xref: /openbmc/linux/arch/x86/math-emu/reg_round.S (revision e3b9f1e8)
1/* SPDX-License-Identifier: GPL-2.0 */
2	.file "reg_round.S"
3/*---------------------------------------------------------------------------+
4 |  reg_round.S                                                              |
5 |                                                                           |
6 | Rounding/truncation/etc for FPU basic arithmetic functions.               |
7 |                                                                           |
8 | Copyright (C) 1993,1995,1997                                              |
9 |                       W. Metzenthen, 22 Parker St, Ormond, Vic 3163,      |
10 |                       Australia.  E-mail billm@suburbia.net               |
11 |                                                                           |
12 | This code has four possible entry points.                                 |
13 | The following must be entered by a jmp instruction:                       |
14 |   fpu_reg_round, fpu_reg_round_sqrt, and fpu_Arith_exit.                  |
15 |                                                                           |
16 | The FPU_round entry point is intended to be used by C code.               |
17 | From C, call as:                                                          |
18 |  int FPU_round(FPU_REG *arg, unsigned int extent, unsigned int control_w) |
19 |                                                                           |
20 |    Return value is the tag of the answer, or-ed with FPU_Exception if     |
21 |    one was raised, or -1 on internal error.                               |
22 |                                                                           |
23 | For correct "up" and "down" rounding, the argument must have the correct  |
24 | sign.                                                                     |
25 |                                                                           |
26 +---------------------------------------------------------------------------*/
27
28/*---------------------------------------------------------------------------+
29 | Four entry points.                                                        |
30 |                                                                           |
31 | Needed by both the fpu_reg_round and fpu_reg_round_sqrt entry points:     |
32 |  %eax:%ebx  64 bit significand                                            |
33 |  %edx       32 bit extension of the significand                           |
34 |  %edi       pointer to an FPU_REG for the result to be stored             |
35 |  stack      calling function must have set up a C stack frame and         |
36 |             pushed %esi, %edi, and %ebx                                   |
37 |                                                                           |
38 | Needed just for the fpu_reg_round_sqrt entry point:                       |
39 |  %cx  A control word in the same format as the FPU control word.          |
40 | Otherwise, PARAM4 must give such a value.                                 |
41 |                                                                           |
42 |                                                                           |
43 | The significand and its extension are assumed to be exact in the          |
44 | following sense:                                                          |
45 |   If the significand by itself is the exact result then the significand   |
46 |   extension (%edx) must contain 0, otherwise the significand extension    |
47 |   must be non-zero.                                                       |
48 |   If the significand extension is non-zero then the significand is        |
49 |   smaller than the magnitude of the correct exact result by an amount     |
50 |   greater than zero and less than one ls bit of the significand.          |
51 |   The significand extension is only required to have three possible       |
52 |   non-zero values:                                                        |
53 |       less than 0x80000000  <=> the significand is less than 1/2 an ls    |
54 |                                 bit smaller than the magnitude of the     |
55 |                                 true exact result.                        |
56 |         exactly 0x80000000  <=> the significand is exactly 1/2 an ls bit  |
57 |                                 smaller than the magnitude of the true    |
58 |                                 exact result.                             |
59 |    greater than 0x80000000  <=> the significand is more than 1/2 an ls    |
60 |                                 bit smaller than the magnitude of the     |
61 |                                 true exact result.                        |
62 |                                                                           |
63 +---------------------------------------------------------------------------*/
64
65/*---------------------------------------------------------------------------+
66 |  The code in this module has become quite complex, but it should handle   |
67 |  all of the FPU flags which are set at this stage of the basic arithmetic |
68 |  computations.                                                            |
69 |  There are a few rare cases where the results are not set identically to  |
70 |  a real FPU. These require a bit more thought because at this stage the   |
71 |  results of the code here appear to be more consistent...                 |
72 |  This may be changed in a future version.                                 |
73 +---------------------------------------------------------------------------*/
74
75
76#include "fpu_emu.h"
77#include "exception.h"
78#include "control_w.h"
79
80/* Flags for FPU_bits_lost */
81#define	LOST_DOWN	$1
82#define	LOST_UP		$2
83
84/* Flags for FPU_denormal */
85#define	DENORMAL	$1
86#define	UNMASKED_UNDERFLOW $2
87
88
89#ifndef NON_REENTRANT_FPU
90/*	Make the code re-entrant by putting
91	local storage on the stack: */
92#define FPU_bits_lost	(%esp)
93#define FPU_denormal	1(%esp)
94
95#else
96/*	Not re-entrant, so we can gain speed by putting
97	local storage in a static area: */
98.data
99	.align 4,0
100FPU_bits_lost:
101	.byte	0
102FPU_denormal:
103	.byte	0
104#endif /* NON_REENTRANT_FPU */
105
106
107.text
108.globl fpu_reg_round
109.globl fpu_Arith_exit
110
111/* Entry point when called from C */
112ENTRY(FPU_round)
113	pushl	%ebp
114	movl	%esp,%ebp
115	pushl	%esi
116	pushl	%edi
117	pushl	%ebx
118
119	movl	PARAM1,%edi
120	movl	SIGH(%edi),%eax
121	movl	SIGL(%edi),%ebx
122	movl	PARAM2,%edx
123
124fpu_reg_round:			/* Normal entry point */
125	movl	PARAM4,%ecx
126
127#ifndef NON_REENTRANT_FPU
128	pushl	%ebx		/* adjust the stack pointer */
129#endif /* NON_REENTRANT_FPU */
130
131#ifdef PARANOID
132/* Cannot use this here yet */
133/*	orl	%eax,%eax */
134/*	jns	L_entry_bugged */
135#endif /* PARANOID */
136
137	cmpw	EXP_UNDER,EXP(%edi)
138	jle	L_Make_denorm			/* The number is a de-normal */
139
140	movb	$0,FPU_denormal			/* 0 -> not a de-normal */
141
142Denorm_done:
143	movb	$0,FPU_bits_lost		/* No bits yet lost in rounding */
144
145	movl	%ecx,%esi
146	andl	CW_PC,%ecx
147	cmpl	PR_64_BITS,%ecx
148	je	LRound_To_64
149
150	cmpl	PR_53_BITS,%ecx
151	je	LRound_To_53
152
153	cmpl	PR_24_BITS,%ecx
154	je	LRound_To_24
155
156#ifdef PECULIAR_486
157/* With the precision control bits set to 01 "(reserved)", a real 80486
158   behaves as if the precision control bits were set to 11 "64 bits" */
159	cmpl	PR_RESERVED_BITS,%ecx
160	je	LRound_To_64
161#ifdef PARANOID
162	jmp	L_bugged_denorm_486
163#endif /* PARANOID */
164#else
165#ifdef PARANOID
166	jmp	L_bugged_denorm	/* There is no bug, just a bad control word */
167#endif /* PARANOID */
168#endif /* PECULIAR_486 */
169
170
171/* Round etc to 24 bit precision */
172LRound_To_24:
173	movl	%esi,%ecx
174	andl	CW_RC,%ecx
175	cmpl	RC_RND,%ecx
176	je	LRound_nearest_24
177
178	cmpl	RC_CHOP,%ecx
179	je	LCheck_truncate_24
180
181	cmpl	RC_UP,%ecx		/* Towards +infinity */
182	je	LUp_24
183
184	cmpl	RC_DOWN,%ecx		/* Towards -infinity */
185	je	LDown_24
186
187#ifdef PARANOID
188	jmp	L_bugged_round24
189#endif /* PARANOID */
190
191LUp_24:
192	cmpb	SIGN_POS,PARAM5
193	jne	LCheck_truncate_24	/* If negative then  up==truncate */
194
195	jmp	LCheck_24_round_up
196
197LDown_24:
198	cmpb	SIGN_POS,PARAM5
199	je	LCheck_truncate_24	/* If positive then  down==truncate */
200
201LCheck_24_round_up:
202	movl	%eax,%ecx
203	andl	$0x000000ff,%ecx
204	orl	%ebx,%ecx
205	orl	%edx,%ecx
206	jnz	LDo_24_round_up
207	jmp	L_Re_normalise
208
209LRound_nearest_24:
210	/* Do rounding of the 24th bit if needed (nearest or even) */
211	movl	%eax,%ecx
212	andl	$0x000000ff,%ecx
213	cmpl	$0x00000080,%ecx
214	jc	LCheck_truncate_24	/* less than half, no increment needed */
215
216	jne	LGreater_Half_24	/* greater than half, increment needed */
217
218	/* Possibly half, we need to check the ls bits */
219	orl	%ebx,%ebx
220	jnz	LGreater_Half_24	/* greater than half, increment needed */
221
222	orl	%edx,%edx
223	jnz	LGreater_Half_24	/* greater than half, increment needed */
224
225	/* Exactly half, increment only if 24th bit is 1 (round to even) */
226	testl	$0x00000100,%eax
227	jz	LDo_truncate_24
228
229LGreater_Half_24:			/* Rounding: increment at the 24th bit */
230LDo_24_round_up:
231	andl	$0xffffff00,%eax	/* Truncate to 24 bits */
232	xorl	%ebx,%ebx
233	movb	LOST_UP,FPU_bits_lost
234	addl	$0x00000100,%eax
235	jmp	LCheck_Round_Overflow
236
237LCheck_truncate_24:
238	movl	%eax,%ecx
239	andl	$0x000000ff,%ecx
240	orl	%ebx,%ecx
241	orl	%edx,%ecx
242	jz	L_Re_normalise		/* No truncation needed */
243
244LDo_truncate_24:
245	andl	$0xffffff00,%eax	/* Truncate to 24 bits */
246	xorl	%ebx,%ebx
247	movb	LOST_DOWN,FPU_bits_lost
248	jmp	L_Re_normalise
249
250
251/* Round etc to 53 bit precision */
252LRound_To_53:
253	movl	%esi,%ecx
254	andl	CW_RC,%ecx
255	cmpl	RC_RND,%ecx
256	je	LRound_nearest_53
257
258	cmpl	RC_CHOP,%ecx
259	je	LCheck_truncate_53
260
261	cmpl	RC_UP,%ecx		/* Towards +infinity */
262	je	LUp_53
263
264	cmpl	RC_DOWN,%ecx		/* Towards -infinity */
265	je	LDown_53
266
267#ifdef PARANOID
268	jmp	L_bugged_round53
269#endif /* PARANOID */
270
271LUp_53:
272	cmpb	SIGN_POS,PARAM5
273	jne	LCheck_truncate_53	/* If negative then  up==truncate */
274
275	jmp	LCheck_53_round_up
276
277LDown_53:
278	cmpb	SIGN_POS,PARAM5
279	je	LCheck_truncate_53	/* If positive then  down==truncate */
280
281LCheck_53_round_up:
282	movl	%ebx,%ecx
283	andl	$0x000007ff,%ecx
284	orl	%edx,%ecx
285	jnz	LDo_53_round_up
286	jmp	L_Re_normalise
287
288LRound_nearest_53:
289	/* Do rounding of the 53rd bit if needed (nearest or even) */
290	movl	%ebx,%ecx
291	andl	$0x000007ff,%ecx
292	cmpl	$0x00000400,%ecx
293	jc	LCheck_truncate_53	/* less than half, no increment needed */
294
295	jnz	LGreater_Half_53	/* greater than half, increment needed */
296
297	/* Possibly half, we need to check the ls bits */
298	orl	%edx,%edx
299	jnz	LGreater_Half_53	/* greater than half, increment needed */
300
301	/* Exactly half, increment only if 53rd bit is 1 (round to even) */
302	testl	$0x00000800,%ebx
303	jz	LTruncate_53
304
305LGreater_Half_53:			/* Rounding: increment at the 53rd bit */
306LDo_53_round_up:
307	movb	LOST_UP,FPU_bits_lost
308	andl	$0xfffff800,%ebx	/* Truncate to 53 bits */
309	addl	$0x00000800,%ebx
310	adcl	$0,%eax
311	jmp	LCheck_Round_Overflow
312
313LCheck_truncate_53:
314	movl	%ebx,%ecx
315	andl	$0x000007ff,%ecx
316	orl	%edx,%ecx
317	jz	L_Re_normalise
318
319LTruncate_53:
320	movb	LOST_DOWN,FPU_bits_lost
321	andl	$0xfffff800,%ebx	/* Truncate to 53 bits */
322	jmp	L_Re_normalise
323
324
325/* Round etc to 64 bit precision */
326LRound_To_64:
327	movl	%esi,%ecx
328	andl	CW_RC,%ecx
329	cmpl	RC_RND,%ecx
330	je	LRound_nearest_64
331
332	cmpl	RC_CHOP,%ecx
333	je	LCheck_truncate_64
334
335	cmpl	RC_UP,%ecx		/* Towards +infinity */
336	je	LUp_64
337
338	cmpl	RC_DOWN,%ecx		/* Towards -infinity */
339	je	LDown_64
340
341#ifdef PARANOID
342	jmp	L_bugged_round64
343#endif /* PARANOID */
344
345LUp_64:
346	cmpb	SIGN_POS,PARAM5
347	jne	LCheck_truncate_64	/* If negative then  up==truncate */
348
349	orl	%edx,%edx
350	jnz	LDo_64_round_up
351	jmp	L_Re_normalise
352
353LDown_64:
354	cmpb	SIGN_POS,PARAM5
355	je	LCheck_truncate_64	/* If positive then  down==truncate */
356
357	orl	%edx,%edx
358	jnz	LDo_64_round_up
359	jmp	L_Re_normalise
360
361LRound_nearest_64:
362	cmpl	$0x80000000,%edx
363	jc	LCheck_truncate_64
364
365	jne	LDo_64_round_up
366
367	/* Now test for round-to-even */
368	testb	$1,%bl
369	jz	LCheck_truncate_64
370
371LDo_64_round_up:
372	movb	LOST_UP,FPU_bits_lost
373	addl	$1,%ebx
374	adcl	$0,%eax
375
376LCheck_Round_Overflow:
377	jnc	L_Re_normalise
378
379	/* Overflow, adjust the result (significand to 1.0) */
380	rcrl	$1,%eax
381	rcrl	$1,%ebx
382	incw	EXP(%edi)
383	jmp	L_Re_normalise
384
385LCheck_truncate_64:
386	orl	%edx,%edx
387	jz	L_Re_normalise
388
389LTruncate_64:
390	movb	LOST_DOWN,FPU_bits_lost
391
392L_Re_normalise:
393	testb	$0xff,FPU_denormal
394	jnz	Normalise_result
395
396L_Normalised:
397	movl	TAG_Valid,%edx
398
399L_deNormalised:
400	cmpb	LOST_UP,FPU_bits_lost
401	je	L_precision_lost_up
402
403	cmpb	LOST_DOWN,FPU_bits_lost
404	je	L_precision_lost_down
405
406L_no_precision_loss:
407	/* store the result */
408
409L_Store_significand:
410	movl	%eax,SIGH(%edi)
411	movl	%ebx,SIGL(%edi)
412
413	cmpw	EXP_OVER,EXP(%edi)
414	jge	L_overflow
415
416	movl	%edx,%eax
417
418	/* Convert the exponent to 80x87 form. */
419	addw	EXTENDED_Ebias,EXP(%edi)
420	andw	$0x7fff,EXP(%edi)
421
422fpu_reg_round_signed_special_exit:
423
424	cmpb	SIGN_POS,PARAM5
425	je	fpu_reg_round_special_exit
426
427	orw	$0x8000,EXP(%edi)	/* Negative sign for the result. */
428
429fpu_reg_round_special_exit:
430
431#ifndef NON_REENTRANT_FPU
432	popl	%ebx		/* adjust the stack pointer */
433#endif /* NON_REENTRANT_FPU */
434
435fpu_Arith_exit:
436	popl	%ebx
437	popl	%edi
438	popl	%esi
439	leave
440	ret
441
442
443/*
444 * Set the FPU status flags to represent precision loss due to
445 * round-up.
446 */
447L_precision_lost_up:
448	push	%edx
449	push	%eax
450	call	set_precision_flag_up
451	popl	%eax
452	popl	%edx
453	jmp	L_no_precision_loss
454
455/*
456 * Set the FPU status flags to represent precision loss due to
457 * truncation.
458 */
459L_precision_lost_down:
460	push	%edx
461	push	%eax
462	call	set_precision_flag_down
463	popl	%eax
464	popl	%edx
465	jmp	L_no_precision_loss
466
467
468/*
469 * The number is a denormal (which might get rounded up to a normal)
470 * Shift the number right the required number of bits, which will
471 * have to be undone later...
472 */
473L_Make_denorm:
474	/* The action to be taken depends upon whether the underflow
475	   exception is masked */
476	testb	CW_Underflow,%cl		/* Underflow mask. */
477	jz	Unmasked_underflow		/* Do not make a denormal. */
478
479	movb	DENORMAL,FPU_denormal
480
481	pushl	%ecx		/* Save */
482	movw	EXP_UNDER+1,%cx
483	subw	EXP(%edi),%cx
484
485	cmpw	$64,%cx	/* shrd only works for 0..31 bits */
486	jnc	Denorm_shift_more_than_63
487
488	cmpw	$32,%cx	/* shrd only works for 0..31 bits */
489	jnc	Denorm_shift_more_than_32
490
491/*
492 * We got here without jumps by assuming that the most common requirement
493 *   is for a small de-normalising shift.
494 * Shift by [1..31] bits
495 */
496	addw	%cx,EXP(%edi)
497	orl	%edx,%edx	/* extension */
498	setne	%ch		/* Save whether %edx is non-zero */
499	xorl	%edx,%edx
500	shrd	%cl,%ebx,%edx
501	shrd	%cl,%eax,%ebx
502	shr	%cl,%eax
503	orb	%ch,%dl
504	popl	%ecx
505	jmp	Denorm_done
506
507/* Shift by [32..63] bits */
508Denorm_shift_more_than_32:
509	addw	%cx,EXP(%edi)
510	subb	$32,%cl
511	orl	%edx,%edx
512	setne	%ch
513	orb	%ch,%bl
514	xorl	%edx,%edx
515	shrd	%cl,%ebx,%edx
516	shrd	%cl,%eax,%ebx
517	shr	%cl,%eax
518	orl	%edx,%edx		/* test these 32 bits */
519	setne	%cl
520	orb	%ch,%bl
521	orb	%cl,%bl
522	movl	%ebx,%edx
523	movl	%eax,%ebx
524	xorl	%eax,%eax
525	popl	%ecx
526	jmp	Denorm_done
527
528/* Shift by [64..) bits */
529Denorm_shift_more_than_63:
530	cmpw	$64,%cx
531	jne	Denorm_shift_more_than_64
532
533/* Exactly 64 bit shift */
534	addw	%cx,EXP(%edi)
535	xorl	%ecx,%ecx
536	orl	%edx,%edx
537	setne	%cl
538	orl	%ebx,%ebx
539	setne	%ch
540	orb	%ch,%cl
541	orb	%cl,%al
542	movl	%eax,%edx
543	xorl	%eax,%eax
544	xorl	%ebx,%ebx
545	popl	%ecx
546	jmp	Denorm_done
547
548Denorm_shift_more_than_64:
549	movw	EXP_UNDER+1,EXP(%edi)
550/* This is easy, %eax must be non-zero, so.. */
551	movl	$1,%edx
552	xorl	%eax,%eax
553	xorl	%ebx,%ebx
554	popl	%ecx
555	jmp	Denorm_done
556
557
558Unmasked_underflow:
559	movb	UNMASKED_UNDERFLOW,FPU_denormal
560	jmp	Denorm_done
561
562
563/* Undo the de-normalisation. */
564Normalise_result:
565	cmpb	UNMASKED_UNDERFLOW,FPU_denormal
566	je	Signal_underflow
567
568/* The number must be a denormal if we got here. */
569#ifdef PARANOID
570	/* But check it... just in case. */
571	cmpw	EXP_UNDER+1,EXP(%edi)
572	jne	L_norm_bugged
573#endif /* PARANOID */
574
575#ifdef PECULIAR_486
576	/*
577	 * This implements a special feature of 80486 behaviour.
578	 * Underflow will be signalled even if the number is
579	 * not a denormal after rounding.
580	 * This difference occurs only for masked underflow, and not
581	 * in the unmasked case.
582	 * Actual 80486 behaviour differs from this in some circumstances.
583	 */
584	orl	%eax,%eax		/* ms bits */
585	js	LPseudoDenormal		/* Will be masked underflow */
586#else
587	orl	%eax,%eax		/* ms bits */
588	js	L_Normalised		/* No longer a denormal */
589#endif /* PECULIAR_486 */
590
591	jnz	LDenormal_adj_exponent
592
593	orl	%ebx,%ebx
594	jz	L_underflow_to_zero	/* The contents are zero */
595
596LDenormal_adj_exponent:
597	decw	EXP(%edi)
598
599LPseudoDenormal:
600	testb	$0xff,FPU_bits_lost	/* bits lost == underflow */
601	movl	TAG_Special,%edx
602	jz	L_deNormalised
603
604	/* There must be a masked underflow */
605	push	%eax
606	pushl	EX_Underflow
607	call	EXCEPTION
608	popl	%eax
609	popl	%eax
610	movl	TAG_Special,%edx
611	jmp	L_deNormalised
612
613
614/*
615 * The operations resulted in a number too small to represent.
616 * Masked response.
617 */
618L_underflow_to_zero:
619	push	%eax
620	call	set_precision_flag_down
621	popl	%eax
622
623	push	%eax
624	pushl	EX_Underflow
625	call	EXCEPTION
626	popl	%eax
627	popl	%eax
628
629/* Reduce the exponent to EXP_UNDER */
630	movw	EXP_UNDER,EXP(%edi)
631	movl	TAG_Zero,%edx
632	jmp	L_Store_significand
633
634
635/* The operations resulted in a number too large to represent. */
636L_overflow:
637	addw	EXTENDED_Ebias,EXP(%edi)	/* Set for unmasked response. */
638	push	%edi
639	call	arith_overflow
640	pop	%edi
641	jmp	fpu_reg_round_signed_special_exit
642
643
644Signal_underflow:
645	/* The number may have been changed to a non-denormal */
646	/* by the rounding operations. */
647	cmpw	EXP_UNDER,EXP(%edi)
648	jle	Do_unmasked_underflow
649
650	jmp	L_Normalised
651
652Do_unmasked_underflow:
653	/* Increase the exponent by the magic number */
654	addw	$(3*(1<<13)),EXP(%edi)
655	push	%eax
656	pushl	EX_Underflow
657	call	EXCEPTION
658	popl	%eax
659	popl	%eax
660	jmp	L_Normalised
661
662
663#ifdef PARANOID
664#ifdef PECULIAR_486
665L_bugged_denorm_486:
666	pushl	EX_INTERNAL|0x236
667	call	EXCEPTION
668	popl	%ebx
669	jmp	L_exception_exit
670#else
671L_bugged_denorm:
672	pushl	EX_INTERNAL|0x230
673	call	EXCEPTION
674	popl	%ebx
675	jmp	L_exception_exit
676#endif /* PECULIAR_486 */
677
678L_bugged_round24:
679	pushl	EX_INTERNAL|0x231
680	call	EXCEPTION
681	popl	%ebx
682	jmp	L_exception_exit
683
684L_bugged_round53:
685	pushl	EX_INTERNAL|0x232
686	call	EXCEPTION
687	popl	%ebx
688	jmp	L_exception_exit
689
690L_bugged_round64:
691	pushl	EX_INTERNAL|0x233
692	call	EXCEPTION
693	popl	%ebx
694	jmp	L_exception_exit
695
696L_norm_bugged:
697	pushl	EX_INTERNAL|0x234
698	call	EXCEPTION
699	popl	%ebx
700	jmp	L_exception_exit
701
702L_entry_bugged:
703	pushl	EX_INTERNAL|0x235
704	call	EXCEPTION
705	popl	%ebx
706L_exception_exit:
707	mov	$-1,%eax
708	jmp	fpu_reg_round_special_exit
709#endif /* PARANOID */
710
711ENDPROC(FPU_round)
712