ARM fixed-point support [6/6]: target-specific parts

On Fri, 13 May 2011 17:31:18 +0100
Julian Brown <julian@codesourcery.com> wrote:

> On Fri, 13 May 2011 14:54:47 +0000 (UTC)
> "Joseph S. Myers" <joseph@codesourcery.com> wrote:
> 
> > On Fri, 13 May 2011, Julian Brown wrote:
> > 
> > > although Joseph's comments have (hopefully) all been addressed.
> > > One must now configure with "--enable-fixed-point" to enable the
> > > fixed-point support.
> > 
> > No, that's not how this configure logic works; it enables fixed
> > point by default for the given targets.  The bit that enables it by
> > default for MIPS (and that you're changing to cover ARM as well) is
> > the code that is called if the user *doesn't* pass
> > --enable-fixed-point or --disable-fixed-point.
> 
> Yes, you're quite right: I had forgotten to run autoconf after
> changing configure.ac whilst testing. Sorry for the misinformation!

This is a new version of the patch, which fixes a couple of issues
discovered on our internal branch since the last version was posted.
This version handles "long long" fixed point types in the same way as
32-bit MIPS: basically the sizes of those types are clamped to 64 bits.

This means the testsuite changes in:

http://gcc.gnu.org/ml/gcc-patches/2011-05/msg00969.html

are no longer required.

Another couple of problems have also been fixed, relating to big-endian
mode. One was the use of simplify_gen_subreg with a hard-wired offset in
a several of the insn expanders, where I should have just used
gen_highpart -- which works in either endianness, and makes the intent
clearer anyway.

The other issue is ABI-related. My previous assertion that fixed-point
types (smaller than word size) should be passed to and returned from
functions "like integers" doesn't really work out very well after all
in big-endian mode, so needs re-thinking. The problem is basically that
such fixed-point types are incorrectly padded in big-endian mode, so
the data ends up in a different part of the passed word than expected.

The way chars and shorts is handled is as follows: when those types are
passed to a function, they are promoted to word-size ints, which is a
nop-op: the values can simply be reinterpreted as ints, and that works
fine in either endianness. We might hope to be able to do the same thing
for smaller-than-word-size fixed-point types, but unfortunately we
can't: e.g. a "short fract" or "fract" can't just be reinterpreted as a
"long fract" (the 32-bit fractional type), because the fixed-point is
in a different place for the wider type (and performing conversions
would itself require libcalls at present). It also doesn't work to
promote small fixed-point types to SImode: as well as being
conceptually messy, we hit ICEs very quickly attempting that.

So instead, I think that small fixed-point types should be handled the
same way as small aggregates or small complex-valued types. This means
effectively no change in little-endian mode, but in big-endian mode:

* for register arguments, fixed-point types are passed in the
  most-significant end of the register.

* for stack arguments, fixed-point types are passed in lower memory
  addresses.

* for return values, fixed-point types are again in the
  most-significant end of the return register (r0).

This is quite easy to set up in the backend code, but unfortunately I
hit a snag in libcall expansion in calls.c: padding of arguments and
return values is not properly supported, as it is for regular function
calls. I've implemented that, and will post it as a separate patch.

OK to apply? Tested alongside the rest of the patch series, in both big
& little-endian mode.

Julian

ChangeLog

    gcc/
    * configure.ac (fixed-point): Add ARM support.
    * configure: Regenerate.
    * config/arm/arm.c (arm_fixed_mode_set): New struct.
    (arm_set_fixed_optab_libfunc): New.
    (arm_set_fixed_conv_libfunc): New.
    (arm_init_libfuncs): Initialise fixed-point helper libfuncs with
    ARM-specific names.
    (aapcs_libcall_value): Return sub-word-size fixed-point libcall
    return values in SImode.
    (arm_return_in_msb): Return fixed-point types in the msb.
    (arm_pad_reg_upwards, arm_pad_arg_upwards): Pad fixed-point types
    upwards.
    (arm_scalar_mode_supported_p): Support fixed-point modes.
    (arm_vector_mode_supported_p): Support vector fixed-point modes.
    * config/arm/arm.h (SHORT_FRACT_TYPE_SIZE, FRACT_TYPE_SIZE)
    (LONG_FRACT_TYPE_SIZE, LONG_LONG_FRACT_TYPE_SIZE)
    (SHORT_ACCUM_TYPE_SIZE, ACCUM_TYPE_SIZE, LONG_ACCUM_TYPE_SIZE)
    (LONG_LONG_ACCUM_TYPE_SIZE, MAX_FIXED_MODE_SIZE): Define.
    * config/arm/arm-modes.def (FRACT, UFRACT, ACCUM, UACCUM): Declare
    vector modes.
    * config/arm/predicates.md (sat_shift_operator): New predicate.
    * config/arm/arm-fixed.md: New.
    * config/arm/arm.md: Include arm-fixed.md.
    * config/arm/t-arm (MD_INCLUDES): Add arm-fixed.md.
    * config/arm/fixed-bit-machine.h: New.

    libgcc/
    * config.host (arm*-*-linux*, arm*-*-uclinux*, arm*-*-eabi*)
    (arm*-*-symbianelf*): Add arm/t-fixed-point makefile fragment.
    target-specific or generic fixed_bit_machine_header to decorate the
    names of fixed-point helper functions.
    * config/arm/t-fixed-point: New.

    gcc/testsuite/
    * gcc.target/arm/fixed-point-exec.c: New test.

ARM fixed-point support [6/6]: target-specific parts

Commit Message

Comments

Patch