object splitting is not done under sparc solaris with gcc-4.2.2

[maeder]

I've used gcc-4.2.2 and ghc-6.8.1 to build ghc-6.8.2. For some reason object splitting is not done although for ghc-6.8.1 it worked fine.

-bash-3.00$ ll hello
-rwxr-xr-x   1 maeder   wimi     3907865 May 23 11:38 hello

-bash-3.00$ ghc --info
 [("Project name","The Glorious Glasgow Haskell Compilation System")
 ,("Project version","6.8.2")
 ,("Booter version","6.8.1")
 ,("Stage","2")
 ,("Interface file version","6")
 ,("Have interpreter","YES")
 ,("Object splitting","YES")
 ,("Have native code generator","NO")
 ,("Support SMP","YES")
 ,("Unregisterised","NO")
 ,("Tables next to code","YES")
 ,("Win32 DLLs","")
 ,("RTS ways"," debug  thr thr_p thr_debug")
 ,("Leading underscore","NO")
 ]

[maeder]

despite -split-objs the output of ar -t libraries/base/dist/build/libHSbase-3.0.1.0.a is

PrelIOUtils.o
WCsubst.o
Win32Utils.o
consUtils.o
dirUtils.o
inputReady.o
longlong.o
selectUtils.o
Concurrent_stub.o
Generics__1.o
Aliases__1.o
Basics__1.o
Instances__1.o
Schemes__1.o
Text__1.o
Twins__1.o
Concurrent__1.o
Arr__1.o
Base__1.o
Conc__1.o
ConsoleHandler__1.o
Dotnet__1.o
Enum__1.o
Environment__1.o
Err__1.o
Exception__1.o
Exts__1.o
Float__1.o
ForeignPtr__1.o
Handle__1.o
IO__1.o
IOBase__1.o
Int__1.o
List__1.o
Num__1.o
PArr__1.o
Pack__1.o
PrimopWrappers__1.o
Ptr__1.o
Read__1.o
Real__1.o
ST__1.o
STRef__1.o
Show__1.o
Stable__1.o
Storable__1.o
TopHandler__1.o
Unicode__1.o
Weak__1.o
Word__1.o
Timeout__1.o
Applicative__1.o
Arrow__1.o
Concurrent__1.o
Chan__1.o
MVar__1.o
QSem__1.o
QSemN__1.o
SampleVar__1.o
Exception__1.o
Monad__1.o
Fix__1.o
Instances__1.o
ST__1.o
Lazy__1.o
Strict__1.o
Bits__1.o
Bool__1.o
Char__1.o
Complex__1.o
Dynamic__1.o
Either__1.o
Eq__1.o
Fixed__1.o
Foldable__1.o
Function__1.o
HashTable__1.o
IORef__1.o
Int__1.o
Ix__1.o
List__1.o
Maybe__1.o
Monoid__1.o
Ord__1.o
Ratio__1.o
STRef__1.o
Lazy__1.o
Strict__1.o
String__1.o
Traversable__1.o
Tuple__1.o
Typeable__1.o
Unique__1.o
Version__1.o
Word__1.o
Trace__1.o
Foreign__1.o
C__1.o
Error__1.o
String__1.o
Types__1.o
ForeignPtr__1.o
Marshal__1.o
Alloc__1.o
Array__1.o
Error__1.o
Pool__1.o
Utils__1.o
Ptr__1.o
StablePtr__1.o
Storable__1.o
Numeric__1.o
Prelude__1.o
GetOpt__1.o
CPUTime__1.o
Environment__1.o
Exit__1.o
IO__1.o
Error__1.o
Unsafe__1.o
Info__1.o
Mem__1.o
StableName__1.o
Weak__1.o
Internals__1.o
Types__1.o
ReadP__1.o
ReadPrec__1.o
Printf__1.o
Read__1.o
Lex__1.o
Show__1.o
Functions__1.o
Coerce__1.o

[simonmar]

What build settings are you using, if any?

[maeder]

I did not explicitly pass -fvia-C (or switched off native code generation). my mk/build.mk contains:

BIN_DIST=1
Project=Ghc
INSTALL_PROGRAM = $(INSTALL) -s -m 755
XMLDocWays = html
HADDOCK_DOCS = YES
SRC_HC_OPTS += -optc-mcpu=ultrasparc -opta-mcpu=ultrasparc
SplitObjs = YES

where SplitObjs = YES was omitted initially (without difference).

[maeder]

I may have built the old ghc-6.8.1 with on older gcc version (maybe gcc-4.0.3)

[maeder]

temp files created by: ghc --make -split-objs -keep-tmp-files HughesPJ.hs

[maeder]

object splitting works with gcc-3.4.4 (I'll attach tmp files). I've called

ghc --make -no-recomp -split-objs -optc-mcpu=ultrasparc -opta-mcpu=ultrasparc -keep-tmp-files HughesPJ.hs

Without the ultrasparc opts gcc-3.4.4 complained about Architecture mismatch

(Requires v9|v9a|v9b; requested architecture is v8.)

[maeder]

successful object splitting

[maeder]

This is also still a problem for the release candidate ghc-6.8.2.20080603

[igloo]

If you look at the .raw_s files you can see that gcc 4.2.2 has floated all the __stg_split_marker's to the top. On other platforms -fno-unit-at-a-time or -fno-toplevel-reorder have fixed this; can you find a flag to fix it for you?

For experimenting, you can compile something with ghc -v -keep-tmp-files and then rerun the gcc command that ghc runs with different flags.

[maeder]

Right, -fno-toplevel-reorder alone does the trick!

cc1: error: unrecognized command line option "-fno-unit-at-atime"

[igloo]

OK, great, should work in 6.8 branch and HEAD now, then.