On 24/07/2009 05:04, Ben Lippmeier wrote:
Simon Marlow wrote:
I'd put money on it being due to this in configure.ac:
tr -d '\r'
Thanks, I installed a fresh copy of GNU coreutils which fixed it.
I'm still trying to get the head working on sparc/solaris again though,
it seems the new "foreign import prim" support doesn't know about sparc
register windows.
A disassembly of
libraries/integer-gmp/dist-install/build/cbits/gmp-wrappers.o says:
00000000 <integer_cmm_int2Integerzh>:
0: 9d e3 bf 90 save %sp, -112, %sp -- push register window
4: a0 10 00 11 mov %l1, %l0 -- try to store R1 to %l0
8: 84 06 e0 0c add %i3, 0xc, %g2 -- try to add 0xc to Hp
c: c2 06 60 5c ld [ %i1 + 0x5c ], %g1 -- try to load HpLim into %g1
10: 80 a0 80 01 cmp %g2, %g1 -- compare Hp+0xc with HpLim
This function is being imported with:
foreign import prim "integer_cmm_int2Integerzh" int2Integer#
:: Int# -> (# Int#, ByteArray# #)
However, it's being called with:
0x0034af10 <sOs_info+172>: st %g1, [ %i0 + -4 ]
0x0034af14 <sOs_info+176>: add %i0, -4, %i0
0x0034af18 <sOs_info+180>: call 0x344cb0 <integer_cmm_int2Integerzh>
0x0034af1c <sOs_info+184>: add %i3, -8, %i3
The problem is that the "save" instruction at the start of
integer_cmm_integer2Intzh pushes the SPARC register window down. Before
the save, register %i1 (the BaseReg) contains a pointer to the STG
register table, and %o1 contains the junk value 0x28. After the save,
%i1 now contains 0x28, so the ld [ %i1 + 0x5c ], %g1 instruction fails.
What calling convention do we want for "foreign import prim" on sparc,
considering primops have to access the STG regs? I suppose we could copy
all the STG regs stored in %i0-%i5 to %o0-%i5 before the save. Or maybe
it'd be better make the cmm primop do the copy if it needs to push the
register window?
I'm not sure what has happened exactly. The cmm code in the integer-gmp
package is supposed to be compiled in the same way as it was when it was
in the RTS, which should yield code that doesn't mess with register
windows at all.
The calling convention for a primop is similar to a Haskell function,
except that all the args are in R1-R8 F1-F4 etc. rather than using a
mixture of registers and the stack. These foreign import prim things
are just like ordinary primops, we haven't invented a new calling
convention or anything.
So how is gmp-wrappers.cmm being compiled?
Cheers,
Simon
_______________________________________________
Cvs-ghc mailing list
[email protected]
http://www.haskell.org/mailman/listinfo/cvs-ghc
