xref: /linux/arch/powerpc/include/asm/sfp-machine.h (revision 2697b79a469b68e3ad3640f55284359c1396278d)
1 /* Machine-dependent software floating-point definitions.  PPC version.
2    Copyright (C) 1997 Free Software Foundation, Inc.
3    This file is part of the GNU C Library.
4 
5    The GNU C Library is free software; you can redistribute it and/or
6    modify it under the terms of the GNU Library General Public License as
7    published by the Free Software Foundation; either version 2 of the
8    License, or (at your option) any later version.
9 
10    The GNU C Library is distributed in the hope that it will be useful,
11    but WITHOUT ANY WARRANTY; without even the implied warranty of
12    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13    Library General Public License for more details.
14 
15    You should have received a copy of the GNU Library General Public
16    License along with the GNU C Library; see the file COPYING.LIB.  If
17    not, write to the Free Software Foundation, Inc.,
18    59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
19 
20    Actually, this is a PPC (32bit) version, written based on the
21    i386, sparc, and sparc64 versions, by me,
22    Peter Maydell (pmaydell@chiark.greenend.org.uk).
23    Comments are by and large also mine, although they may be inaccurate.
24 
25    In picking out asm fragments I've gone with the lowest common
26    denominator, which also happens to be the hardware I have :->
27    That is, a SPARC without hardware multiply and divide.
28  */
29 
30 /* basic word size definitions */
31 #define _FP_W_TYPE_SIZE		32
32 #define _FP_W_TYPE		unsigned int
33 #define _FP_WS_TYPE		signed int
34 #define _FP_I_TYPE		int
35 
36 #define __ll_B			((UWtype) 1 << (W_TYPE_SIZE / 2))
37 #define __ll_lowpart(t)		((UWtype) (t) & (__ll_B - 1))
38 #define __ll_highpart(t)	((UWtype) (t) >> (W_TYPE_SIZE / 2))
39 
40 /* You can optionally code some things like addition in asm. For
41  * example, i386 defines __FP_FRAC_ADD_2 as asm. If you don't
42  * then you get a fragment of C code [if you change an #ifdef 0
43  * in op-2.h] or a call to add_ssaaaa (see below).
44  * Good places to look for asm fragments to use are gcc and glibc.
45  * gcc's longlong.h is useful.
46  */
47 
48 /* We need to know how to multiply and divide. If the host word size
49  * is >= 2*fracbits you can use FP_MUL_MEAT_n_imm(t,R,X,Y) which
50  * codes the multiply with whatever gcc does to 'a * b'.
51  * _FP_MUL_MEAT_n_wide(t,R,X,Y,f) is used when you have an asm
52  * function that can multiply two 1W values and get a 2W result.
53  * Otherwise you're stuck with _FP_MUL_MEAT_n_hard(t,R,X,Y) which
54  * does bitshifting to avoid overflow.
55  * For division there is FP_DIV_MEAT_n_imm(t,R,X,Y,f) for word size
56  * >= 2*fracbits, where f is either _FP_DIV_HELP_imm or
57  * _FP_DIV_HELP_ldiv (see op-1.h).
58  * _FP_DIV_MEAT_udiv() is if you have asm to do 2W/1W => (1W, 1W).
59  * [GCC and glibc have longlong.h which has the asm macro udiv_qrnnd
60  * to do this.]
61  * In general, 'n' is the number of words required to hold the type,
62  * and 't' is either S, D or Q for single/double/quad.
63  *           -- PMM
64  */
65 /* Example: SPARC64:
66  * #define _FP_MUL_MEAT_S(R,X,Y)	_FP_MUL_MEAT_1_imm(S,R,X,Y)
67  * #define _FP_MUL_MEAT_D(R,X,Y)	_FP_MUL_MEAT_1_wide(D,R,X,Y,umul_ppmm)
68  * #define _FP_MUL_MEAT_Q(R,X,Y)	_FP_MUL_MEAT_2_wide(Q,R,X,Y,umul_ppmm)
69  *
70  * #define _FP_DIV_MEAT_S(R,X,Y)	_FP_DIV_MEAT_1_imm(S,R,X,Y,_FP_DIV_HELP_imm)
71  * #define _FP_DIV_MEAT_D(R,X,Y)	_FP_DIV_MEAT_1_udiv(D,R,X,Y)
72  * #define _FP_DIV_MEAT_Q(R,X,Y)	_FP_DIV_MEAT_2_udiv_64(Q,R,X,Y)
73  *
74  * Example: i386:
75  * #define _FP_MUL_MEAT_S(R,X,Y)   _FP_MUL_MEAT_1_wide(S,R,X,Y,_i386_mul_32_64)
76  * #define _FP_MUL_MEAT_D(R,X,Y)   _FP_MUL_MEAT_2_wide(D,R,X,Y,_i386_mul_32_64)
77  *
78  * #define _FP_DIV_MEAT_S(R,X,Y)   _FP_DIV_MEAT_1_udiv(S,R,X,Y,_i386_div_64_32)
79  * #define _FP_DIV_MEAT_D(R,X,Y)   _FP_DIV_MEAT_2_udiv_64(D,R,X,Y)
80  */
81 
82 #define _FP_MUL_MEAT_S(R,X,Y)   _FP_MUL_MEAT_1_wide(_FP_WFRACBITS_S,R,X,Y,umul_ppmm)
83 #define _FP_MUL_MEAT_D(R,X,Y)   _FP_MUL_MEAT_2_wide(_FP_WFRACBITS_D,R,X,Y,umul_ppmm)
84 
85 #define _FP_DIV_MEAT_S(R,X,Y)	_FP_DIV_MEAT_1_udiv_norm(S,R,X,Y)
86 #define _FP_DIV_MEAT_D(R,X,Y)	_FP_DIV_MEAT_2_udiv(D,R,X,Y)
87 
88 /* These macros define what NaN looks like. They're supposed to expand to
89  * a comma-separated set of 32bit unsigned ints that encode NaN.
90  */
91 #define _FP_NANFRAC_S		((_FP_QNANBIT_S << 1) - 1)
92 #define _FP_NANFRAC_D		((_FP_QNANBIT_D << 1) - 1), -1
93 #define _FP_NANFRAC_Q		((_FP_QNANBIT_Q << 1) - 1), -1, -1, -1
94 #define _FP_NANSIGN_S		0
95 #define _FP_NANSIGN_D		0
96 #define _FP_NANSIGN_Q		0
97 
98 #define _FP_KEEPNANFRACP 1
99 
100 #ifdef FP_EX_BOOKE_E500_SPE
101 #define FP_EX_INEXACT		(1 << 21)
102 #define FP_EX_INVALID		(1 << 20)
103 #define FP_EX_DIVZERO		(1 << 19)
104 #define FP_EX_UNDERFLOW		(1 << 18)
105 #define FP_EX_OVERFLOW		(1 << 17)
106 #define FP_INHIBIT_RESULTS	0
107 
108 #define __FPU_FPSCR	(current->thread.spefscr)
109 #define __FPU_ENABLED_EXC		\
110 ({					\
111 	(__FPU_FPSCR >> 2) & 0x1f;	\
112 })
113 #else
114 /* Exception flags.  We use the bit positions of the appropriate bits
115    in the FPSCR, which also correspond to the FE_* bits.  This makes
116    everything easier ;-).  */
117 #define FP_EX_INVALID         (1 << (31 - 2))
118 #define FP_EX_INVALID_SNAN	EFLAG_VXSNAN
119 #define FP_EX_INVALID_ISI	EFLAG_VXISI
120 #define FP_EX_INVALID_IDI	EFLAG_VXIDI
121 #define FP_EX_INVALID_ZDZ	EFLAG_VXZDZ
122 #define FP_EX_INVALID_IMZ	EFLAG_VXIMZ
123 #define FP_EX_OVERFLOW        (1 << (31 - 3))
124 #define FP_EX_UNDERFLOW       (1 << (31 - 4))
125 #define FP_EX_DIVZERO         (1 << (31 - 5))
126 #define FP_EX_INEXACT         (1 << (31 - 6))
127 
128 #define __FPU_FPSCR	(current->thread.fp_state.fpscr)
129 
130 /* We only actually write to the destination register
131  * if exceptions signalled (if any) will not trap.
132  */
133 #define __FPU_ENABLED_EXC \
134 ({						\
135 	(__FPU_FPSCR >> 3) & 0x1f;	\
136 })
137 
138 #endif
139 
140 /*
141  * If one NaN is signaling and the other is not,
142  * we choose that one, otherwise we choose X.
143  */
144 #define _FP_CHOOSENAN(fs, wc, R, X, Y, OP)			\
145   do {								\
146     if ((_FP_FRAC_HIGH_RAW_##fs(Y) & _FP_QNANBIT_##fs)		\
147 	&& !(_FP_FRAC_HIGH_RAW_##fs(X) & _FP_QNANBIT_##fs))	\
148       {								\
149 	R##_s = X##_s;						\
150 	_FP_FRAC_COPY_##wc(R,X);				\
151       }								\
152     else							\
153       {								\
154 	R##_s = Y##_s;						\
155 	_FP_FRAC_COPY_##wc(R,Y);				\
156       }								\
157     R##_c = FP_CLS_NAN;						\
158   } while (0)
159 
160 
161 #include <linux/kernel.h>
162 #include <linux/sched.h>
163 
164 #define __FPU_TRAP_P(bits) \
165 	((__FPU_ENABLED_EXC & (bits)) != 0)
166 
167 #define __FP_PACK_S(val,X)			\
168 ({  int __exc = _FP_PACK_CANONICAL(S,1,X);	\
169     if(!__exc || !__FPU_TRAP_P(__exc))		\
170         _FP_PACK_RAW_1_P(S,val,X);		\
171     __exc;					\
172 })
173 
174 #define __FP_PACK_D(val,X)			\
175    do {									\
176 	_FP_PACK_CANONICAL(D, 2, X);					\
177 	if (!FP_CUR_EXCEPTIONS || !__FPU_TRAP_P(FP_CUR_EXCEPTIONS))	\
178 		_FP_PACK_RAW_2_P(D, val, X);				\
179    } while (0)
180 
181 #define __FP_PACK_DS(val,X)							\
182    do {										\
183 	   FP_DECL_S(__X);							\
184 	   FP_CONV(S, D, 1, 2, __X, X);						\
185 	   _FP_PACK_CANONICAL(S, 1, __X);					\
186 	   if (!FP_CUR_EXCEPTIONS || !__FPU_TRAP_P(FP_CUR_EXCEPTIONS)) {	\
187 		   _FP_UNPACK_CANONICAL(S, 1, __X);				\
188 		   FP_CONV(D, S, 2, 1, X, __X);					\
189 		   _FP_PACK_CANONICAL(D, 2, X);					\
190 		   if (!FP_CUR_EXCEPTIONS || !__FPU_TRAP_P(FP_CUR_EXCEPTIONS))	\
191 		   _FP_PACK_RAW_2_P(D, val, X);					\
192 	   }									\
193    } while (0)
194 
195 /* Obtain the current rounding mode. */
196 #define FP_ROUNDMODE			\
197 ({					\
198 	__FPU_FPSCR & 0x3;		\
199 })
200 
201 /* the asm fragments go here: all these are taken from glibc-2.0.5's
202  * stdlib/longlong.h
203  */
204 
205 #include <linux/types.h>
206 #include <asm/byteorder.h>
207 
208 /* add_ssaaaa is used in op-2.h and should be equivalent to
209  * #define add_ssaaaa(sh,sl,ah,al,bh,bl) (sh = ah+bh+ (( sl = al+bl) < al))
210  * add_ssaaaa(high_sum, low_sum, high_addend_1, low_addend_1,
211  * high_addend_2, low_addend_2) adds two UWtype integers, composed by
212  * HIGH_ADDEND_1 and LOW_ADDEND_1, and HIGH_ADDEND_2 and LOW_ADDEND_2
213  * respectively.  The result is placed in HIGH_SUM and LOW_SUM.  Overflow
214  * (i.e. carry out) is not stored anywhere, and is lost.
215  */
216 #define add_ssaaaa(sh, sl, ah, al, bh, bl) \
217   do {									\
218     if (__builtin_constant_p (bh) && (bh) == 0)				\
219       __asm__ ("add%I4c %1,%3,%4\n\taddze %0,%2"		\
220 	     : "=r" (sh), "=&r" (sl) : "r" (ah), "%r" (al), "rI" (bl));\
221     else if (__builtin_constant_p (bh) && (bh) == ~(USItype) 0)		\
222       __asm__ ("add%I4c %1,%3,%4\n\taddme %0,%2"		\
223 	     : "=r" (sh), "=&r" (sl) : "r" (ah), "%r" (al), "rI" (bl));\
224     else								\
225       __asm__ ("add%I5c %1,%4,%5\n\tadde %0,%2,%3"		\
226 	     : "=r" (sh), "=&r" (sl)					\
227 	     : "%r" (ah), "r" (bh), "%r" (al), "rI" (bl));		\
228   } while (0)
229 
230 /* sub_ddmmss is used in op-2.h and udivmodti4.c and should be equivalent to
231  * #define sub_ddmmss(sh, sl, ah, al, bh, bl) (sh = ah-bh - ((sl = al-bl) > al))
232  * sub_ddmmss(high_difference, low_difference, high_minuend, low_minuend,
233  * high_subtrahend, low_subtrahend) subtracts two two-word UWtype integers,
234  * composed by HIGH_MINUEND_1 and LOW_MINUEND_1, and HIGH_SUBTRAHEND_2 and
235  * LOW_SUBTRAHEND_2 respectively.  The result is placed in HIGH_DIFFERENCE
236  * and LOW_DIFFERENCE.  Overflow (i.e. carry out) is not stored anywhere,
237  * and is lost.
238  */
239 #define sub_ddmmss(sh, sl, ah, al, bh, bl) \
240   do {									\
241     if (__builtin_constant_p (ah) && (ah) == 0)				\
242       __asm__ ("subf%I3c %1,%4,%3\n\tsubfze %0,%2"	\
243 	       : "=r" (sh), "=&r" (sl) : "r" (bh), "rI" (al), "r" (bl));\
244     else if (__builtin_constant_p (ah) && (ah) == ~(USItype) 0)		\
245       __asm__ ("subf%I3c %1,%4,%3\n\tsubfme %0,%2"	\
246 	       : "=r" (sh), "=&r" (sl) : "r" (bh), "rI" (al), "r" (bl));\
247     else if (__builtin_constant_p (bh) && (bh) == 0)			\
248       __asm__ ("subf%I3c %1,%4,%3\n\taddme %0,%2"		\
249 	       : "=r" (sh), "=&r" (sl) : "r" (ah), "rI" (al), "r" (bl));\
250     else if (__builtin_constant_p (bh) && (bh) == ~(USItype) 0)		\
251       __asm__ ("subf%I3c %1,%4,%3\n\taddze %0,%2"		\
252 	       : "=r" (sh), "=&r" (sl) : "r" (ah), "rI" (al), "r" (bl));\
253     else								\
254       __asm__ ("subf%I4c %1,%5,%4\n\tsubfe %0,%3,%2"	\
255 	       : "=r" (sh), "=&r" (sl)					\
256 	       : "r" (ah), "r" (bh), "rI" (al), "r" (bl));		\
257   } while (0)
258 
259 /* asm fragments for mul and div */
260 
261 /* umul_ppmm(high_prod, low_prod, multipler, multiplicand) multiplies two
262  * UWtype integers MULTIPLER and MULTIPLICAND, and generates a two UWtype
263  * word product in HIGH_PROD and LOW_PROD.
264  */
265 #define umul_ppmm(ph, pl, m0, m1) \
266   do {									\
267     USItype __m0 = (m0), __m1 = (m1);					\
268     __asm__ ("mulhwu %0,%1,%2" : "=r" (ph) : "%r" (m0), "r" (m1));	\
269     (pl) = __m0 * __m1;							\
270   } while (0)
271 
272 /* udiv_qrnnd(quotient, remainder, high_numerator, low_numerator,
273  * denominator) divides a UDWtype, composed by the UWtype integers
274  * HIGH_NUMERATOR and LOW_NUMERATOR, by DENOMINATOR and places the quotient
275  * in QUOTIENT and the remainder in REMAINDER.  HIGH_NUMERATOR must be less
276  * than DENOMINATOR for correct operation.  If, in addition, the most
277  * significant bit of DENOMINATOR must be 1, then the pre-processor symbol
278  * UDIV_NEEDS_NORMALIZATION is defined to 1.
279  */
280 #define udiv_qrnnd(q, r, n1, n0, d) \
281   do {									\
282     UWtype __d1, __d0, __q1, __q0;					\
283     UWtype __r1, __r0, __m;						\
284     __d1 = __ll_highpart (d);						\
285     __d0 = __ll_lowpart (d);						\
286 									\
287     __r1 = (n1) % __d1;							\
288     __q1 = (n1) / __d1;							\
289     __m = (UWtype) __q1 * __d0;						\
290     __r1 = __r1 * __ll_B | __ll_highpart (n0);				\
291     if (__r1 < __m)							\
292       {									\
293 	__q1--, __r1 += (d);						\
294 	if (__r1 >= (d)) /* i.e. we didn't get carry when adding to __r1 */\
295 	  if (__r1 < __m)						\
296 	    __q1--, __r1 += (d);					\
297       }									\
298     __r1 -= __m;							\
299 									\
300     __r0 = __r1 % __d1;							\
301     __q0 = __r1 / __d1;							\
302     __m = (UWtype) __q0 * __d0;						\
303     __r0 = __r0 * __ll_B | __ll_lowpart (n0);				\
304     if (__r0 < __m)							\
305       {									\
306 	__q0--, __r0 += (d);						\
307 	if (__r0 >= (d))						\
308 	  if (__r0 < __m)						\
309 	    __q0--, __r0 += (d);					\
310       }									\
311     __r0 -= __m;							\
312 									\
313     (q) = (UWtype) __q1 * __ll_B | __q0;				\
314     (r) = __r0;								\
315   } while (0)
316 
317 #define UDIV_NEEDS_NORMALIZATION 1
318 
319 #define abort()								\
320 	return 0
321 
322 #ifdef __BIG_ENDIAN
323 #define __BYTE_ORDER __BIG_ENDIAN
324 #else
325 #define __BYTE_ORDER __LITTLE_ENDIAN
326 #endif
327 
328 /* Exception flags. */
329 #define EFLAG_INVALID		(1 << (31 - 2))
330 #define EFLAG_OVERFLOW		(1 << (31 - 3))
331 #define EFLAG_UNDERFLOW		(1 << (31 - 4))
332 #define EFLAG_DIVZERO		(1 << (31 - 5))
333 #define EFLAG_INEXACT		(1 << (31 - 6))
334 
335 #define EFLAG_VXSNAN		(1 << (31 - 7))
336 #define EFLAG_VXISI		(1 << (31 - 8))
337 #define EFLAG_VXIDI		(1 << (31 - 9))
338 #define EFLAG_VXZDZ		(1 << (31 - 10))
339 #define EFLAG_VXIMZ		(1 << (31 - 11))
340 #define EFLAG_VXVC		(1 << (31 - 12))
341 #define EFLAG_VXSOFT		(1 << (31 - 21))
342 #define EFLAG_VXSQRT		(1 << (31 - 22))
343 #define EFLAG_VXCVI		(1 << (31 - 23))
344