Resending this mail cause the damn gmail fscked up my previous
one...


Hi again,

After three days of puking blood while wandering inside the GHC
build system, I think I've found a possible solution to this messy
problem...

First, the "unknown symbol `__guard_local'" error is an entirely
different beast than the random crashes at runtime with libgmp.a.

Good news is kili's "SymE_NeedsProto" workaround effectively fixes
it.

On the other hand, the root cause of segmentation faults at runtime
is that ghc requires a patched gmp library with customized memory
allocation functions.

Quoting from
http://hackage.haskell.org/trac/ghc/wiki/ReplacingGMPNotes/TheCurrentGMPImplementation

---8<---

   The most insidious and unique feature of the GHC implementation with
   GMP is memory management. GHC uses the special memory functions in GMP
   ( mp_set_memory_functions( (*alloc)(), (*realloc)(), (*dealloc)() ) to
   let GMP use GHC's own garbage collector. The memory functions are
   implemented in rts/sm/Storage.c as:
          static void* stgAllocForGMP (size_t size_in_bytes);
          static void  stgDeallocForGMP (void *ptr STG_UNUSED,
                                         size_t size STG_UNUSED);
          static void* stgReallocForGMP (void *ptr,
                                         size_t old_size,
                                         size_t new_size);

   These special allocation functions bring most GMP memory usage into the
   GHC heap but do not seem that efficient otherwise. (I could be wrong
   --PDT.) Allocation uses the internal allocate() interface, so no
   garbage collection is performed during a GMP operation. Note that GMP
   operations may use these functions to allocate more memory themselves,
   after being called from Haskell code. The memory allocated is a simple
   array of words (W_), rounded up to a whole number.

---8<---

And indeed:

$ nm 
/usr/ports/pobj/ghc-7.4.2-no_doc/ghc-7.4.2/libraries/integer-gmp/gmp/libgmp.a | 
grep stg
         U stgAllocForGMP
         U stgDeallocForGMP
         U stgReallocForGMP

But since /usr/local/lib/libgmp.a comes from a vanilla gmp-5.0.2,
we're totally belly-up!

As a proof, try this:

$ cd /usr/ports/devel/gmp
$ make clean patch
$ cd /usr/ports/pobj/gmp-5.0.2/gmp-5.0.2/
$ patch -NEp1 -i 
/usr/ports/pobj/ghc-7.4.2-no_doc/ghc-7.4.2/libraries/integer-gmp/gmp/tarball/patch
$ cd -
$ make fake
$ sudo cp /usr/ports/pobj/gmp-5.0.2/fake-amd64/usr/local/lib/libgmp.a 
/usr/local/lib/libgmp.a

and then repeat the fibonacci test:

$ ghci
GHCi, version 7.4.2: http://www.haskell.org/ghc/  :? for help
Loading package ghc-prim ... linking ... done.
Loading package integer-gmp ... linking ... done.
Loading package base ... linking ... done.
Prelude> let fibs = 0 : 1 : zipWith (+) fibs (tail fibs)
Prelude> let fib n = fibs!!n
Prelude> fib 10000
33644764876431783266621612005107543310302148460680063906564769974680081442166662368155595513633734025582065332680836159373734790483865268263040892463056431887354544369559827491606602099884183933864652731300088830269235673613135117579297437854413752130520504347701602264758318906527890855154366159582987279682987510631200575428783453215515103870818298969791613127856265033195487140214287532698187962046936097879900350962302291026368131493195275630227837628441540360584402572114334961180023091208287046088923962328835461505776583271252546093591128203925285393434620904245248929403901706233888991085841065183173360437470737908552631764325733993712871937587746897479926305837065742830161637408969178426378624212835258112820516370298089332099905707920064367426202389783111470054074998459250360633560933883831923386783056136435351892133279732908133732642652633989763922723407882928177953580570993691049175470808931841056146322338217465637321248226383092103297701648054726243842374862411453093812206564914032751086643394517512161526545361333111314042436854805106765843493523836959653428071768775328348234345557366719731392746273629108210679280784718035329131176778924659089938635459327894523777674406192240337638674004021330343297496902028328145933418826817683893072003634795623117103101291953169794607632737589253530772552375943788434504067715555779056450443016640119462580972216729758615026968443146952034614932291105970676243268515992834709891284706740862008587135016260312071903172086094081298321581077282076353186624611278245537208532365305775956430072517744315051539600905168603220349163222640885248852433158051534849622434848299380905070483482449327453732624567755879089187190803662058009594743150052402532709746995318770724376825907419939632265984147498193609285223945039707165443156421328157688908058783183404917434556270520223564846495196112460268313970975069382648706613264507665074611512677522748621598642530711298441182622661057163515069260029861704945425047491378115154139941550671256271197133252763631939606902895650288268608362241082050562430701794976171121233066073310059947366875
Prelude>
Leaving GHCi.
$

How to properly fix this problem (ports-wise speaking)?

A working solution is to use the gmp library that comes bundled
with GHC.

While doing a build with my diff, for mysterious causes I caught
another error:

---8<---

"inplace/bin/ghc-stage1" -o utils/hsc2hs/dist-install/build/tmp/hsc2hs   -H32m 
-O    -hide-all-packages -i -iutils/hsc2hs/. -iutils/hsc2hs/dist-install/build 
-iutils/hsc2hs/dist-install/build/autogen -Iutils/hsc2hs/dist-install/build 
-Iutils/hsc2hs/dist-install/build/autogen     -optP-include 
-optPutils/hsc2hs/dist-install/build/autogen/cabal_macros.h -package 
base-4.5.1.0 -package containers-0.4.2.1 -package directory-1.1.0.2 -package 
process-1.1.0.1  -XHaskell98 -XCPP -XForeignFunctionInterface  
-no-user-package-conf -rtsopts     -odir utils/hsc2hs/dist-install/build -hidir 
utils/hsc2hs/dist-install/build -stubdir utils/hsc2hs/dist-install/build -hisuf 
hi -osuf  o -hcsuf hc   utils/hsc2hs/dist-install/build/Main.o 
utils/hsc2hs/dist-install/build/HSCParser.o 
utils/hsc2hs/dist-install/build/DirectCodegen.o 
utils/hsc2hs/dist-install/build/CrossCodegen.o 
utils/hsc2hs/dist-install/build/UtilsCodegen.o 
utils/hsc2hs/dist-install/build/Common.o utils/hsc2hs/dist-install/build/C.o 
utils/hsc2hs/dist-install/build/Flags.o 
utils/hsc2hs/dist-install/build/Paths_hsc2hs.o
/usr/ports/pobj/ghc-7.4.2-no_doc/ghc-7.4.2/rts/dist/build/libHSrts.a(RtsFlags.o):
 In function `copyArg':
RtsFlags.c:(.text+0x2bd): warning: warning: strcpy() is almost always misused, 
please use strlcpy()
/usr/ports/pobj/ghc-7.4.2-no_doc/ghc-7.4.2/rts/dist/build/libHSrts.a(RtsUtils.o):
 In function `showStgWord64':
RtsUtils.c:(.text+0x329): warning: warning: sprintf() is often misused, please 
use snprintf()
/usr/ports/pobj/ghc-7.4.2-no_doc/ghc-7.4.2/rts/dist/build/libHSrts.a(ffi64.o): 
In function `ffi_call':

../src/x86/ffi64.c:486:0:  undefined reference to `ffi_call_unix64'
/usr/ports/pobj/ghc-7.4.2-no_doc/ghc-7.4.2/rts/dist/build/libHSrts.a(ffi64.o): 
In function `ffi_prep_closure_loc':

../src/x86/ffi64.c:512:0:
     undefined reference to `ffi_closure_unix64'

---8<---

So I fixed this too...

Here's the diff (attached), comments are highly appreciated.

Ciao,
David

Attachment: ghc-fix.diff.gz
Description: Binary data

Reply via email to