Hello all,
I've just pinned down a nasty bug in the new code generator (not my code,
whoo!) that was tickled by my aggressive optimization pass. In the old
code generator (CgUtils.hs), we have the following section for caller saves:
callerSaves :: GlobalReg -> Bool
#ifdef CALLER_SAVES_Base
callerSaves BaseReg = True
#endif
#ifdef CALLER_SAVES_R1
callerSaves (VanillaReg 1 _) = True
#endif
#ifdef CALLER_SAVES_R2
callerSaves (VanillaReg 2 _) = True
#endif
#ifdef CALLER_SAVES_R3
callerSaves (VanillaReg 3 _) = True
#endif
However, the checks for CALLER_SAVES_RX are conspicuously missing from
the new code generator (StgCmmUtils.hs, I believe it was added in
commit 176fa33f17dd78355cc572e006d2ab26898e2c69). I'm pretty sure we
should be saving these registers in the new world order... (and indeed,
validate on the new code generator crashes because my inlining pass is
to aggressive when it comes to things like R3).
But, I kind of suspect this is *not* actually where we want to do these caller
saves. A result of arranging things this early is that the hack in
cmm/CmmOpt.hs which avoids inlining into calls for fear of interfering with the
native calling convention. Since we're moving all of this machinery into
Hoopl-land, this hack will now have to infect loads and loads of other places;
in particular, I can no longer treat unsafe foreign calls as "fat machine
instructions". I cannot safely do a call f(2,R3) on x86-64 since R3 will
get clobbered with 2 before I manage to assign the third argument register (R4).
This sucks, and leads to bugs like the one above.
As the comment in cmm/CmmOpt.hs notes, to fix this we'll have to teach the
backends how to work around caller-saves (this is "more correct" anyway, since
caller saves varies from arch to arch), but I don't think I'd really trust
myself to get it right. In the mean time, "unsafe foreign calls" are
just fat machine instruction is a myth myth myth. Be careful.
Cheers,
Edward
_______________________________________________
Cvs-ghc mailing list
[email protected]
http://www.haskell.org/mailman/listinfo/cvs-ghc
