fp_util.s

Linux内核源代码 为压缩文件 是<<Linux内核>>一书中的源代码

/*
 * fp_util.S
 *
 * Copyright Roman Zippel, 1997.  All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 * 1. Redistributions of source code must retain the above copyright
 *    notice, and the entire permission notice in its entirety,
 *    including the disclaimer of warranties.
 * 2. Redistributions in binary form must reproduce the above copyright
 *    notice, this list of conditions and the following disclaimer in the
 *    documentation and/or other materials provided with the distribution.
 * 3. The name of the author may not be used to endorse or promote
 *    products derived from this software without specific prior
 *    written permission.
 *
 * ALTERNATIVELY, this product may be distributed under the terms of
 * the GNU Public License, in which case the provisions of the GPL are
 * required INSTEAD OF the above restrictions.  (This clause is
 * necessary due to a potential bad interaction between the GPL and
 * the restrictions contained in a BSD-style copyright.)
 *
 * THIS SOFTWARE IS PROVIDED ``AS IS'' AND ANY EXPRESS OR IMPLIED
 * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
 * DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT,
 * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
 * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
 * OF THE POSSIBILITY OF SUCH DAMAGE.
 */

#include <linux/config.h>
#include "fp_emu.h"

/*
 * Here are lots of conversion and normalization functions mainly
 * used by fp_scan.S
 * Note that these functions are optimized for "normal" numbers,
 * these are handled first and exit as fast as possible, this is
 * especially important for fp_normalize_ext/fp_conv_ext2ext, as
 * it's called very often.
 * The register usage is optimized for fp_scan.S and which register
 * is currently at that time unused, be careful if you want change
 * something here. %d0 and %d1 is always usable, sometimes %d2 (or
 * only the lower half) most function have to return the %a0
 * unmodified, so that the caller can immediatly reuse it.
 */

	.globl	fp_ill, fp_end

	| exits from fp_scan:
	| illegal instruction
fp_ill:
	printf	,"fp_illegal\n"
	rts
	| completed instruction
fp_end:
	tst.l	(TASK_MM-8,%a2)
	jmi	1f
	tst.l	(TASK_MM-4,%a2)
	jmi	1f
	tst.l	(TASK_MM,%a2)
	jpl	2f
1:	printf	,"oops:%p,%p,%p\n",3,%a2@(TASK_MM-8),%a2@(TASK_MM-4),%a2@(TASK_MM)
2:	clr.l	%d0
	rts

	.globl	fp_conv_long2ext, fp_conv_single2ext
	.globl	fp_conv_double2ext, fp_conv_ext2ext
	.globl	fp_normalize_ext, fp_normalize_double
	.globl	fp_normalize_single, fp_normalize_single_fast
	.globl	fp_conv_ext2double, fp_conv_ext2single
	.globl	fp_conv_ext2long, fp_conv_ext2short
	.globl	fp_conv_ext2byte
	.globl	fp_finalrounding_single, fp_finalrounding_single_fast
	.globl	fp_finalrounding_double
	.globl	fp_finalrounding, fp_finaltest, fp_final

/*
 * First several conversion functions from a source operand
 * into the extended format. Note, that only fp_conv_ext2ext
 * normalizes the number and is always called after the other
 * conversion functions, which only move the information into
 * fp_ext structure.
 */

	| fp_conv_long2ext:
	|
	| args:	%d0 = source (32-bit long)
	|	%a0 = destination (ptr to struct fp_ext)

fp_conv_long2ext:
	printf	PCONV,"l2e: %p -> %p(",2,%d0,%a0
	clr.l	%d1			| sign defaults to zero
	tst.l	%d0
	jeq	fp_l2e_zero		| is source zero?
	jpl	1f			| positive?
	moveq	#1,%d1
	neg.l	%d0
1:	swap	%d1
	move.w	#0x3fff+31,%d1
	move.l	%d1,(%a0)+		| set sign / exp
	move.l	%d0,(%a0)+		| set mantissa
	clr.l	(%a0)
	subq.l	#8,%a0			| restore %a0
	printx	PCONV,%a0@
	printf	PCONV,")\n"
	rts
	| source is zero
fp_l2e_zero:
	clr.l	(%a0)+
	clr.l	(%a0)+
	clr.l	(%a0)
	subq.l	#8,%a0
	printx	PCONV,%a0@
	printf	PCONV,")\n"
	rts

	| fp_conv_single2ext
	| args:	%d0 = source (single-precision fp value)
	|	%a0 = dest (struct fp_ext *)

fp_conv_single2ext:
	printf	PCONV,"s2e: %p -> %p(",2,%d0,%a0
	move.l	%d0,%d1
	lsl.l	#8,%d0			| shift mantissa
	lsr.l	#8,%d1			| exponent / sign
	lsr.l	#7,%d1
	lsr.w	#8,%d1
	jeq	fp_s2e_small		| zero / denormal?
	cmp.w	#0xff,%d1		| NaN / Inf?
	jeq	fp_s2e_large
	bset	#31,%d0			| set explizit bit
	add.w	#0x3fff-0x7f,%d1	| re-bias the exponent.
9:	move.l	%d1,(%a0)+		| fp_ext.sign, fp_ext.exp
	move.l	%d0,(%a0)+		| high lword of fp_ext.mant
	clr.l	(%a0)			| low lword = 0
	subq.l	#8,%a0
	printx	PCONV,%a0@
	printf	PCONV,")\n"
	rts
	| zeros and denormalized
fp_s2e_small:
	| exponent is zero, so explizit bit is already zero too
	tst.l	%d0
	jeq	9b
	move.w	#0x4000-0x7f,%d1
	jra	9b
	| infinities and NAN
fp_s2e_large:
	bclr	#31,%d0			| clear explizit bit
	move.w	#0x7fff,%d1
	jra	9b

fp_conv_double2ext:
#ifdef FPU_EMU_DEBUG
	getuser.l %a1@(0),%d0,fp_err_ua2,%a1
	getuser.l %a1@(4),%d1,fp_err_ua2,%a1
	printf	PCONV,"d2e: %p%p -> %p(",3,%d0,%d1,%a0
#endif
	getuser.l (%a1)+,%d0,fp_err_ua2,%a1
	move.l	%d0,%d1
	lsl.l	#8,%d0			| shift high mantissa
	lsl.l	#3,%d0
	lsr.l	#8,%d1			| exponent / sign
	lsr.l	#7,%d1
	lsr.w	#5,%d1
	jeq	fp_d2e_small		| zero / denormal?
	cmp.w	#0x7ff,%d1		| NaN / Inf?
	jeq	fp_d2e_large
	bset	#31,%d0			| set explizit bit
	add.w	#0x3fff-0x3ff,%d1	| re-bias the exponent.
9:	move.l	%d1,(%a0)+		| fp_ext.sign, fp_ext.exp
	move.l	%d0,(%a0)+
	getuser.l (%a1)+,%d0,fp_err_ua2,%a1
	move.l	%d0,%d1
	lsl.l	#8,%d0
	lsl.l	#3,%d0
	move.l	%d0,(%a0)
	moveq	#21,%d0
	lsr.l	%d0,%d1
	or.l	%d1,-(%a0)
	subq.l	#4,%a0
	printx	PCONV,%a0@
	printf	PCONV,")\n"
	rts
	| zeros and denormalized
fp_d2e_small:
	| exponent is zero, so explizit bit is already zero too
	tst.l	%d0
	jeq	9b
	move.w	#0x4000-0x3ff,%d1
	jra	9b
	| infinities and NAN
fp_d2e_large:
	bclr	#31,%d0			| clear explizit bit
	move.w	#0x7fff,%d1
	jra	9b

	| fp_conv_ext2ext:
	| originally used to get longdouble from userspace, now it's
	| called before arithmetic operations to make sure the number
	| is normalized [maybe rename it?].
	| args:	%a0 = dest (struct fp_ext *)
	| returns 0 in %d0 for a NaN, otherwise 1

fp_conv_ext2ext:
	printf	PCONV,"e2e: %p(",1,%a0
	printx	PCONV,%a0@
	printf	PCONV,"), "
	move.l	(%a0)+,%d0
	cmp.w	#0x7fff,%d0		| Inf / NaN?
	jeq	fp_e2e_large
	move.l	(%a0),%d0
	jpl	fp_e2e_small		| zero / denorm?
	| The high bit is set, so normalization is irrelevant.
fp_e2e_checkround:
	subq.l	#4,%a0
#ifdef CONFIG_M68KFPU_EMU_EXTRAPREC
	move.b	(%a0),%d0
	jne	fp_e2e_round
#endif
	printf	PCONV,"%p(",1,%a0
	printx	PCONV,%a0@
	printf	PCONV,")\n"
	moveq	#1,%d0
	rts
#ifdef CONFIG_M68KFPU_EMU_EXTRAPREC
fp_e2e_round:
	fp_set_sr FPSR_EXC_INEX2
	clr.b	(%a0)
	move.w	(FPD_RND,FPDATA),%d2
	jne	fp_e2e_roundother	| %d2 == 0, round to nearest
	tst.b	%d0			| test guard bit
	jpl	9f			| zero is closer
	btst	#0,(11,%a0)		| test lsb bit
	jne	fp_e2e_doroundup	| round to infinity
	lsl.b	#1,%d0			| check low bits
	jeq	9f			| round to zero
fp_e2e_doroundup:
	addq.l	#1,(8,%a0)
	jcc	9f
	addq.l	#1,(4,%a0)
	jcc	9f
	move.w	#0x8000,(4,%a0)
	addq.w	#1,(2,%a0)
9:	printf	PNORM,"%p(",1,%a0
	printx	PNORM,%a0@
	printf	PNORM,")\n"
	rts
fp_e2e_roundother:
	subq.w	#2,%d2
	jcs	9b			| %d2 < 2, round to zero
	jhi	1f			| %d2 > 2, round to +infinity
	tst.b	(1,%a0)			| to -inf
	jne	fp_e2e_doroundup	| negative, round to infinity
	jra	9b			| positive, round to zero
1:	tst.b	(1,%a0)			| to +inf
	jeq	fp_e2e_doroundup	| positive, round to infinity
	jra	9b			| negative, round to zero
#endif
	| zeros and subnormals:
	| try to normalize these anyway.
fp_e2e_small:
	jne	fp_e2e_small1		| high lword zero?
	move.l	(4,%a0),%d0
	jne	fp_e2e_small2
#ifdef CONFIG_M68KFPU_EMU_EXTRAPREC
	clr.l	%d0
	move.b	(-4,%a0),%d0
	jne	fp_e2e_small3
#endif
	| Genuine zero.
	clr.w	-(%a0)
	subq.l	#2,%a0
	printf	PNORM,"%p(",1,%a0
	printx	PNORM,%a0@
	printf	PNORM,")\n"
	moveq	#1,%d0
	rts
	| definitely subnormal, need to shift all 64 bits
fp_e2e_small1:
	bfffo	%d0{#0,#32},%d1
	move.w	-(%a0),%d2
	sub.w	%d1,%d2
	jcc	1f
	| Pathologically small, denormalize.
	add.w	%d2,%d1
	clr.w	%d2
1:	move.w	%d2,(%a0)+
	move.w	%d1,%d2
	jeq	fp_e2e_checkround
	| fancy 64-bit double-shift begins here
	lsl.l	%d2,%d0
	move.l	%d0,(%a0)+
	move.l	(%a0),%d0
	move.l	%d0,%d1
	lsl.l	%d2,%d0
	move.l	%d0,(%a0)
	neg.w	%d2
	and.w	#0x1f,%d2
	lsr.l	%d2,%d1
	or.l	%d1,-(%a0)
#ifdef CONFIG_M68KFPU_EMU_EXTRAPREC
fp_e2e_extra1:
	clr.l	%d0
	move.b	(-4,%a0),%d0
	neg.w	%d2
	add.w	#24,%d2
	jcc	1f
	clr.b	(-4,%a0)
	lsl.l	%d2,%d0
	or.l	%d0,(4,%a0)
	jra	fp_e2e_checkround
1:	addq.w	#8,%d2
	lsl.l	%d2,%d0
	move.b	%d0,(-4,%a0)
	lsr.l	#8,%d0
	or.l	%d0,(4,%a0)
#endif
	jra	fp_e2e_checkround
	| pathologically small subnormal
fp_e2e_small2:
	bfffo	%d0{#0,#32},%d1
	add.w	#32,%d1
	move.w	-(%a0),%d2
	sub.w	%d1,%d2
	jcc	1f
	| Beyond pathologically small, denormalize.
	add.w	%d2,%d1
	clr.w	%d2
1:	move.w	%d2,(%a0)+
	ext.l	%d1
	jeq	fp_e2e_checkround
	clr.l	(4,%a0)
	sub.w	#32,%d2
	jcs	1f
	lsl.l	%d1,%d0			| lower lword needs only to be shifted
	move.l	%d0,(%a0)		| into the higher lword
#ifdef CONFIG_M68KFPU_EMU_EXTRAPREC
	clr.l	%d0
	move.b	(-4,%a0),%d0
	clr.b	(-4,%a0)
	neg.w	%d1
	add.w	#32,%d1
	bfins	%d0,(%a0){%d1,#8}
#endif
	jra	fp_e2e_checkround
1:	neg.w	%d1			| lower lword is splitted between
	bfins	%d0,(%a0){%d1,#32}	| higher and lower lword
#ifndef CONFIG_M68KFPU_EMU_EXTRAPREC
	jra	fp_e2e_checkround
#else
	move.w	%d1,%d2
	jra	fp_e2e_extra1
	| These are extremely small numbers, that will mostly end up as zero
	| anyway, so this is only important for correct rounding.
fp_e2e_small3:
	bfffo	%d0{#24,#8},%d1
	add.w	#40,%d1
	move.w	-(%a0),%d2
	sub.w	%d1,%d2
	jcc	1f
	| Pathologically small, denormalize.
	add.w	%d2,%d1
	clr.w	%d2
1:	move.w	%d2,(%a0)+
	ext.l	%d1
	jeq	fp_e2e_checkround
	cmp.w	#8,%d1
	jcs	2f
1:	clr.b	(-4,%a0)
	sub.w	#64,%d1
	jcs	1f
	add.w	#24,%d1
	lsl.l	%d1,%d0
	move.l	%d0,(%a0)
	jra	fp_e2e_checkround
1:	neg.w	%d1
	bfins	%d0,(%a0){%d1,#8}
	jra	fp_e2e_checkround
2:	lsl.l	%d1,%d0
	move.b	%d0,(-4,%a0)
	lsr.l	#8,%d0
	move.b	%d0,(7,%a0)
	jra	fp_e2e_checkround
#endif
1:	move.l	%d0,%d1			| lower lword is splitted between
	lsl.l	%d2,%d0			| higher and lower lword
	move.l	%d0,(%a0)
	move.l	%d1,%d0
	neg.w	%d2
	add.w	#32,%d2
	lsr.l	%d2,%d0
	move.l	%d0,-(%a0)
	jra	fp_e2e_checkround
	| Infinities and NaNs
fp_e2e_large:
	move.l	(%a0)+,%d0
	jne	3f
1:	tst.l	(%a0)
	jne	4f
	moveq	#1,%d0
2:	subq.l	#8,%a0
	printf	PCONV,"%p(",1,%a0
	printx	PCONV,%a0@
	printf	PCONV,")\n"
	rts
	| we have maybe a NaN, shift off the highest bit
3:	lsl.l	#1,%d0
	jeq	1b
	| we have a NaN, clear the return value
4:	clrl	%d0
	jra	2b


/*
 * Normalization functions.  Call these on the output of general
 * FP operators, and before any conversion into the destination
 * formats. fp_normalize_ext has always to be called first, the
 * following conversion functions expect an already normalized
 * number.
 */

	| fp_normalize_ext:
	| normalize an extended in extended (unpacked) format, basically
	| it does the same as fp_conv_ext2ext, additionally it also does
	| the necessary postprocessing checks.
	| args:	%a0 (struct fp_ext *)
	| NOTE: it does _not_ modify %a0/%a1 and the upper word of %d2

fp_normalize_ext:
	printf	PNORM,"ne: %p(",1,%a0
	printx	PNORM,%a0@
	printf	PNORM,"), "
	move.l	(%a0)+,%d0
	cmp.w	#0x7fff,%d0		| Inf / NaN?
	jeq	fp_ne_large
	move.l	(%a0),%d0
	jpl	fp_ne_small		| zero / denorm?
	| The high bit is set, so normalization is irrelevant.
fp_ne_checkround:
	subq.l	#4,%a0
#ifdef CONFIG_M68KFPU_EMU_EXTRAPREC
	move.b	(%a0),%d0
	jne	fp_ne_round
#endif
	printf	PNORM,"%p(",1,%a0
	printx	PNORM,%a0@
	printf	PNORM,")\n"
	rts
#ifdef CONFIG_M68KFPU_EMU_EXTRAPREC
fp_ne_round:
	fp_set_sr FPSR_EXC_INEX2
	clr.b	(%a0)
	move.w	(FPD_RND,FPDATA),%d2
	jne	fp_ne_roundother	| %d2 == 0, round to nearest
	tst.b	%d0			| test guard bit
	jpl	9f			| zero is closer
	btst	#0,(11,%a0)		| test lsb bit
	jne	fp_ne_doroundup		| round to infinity
	lsl.b	#1,%d0			| check low bits
	jeq	9f			| round to zero
fp_ne_doroundup:
	addq.l	#1,(8,%a0)
	jcc	9f
	addq.l	#1,(4,%a0)
	jcc	9f
	addq.w	#1,(2,%a0)
	move.w	#0x8000,(4,%a0)
9:	printf	PNORM,"%p(",1,%a0
	printx	PNORM,%a0@
	printf	PNORM,")\n"
	rts
fp_ne_roundother:
	subq.w	#2,%d2
	jcs	9b			| %d2 < 2, round to zero
	jhi	1f			| %d2 > 2, round to +infinity
	tst.b	(1,%a0)			| to -inf
	jne	fp_ne_doroundup		| negative, round to infinity