[ARM] ISO/IEC TR 18037 fixed-point support

Hi,

This patch implements the ARM target-specific parts required to support
fixed-point operations as specified in the draft ISO/IEC technical
report TR 18037:

http://www.open-std.org/jtc1/sc22/wg14/www/docs/n1275.pdf

Only basic arithmetic operations are implemented (addition, subtraction,
multiplication, on various data types, with saturating and
non-saturating variants where appropriate). Other functionality is
provided by support functions in libgcc (type conversions,
comparisons, and so on). Although more efficient versions of some of
these could be implemented inline, it seems to me they would mostly use
regular integer operations: I'm not at all sure that building that into
each target rather than teaching generic code how to do it is a
sensible way to go.

The patch implements:

  * Addition and subtraction for fract and accum values (short,
    regular, long).

  * Addition and subtraction for word-size vectors of fract and accum
    values (both saturating and non-saturating).  (It's a little unclear
    how useful these are.)

  * Multiplication of fract values: short, regular and long. (There's
    no point in implementing saturating versions of these, AFAICT).

  * Multiplication of accum values: unsigned and signed, saturating and
    non-saturating. Short and regular only, no long.

Those seem to be the variations which can be implemented using
efficient or fairly-efficient instruction sequences. Most of the
fancier ARM multiplication instructions don't seem very useful for
these kinds of operations, and so are still unsupported.

I've not altered the "default" sizes for numbers of fractional &
integral bits for the fixed-point types: they seem to map onto ARM
instructions as well as anything else would.

I've added a single (execution) test case to exercise the new
operations a little. Additionally the generic fixed-point.exp tests
pass with the patch.

The instructions using "APSR_nzcvq" require a recent-enough version of
binutils to include the following patch to assemble emitted code:

http://sourceware.org/ml/binutils/2011-04/msg00021.html

There's currently no support for "long long" fixed-point types, since
those imply support for TImode arithmetic, which isn't sensible on ARM.
I've modified tests to make optional the requirement for long long
support.

A note about ARM EABI compliance: nothing is specified in the EABI for
fixed-point types. Vendor-specific helper functions are supposed to go
in their own namespace, so the patch renames the fixed-point helper
functions from "__foo" to "__gnu_foo", and causes them to be linked
statically with libgcc by not giving them versioned symbols: at some
later date if/when fixed-point helpers are added to the ARM EABI, those
helpers can replace the __gnu_foo ones, be given suitably versioned
symbols, etc.

Non-saturating sub-word-sized fixed-point quantities are never
sign/zero-extended after being operated on, including at ABI
boundaries. This means that bits outside the size of the type are
undefined when passing them to/returning them from functions. Again,
this behaviour may need to be tightened up if/when fixed-point support
gets added to the ARM EABI. Otherwise, fixed-point values are passed and
returned like integers.

Tested with cross to ARM Linux. OK to apply?

Thanks,

Julian

ChangeLog

    gcc/
    * optabs.c (prepare_cmp_insn): Use correct biasing for fixed-point
    comparison helpers.
    * final.c (output_addr_const): Print fixed-point constants as
    decimal not hex (avoiding an assembler overflow warning for
    negative byte constants).
    * tree-ssa-sccvn.c (copy_reference_ops_from_ref): Support FIXED_CST.
    * config/arm/arm.c (arm_fixed_point_supported_p): New.
    (TARGET_FIXED_POINT_SUPPORTED_P): Define target macro using above.
    (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.
    (arm_scalar_mode_supported_p): Support fixed-point modes.
    (arm_vector_mode_supported_p): Support vector fixed-point modes.
    * config/arm/arm-modes.def (FRACT, UFRACT, ACCUM, UACCUM): Declare
    vector modes.
    * predicates.md (sat_shift_operator): New predicate.
    * config/arm/arm-fixed.md: New.
    * config/arm/arm.md: Include arm-fixed.md.
    * config/arm/fixed-bit-machine.h: New.
    * config/fixed-bit.h (fixed-bit-machine.h): Include.
    (DECORATE_FIXED_NAME): Conditionally define if undefined. Use
    throughout file to decorate function names.
    * config/fixed-bit-mach-generic.h: New.

    libgcc/
    * config.host (arm*-*-linux*, arm*-*-uclinux*, arm*-*-eabi*)
    (arm*-*-nucleuseabi*, arm*-*-symbianelf*): Add arm/t-fixed-point
    makefile fragment.
    * Makefile.in (FIXED_BIT_MACHINE): New rule. Include
    target-specific or generic fixed_bit_machine_header to decorate the
    names of fixed-point helper functions.
    * config/arm/t-fixed-point: New.
    * fixed-obj.mk ($(o)_s$(objext)): Depend on $(FIXED_BIT_MACHINE).

    gcc/testsuite/
    * gcc.target/arm/fixed-point-exec.c: New test.
    * gcc.dg/fixed-point/func-scalar.c: Include longlong-support.h.
    Avoid long long tests on targets which don't support them.
    * gcc.dg/fixed-point/unary.c: Likewise.
    * gcc.dg/fixed-point/muldiv-warning.c: Likewise.
    * gcc.dg/fixed-point/operator-logical.c: Likewise.
    * gcc.dg/fixed-point/func-array.c: Likewise.
    * gcc.dg/fixed-point/func-struct.c: Likewise.
    * gcc.dg/fixed-point/func-vararg-mixed.c: Likewise.
    * gcc.dg/fixed-point/binary.c: Likewise.
    * gcc.dg/fixed-point/comma.c: Likewise.
    * gcc.dg/fixed-point/func-mixed.c: Likewise.
    * gcc.dg/fixed-point/func-vararg.c: Likewise.
    * gcc.dg/fixed-point/operator-cond.c: Likewise.
    * gcc.dg/fixed-point/loop-index.c: Likewise.
    * gcc.dg/fixed-point/convert.c: Likewise.
    * gcc.dg/fixed-point/allop-const.c: Likewise.
    * gcc.dg/fixed-point/call-by-value.c: Likewise.
    * gcc.dg/fixed-point/longlong-support.h: New.

[ARM] ISO/IEC TR 18037 fixed-point support

Commit Message

Comments

Patch