use it too. Furthermore the Cmm data type would be the perfect place to spot self-tail calls, eliminate memory operations etc -- all the things that John has been mentioning in his message. That is, a C-- ==> C-- optimiser would be a nice independent project. My reason for mentioning all this now, when it's still partly vapourware, is to see if anyone else is interested. In particular, perhaps a C-- optimiser for jhc would be usable in GHC and vice versa. Simon | -----Original Message----- | From: glasgow-haskell-users-bounces(a)haskell.org [mailto:glasgow-haskell-users- | bounces(a)haskell.org] On Behalf Of John Meacham | Sent: 27 October 2005 00:37 | To: glasgow-haskell-users(a)haskell.org | Subject: Re: jhc vs ghc and the surprising result involving ghc generatedassembly. | | On Wed, Oct 26, 2005 at 12:24:14PM -0400, Jan-Willem Maessen wrote: | > Nice analysis. I indeed found with phc that shadow stack references | > absolutely killed performance, and I aggressively cached stack | > locations in locals, spilling to stack only when GC information | > needed to be accurate. [There was a giant infrastructure to save | > only live data to stack, but we won't go into that now as it was the | > source of almost all the codegen bugs...] | | phc? there is another haskell compiler out there? | | > This makes a big difference. The phc compiler even put comments in | > the code so that I could figure out what came from where. | | yeah, that is something I would like to add. Unfortunatly I wasn't | forward thinking enough to put annotation points everywhere I might | eventually need them so I will have to go through and do that at some | point :) | | > I'm impressed that gcc found this. It's definitely living a bit | > dangerously, and your suggestions below for self tail call handling | > are the ones I found most effective. (They also allowed me to bypass | > some prologue garbage, since phc used a one-C-function-per-Haskell- | > function model with internal resumption points.) Non-self tail calls | > I was careful to compile to: | > return f(...); | > I expect from the above that gcc does better at spotting tail calls now. | [..rest cut..]

4 3

jhc vs ghc and the surprising result involving ghc generated assembly.
by John Meacham 02 Nov '05

02 Nov '05

(I apologize in advance if this message seems self congradulatory, but after a long time of being disheartened by jhc only having marginal gains over ghc, I am finally seeing some substantial benefits, many of which are the result of optimizations that can actually be ported back to ghc) So, what started out as a simple pat on my own back for getting strictness and CPR finally working turned into an adventure in assembly language. An incidental comparason between jhc and ghc's output completly surprised me, for tight mathematical inner loops jhc is producing code that runs 3-7x faster than ghc. This surprised me because I expected them to be identical, they determine the same strictness and produce the same worker/wrapper split, but due to a quirk of how ghc generates C code, it ends up producing an ultimatly slower result. fortunatly, I think the problem is easy to mitigate. (and jhc loses its lead once again :) ) (all the following snippets are unedited/unreformatted output of the programs specified except for the removal of some profiling tags) the motivating example is the plain old basic accumulating factorial function: fac :: Int -> Int -> Int fac 1 r = r fac n r = fac (n - 1) (n*r) the strictness info is correctly infered that fac is strict in all its arguments and returns a CAF so it is translated into a nice and speedy version that takes unboxed arguments and returns unboxed results. here is the C code that jhc generates. (As an aside, I am very proud of how readable and how much structure the jhc generated C code preserves of the original haskell. it's a small thing, perhaps only implementors appreciate it, but I am glad I spent the time needed to do so.) /* fW@.fMain.fac */ static int fWXAXDfMainXDfac(int v94, int v95) { int v96; int v97; int v98; int v99; switch (v94) { case 1: return v95; break; default: v96 = v94; v97 = (int)(v94 - 1); v98 = (int)(v94 * v95); v99 = fWXAXDfMainXDfac(v97, v98); return v99; break; } } notice that besides being a bit verbose and using a tailcall, this is exactly what A C programmer would write. In fact, the generated assembly is quite nice and I think perhaps provably optimal even compared to hand-coded assembly :) jhc assembly output: fWXAXDfMainXDfac: .L108: cmpl $1, %edi je .L110 imull %edi, %esi decl %edi jmp .L108 .L110: movl %esi, %eax ret notice, a tiny inner loop, 4 instructions and one conditional jump. no memory acceses whatsoever. now, lets look at what ghc does with the same function: the generated C code: EI_(Main_zdwfac_info); FN_(Main_zdwfac_entry) { W_ _s27g; W_ _s27i; W_ _s27l; FB_ _s27g = *Sp; if (_s27g != 0x1UL) goto _c282; R1.p = (P_)(Sp[1]); Sp=Sp+2; JMP_(*Sp); _c282: _s27l = _s27g * (Sp[1]); _s27i = _s27g - 0x1UL; Sp[1] = _s27l; *Sp = _s27i; JMP_((W_)&Main_zdwfac_info); FE_ } it looks complicated, but what it effectivly does is pop the arguments off the stack, run the code but with an explicit jump rather than recursion and push the result back onto the stack. other than the stack stuff at the beginnig and end, we would expect this to get compiled to roughly the same assembly as the jhc version with a nice tight inner loop and just some stack futzing boilerplate. and now the generated assembly. Main_zdwfac_info: .text .align 8 .text movq (%rbp), %rdx cmpq $1, %rdx jne .L2 movq 8(%rbp), %r13 leaq 16(%rbp), %rbp movq (%rbp), %rax .L4: jmp *%rax .L2: movq %rdx, %rax imulq 8(%rbp), %rax movq %rax, 8(%rbp) leaq -1(%rdx), %rax movq %rax, (%rbp) movl $Main_zdwfac_info, %eax jmp .L4 ack! lets count what happens on each iteration of the loop: we have 5 (!) memory acceses and two jumps, one of them being indirect! in fact, each time through the loop, it loads the same values into the same registers. this is really bad compared to the jhc assembly. the basic issue is an interaction between ghc's use of global registers, a stack, and indirect calls. an indirect jump is very expensive. modern processors are pipelined, having many instructions in the queue at once, looking ahead and beginning to evaluate what is coming up. when an indirect jump is encountered, the CPU has no choice but to flush the whole pipeline because it has no idea where it goes, even with conditional branches cpus can predict with good accuracy which branch will be taken and at worst, it discards the pipeline, but with an indirect jump, there is no chance of it having a full pipeline. However the major issue is the following. %rbp is the global stack pointer pointing to the STG stack, since it is global it can be modified from anywhere, since gcc can't know if the function it jumps to modified said register it has no choice but to load all values relative to it every time through the loop because it may have changed. furthermore gotos and labels are very problematic for gcc to optimize around. for various tiresome reasons gcc cannot perform (most) code motion optimizations across explicit labels and gotos, especially when they deal with the global register variables and memory stores and loads. since these are arguably some of the most important optimizations, this is quite bad for the generated assembly. loop: if () goto loop; is not equivalent to a do-while loop, loop invarients cannot be hoisted out of the above for instance (except in some cases... it is all quite tricky and we want gcc to have as much freedom as possible). all in all, this makes the code generated by gcc compiling something generated by ghc not very good at all. there are a couple of things we can do to mitigate these problems: get rid of indirect jumps whenever possible. use C control constructs rather than gotos. A couple simple rules seem to help greatly. * turn anything of the form JMP_((W_)&self) where self is oneself into a goto that gotos a label at the beginning of the function. * do simple pattern matthing on the basic blocks to recognize where C control constructs can be placed. for instance turn if (x) { goto y; } blah.. baz.. JMP_(foo) into if (x) { goto y; } else { blah.. baz.. JMP_(foo) } extending the else to after the next jump or goto. * getting stack dereferences out of your loops. manually performing the first two optimizations mentioned above we get: EI_(Main_zdwfac_info); FN_(Main_zdwfac_entry) { W_ _s27g; W_ _s27i; W_ _s27l; FB_ fac_entry: _s27g = *Sp; if (_s27g != 0x1UL) { goto _c282; } else { R1.p = (P_)(Sp[1]); Sp=Sp+2; JMP_(*Sp); } _c282: _s27l = _s27g * (Sp[1]); _s27i = _s27g - 0x1UL; Sp[1] = _s27l; *Sp = _s27i; goto fac_entry; FE_ } and this produces the assembly: Main_zdwfac_info: .text .align 8 .text movq %rbp, %rdx movq (%rbp), %rcx cmpq $1, %rcx je .L3 .L6: movq %rcx, %rax imulq 8(%rdx), %rax movq %rax, 8(%rdx) leaq -1(%rcx), %rax movq %rax, (%rbp) .L4: movq %rbp, %rdx movq %rax, %rcx cmpq $1, %rax jne .L6 .L3: movq 8(%rdx), %r13 leaq 16(%rdx), %rbp jmp *(%rbp) we still have some unnecesarry memory accesses, but the indirect jump and the spurious jump are gone and we have less instructions in the main loop making this code noticably faster. in order to get rid of the unessesary memory accesess, we need to either 1. fully convert it to use C control constructs, so gcc will do it for us. (code motion and loop invarient being inhibited again by gotos) or 2. do it ourselves by analyzing when the consumer of what we are putting on the stack is ourselves. the first is greatly preferable but not always possible. These should be straightforward to implement in the C code generator. it also suggests we might want to try to use the native C calling convention on leaf nodes that deal with unboxed values (so we get register passing and return values for free) or taking groups of mutually recursive functions and turning them all into one function with explicit jumps between them. to show they are actually optimizing the same function, here are both of their core representaions, other than syntax, they are identical: jhc core: (uses unicode, utf8 formatted) W@.fMain.fac∷int → int → int = λx9216∷int.λx9222∷int.(case x9216 of 1 → x9222 ; x9238∷case <(int)-(int,int) x9216 1∷int> of x9252∷int → case <(int)*(int,int) x9216 x9222∷int> of x34∷int → case W@.fMain.fac x9252 x34 of x9208∷int → x9208;;;;;) ghc core: %rec {zdwfac :: GHCziPrim.Intzh -> GHCziPrim.Intzh -> GHCziPrim.Intzh = \ (ww::GHCziPrim.Intzh) (ww1::GHCziPrim.Intzh) -> %case (GHCziPrim.Intzh) ww %of (ds::GHCziPrim.Intzh) {%_ -> Main.zdwfac (GHCziPrim.zmzh ds (1::GHCziPrim.Intzh)) (GHCziPrim.ztzh ds ww1); (1::GHCziPrim.Intzh) -> ww1}}; some random notes: the 3x-7x factor was tested on an i386, on x86_64 the margin is much much greater for reasons that are still unclear. while testing this I noticed that jhc and ghc have dramatically different results on x86_64 pretty much across the board, if their programs take comparable amounts of time on i386, the jhc one will run twice as fast as the ghc one on x86_64. I think ghc must be tickling the x86_64 in a particularly bad way for this dramatic of an effect to be observed. I will poke around the generated assembly of ghc some more and see if I can find the culprit. (this is also online at: http://repetae.net/jhc-vs-ghc-assembly.txt ) John -- John Meacham - ⑆repetae.net⑆john⑈

5 10

RE: jhc vs ghc and the surprising result involving ghc generatedassembly.
by Simon Marlow 01 Nov '05

01 Nov '05

John, this is great stuff. There's clearly a lot we can do to improve GHC's code, at least for fac :-) (I'd be really interested in any numbers you have for larger programs too, eg. the nofib suite). > and now the generated assembly. > > Main_zdwfac_info: > .text > .align 8 > .text > movq (%rbp), %rdx > cmpq $1, %rdx > jne .L2 > movq 8(%rbp), %r13 > leaq 16(%rbp), %rbp > movq (%rbp), %rax > .L4: > jmp *%rax gcc started generating this rubbish around version 3.4, if I recall correctly. I've tried disabling various optimisations, but can't seem to convince gcc not to generate the extra jump. You don't get this from the native code generator, BTW. Actually, our NCG beats gcc in most cases on x86_64, I believe. > .L2: > movq %rdx, %rax > imulq 8(%rbp), %rax > movq %rax, 8(%rbp) > leaq -1(%rdx), %rax > movq %rax, (%rbp) > movl $Main_zdwfac_info, %eax > jmp .L4 [snip] > A couple simple rules seem to help greatly. > > * turn anything of the form JMP_((W_)&self) where self is oneself into a goto > that gotos a label at the beginning of the function. Sounds like a good idea. In fact, we should do all this as a C-- -> C-- optimisation, as Simon PJ pointed out (and I mentioned on IRC yesterday). one small caveat: our register allocator can't handle loops at the moment, so any optimisations that generated actual loops would have to be disabled for the NCG for now. > * do simple pattern matthing on the basic blocks to recognize where C control > constructs can be placed. > > for instance turn > > if (x) { goto y; } > blah.. > baz.. > JMP_(foo) > > into > > if (x) { goto y; } else { > blah.. > baz.. > JMP_(foo) > } I don't actually understand why this helps, but it's easy enough to do. > * getting stack dereferences out of your loops. A bit more tricky, but could still be done as C-- to C--. Note that GHC's back end is really aimed at producing good code when there are registers available for passing arguments - this isn't true on x86 or x86_64 at the moment, though. > 1. fully convert it to use C control constructs, so gcc will do it for us. > (code motion and loop invarient being inhibited again by gotos) Ok. Jan's do-while idea seems like a cheap way to achieve this. > These should be straightforward to implement in the C code generator. it also > suggests we might want to try to use the native C calling convention on leaf > nodes that deal with unboxed values (so we get register passing and return > values for free) or taking groups of mutually recursive functions and turning > them all into one function with explicit jumps between them. Using the real C calling convention brings some problems - we'd have to use setjmp/longjmp to pass control back to the scheduler or garbage collector. Cheers, Simon

4 4

ghc-6.4.1: panic - Prelude.last: empty list
by Einar Karttunen 30 Oct '05

30 Oct '05

Hello I am having problems with GHC 6.4.1 dying with the message: ghc-6.4.1: panic! (the `impossible' happened, GHC version 6.4.1): Prelude.last: empty list The source is nontrivial and contains template haskell. The error seems to appear after byte code generation. The log of the compile with -v9 is at the address below: http://cs.helsinki.fi/u/ekarttun/haskell/log.txt Is there anything else I should do to isolate the error better? - Einar Karttunen

2 1

Error compiling using Network module
by Jorge Guerra 28 Oct '05

28 Oct '05

Hello to all, I have recently began using ghc, in particular the network module and I'm having problems with it. The following code ilustrates my problem, is very simple all I try to do is connect to a socket in my local machine send a request and print the response. -- BEGIN CODE import IO import Network host = "127.0.0.1" port = PortNumber 631 request = "GET / HTTP/1.1\nHost: 127.0.0.1:631\n\n" main = withSocketsDo $ do h <- connectTo host port hSetBuffering h LineBuffering hPutStr h request getResponse h hClose h getResponse h = do b <- hIsEOF h if b == False then do line <- hGetChar h hFlush h putChar line getResponse h else putStrLn "End of Input" -- END CODE When I try to compile it using the command: ghc -o socketTest socketTest.hs The following error appears: socketTest.o(.text+0xe2): In function `r2dL_info': : undefined reference to `NetworkziSocket_zdfNumPortNumber_closure' socketTest.o(.text+0x7c1): In function `s2f8_info': : undefined reference to `Network_connectTo_closure' socketTest.o(.text+0x8c3): In function `Main_main_info': : undefined reference to `NetworkziSocket_withSocketsDo_closure' socketTest.o(.text+0x971): In function `__stginit_Main_': : undefined reference to `__stginit_Network_' socketTest.o(.data+0x20): In function `r2dN_closure': : undefined reference to `Network_PortNumber_static_info' socketTest.o(.rodata+0x10): In function `r2dL_srt': : undefined reference to `NetworkziSocket_zdfNumPortNumber_closure' socketTest.o(.rodata+0x74): In function `s2fJ_srt': : undefined reference to `Network_connectTo_closure' socketTest.o(.rodata+0x9c): In function `Main_main_srt': : undefined reference to `NetworkziSocket_withSocketsDo_closure' collect2: ld returned 1 exit status I'm using ghc-6.4.1 in Gentoo Linux 2005.0 running kernel 2.6.13 and gcc-3.3.6 I don't know if is an error in my code or in ghc. Any sugestions? Thanks for your help -- Jorge E. Guerra D. Departamento de Computación y Tecnología de la Información Universidad Simón Bolívar

2 1

RE: Segfaulting programs with GHC 6.4.1
by Simon Marlow 28 Oct '05

28 Oct '05

On 24 October 2005 17:23, John Goerzen wrote: > 9. Now run dist/build/gopherbot. > > You should see it start to download documents, and crash after a > few minutes. I have it running on my laptop, but I'm behind a firewall right now and I guess it doesn't understand the http_proxy environment variable (is there a way to tell it to use a proxy?). At least, it seems to be complaining about timeouts on its connections. I'll try later outside the firewall. Cheers, Simon

2 1

RE: Default name of target executable
by Simon Marlow 28 Oct '05

28 Oct '05

On 24 October 2005 21:45, Tomasz Zielonka wrote: > On 10/24/05, Tomasz Zielonka <tomasz.zielonka(a)gmail.com> wrote: >> Here is the patch. > > Again, with a small bugfix in docs. Thanks; now committed. Cheers, Simon

1 0

[ ghc-Feature Requests-1340203 ] Debug.Trace.trace should work on Show
by SourceForge.net 27 Oct '05

27 Oct '05

Feature Requests item #1340203, was opened at 2005-10-28 04:38 Message generated for change (Settings changed) made by jch You can respond by visiting: https://sourceforge.net/tracker/?func=detail&atid=358032&aid=1340203&group_… Please note that this message will contain a full copy of the comment thread, including the initial issue submission, for this request, not just the latest update. Category: None Group: None Status: Open >Priority: 2 Submitted By: Juliusz Chroboczek (jch) Assigned to: Nobody/Anonymous (nobody) Summary: Debug.Trace.trace should work on Show Initial Comment: Debug.Trace.trace has type trace :: String -> a -> a which forces me to insert calls to show. Could it be generalised to trace :: Show a => a -> b -> b Thanks, Juliusz Chroboczek ---------------------------------------------------------------------- You can respond by visiting: https://sourceforge.net/tracker/?func=detail&atid=358032&aid=1340203&group_…

1 0