[PATCH, i386]: Review FLOAT_MODE_P macro for i386 target

["Uros Bizjak" <ubizjak at gmail dot com>]

Hello!

This is the second part of FLOAT_MODE_P patch. This patch implements following:

- check TARGET_80387 in X87_FLOAT_MODE_P, this removes a bunch of
TARGET_80387 checks from i386.md and i386.c
- cleanups function_value_32() w.r.t. floating point return values.
Now gcc correctly generates scalar return value in xmm0 for "-msse2
-mregparm=sse -mno-80387".
- uses SCALAR_FLOAT_MODE_P macro in ix86_expand_* functions. This is
due to the fact that X87_FLOAT_MODE_P now disables all modes for
!TARGET_80387 but we still want to expand for SSE float modes.
However, correct FP modes are already filtered out by expanders and
insn patterns in i386.md. An assert checks that no decimal FP sneaks
in.

- target RTX costs. We don't model any SSE FP operation (either scalar
or vector) cost value and  insn RTX costs are computed based on x87
costs. As a RFC, x87/SSE/vecSSE modes are separated in this patch, but
are just pointing to x87 costs.  We need new fields in target cost
structure to cover SSE and vecSSE instructions.

Patch is bootstrapped on i686-pc-linux-gnu and regression tested for
all defaul languages. If there are no comments, I plan to commit this
patch to mainline in a couple of days.

2007-04-11  Uros Bizjak  <ubizjak@gmail.com>

	* config/i386/i386.h (X87_FLOAT_MODE_P): Check for TARGET_80387.
	* config/i386/i386.md (*cmpfp0): Remove check for TARGET_80387, this
	check is now implied in X87_FLOAT_MODE_P.
	(*cmpfp_u, *cmpfp_<mode>, *cmpfp_i_i387): Ditto.
	(*cmpfp_iu_387, fix_trunc<mode>_fisttp_i386_1): Ditto.
	(fix_trunc<mode>_i386_fisttp): Ditto.
	(fix_trunc<mode>_i387_fisttp_with_temp): Ditto.
	(*fix_trunc<mode>_i387_1, fix_truncdi_i387): Ditto.
	(fix_truncdi_i387_with_temp, fix_trunc<mode>_i387): Ditto.
	(fix_trunc<mode>_i387_with_temp, *fp_jcc_1_387): Ditto.
	(*fp_jcc_2_387, *fp_jcc_5_387, *fp_jcc_6_387): Ditto.
	(*fp_jcc_7_387, *fp_jcc_8<mode>_387): Ditto.
	(unnamed_splitters): Ditto.
	* config/i386/i386.c (function_value_32): Generate FIRST_FLOAT_REG
	for X87_FLOAT_MODE_P mode.  Override FIRST_FLOAT_REG with
	FIRST_SSE_REG for local functions when SSE math is enabled or
	for functions with sseregparm attribute.
	(standard_80387_constant_p): Return -1 if mode is not
	X87_FLOAT_MODE_P.
	(ix86_cc_mode): Assert that scalar mode is not DECIMAL_FLOAT_MODE_P.
	(ix86_expand_compare): Ditto.
	(ix86_expand_carry_flag_compare): Ditto.
	(ix86_expand_int_movcc): Check for SCALAR_FLOAT_MODE_P instead
	of FLOAT_MODE_P for cmp_mode and assert that cmp_mode is not
	DECIMAL_FLOAT_MODE_P.
	(ix86_preferred_output_reload_class): Use X87_FLOAT_MODE_P instead
	of SCALAR_FLOAT_MODE_P.
	(ix86_rtx_costs) [PLUS] Remove FLOAT_MODE_P and fall through to ...
	[MINUS]: ... here.  Add SSE_FLOAT_MODE_P and X87_FLOAT_MODE_P
	checks before FLOAT_MODE_P.
	[MULT]: Add SSE_FLOAT_MODE_P and X87_FLOAT_MODE_P checks
	before FLOAT_MODE_P.
	[DIV]: Ditto.
	[NEG]: Ditto.
	[ABS]: Ditto.
	[SQRT]: Ditto.
	[FLOAT_EXTEND]: Use SSE_FLOAT_MODE_P.

Uros.

Attachment: i386-fp-mode.diff
Description: Binary data