gcc version to use on sparc/solaris

[Duncan Coutts]

On Fri, 2009-01-02 at 21:06 +1100, Ben Lippmeier wrote:

> I'm running into some huge compile times that I'm hoping someone will 
> have some suggestions about. When compiling Parser.hs the intermediate 
> .hc file is 4MB big, and is taking GCC 4.2.1 more than 2 hours to get 
> through.
> 
> Another attempted compile of GHC 6.8.3 is stuck on 
> libraries/template-haskell/Language/Haskell/TH/Syntax.hs with a 4MB 
> intermediate .hc file.

Here is what I've discovered...

I built four versions of gcc and used them to build ghc-6.8.3. I
selected the last point release of the last four major gcc releases:
gcc-4.0.4
gcc-4.1.2
gcc-4.2.4
gcc-4.3.2

Summary: gcc-4.0.4 or gcc-4.1.2 seems to be the best choice at the
moment for ghc on sparc/solaris.


gcc-4.0.4
=========

This version passes all expected tests on the gcc testsuite. Using it to
build ghc-6.8.3 works fine and it does not take too long.

The ghc testsuite result looks like:

    2128 total tests, which gave rise to
    8040 test cases, of which
       1 caused framework failures
    1561 were skipped

    6130 expected passes
     281 expected failures
       0 unexpected passes
      48 unexpected failures

Unexpected failures:
   1861(normal,optc,hpc,profc,threaded1,threaded2)
   barton-mangler-bug(normal,hpc,threaded1)
   cabal01(normal)
   conc042(threaded2)
   conc043(threaded2)
   conc044(threaded2)
   conc045(threaded2)
   conflicting_flags(normal)
   enum01(hpc)
   ffi017(ghci)
   ffi019(normal,optc,hpc,profc,ghci,threaded1,threaded2)
   ghciprog004(normal)
   hClose002(normal,optc,hpc,profc,ghci,threaded1,threaded2)
   joao-circular(normal,optc,hpc,profc,threaded1,threaded2)
   num012(normal,optc,hpc,profc,ghci,threaded1,threaded2)
   seward-space-leak(ghci)
   testblockalloc(normal,threaded1)

Some fail in fairly simple ways:

cabal01 fails because Solaris tar does not support -C

conflicting_flags fails due to a difference in output:
-ghc: cannot use `-V' with `--print-libdir'
+ghc-6.8.3: cannot use `-V' with `--print-libdir'

hClose002 fails with a different error:
-Left hClose: invalid argument (Bad file descriptor)
+Left hClose: invalid argument (Bad file number)

1861 fails because ghc uses C99 INFINITY macro without using C99 and
Solaris header files are more strict about standards than other systems.
http://hackage.haskell.org/trac/ghc/ticket/2929


Some look a bit more scary:

testblockalloc fails with "out of memory (requested 8388608 bytes)"

num012 produces different output:
--2147483648
+2147483647
looks like overflow, so definitely something wrong.

A number of tests timeout:
seward-space-leak(ghci)
joao-circular(all)
barton-mangler-bug(all)

A few tests get bus errors:
conc042, conc043, conc044, conc045
ghciprog004
ffi019(all)


gcc-4.1.2
=========

This version passes all expected tests on the gcc testsuite. Using it to
build ghc-6.8.3 works fine and it does not take too long.

The ghc testsuite result looks like:

    2128 total tests, which gave rise to
    8040 test cases, of which
       1 caused framework failures
    1561 were skipped

    6125 expected passes
     281 expected failures
       0 unexpected passes
      49 unexpected failures

Unexpected failures:
   1861(normal,optc,hpc,profc,threaded1,threaded2)
   barton-mangler-bug(normal,optc,hpc,threaded1,threaded2)
   cabal01(normal)
   conc042(threaded2)
   conc043(threaded2)
   conc044(threaded2)
   conc045(threaded2)
   conflicting_flags(normal)
   ffi017(ghci)
   ffi019(normal,optc,hpc,profc,ghci,threaded1,threaded2)
   ghciprog004(normal)
   hClose002(normal,optc,hpc,profc,ghci,threaded1,threaded2)
   joao-circular(normal,optc,hpc,profc,threaded1,threaded2)
   num012(normal,optc,hpc,profc,ghci,threaded1,threaded2)
   seward-space-leak(ghci)
   testblockalloc(normal,threaded1)

The only difference is that enum01(hpc) does not fail and
barton-mangler-bug fails in two more ways (optc, threaded2)

gcc-4.2.4
=========

This fails 4 tests on the gcc testsuite.

This is the version that gives us massive compile times for larger .hc
files (8 hours+ for Syntax.hc). I thought that using -fno-unit-at-a-time
would help but ghc is already using that so no luck.

It does eventually build but takes nearly 2 days. No testsuite results
yet. I'm not sure I can be bothered.

gcc-4.3.2
=========

This fails 50 tests on the gcc testsuite.

Does not work with ghc-6.8.3 at all, presumably due to the evil mangler.
This manifests itself as ./configure failing when it tries to compile
and run utils/pwd/pwd:

checking for path to top of build tree... 
/tmp/ghc15055_0/ghc15055_0.hc: In function 'sqD_ret':

/tmp/ghc15055_0/ghc15055_0.hc:95:0:
     note: if this code is reached, the program will abort

[...]
/tmp/ghc15055_0/ghc15055_0.hc: In function '__stginit_Main':

/tmp/ghc15055_0/ghc15055_0.hc:750:0:
     note: if this code is reached, the program will abort

configure: error: cannot determine current directory

Indeed it does abort:
$ utils/pwd/pwd
Illegal Instruction (core dumped)


Duncan