With further poking, I think the new codegen is actually tickling
an existing bug in the native code generator optimizations, since
the cmmz output looks ok:
cSH:
_sQR::I32 = I32[_sRi::I32 + 3]; // CmmAssign
_sQS::I32 = I32[_sRi::I32 + 7]; // CmmAssign
_sQT::I32 = I32[_sRi::I32 + 11]; // CmmAssign
_sQU::I32 = I32[_sRi::I32 + 15]; // CmmAssign
(_sR1::I32) = call "ccall" arg hints: [`signed',
PtrHint,] result hints: []
(_sQR::I32 & (-4))(_sQS::I32, _sQT::I32, _sQU::I32); // CmmUnsafeForeignCall
And the only change is that in the original code generator,
the assignment to _sQR is elided.
_cSn::I32 = I32[R1 + 7];
_cSp::I32 = I32[R1 + 11];
_cSr::I32 = I32[R1 + 15];
(_sR1::I32,) = foreign "ccall"
I32[R1 + 3]((_cSn::I32, `signed'), (_cSp::I32, PtrHint),
(_cSr::I32,))[_unsafe_call_];
I further verified that there was no problem if I used -fvia-C.
On closer inspection, the fact that _sQR is referenced nowhere
in this dump should have raised alarms (I think the register
allocater happened to assign it to the same register as
_sRi, which is why the assembly looked vaguely plausible.)
I'm still not sure where a fix might lie, but if I take another
crack at it tomorrow I will probably figure it out.
Cheers,
Edward
Excerpts from Edward Z. Yang's message of Wed Jan 12 17:10:11 -0500 2011:
> I appear to have tracked down the bug for ffi021: the new
> code generator doesn't appear to clear the tag bit for the
> pointer to heap before:
>
> // outOfLine should follow:
> (_sR1::I32,) = foreign "ccall"
> _sQR::I32((I32[_sRi::I32 + 7], `signed'),
> (I32[_sRi::I32 + 11], PtrHint),
> (I32[_sRi::I32 + 15],))[_unsafe_call_];
> // emitReturn: Sequel: Assign
> ;
>
> (gdb) disas
> Dump of assembler code for function sRi_info:
> => 0x0804aa6c <+0>: mov %esi,%eax
> 0x0804aa6e <+2>: lea 0x0(%ebp),%ecx
> 0x0804aa71 <+5>: cmp 0x54(%ebx),%ecx
> 0x0804aa74 <+8>: jb 0x804aab3 <sRi_info+71>
> 0x0804aa76 <+10>: add $0x4,%ebp
> 0x0804aa79 <+13>: add $0x8,%edi
> 0x0804aa7c <+16>: cmp 0x5c(%ebx),%edi
> 0x0804aa7f <+19>: ja 0x804aaa4 <sRi_info+56>
> 0x0804aa81 <+21>: pushl 0xf(%eax)
> 0x0804aa84 <+24>: pushl 0xb(%eax)
> 0x0804aa87 <+27>: pushl 0x7(%eax)
> 0x0804aa8a <+30>: call *%eax
>
> The pushes to the stack properly untag eax, but then we just
> call the tagged pointer, which seems pretty wrong to me. Here is
> the old C--:
>
> (_sR1::I32,) = foreign "ccall"
> I32[R1 + 3]((_cSc::I32, `signed'), (_cSe::I32, PtrHint),
> (_cSg::I32,))[_unsafe_call_];
>
> Unfortunately, I can't figure out where this +3 is supposed to
> be happening, so I don't have a patch. Some guidance here would
> be appreciated.
>
> Cheers,
> Edward
_______________________________________________
Glasgow-haskell-users mailing list
Glasgow-haskell-users@haskell.org
http://www.haskell.org/mailman/listinfo/glasgow-haskell-users
