This is the mail archive of the
binutils@sourceware.org
mailing list for the binutils project.
RE: [PATCHv4] Add support for O32 FPXX ABI
- From: Matthew Fortune <Matthew dot Fortune at imgtec dot com>
- To: Richard Sandiford <rdsandiford at googlemail dot com>
- Cc: "binutils at sourceware dot org" <binutils at sourceware dot org>, "Moore, Catherine" <Catherine_Moore at mentor dot com>, "macro at codesourcery dot com" <macro at codesourcery dot com>, "Joseph Myers (joseph at codesourcery dot com)" <joseph at codesourcery dot com>
- Date: Mon, 28 Jul 2014 12:04:53 +0000
- Subject: RE: [PATCHv4] Add support for O32 FPXX ABI
- Authentication-results: sourceware.org; auth=none
- References: <6D39441BF12EF246A7ABCE6654B0235320EB4042 at LEMAIL01 dot le dot imgtec dot org> <87fvhnma8u dot fsf at talisman dot default>
Richard Sandiford <rdsandiford@googlemail.com> writes:
> Matthew Fortune <Matthew.Fortune@imgtec.com> writes:
> > @@ -258,6 +258,10 @@ struct mips_set_options
> > Changed by .set singlefloat or .set doublefloat, command-line
> options
> > -msingle-float or -mdouble-float. The default is false. */
> > bfd_boolean single_float;
> > +
> > + /* 1 if single-precision operations on odd-numbered registers are
> > + not allowed (even if supported by ISA_HAS_ODD_SINGLE_FPR). */
> > + int nooddspreg;
> > };
>
> I think "oddspreg" would be better, partly for consistency with the
> ISA_HAS_* macro and partly to avoid the double negative in things like:
>
> bfd_boolean oddspreg = (ISA_HAS_ODD_SINGLE_FPR (mips_opts.isa,
> mips_opts.arch)
> || FPR_SIZE == 64)
> && !mips_opts.nooddspreg;
OK. There was a reason this worked better as nooddspreg to start with but
it appears long since gone.
> > - /* Single-precision coprocessor loads and moves are OK too. */
> > + /* Single-precision coprocessor loads and moves are OK for 32-bit
> registers,
> > + otherwise it depends on oddspreg. */
> > if ((insn->pinfo & FP_S)
> > - && (insn->pinfo & (INSN_COPROC_MEMORY_DELAY | INSN_STORE_MEMORY
> > - | INSN_LOAD_COPROC_DELAY | INSN_COPROC_MOVE_DELAY)))
> > + && (insn->pinfo & INSN_FP_32_MOVE))
> > + return FPR_SIZE == 32 || oddspreg;
>
> Why do you need a new flag for this?
There are two reasons for this. The lwxc1 instruction is just an
INSN_LOAD_MEMORY. I could however just add that to the list of flags to
check (and remove INSN_COPROC_MEMORY_DELAY if the LOAD is added)... but
the bigger reason is that the micromips mtc1/mfc1 instructions are not
INSN_COPROC_MOVE_DELAY so there is no flag to detect them. This has never
been a problem before as there has always been 32 single precision
registers for micromips but now that may be restricted to 16 depending on
ABI. I could just attach a new flag to the micromips instructions which
lack any other flag but it seems clearer to attach it to the others too.
The only other option is to add INSN_COPROC_MOVE_DELAY to the mtc1/mfc1
instructions in micromips even though that is something of a lie. What do
you think is cleanest?
I would also like to get rid of all the ctc1/cfc1/cttc1/cftc1 instructions
that allow the use of floating point register names: $f0. The problem with
these is that they don't actually write floating point registers but they
will interact with the oddspreg logic as their operands have type
OP_REG_FP. Anything relying on ctc1 $0, $f[0-31] is probably expecting the
wrong thing to happen anyway. If that's OK I'll do a separate patch?
Thanks,
Matthew