RMAC has needed a struct for fixups for some time, and now it has one.
All of the credit for pushing (and patching!) in that direction goes to
ggn; blame for the way it's implemented goes to me. There's still room
for improvement; but for now, this should leave us in much better shape.
Now at v1.12.0.
Some of the following changes are ggn's, and some are mine:
- When tokenizing floats we need to store them using a double pointer
- PTR union needed a (double *)
- Major changes to float depositing in eagen0.c
- Reverted the changes in expr.c so at least floats are processed by
expr() and friends
- SYM svalue needs to be 64 bits
- When EQUing a float symbol don't chop off the upper 32 bits from eval
- Added fltpoint.{c,h} in order to properly create IEEE-754 floating
point and Motorola extended numbers
- Fixed float evaluations in evexpr()
- Fixed floating point depositions in direct.c (in d_dc())
- Upped the BSD image limit in object.c to 8MB for crazy people making
6MB Jaguar ROMs (will need a real fix at some point)
- New command line switch -m[cpu] enables you to switch to a different cpu from the command line
- BYTESWAP64 macro fix
- New optimisation flags o6 and o7 turn null branches to NOP and convert clr.l Dx to moveq #0,Dx
- Remove tons of FPU/PMMU codegen functions in mech.c. They are now handled in a more smart way. Condition codes now stored in 68ktab - updated 68ktab to reflect that
The last commit had gone a bit overboard with the 32 vs 64 bit token
changes; this has been rectified. There's still a ways to go with the
floating point code, but this should be stable for now. Version now at
1.10.1.
The float changes will need some going over to ensure that we don't end
up with what we had when pointers were shoved into the token stream
willy-nilly.
Version bump to 1.9.0. Right now the only thing that supports it is
dc.d; now that the infrastructure supports it, anything else that needs
64-bit support (such as FPUs and the like) can be done very easily now.
- Macro'd the FPU code generators.
- Fix some bfxxxx problems including {offset,width} parsing and valid <ea> modes.
- Add symbol checks for <ea> parsing
- Try at fixing k factor for fmove - had to change {} parsing in general so bfxxxx code was also changed.
- Fixed cinvl/p/a instructions. Converted all trapcc functions into macros.
- pmove/pmovefd fixed and verified against devpac.
- pack/unpk implemented and tested
- Fix left hand side of FPU register lists.
- Fix some 020 move parsing errors.
- Change tokenizer strategy to not munch forced .b/.w/.l extensions after constants
- Get rid of probably redundant code in m_move030.
- ploadr/ploadw implemented. Also added extra needed addressing mode to the mask for pmove which is used by the ploads too.
- Some pmove fixes.
- Added cpdbcc instructions.
- pflush/pflusha implemented for 68030/68040.
- Fix warning for \! inside string - we might be evaluating a macro so it's legal.
- Handle (Dn[.w][*scale],label[.l]) ea case.
- Enable scale value in ea to be expression (evaluated) besides constant.
- Fix for ([address[.wl]]). It seems that undefined symbols were not stored properly in eagen0.s (it used aNexpr instead of aNbexpr when storing base displacement).
- Fixes for fmove freg,freg (no size suffix) and ftst freg (no size suffix).
- fmovem now defaults to .x
- Fixed muls.l
ggn deserves most of the credit for this, as my job was going through
and tossing out the stuff that wasn't needed. ;-) There might be some
ELFish things that still need fixing; time, as usual, will tell.
Turns out the tokenizer would not properly tokenize DOTx constructs
unless they were hanging off the end of a symbol. This should fix that
once and for all.
For some reason, there was code in several places that marked fixups/symbols
as belonging to a RISC section when it was clearly not the case. As a result,
it caused serious problems by reversing words in 68K sections just because a
symbol had been seen in a MOVEI # statement in a RISC section. Probably not
the last nasty surprise in this pile of spaghetti. :-/
Basically, expr() was looking at the token following the one it was
looking at and bypassing the longer parse path if it found an EOL token
there. Problem is, some tokens have follow on values and so can be
considered compound tokens. In this case, the EOL token codes to 101,
having a constant with a value of 101 will not evaluate correctly in
this case as the CONST token is a compound token.
The short of this is that making assumptions is BAD! Don't do it! It
WILL come around to bite you in the ass eventually, in the form of
subtle bugs that are difficult to chase down. Assume nothing!
ggn