Re: [PATCH, 3 of 5], Add suport for PowerPC IEEE 128-bit floating point

[Michael Meissner <meissner at linux dot vnet dot ibm dot com>]

These patches are the machine independent changes needed to add IEEE 128-bit
floating point to the PowerPC.  The change is to add a new mode macro,
SPECIAL_FLOATING_MODE, that is similar to FRACTIONAL_FLOATING_MODE, except that
the normal widening functions skip SPECIAL_FLOATING_MODE's modes.  In
particular, on the PowerPC, the existing long double functions are faster, than
the IEEE 128-bit emulation functions, so it is desirable that double widens to
long double instead of __float128.

As a consequence of the way I implemented SPECIAL_FLOATING_MODE, some of the
changes involved code that starts at the narrowest type, and goes up by
widening types, until an appropriate mode is found, or we find the mode
itself.  I needed to check whether the wider mode was VOIDmode as well as the
original type.

2014-07-15  Michael Meissner  <meissner@linux.vnet.ibm.com>

	* cse.c (cse_insn): When going through the list of smaller types,
	exit the loop if the type is VOIDmode in addition to the type we
	are looking at, to allow special modes that aren't normally
	considered when widening types.  This is used by the PowerPC, so
	that floating point is widened from double to long double, and
	does not get converted to the slower __float128 (IEEE 128-bit
	floating point).
	* dse.c (find_shift_sequence): Likewise.
	* combine.c (simplify_comparison): Likewise.
	* rtlanal.c (init_num_sign_bit_copies_in_rep): Likewise.

	* machmode.h (GET_MODE_WIDER_MODE_SPECIAL): New macro to get the
	wider types, including special modes that are normally skipped in
	looking for a wider type.

	* machmode.def (SPECIAL_FLOAT_MODE): Document the use of
	SPECIAL_FLOAT_MODE in comments.

	* expr.c (init_expr_target): Use the GET_MODE_WIDER_MODE_SPECIAL
	macro to include special modes in doing the initialization.
	(compress_float_constant): Likewise.

	* genmodes.c (struct mode_data): Add support for
	SPECIAL_FLOAT_MODE that creates floating point modes that are
	normally skipped in widening types, but are available if the user
	explicitly uses the type.
	(complete_mode): Likewise.
	(FLOAT_MODE): Likewise.
	(FRACTIONAL_FLOAT_MODE): Likewise.
	(SPECIAL_FLOAT_MODE): Likewise.
	(FLOAT_MODE_INTERNAL): Likewise.
	(make_float_mode): Likewise.
	(emit_mode_wider): Likewise.

	* emit-rtl.c (init_derived_machine_modes): Use the
	GET_MODE_WIDER_MODE_SPECIAL macro to include special modes in
	doing the initialization.
	(init_emit_once): Likewise.

-- 
Michael Meissner, IBM
IBM, M/S 2506R, 550 King Street, Littleton, MA 01460-6245, USA
email: meissner@linux.vnet.ibm.com, phone: +1 (978) 899-4797

Attachment: gcc-power8.patch115e
Description: Text document