1 |
282 |
jeremybenn |
;; Copyright (C) 2007, 2008, 2009 Free Software Foundation, Inc.
|
2 |
|
|
;;
|
3 |
|
|
;; This file is part of GCC.
|
4 |
|
|
;;
|
5 |
|
|
;; GCC is free software; you can redistribute it and/or modify
|
6 |
|
|
;; it under the terms of the GNU General Public License as published by
|
7 |
|
|
;; the Free Software Foundation; either version 3, or (at your option)
|
8 |
|
|
;; any later version.
|
9 |
|
|
;;
|
10 |
|
|
;; GCC 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
|
13 |
|
|
;; GNU General Public License for more details.
|
14 |
|
|
;;
|
15 |
|
|
;; You should have received a copy of the GNU General Public License
|
16 |
|
|
;; along with GCC; see the file COPYING3. If not see
|
17 |
|
|
;; .
|
18 |
|
|
|
19 |
|
|
;; For the internal conditional math routines:
|
20 |
|
|
|
21 |
|
|
;; operand 0 is always the result
|
22 |
|
|
;; operand 1 is always the predicate
|
23 |
|
|
;; operand 2, 3, and sometimes 4 are the input values.
|
24 |
|
|
;; operand 4 or 5 is the floating point status register to use.
|
25 |
|
|
;; operand 5 or 6 is the rounding to do. (0 = single, 1 = double, 2 = none)
|
26 |
|
|
;;
|
27 |
|
|
;; addrf3_cond - F0 = F2 + F3
|
28 |
|
|
;; subrf3_cond - F0 = F2 - F3
|
29 |
|
|
;; mulrf3_cond - F0 = F2 * F3
|
30 |
|
|
;; nmulrf3_cond - F0 = - (F2 * F3)
|
31 |
|
|
;; m1addrf4_cond - F0 = (F2 * F3) + F4
|
32 |
|
|
;; m1subrf4_cond - F0 = (F2 * F3) - F4
|
33 |
|
|
;; m2addrf4_cond - F0 = F2 + (F3 * F4)
|
34 |
|
|
;; m2subrf4_cond - F0 = F2 - (F3 * F4)
|
35 |
|
|
|
36 |
|
|
;; Basic plus/minus/mult operations
|
37 |
|
|
|
38 |
|
|
(define_insn "addrf3_cond"
|
39 |
|
|
[(set (match_operand:RF 0 "fr_register_operand" "=f,f")
|
40 |
|
|
(if_then_else:RF (ne:RF (match_operand:BI 1 "register_operand" "c,c")
|
41 |
|
|
(const_int 0))
|
42 |
|
|
(plus:RF
|
43 |
|
|
(match_operand:RF 2 "fr_reg_or_fp01_operand" "fG,fG")
|
44 |
|
|
(match_operand:RF 3 "fr_reg_or_fp01_operand" "fG,fG"))
|
45 |
|
|
(match_operand:RF 4 "fr_reg_or_0_operand" "0,H")))
|
46 |
|
|
(use (match_operand:SI 5 "const_int_operand" ""))
|
47 |
|
|
(use (match_operand:SI 6 "const_int_operand" ""))]
|
48 |
|
|
""
|
49 |
|
|
"(%1) fadd%R6.s%5 %0 = %F2, %F3"
|
50 |
|
|
[(set_attr "itanium_class" "fmac")
|
51 |
|
|
(set_attr "predicable" "no")])
|
52 |
|
|
|
53 |
|
|
(define_insn "subrf3_cond"
|
54 |
|
|
[(set (match_operand:RF 0 "fr_register_operand" "=f,f")
|
55 |
|
|
(if_then_else:RF (ne:RF (match_operand:BI 1 "register_operand" "c,c")
|
56 |
|
|
(const_int 0))
|
57 |
|
|
(minus:RF
|
58 |
|
|
(match_operand:RF 2 "fr_reg_or_fp01_operand" "fG,fG")
|
59 |
|
|
(match_operand:RF 3 "fr_reg_or_fp01_operand" "fG,fG"))
|
60 |
|
|
(match_operand:RF 4 "fr_reg_or_0_operand" "0,H")))
|
61 |
|
|
(use (match_operand:SI 5 "const_int_operand" ""))
|
62 |
|
|
(use (match_operand:SI 6 "const_int_operand" ""))]
|
63 |
|
|
""
|
64 |
|
|
"(%1) fsub%R6.s%5 %0 = %F2, %F3"
|
65 |
|
|
[(set_attr "itanium_class" "fmac")
|
66 |
|
|
(set_attr "predicable" "no")])
|
67 |
|
|
|
68 |
|
|
(define_insn "mulrf3_cond"
|
69 |
|
|
[(set (match_operand:RF 0 "fr_register_operand" "=f,f")
|
70 |
|
|
(if_then_else:RF (ne:RF (match_operand:BI 1 "register_operand" "c,c")
|
71 |
|
|
(const_int 0))
|
72 |
|
|
(mult:RF
|
73 |
|
|
(match_operand:RF 2 "fr_reg_or_fp01_operand" "fG,fG")
|
74 |
|
|
(match_operand:RF 3 "fr_reg_or_fp01_operand" "fG,fG"))
|
75 |
|
|
(match_operand:RF 4 "fr_reg_or_0_operand" "0,H")))
|
76 |
|
|
(use (match_operand:SI 5 "const_int_operand" ""))
|
77 |
|
|
(use (match_operand:SI 6 "const_int_operand" ""))]
|
78 |
|
|
""
|
79 |
|
|
"(%1) fmpy%R6.s%5 %0 = %F2, %F3"
|
80 |
|
|
[(set_attr "itanium_class" "fmac")
|
81 |
|
|
(set_attr "predicable" "no")])
|
82 |
|
|
|
83 |
|
|
;; neg-mult operation
|
84 |
|
|
|
85 |
|
|
(define_insn "nmulrf3_cond"
|
86 |
|
|
[(set (match_operand:RF 0 "fr_register_operand" "=f,f")
|
87 |
|
|
(if_then_else:RF (ne:RF (match_operand:BI 1 "register_operand" "c,c")
|
88 |
|
|
(const_int 0))
|
89 |
|
|
(neg:RF (mult:RF
|
90 |
|
|
(match_operand:RF 2 "fr_reg_or_fp01_operand" "fG,fG")
|
91 |
|
|
(match_operand:RF 3 "fr_reg_or_fp01_operand" "fG,fG")))
|
92 |
|
|
(match_operand:RF 4 "fr_reg_or_0_operand" "0,H")))
|
93 |
|
|
(use (match_operand:SI 5 "const_int_operand" ""))
|
94 |
|
|
(use (match_operand:SI 6 "const_int_operand" ""))]
|
95 |
|
|
""
|
96 |
|
|
"(%1) fnmpy%R6.s%5 %0 = %F2, %F3"
|
97 |
|
|
[(set_attr "itanium_class" "fmac")
|
98 |
|
|
(set_attr "predicable" "no")])
|
99 |
|
|
|
100 |
|
|
;; add-mult/sub-mult operations (mult as op1)
|
101 |
|
|
|
102 |
|
|
(define_insn "m1addrf4_cond"
|
103 |
|
|
[(set (match_operand:RF 0 "fr_register_operand" "=f,f")
|
104 |
|
|
(if_then_else:RF (ne:RF (match_operand:BI 1 "register_operand" "c,c")
|
105 |
|
|
(const_int 0))
|
106 |
|
|
(plus:RF
|
107 |
|
|
(mult:RF
|
108 |
|
|
(match_operand:RF 2 "fr_reg_or_fp01_operand" "fG,fG")
|
109 |
|
|
(match_operand:RF 3 "fr_reg_or_fp01_operand" "fG,fG"))
|
110 |
|
|
(match_operand:RF 4 "fr_reg_or_fp01_operand" "fG,fG"))
|
111 |
|
|
(match_operand:RF 5 "fr_reg_or_0_operand" "0,H")))
|
112 |
|
|
(use (match_operand:SI 6 "const_int_operand" ""))
|
113 |
|
|
(use (match_operand:SI 7 "const_int_operand" ""))]
|
114 |
|
|
""
|
115 |
|
|
"(%1) fma%R7.s%6 %0 = %F2, %F3, %F4"
|
116 |
|
|
[(set_attr "itanium_class" "fmac")
|
117 |
|
|
(set_attr "predicable" "no")])
|
118 |
|
|
|
119 |
|
|
(define_insn "m1subrf4_cond"
|
120 |
|
|
[(set (match_operand:RF 0 "fr_register_operand" "=f,f")
|
121 |
|
|
(if_then_else:RF (ne:RF (match_operand:BI 1 "register_operand" "c,c")
|
122 |
|
|
(const_int 0))
|
123 |
|
|
(minus:RF
|
124 |
|
|
(mult:RF
|
125 |
|
|
(match_operand:RF 2 "fr_reg_or_fp01_operand" "fG,fG")
|
126 |
|
|
(match_operand:RF 3 "fr_reg_or_fp01_operand" "fG,fG"))
|
127 |
|
|
(match_operand:RF 4 "fr_reg_or_fp01_operand" "fG,fG"))
|
128 |
|
|
(match_operand:RF 5 "fr_reg_or_0_operand" "0,H")))
|
129 |
|
|
(use (match_operand:SI 6 "const_int_operand" ""))
|
130 |
|
|
(use (match_operand:SI 7 "const_int_operand" ""))]
|
131 |
|
|
""
|
132 |
|
|
"(%1) fms%R7.s%6 %0 = %F2, %F3, %F4"
|
133 |
|
|
[(set_attr "itanium_class" "fmac")
|
134 |
|
|
(set_attr "predicable" "no")])
|
135 |
|
|
|
136 |
|
|
;; add-mult/sub-mult operations (mult as op2)
|
137 |
|
|
|
138 |
|
|
(define_insn "m2addrf4_cond"
|
139 |
|
|
[(set (match_operand:RF 0 "fr_register_operand" "=f,f")
|
140 |
|
|
(if_then_else:RF (ne:RF (match_operand:BI 1 "register_operand" "c,c")
|
141 |
|
|
(const_int 0))
|
142 |
|
|
(plus:RF
|
143 |
|
|
(match_operand:RF 2 "fr_reg_or_fp01_operand" "fG,fG")
|
144 |
|
|
(mult:RF
|
145 |
|
|
(match_operand:RF 3 "fr_reg_or_fp01_operand" "fG,fG")
|
146 |
|
|
(match_operand:RF 4 "fr_reg_or_fp01_operand" "fG,fG")))
|
147 |
|
|
(match_operand:RF 5 "fr_reg_or_0_operand" "0,H")))
|
148 |
|
|
(use (match_operand:SI 6 "const_int_operand" ""))
|
149 |
|
|
(use (match_operand:SI 7 "const_int_operand" ""))]
|
150 |
|
|
""
|
151 |
|
|
"(%1) fma%R7.s%6 %0 = %F3, %F4, %F2"
|
152 |
|
|
[(set_attr "itanium_class" "fmac")
|
153 |
|
|
(set_attr "predicable" "no")])
|
154 |
|
|
|
155 |
|
|
(define_insn "m2subrf4_cond"
|
156 |
|
|
[(set (match_operand:RF 0 "fr_register_operand" "=f,f")
|
157 |
|
|
(if_then_else:RF (ne:RF (match_operand:BI 1 "register_operand" "c,c")
|
158 |
|
|
(const_int 0))
|
159 |
|
|
(minus:RF
|
160 |
|
|
(match_operand:RF 2 "fr_reg_or_fp01_operand" "fG,fG")
|
161 |
|
|
(mult:RF
|
162 |
|
|
(match_operand:RF 3 "fr_reg_or_fp01_operand" "fG,fG")
|
163 |
|
|
(match_operand:RF 4 "fr_reg_or_fp01_operand" "fG,fG")))
|
164 |
|
|
(match_operand:RF 5 "fr_reg_or_0_operand" "0,H")))
|
165 |
|
|
(use (match_operand:SI 6 "const_int_operand" ""))
|
166 |
|
|
(use (match_operand:SI 7 "const_int_operand" ""))]
|
167 |
|
|
""
|
168 |
|
|
"(%1) fnma%R7.s%6 %0 = %F3, %F4, %F2"
|
169 |
|
|
[(set_attr "itanium_class" "fmac")
|
170 |
|
|
(set_attr "predicable" "no")])
|
171 |
|
|
|
172 |
|
|
;; Conversions to/from RF and SF/DF/XF
|
173 |
|
|
;; These conversions should not generate any code but make it possible
|
174 |
|
|
;; for all the instructions used to implement floating point division
|
175 |
|
|
;; to be written for RFmode only and to not have to handle multiple
|
176 |
|
|
;; modes or to have to handle a register in more than one mode.
|
177 |
|
|
|
178 |
|
|
(define_mode_iterator SDX_F [SF DF XF])
|
179 |
|
|
|
180 |
|
|
(define_insn "extendrf2"
|
181 |
|
|
[(set (match_operand:RF 0 "fr_register_operand" "=f")
|
182 |
|
|
(float_extend:RF (match_operand:SDX_F 1 "fr_reg_or_fp01_operand" "fG")))]
|
183 |
|
|
""
|
184 |
|
|
"#"
|
185 |
|
|
[(set_attr "itanium_class" "fmisc")
|
186 |
|
|
(set_attr "predicable" "yes")])
|
187 |
|
|
|
188 |
|
|
(define_split
|
189 |
|
|
[(set (match_operand:RF 0 "fr_register_operand" "")
|
190 |
|
|
(float_extend:RF (match_operand:SDX_F 1 "fr_reg_or_fp01_operand" "")))]
|
191 |
|
|
"reload_completed"
|
192 |
|
|
[(set (match_dup 0) (match_dup 2))]
|
193 |
|
|
{
|
194 |
|
|
if (operands[1] == CONST0_RTX (mode))
|
195 |
|
|
operands[2] = gen_rtx_REG (RFmode, FR_REG (0));
|
196 |
|
|
else if (operands[1] == CONST1_RTX (mode))
|
197 |
|
|
operands[2] = gen_rtx_REG (RFmode, FR_REG (1));
|
198 |
|
|
else
|
199 |
|
|
operands[2] = gen_rtx_REG (RFmode, REGNO (operands[1]));
|
200 |
|
|
})
|
201 |
|
|
|
202 |
|
|
|
203 |
|
|
(define_insn "truncrf2"
|
204 |
|
|
[(set (match_operand:SDX_F 0 "fr_register_operand" "=f")
|
205 |
|
|
(float_truncate:SDX_F (match_operand:RF 1 "fr_reg_or_fp01_operand" "fG")))]
|
206 |
|
|
""
|
207 |
|
|
"#"
|
208 |
|
|
[(set_attr "itanium_class" "fmisc")
|
209 |
|
|
(set_attr "predicable" "yes")])
|
210 |
|
|
|
211 |
|
|
(define_split
|
212 |
|
|
[(set (match_operand:SDX_F 0 "fr_register_operand" "")
|
213 |
|
|
(float_truncate:SDX_F (match_operand:RF 1 "fr_reg_or_fp01_operand" "")))]
|
214 |
|
|
"reload_completed"
|
215 |
|
|
[(set (match_dup 0) (match_dup 2))]
|
216 |
|
|
{
|
217 |
|
|
if (operands[1] == CONST0_RTX (RFmode))
|
218 |
|
|
operands[2] = gen_rtx_REG (mode, FR_REG (0));
|
219 |
|
|
else if (operands[1] == CONST1_RTX (RFmode))
|
220 |
|
|
operands[2] = gen_rtx_REG (mode, FR_REG (1));
|
221 |
|
|
else
|
222 |
|
|
operands[2] = gen_rtx_REG (mode, REGNO (operands[1]));
|
223 |
|
|
})
|
224 |
|
|
|
225 |
|
|
;; Float to integer truncations using an alternative status register.
|
226 |
|
|
|
227 |
|
|
(define_insn "fix_truncrfdi2_alts"
|
228 |
|
|
[(set (match_operand:DI 0 "fr_register_operand" "=f")
|
229 |
|
|
(fix:DI (match_operand:RF 1 "fr_register_operand" "f")))
|
230 |
|
|
(use (match_operand:SI 2 "const_int_operand" ""))]
|
231 |
|
|
""
|
232 |
|
|
"fcvt.fx.trunc.s%2 %0 = %1"
|
233 |
|
|
[(set_attr "itanium_class" "fcvtfx")])
|
234 |
|
|
|
235 |
|
|
(define_insn "fixuns_truncrfdi2_alts"
|
236 |
|
|
[(set (match_operand:DI 0 "fr_register_operand" "=f")
|
237 |
|
|
(unsigned_fix:DI (match_operand:RF 1 "fr_register_operand" "f")))
|
238 |
|
|
(use (match_operand:SI 2 "const_int_operand" ""))]
|
239 |
|
|
""
|
240 |
|
|
"fcvt.fxu.trunc.s%2 %0 = %1"
|
241 |
|
|
[(set_attr "itanium_class" "fcvtfx")])
|
242 |
|
|
|
243 |
|
|
(define_insn "setf_exp_rf"
|
244 |
|
|
[(set (match_operand:RF 0 "fr_register_operand" "=f")
|
245 |
|
|
(unspec:RF [(match_operand:DI 1 "register_operand" "r")]
|
246 |
|
|
UNSPEC_SETF_EXP))]
|
247 |
|
|
""
|
248 |
|
|
"setf.exp %0 = %1"
|
249 |
|
|
[(set_attr "itanium_class" "frfr")])
|
250 |
|
|
|
251 |
|
|
;; Reciprocal approximation
|
252 |
|
|
|
253 |
|
|
(define_insn "recip_approx_rf"
|
254 |
|
|
[(set (match_operand:RF 0 "fr_register_operand" "=f")
|
255 |
|
|
(unspec:RF [(match_operand:RF 1 "fr_reg_or_fp01_operand" "fG")
|
256 |
|
|
(match_operand:RF 2 "fr_reg_or_fp01_operand" "fG")]
|
257 |
|
|
UNSPEC_FR_RECIP_APPROX_RES))
|
258 |
|
|
(set (match_operand:BI 3 "register_operand" "=c")
|
259 |
|
|
(unspec:BI [(match_dup 1) (match_dup 2)] UNSPEC_FR_RECIP_APPROX))
|
260 |
|
|
(use (match_operand:SI 4 "const_int_operand" ""))]
|
261 |
|
|
""
|
262 |
|
|
"frcpa.s%4 %0, %3 = %F1, %F2"
|
263 |
|
|
[(set_attr "itanium_class" "fmisc")
|
264 |
|
|
(set_attr "predicable" "no")])
|
265 |
|
|
|
266 |
|
|
;; Single precision floating point division
|
267 |
|
|
|
268 |
|
|
(define_expand "divsf3"
|
269 |
|
|
[(set (match_operand:SF 0 "fr_register_operand" "")
|
270 |
|
|
(div:SF (match_operand:SF 1 "fr_reg_or_fp01_operand" "")
|
271 |
|
|
(match_operand:SF 2 "fr_reg_or_fp01_operand" "")))]
|
272 |
|
|
"TARGET_INLINE_FLOAT_DIV"
|
273 |
|
|
{
|
274 |
|
|
rtx insn;
|
275 |
|
|
if (TARGET_INLINE_FLOAT_DIV == INL_MIN_LAT)
|
276 |
|
|
insn = gen_divsf3_internal_lat (operands[0], operands[1], operands[2]);
|
277 |
|
|
else
|
278 |
|
|
insn = gen_divsf3_internal_thr (operands[0], operands[1], operands[2]);
|
279 |
|
|
emit_insn (insn);
|
280 |
|
|
DONE;
|
281 |
|
|
})
|
282 |
|
|
|
283 |
|
|
;; Single precision floating point division (maximum throughput algorithm).
|
284 |
|
|
|
285 |
|
|
(define_expand "divsf3_internal_thr"
|
286 |
|
|
[(set (match_operand:SF 0 "fr_register_operand" "")
|
287 |
|
|
(div:SF (match_operand:SF 1 "fr_reg_or_fp01_operand" "")
|
288 |
|
|
(match_operand:SF 2 "fr_reg_or_fp01_operand" "")))]
|
289 |
|
|
"TARGET_INLINE_FLOAT_DIV"
|
290 |
|
|
{
|
291 |
|
|
rtx y = gen_reg_rtx (RFmode);
|
292 |
|
|
rtx a = gen_reg_rtx (RFmode);
|
293 |
|
|
rtx b = gen_reg_rtx (RFmode);
|
294 |
|
|
rtx e = gen_reg_rtx (RFmode);
|
295 |
|
|
rtx y1 = gen_reg_rtx (RFmode);
|
296 |
|
|
rtx y2 = gen_reg_rtx (RFmode);
|
297 |
|
|
rtx q = gen_reg_rtx (RFmode);
|
298 |
|
|
rtx r = gen_reg_rtx (RFmode);
|
299 |
|
|
rtx q_res = gen_reg_rtx (RFmode);
|
300 |
|
|
rtx cond = gen_reg_rtx (BImode);
|
301 |
|
|
rtx zero = CONST0_RTX (RFmode);
|
302 |
|
|
rtx one = CONST1_RTX (RFmode);
|
303 |
|
|
rtx status0 = CONST0_RTX (SImode);
|
304 |
|
|
rtx status1 = CONST1_RTX (SImode);
|
305 |
|
|
rtx trunc_sgl = CONST0_RTX (SImode);
|
306 |
|
|
rtx trunc_off = CONST2_RTX (SImode);
|
307 |
|
|
|
308 |
|
|
/* Empty conversions to put inputs into RFmode. */
|
309 |
|
|
emit_insn (gen_extendsfrf2 (a, operands[1]));
|
310 |
|
|
emit_insn (gen_extendsfrf2 (b, operands[2]));
|
311 |
|
|
/* y = 1 / b */
|
312 |
|
|
emit_insn (gen_recip_approx_rf (y, a, b, cond, status0));
|
313 |
|
|
/* e = 1 - (b * y) */
|
314 |
|
|
emit_insn (gen_m2subrf4_cond (e, cond, one, b, y, zero, status1, trunc_off));
|
315 |
|
|
/* y1 = y + (y * e) */
|
316 |
|
|
emit_insn (gen_m2addrf4_cond (y1, cond, y, y, e, zero, status1, trunc_off));
|
317 |
|
|
/* y2 = y + (y1 * e) */
|
318 |
|
|
emit_insn (gen_m2addrf4_cond (y2, cond, y, y1, e, zero, status1, trunc_off));
|
319 |
|
|
/* q = single(a * y2) */
|
320 |
|
|
emit_insn (gen_mulrf3_cond (q, cond, a, y2, zero, status1, trunc_sgl));
|
321 |
|
|
/* r = a - (q * b) */
|
322 |
|
|
emit_insn (gen_m2subrf4_cond (r, cond, a, q, b, zero, status1, trunc_off));
|
323 |
|
|
/* Q = single (q + (r * y2)) */
|
324 |
|
|
emit_insn (gen_m2addrf4_cond (q_res, cond, q, r, y2, y, status0, trunc_sgl));
|
325 |
|
|
/* Conversion back into SFmode. */
|
326 |
|
|
emit_insn (gen_truncrfsf2 (operands[0], q_res));
|
327 |
|
|
DONE;
|
328 |
|
|
})
|
329 |
|
|
|
330 |
|
|
;; Single precision floating point division (minimum latency algorithm).
|
331 |
|
|
|
332 |
|
|
(define_expand "divsf3_internal_lat"
|
333 |
|
|
[(set (match_operand:SF 0 "fr_register_operand" "")
|
334 |
|
|
(div:SF (match_operand:SF 1 "fr_reg_or_fp01_operand" "")
|
335 |
|
|
(match_operand:SF 2 "fr_reg_or_fp01_operand" "")))]
|
336 |
|
|
"TARGET_INLINE_FLOAT_DIV"
|
337 |
|
|
{
|
338 |
|
|
rtx y = gen_reg_rtx (RFmode);
|
339 |
|
|
rtx a = gen_reg_rtx (RFmode);
|
340 |
|
|
rtx b = gen_reg_rtx (RFmode);
|
341 |
|
|
rtx e = gen_reg_rtx (RFmode);
|
342 |
|
|
rtx q = gen_reg_rtx (RFmode);
|
343 |
|
|
rtx e1 = gen_reg_rtx (RFmode);
|
344 |
|
|
rtx y1 = gen_reg_rtx (RFmode);
|
345 |
|
|
rtx q1 = gen_reg_rtx (RFmode);
|
346 |
|
|
rtx r = gen_reg_rtx (RFmode);
|
347 |
|
|
rtx q_res = gen_reg_rtx (RFmode);
|
348 |
|
|
rtx cond = gen_reg_rtx (BImode);
|
349 |
|
|
rtx zero = CONST0_RTX (RFmode);
|
350 |
|
|
rtx one = CONST1_RTX (RFmode);
|
351 |
|
|
rtx status0 = CONST0_RTX (SImode);
|
352 |
|
|
rtx status1 = CONST1_RTX (SImode);
|
353 |
|
|
rtx trunc_sgl = CONST0_RTX (SImode);
|
354 |
|
|
rtx trunc_off = CONST2_RTX (SImode);
|
355 |
|
|
|
356 |
|
|
/* Empty conversions to put inputs into RFmode. */
|
357 |
|
|
emit_insn (gen_extendsfrf2 (a, operands[1]));
|
358 |
|
|
emit_insn (gen_extendsfrf2 (b, operands[2]));
|
359 |
|
|
/* y = 1 / b */
|
360 |
|
|
emit_insn (gen_recip_approx_rf (y, a, b, cond, status0));
|
361 |
|
|
/* q = a * y */
|
362 |
|
|
emit_insn (gen_mulrf3_cond (q, cond, a, y, zero, status1, trunc_off));
|
363 |
|
|
/* e = 1 - (b * y) */
|
364 |
|
|
emit_insn (gen_m2subrf4_cond (e, cond, one, b, y, zero, status1, trunc_off));
|
365 |
|
|
/* e1 = e + (e * e) */
|
366 |
|
|
emit_insn (gen_m2addrf4_cond (e1, cond, e, e, e, zero, status1, trunc_off));
|
367 |
|
|
/* q1 = single(q + (q * e1)) */
|
368 |
|
|
emit_insn (gen_m2addrf4_cond (q1, cond, q, q, e1, zero, status1, trunc_sgl));
|
369 |
|
|
/* y1 = y + (y * e1) */
|
370 |
|
|
emit_insn (gen_m2addrf4_cond (y1, cond, y, y, e1, zero, status1, trunc_off));
|
371 |
|
|
/* r = a - (q1 * b) */
|
372 |
|
|
emit_insn (gen_m2subrf4_cond (r, cond, a, q1, b, zero, status1, trunc_off));
|
373 |
|
|
/* Q = single (q1 + (r * y1)) */
|
374 |
|
|
emit_insn (gen_m2addrf4_cond (q_res, cond, q1, r, y1, y, status0, trunc_sgl));
|
375 |
|
|
/* Conversion back into SFmode. */
|
376 |
|
|
emit_insn (gen_truncrfsf2 (operands[0], q_res));
|
377 |
|
|
DONE;
|
378 |
|
|
})
|
379 |
|
|
|
380 |
|
|
;; Double precision floating point division
|
381 |
|
|
|
382 |
|
|
(define_expand "divdf3"
|
383 |
|
|
[(set (match_operand:DF 0 "fr_register_operand" "")
|
384 |
|
|
(div:DF (match_operand:DF 1 "fr_reg_or_fp01_operand" "")
|
385 |
|
|
(match_operand:DF 2 "fr_reg_or_fp01_operand" "")))]
|
386 |
|
|
"TARGET_INLINE_FLOAT_DIV"
|
387 |
|
|
{
|
388 |
|
|
rtx insn;
|
389 |
|
|
if (TARGET_INLINE_FLOAT_DIV == INL_MIN_LAT)
|
390 |
|
|
insn = gen_divdf3_internal_lat (operands[0], operands[1], operands[2]);
|
391 |
|
|
else
|
392 |
|
|
insn = gen_divdf3_internal_thr (operands[0], operands[1], operands[2]);
|
393 |
|
|
emit_insn (insn);
|
394 |
|
|
DONE;
|
395 |
|
|
})
|
396 |
|
|
|
397 |
|
|
;; Double precision floating point division (maximum throughput algorithm).
|
398 |
|
|
|
399 |
|
|
(define_expand "divdf3_internal_thr"
|
400 |
|
|
[(set (match_operand:DF 0 "fr_register_operand" "")
|
401 |
|
|
(div:DF (match_operand:DF 1 "fr_reg_or_fp01_operand" "")
|
402 |
|
|
(match_operand:DF 2 "fr_reg_or_fp01_operand" "")))]
|
403 |
|
|
"TARGET_INLINE_FLOAT_DIV"
|
404 |
|
|
{
|
405 |
|
|
rtx q_res = gen_reg_rtx (RFmode);
|
406 |
|
|
rtx a = gen_reg_rtx (RFmode);
|
407 |
|
|
rtx b = gen_reg_rtx (RFmode);
|
408 |
|
|
rtx y = gen_reg_rtx (RFmode);
|
409 |
|
|
rtx e = gen_reg_rtx (RFmode);
|
410 |
|
|
rtx y1 = gen_reg_rtx (RFmode);
|
411 |
|
|
rtx e1 = gen_reg_rtx (RFmode);
|
412 |
|
|
rtx y2 = gen_reg_rtx (RFmode);
|
413 |
|
|
rtx e2 = gen_reg_rtx (RFmode);
|
414 |
|
|
rtx y3 = gen_reg_rtx (RFmode);
|
415 |
|
|
rtx q = gen_reg_rtx (RFmode);
|
416 |
|
|
rtx r = gen_reg_rtx (RFmode);
|
417 |
|
|
rtx cond = gen_reg_rtx (BImode);
|
418 |
|
|
rtx zero = CONST0_RTX (RFmode);
|
419 |
|
|
rtx one = CONST1_RTX (RFmode);
|
420 |
|
|
rtx status0 = CONST0_RTX (SImode);
|
421 |
|
|
rtx status1 = CONST1_RTX (SImode);
|
422 |
|
|
rtx trunc_dbl = CONST1_RTX (SImode);
|
423 |
|
|
rtx trunc_off = CONST2_RTX (SImode);
|
424 |
|
|
/* Empty conversions to put inputs into RFmode */
|
425 |
|
|
emit_insn (gen_extenddfrf2 (a, operands[1]));
|
426 |
|
|
emit_insn (gen_extenddfrf2 (b, operands[2]));
|
427 |
|
|
/* y = 1 / b */
|
428 |
|
|
emit_insn (gen_recip_approx_rf (y, a, b, cond, status0));
|
429 |
|
|
/* e = 1 - (b * y) */
|
430 |
|
|
emit_insn (gen_m2subrf4_cond (e, cond, one, b, y, zero, status1, trunc_off));
|
431 |
|
|
/* y1 = y + (y * e) */
|
432 |
|
|
emit_insn (gen_m2addrf4_cond (y1, cond, y, y, e, zero, status1, trunc_off));
|
433 |
|
|
/* e1 = e * e */
|
434 |
|
|
emit_insn (gen_mulrf3_cond (e1, cond, e, e, zero, status1, trunc_off));
|
435 |
|
|
/* y2 = y1 + (y1 * e1) */
|
436 |
|
|
emit_insn (gen_m2addrf4_cond (y2, cond, y1, y1, e1, zero, status1, trunc_off));
|
437 |
|
|
/* e2 = e1 * e1 */
|
438 |
|
|
emit_insn (gen_mulrf3_cond (e2, cond, e1, e1, zero, status1, trunc_off));
|
439 |
|
|
/* y3 = y2 + (y2 * e2) */
|
440 |
|
|
emit_insn (gen_m2addrf4_cond (y3, cond, y2, y2, e2, zero, status1, trunc_off));
|
441 |
|
|
/* q = double (a * y3) */
|
442 |
|
|
emit_insn (gen_mulrf3_cond (q, cond, a, y3, zero, status1, trunc_dbl));
|
443 |
|
|
/* r = a - (b * q) */
|
444 |
|
|
emit_insn (gen_m2subrf4_cond (r, cond, a, b, q, zero, status1, trunc_off));
|
445 |
|
|
/* Q = double (q + (r * y3)) */
|
446 |
|
|
emit_insn (gen_m2addrf4_cond (q_res, cond, q, r, y3, y, status0, trunc_dbl));
|
447 |
|
|
/* Conversion back into DFmode */
|
448 |
|
|
emit_insn (gen_truncrfdf2 (operands[0], q_res));
|
449 |
|
|
DONE;
|
450 |
|
|
})
|
451 |
|
|
|
452 |
|
|
;; Double precision floating point division (minimum latency algorithm).
|
453 |
|
|
|
454 |
|
|
(define_expand "divdf3_internal_lat"
|
455 |
|
|
[(set (match_operand:DF 0 "fr_register_operand" "")
|
456 |
|
|
(div:DF (match_operand:DF 1 "fr_reg_or_fp01_operand" "")
|
457 |
|
|
(match_operand:DF 2 "fr_reg_or_fp01_operand" "")))]
|
458 |
|
|
"TARGET_INLINE_FLOAT_DIV"
|
459 |
|
|
{
|
460 |
|
|
rtx q_res = gen_reg_rtx (RFmode);
|
461 |
|
|
rtx a = gen_reg_rtx (RFmode);
|
462 |
|
|
rtx b = gen_reg_rtx (RFmode);
|
463 |
|
|
rtx y = gen_reg_rtx (RFmode);
|
464 |
|
|
rtx e = gen_reg_rtx (RFmode);
|
465 |
|
|
rtx y1 = gen_reg_rtx (RFmode);
|
466 |
|
|
rtx e1 = gen_reg_rtx (RFmode);
|
467 |
|
|
rtx q1 = gen_reg_rtx (RFmode);
|
468 |
|
|
rtx y2 = gen_reg_rtx (RFmode);
|
469 |
|
|
rtx e2 = gen_reg_rtx (RFmode);
|
470 |
|
|
rtx q2 = gen_reg_rtx (RFmode);
|
471 |
|
|
rtx e3 = gen_reg_rtx (RFmode);
|
472 |
|
|
rtx q = gen_reg_rtx (RFmode);
|
473 |
|
|
rtx r1 = gen_reg_rtx (RFmode);
|
474 |
|
|
rtx cond = gen_reg_rtx (BImode);
|
475 |
|
|
rtx zero = CONST0_RTX (RFmode);
|
476 |
|
|
rtx one = CONST1_RTX (RFmode);
|
477 |
|
|
rtx status0 = CONST0_RTX (SImode);
|
478 |
|
|
rtx status1 = CONST1_RTX (SImode);
|
479 |
|
|
rtx trunc_dbl = CONST1_RTX (SImode);
|
480 |
|
|
rtx trunc_off = CONST2_RTX (SImode);
|
481 |
|
|
|
482 |
|
|
/* Empty conversions to put inputs into RFmode */
|
483 |
|
|
emit_insn (gen_extenddfrf2 (a, operands[1]));
|
484 |
|
|
emit_insn (gen_extenddfrf2 (b, operands[2]));
|
485 |
|
|
/* y = 1 / b */
|
486 |
|
|
emit_insn (gen_recip_approx_rf (y, a, b, cond, status0));
|
487 |
|
|
/* e = 1 - (b * y) */
|
488 |
|
|
emit_insn (gen_m2subrf4_cond (e, cond, one, b, y, zero, status1, trunc_off));
|
489 |
|
|
/* q = a * y */
|
490 |
|
|
emit_insn (gen_mulrf3_cond (q, cond, a, y, zero, status1, trunc_off));
|
491 |
|
|
/* e2 = e + (e * e) */
|
492 |
|
|
emit_insn (gen_m2addrf4_cond (e2, cond, e, e, e, zero, status1, trunc_off));
|
493 |
|
|
/* e1 = e * e */
|
494 |
|
|
emit_insn (gen_mulrf3_cond (e1, cond, e, e, zero, status1, trunc_off));
|
495 |
|
|
/* e3 = e + (e1 * e1) */
|
496 |
|
|
emit_insn (gen_m2addrf4_cond (e3, cond, e, e1, e1, zero, status1, trunc_off));
|
497 |
|
|
/* q1 = q + (q * e2) */
|
498 |
|
|
emit_insn (gen_m2addrf4_cond (q1, cond, q, q, e2, zero, status1, trunc_off));
|
499 |
|
|
/* y1 = y + (y * e2) */
|
500 |
|
|
emit_insn (gen_m2addrf4_cond (y1, cond, y, y, e2, zero, status1, trunc_off));
|
501 |
|
|
/* q2 = double(q + (q1 * e3)) */
|
502 |
|
|
emit_insn (gen_m2addrf4_cond (q2, cond, q, q1, e3, zero, status1, trunc_dbl));
|
503 |
|
|
/* y2 = y + (y1 * e3) */
|
504 |
|
|
emit_insn (gen_m2addrf4_cond (y2, cond, y, y1, e3, zero, status1, trunc_off));
|
505 |
|
|
/* r1 = a - (b * q2) */
|
506 |
|
|
emit_insn (gen_m2subrf4_cond (r1, cond, a, b, q2, zero, status1, trunc_off));
|
507 |
|
|
/* Q = double (q2 + (r1 * y2)) */
|
508 |
|
|
emit_insn (gen_m2addrf4_cond (q_res, cond, q2, r1, y2, y, status0, trunc_dbl));
|
509 |
|
|
/* Conversion back into DFmode */
|
510 |
|
|
emit_insn (gen_truncrfdf2 (operands[0], q_res));
|
511 |
|
|
DONE;
|
512 |
|
|
})
|
513 |
|
|
|
514 |
|
|
;; Extended precision floating point division.
|
515 |
|
|
|
516 |
|
|
(define_expand "divxf3"
|
517 |
|
|
[(set (match_operand:XF 0 "fr_register_operand" "")
|
518 |
|
|
(div:XF (match_operand:XF 1 "fr_reg_or_fp01_operand" "")
|
519 |
|
|
(match_operand:XF 2 "fr_reg_or_fp01_operand" "")))]
|
520 |
|
|
"TARGET_INLINE_FLOAT_DIV"
|
521 |
|
|
{
|
522 |
|
|
rtx q_res = gen_reg_rtx (RFmode);
|
523 |
|
|
rtx a = gen_reg_rtx (RFmode);
|
524 |
|
|
rtx b = gen_reg_rtx (RFmode);
|
525 |
|
|
rtx y = gen_reg_rtx (RFmode);
|
526 |
|
|
rtx e = gen_reg_rtx (RFmode);
|
527 |
|
|
rtx y1 = gen_reg_rtx (RFmode);
|
528 |
|
|
rtx e1 = gen_reg_rtx (RFmode);
|
529 |
|
|
rtx q1 = gen_reg_rtx (RFmode);
|
530 |
|
|
rtx y2 = gen_reg_rtx (RFmode);
|
531 |
|
|
rtx e2 = gen_reg_rtx (RFmode);
|
532 |
|
|
rtx y3 = gen_reg_rtx (RFmode);
|
533 |
|
|
rtx e3 = gen_reg_rtx (RFmode);
|
534 |
|
|
rtx e4 = gen_reg_rtx (RFmode);
|
535 |
|
|
rtx q = gen_reg_rtx (RFmode);
|
536 |
|
|
rtx r = gen_reg_rtx (RFmode);
|
537 |
|
|
rtx r1 = gen_reg_rtx (RFmode);
|
538 |
|
|
rtx cond = gen_reg_rtx (BImode);
|
539 |
|
|
rtx zero = CONST0_RTX (RFmode);
|
540 |
|
|
rtx one = CONST1_RTX (RFmode);
|
541 |
|
|
rtx status0 = CONST0_RTX (SImode);
|
542 |
|
|
rtx status1 = CONST1_RTX (SImode);
|
543 |
|
|
rtx trunc_off = CONST2_RTX (SImode);
|
544 |
|
|
|
545 |
|
|
/* Empty conversions to put inputs into RFmode */
|
546 |
|
|
emit_insn (gen_extendxfrf2 (a, operands[1]));
|
547 |
|
|
emit_insn (gen_extendxfrf2 (b, operands[2]));
|
548 |
|
|
/* y = 1 / b */
|
549 |
|
|
emit_insn (gen_recip_approx_rf (y, a, b, cond, status0));
|
550 |
|
|
/* e = 1 - (b * y) */
|
551 |
|
|
emit_insn (gen_m2subrf4_cond (e, cond, one, b, y, zero, status1, trunc_off));
|
552 |
|
|
/* q = a * y */
|
553 |
|
|
emit_insn (gen_mulrf3_cond (q, cond, a, y, zero, status1, trunc_off));
|
554 |
|
|
/* e2 = e + (e * e) */
|
555 |
|
|
emit_insn (gen_m2addrf4_cond (e2, cond, e, e, e, zero, status1, trunc_off));
|
556 |
|
|
/* e1 = e * e */
|
557 |
|
|
emit_insn (gen_mulrf3_cond (e1, cond, e, e, zero, status1, trunc_off));
|
558 |
|
|
/* y1 = y + (y * e2) */
|
559 |
|
|
emit_insn (gen_m2addrf4_cond (y1, cond, y, y, e2, zero, status1, trunc_off));
|
560 |
|
|
/* e3 = e + (e1 * e1) */
|
561 |
|
|
emit_insn (gen_m2addrf4_cond (e3, cond, e, e1, e1, zero, status1, trunc_off));
|
562 |
|
|
/* y2 = y + (y1 * e3) */
|
563 |
|
|
emit_insn (gen_m2addrf4_cond (y2, cond, y, y1, e3, zero, status1, trunc_off));
|
564 |
|
|
/* r = a - (b * q) */
|
565 |
|
|
emit_insn (gen_m2subrf4_cond (r, cond, a, b, q, zero, status1, trunc_off));
|
566 |
|
|
/* e4 = 1 - (b * y2) */
|
567 |
|
|
emit_insn (gen_m2subrf4_cond (e4, cond, one, b, y2, zero, status1, trunc_off));
|
568 |
|
|
/* q1 = q + (r * y2) */
|
569 |
|
|
emit_insn (gen_m2addrf4_cond (q1, cond, q, r, y2, zero, status1, trunc_off));
|
570 |
|
|
/* y3 = y2 + (y2 * e4) */
|
571 |
|
|
emit_insn (gen_m2addrf4_cond (y3, cond, y2, y2, e4, zero, status1, trunc_off));
|
572 |
|
|
/* r1 = a - (b * q1) */
|
573 |
|
|
emit_insn (gen_m2subrf4_cond (r1, cond, a, b, q1, zero, status1, trunc_off));
|
574 |
|
|
/* Q = q1 + (r1 * y3) */
|
575 |
|
|
emit_insn (gen_m2addrf4_cond (q_res, cond, q1, r1, y3, y, status0, trunc_off));
|
576 |
|
|
/* Conversion back into XFmode */
|
577 |
|
|
emit_insn (gen_truncrfxf2 (operands[0], q_res));
|
578 |
|
|
DONE;
|
579 |
|
|
})
|
580 |
|
|
|
581 |
|
|
|
582 |
|
|
;; Integer division operations
|
583 |
|
|
|
584 |
|
|
(define_expand "divsi3"
|
585 |
|
|
[(set (match_operand:SI 0 "register_operand" "")
|
586 |
|
|
(div:SI (match_operand:SI 1 "general_operand" "")
|
587 |
|
|
(match_operand:SI 2 "general_operand" "")))]
|
588 |
|
|
"TARGET_INLINE_INT_DIV"
|
589 |
|
|
{
|
590 |
|
|
rtx op1_rf, op2_rf, op0_rf, op0_di;
|
591 |
|
|
|
592 |
|
|
op0_rf = gen_reg_rtx (RFmode);
|
593 |
|
|
op0_di = gen_reg_rtx (DImode);
|
594 |
|
|
|
595 |
|
|
if (! register_operand (operands[1], SImode))
|
596 |
|
|
operands[1] = force_reg (SImode, operands[1]);
|
597 |
|
|
op1_rf = gen_reg_rtx (RFmode);
|
598 |
|
|
expand_float (op1_rf, operands[1], 0);
|
599 |
|
|
|
600 |
|
|
if (! register_operand (operands[2], SImode))
|
601 |
|
|
operands[2] = force_reg (SImode, operands[2]);
|
602 |
|
|
op2_rf = gen_reg_rtx (RFmode);
|
603 |
|
|
expand_float (op2_rf, operands[2], 0);
|
604 |
|
|
|
605 |
|
|
emit_insn (gen_cond_trap (EQ, operands[2], CONST0_RTX (SImode),
|
606 |
|
|
CONST1_RTX (SImode)));
|
607 |
|
|
|
608 |
|
|
emit_insn (gen_divsi3_internal (op0_rf, op1_rf, op2_rf));
|
609 |
|
|
|
610 |
|
|
emit_insn (gen_fix_truncrfdi2_alts (op0_di, op0_rf, const1_rtx));
|
611 |
|
|
emit_move_insn (operands[0], gen_lowpart (SImode, op0_di));
|
612 |
|
|
DONE;
|
613 |
|
|
})
|
614 |
|
|
|
615 |
|
|
(define_expand "modsi3"
|
616 |
|
|
[(set (match_operand:SI 0 "register_operand" "")
|
617 |
|
|
(mod:SI (match_operand:SI 1 "general_operand" "")
|
618 |
|
|
(match_operand:SI 2 "general_operand" "")))]
|
619 |
|
|
"TARGET_INLINE_INT_DIV"
|
620 |
|
|
{
|
621 |
|
|
rtx op2_neg, op1_di, div;
|
622 |
|
|
|
623 |
|
|
div = gen_reg_rtx (SImode);
|
624 |
|
|
emit_insn (gen_divsi3 (div, operands[1], operands[2]));
|
625 |
|
|
|
626 |
|
|
op2_neg = expand_unop (SImode, neg_optab, operands[2], NULL_RTX, 0);
|
627 |
|
|
|
628 |
|
|
/* This is a trick to get us to reuse the value that we're sure to
|
629 |
|
|
have already copied to the FP regs. */
|
630 |
|
|
op1_di = gen_reg_rtx (DImode);
|
631 |
|
|
convert_move (op1_di, operands[1], 0);
|
632 |
|
|
|
633 |
|
|
emit_insn (gen_maddsi4 (operands[0], div, op2_neg,
|
634 |
|
|
gen_lowpart (SImode, op1_di)));
|
635 |
|
|
DONE;
|
636 |
|
|
})
|
637 |
|
|
|
638 |
|
|
(define_expand "udivsi3"
|
639 |
|
|
[(set (match_operand:SI 0 "register_operand" "")
|
640 |
|
|
(udiv:SI (match_operand:SI 1 "general_operand" "")
|
641 |
|
|
(match_operand:SI 2 "general_operand" "")))]
|
642 |
|
|
"TARGET_INLINE_INT_DIV"
|
643 |
|
|
{
|
644 |
|
|
rtx op1_rf, op2_rf, op0_rf, op0_di;
|
645 |
|
|
|
646 |
|
|
op0_rf = gen_reg_rtx (RFmode);
|
647 |
|
|
op0_di = gen_reg_rtx (DImode);
|
648 |
|
|
|
649 |
|
|
if (! register_operand (operands[1], SImode))
|
650 |
|
|
operands[1] = force_reg (SImode, operands[1]);
|
651 |
|
|
op1_rf = gen_reg_rtx (RFmode);
|
652 |
|
|
expand_float (op1_rf, operands[1], 1);
|
653 |
|
|
|
654 |
|
|
if (! register_operand (operands[2], SImode))
|
655 |
|
|
operands[2] = force_reg (SImode, operands[2]);
|
656 |
|
|
op2_rf = gen_reg_rtx (RFmode);
|
657 |
|
|
expand_float (op2_rf, operands[2], 1);
|
658 |
|
|
|
659 |
|
|
emit_insn (gen_cond_trap (EQ, operands[2], CONST0_RTX (SImode),
|
660 |
|
|
CONST1_RTX (SImode)));
|
661 |
|
|
|
662 |
|
|
emit_insn (gen_divsi3_internal (op0_rf, op1_rf, op2_rf));
|
663 |
|
|
|
664 |
|
|
emit_insn (gen_fixuns_truncrfdi2_alts (op0_di, op0_rf, const1_rtx));
|
665 |
|
|
emit_move_insn (operands[0], gen_lowpart (SImode, op0_di));
|
666 |
|
|
DONE;
|
667 |
|
|
})
|
668 |
|
|
|
669 |
|
|
(define_expand "umodsi3"
|
670 |
|
|
[(set (match_operand:SI 0 "register_operand" "")
|
671 |
|
|
(umod:SI (match_operand:SI 1 "general_operand" "")
|
672 |
|
|
(match_operand:SI 2 "general_operand" "")))]
|
673 |
|
|
"TARGET_INLINE_INT_DIV"
|
674 |
|
|
{
|
675 |
|
|
rtx op2_neg, op1_di, div;
|
676 |
|
|
|
677 |
|
|
div = gen_reg_rtx (SImode);
|
678 |
|
|
emit_insn (gen_udivsi3 (div, operands[1], operands[2]));
|
679 |
|
|
|
680 |
|
|
op2_neg = expand_unop (SImode, neg_optab, operands[2], NULL_RTX, 0);
|
681 |
|
|
|
682 |
|
|
/* This is a trick to get us to reuse the value that we're sure to
|
683 |
|
|
have already copied to the FP regs. */
|
684 |
|
|
op1_di = gen_reg_rtx (DImode);
|
685 |
|
|
convert_move (op1_di, operands[1], 1);
|
686 |
|
|
|
687 |
|
|
emit_insn (gen_maddsi4 (operands[0], div, op2_neg,
|
688 |
|
|
gen_lowpart (SImode, op1_di)));
|
689 |
|
|
DONE;
|
690 |
|
|
})
|
691 |
|
|
|
692 |
|
|
(define_expand "divsi3_internal"
|
693 |
|
|
[(set (match_operand:RF 0 "fr_register_operand" "")
|
694 |
|
|
(float:RF (div:SI (match_operand:RF 1 "fr_register_operand" "")
|
695 |
|
|
(match_operand:RF 2 "fr_register_operand" ""))))]
|
696 |
|
|
"TARGET_INLINE_INT_DIV"
|
697 |
|
|
{
|
698 |
|
|
rtx a = operands[1];
|
699 |
|
|
rtx b = operands[2];
|
700 |
|
|
rtx y = gen_reg_rtx (RFmode);
|
701 |
|
|
rtx e = gen_reg_rtx (RFmode);
|
702 |
|
|
rtx e1 = gen_reg_rtx (RFmode);
|
703 |
|
|
rtx q = gen_reg_rtx (RFmode);
|
704 |
|
|
rtx q1 = gen_reg_rtx (RFmode);
|
705 |
|
|
rtx cond = gen_reg_rtx (BImode);
|
706 |
|
|
rtx zero = CONST0_RTX (RFmode);
|
707 |
|
|
rtx one = CONST1_RTX (RFmode);
|
708 |
|
|
rtx status1 = CONST1_RTX (SImode);
|
709 |
|
|
rtx trunc_off = CONST2_RTX (SImode);
|
710 |
|
|
rtx twon34_exp = gen_reg_rtx (DImode);
|
711 |
|
|
rtx twon34 = gen_reg_rtx (RFmode);
|
712 |
|
|
|
713 |
|
|
/* Load cosntant 2**(-34) */
|
714 |
|
|
emit_move_insn (twon34_exp, GEN_INT (65501));
|
715 |
|
|
emit_insn (gen_setf_exp_rf (twon34, twon34_exp));
|
716 |
|
|
|
717 |
|
|
/* y = 1 / b */
|
718 |
|
|
emit_insn (gen_recip_approx_rf (y, a, b, cond, status1));
|
719 |
|
|
/* e = 1 - (b * y) */
|
720 |
|
|
emit_insn (gen_m2subrf4_cond (e, cond, one, b, y, zero, status1, trunc_off));
|
721 |
|
|
/* q = a * y */
|
722 |
|
|
emit_insn (gen_mulrf3_cond (q, cond, a, y, zero, status1, trunc_off));
|
723 |
|
|
/* q1 = q + (q * e) */
|
724 |
|
|
emit_insn (gen_m2addrf4_cond (q1, cond, q, q, e, zero, status1, trunc_off));
|
725 |
|
|
/* e1 = (2**-34) + (e * e) */
|
726 |
|
|
emit_insn (gen_m2addrf4_cond (e1, cond, twon34, e, e, zero, status1, trunc_off));
|
727 |
|
|
/* q2 = q1 + (e1 * q1) */
|
728 |
|
|
emit_insn (gen_m2addrf4_cond (operands[0], cond, q1, e1, q1, y, status1, trunc_off));
|
729 |
|
|
DONE;
|
730 |
|
|
})
|
731 |
|
|
|
732 |
|
|
(define_expand "divdi3"
|
733 |
|
|
[(set (match_operand:DI 0 "register_operand" "")
|
734 |
|
|
(div:DI (match_operand:DI 1 "general_operand" "")
|
735 |
|
|
(match_operand:DI 2 "general_operand" "")))]
|
736 |
|
|
"TARGET_INLINE_INT_DIV"
|
737 |
|
|
{
|
738 |
|
|
rtx op1_rf, op2_rf, op0_rf;
|
739 |
|
|
|
740 |
|
|
op0_rf = gen_reg_rtx (RFmode);
|
741 |
|
|
|
742 |
|
|
if (! register_operand (operands[1], DImode))
|
743 |
|
|
operands[1] = force_reg (DImode, operands[1]);
|
744 |
|
|
op1_rf = gen_reg_rtx (RFmode);
|
745 |
|
|
expand_float (op1_rf, operands[1], 0);
|
746 |
|
|
|
747 |
|
|
if (! register_operand (operands[2], DImode))
|
748 |
|
|
operands[2] = force_reg (DImode, operands[2]);
|
749 |
|
|
op2_rf = gen_reg_rtx (RFmode);
|
750 |
|
|
expand_float (op2_rf, operands[2], 0);
|
751 |
|
|
|
752 |
|
|
emit_insn (gen_cond_trap (EQ, operands[2], CONST0_RTX (DImode),
|
753 |
|
|
CONST1_RTX (DImode)));
|
754 |
|
|
|
755 |
|
|
if (TARGET_INLINE_INT_DIV == INL_MIN_LAT)
|
756 |
|
|
emit_insn (gen_divdi3_internal_lat (op0_rf, op1_rf, op2_rf));
|
757 |
|
|
else
|
758 |
|
|
emit_insn (gen_divdi3_internal_thr (op0_rf, op1_rf, op2_rf));
|
759 |
|
|
|
760 |
|
|
emit_insn (gen_fix_truncrfdi2_alts (operands[0], op0_rf, const1_rtx));
|
761 |
|
|
DONE;
|
762 |
|
|
})
|
763 |
|
|
|
764 |
|
|
(define_expand "moddi3"
|
765 |
|
|
[(set (match_operand:DI 0 "register_operand" "")
|
766 |
|
|
(mod:SI (match_operand:DI 1 "general_operand" "")
|
767 |
|
|
(match_operand:DI 2 "general_operand" "")))]
|
768 |
|
|
"TARGET_INLINE_INT_DIV"
|
769 |
|
|
{
|
770 |
|
|
rtx op2_neg, div;
|
771 |
|
|
|
772 |
|
|
div = gen_reg_rtx (DImode);
|
773 |
|
|
emit_insn (gen_divdi3 (div, operands[1], operands[2]));
|
774 |
|
|
|
775 |
|
|
op2_neg = expand_unop (DImode, neg_optab, operands[2], NULL_RTX, 0);
|
776 |
|
|
|
777 |
|
|
emit_insn (gen_madddi4 (operands[0], div, op2_neg, operands[1]));
|
778 |
|
|
DONE;
|
779 |
|
|
})
|
780 |
|
|
|
781 |
|
|
(define_expand "udivdi3"
|
782 |
|
|
[(set (match_operand:DI 0 "register_operand" "")
|
783 |
|
|
(udiv:DI (match_operand:DI 1 "general_operand" "")
|
784 |
|
|
(match_operand:DI 2 "general_operand" "")))]
|
785 |
|
|
"TARGET_INLINE_INT_DIV"
|
786 |
|
|
{
|
787 |
|
|
rtx op1_rf, op2_rf, op0_rf;
|
788 |
|
|
|
789 |
|
|
op0_rf = gen_reg_rtx (RFmode);
|
790 |
|
|
|
791 |
|
|
if (! register_operand (operands[1], DImode))
|
792 |
|
|
operands[1] = force_reg (DImode, operands[1]);
|
793 |
|
|
op1_rf = gen_reg_rtx (RFmode);
|
794 |
|
|
expand_float (op1_rf, operands[1], 1);
|
795 |
|
|
|
796 |
|
|
if (! register_operand (operands[2], DImode))
|
797 |
|
|
operands[2] = force_reg (DImode, operands[2]);
|
798 |
|
|
op2_rf = gen_reg_rtx (RFmode);
|
799 |
|
|
expand_float (op2_rf, operands[2], 1);
|
800 |
|
|
|
801 |
|
|
emit_insn (gen_cond_trap (EQ, operands[2], CONST0_RTX (DImode),
|
802 |
|
|
CONST1_RTX (DImode)));
|
803 |
|
|
|
804 |
|
|
if (TARGET_INLINE_INT_DIV == INL_MIN_LAT)
|
805 |
|
|
emit_insn (gen_divdi3_internal_lat (op0_rf, op1_rf, op2_rf));
|
806 |
|
|
else
|
807 |
|
|
emit_insn (gen_divdi3_internal_thr (op0_rf, op1_rf, op2_rf));
|
808 |
|
|
|
809 |
|
|
emit_insn (gen_fixuns_truncrfdi2_alts (operands[0], op0_rf, const1_rtx));
|
810 |
|
|
DONE;
|
811 |
|
|
})
|
812 |
|
|
|
813 |
|
|
(define_expand "umoddi3"
|
814 |
|
|
[(set (match_operand:DI 0 "register_operand" "")
|
815 |
|
|
(umod:DI (match_operand:DI 1 "general_operand" "")
|
816 |
|
|
(match_operand:DI 2 "general_operand" "")))]
|
817 |
|
|
"TARGET_INLINE_INT_DIV"
|
818 |
|
|
{
|
819 |
|
|
rtx op2_neg, div;
|
820 |
|
|
|
821 |
|
|
div = gen_reg_rtx (DImode);
|
822 |
|
|
emit_insn (gen_udivdi3 (div, operands[1], operands[2]));
|
823 |
|
|
|
824 |
|
|
op2_neg = expand_unop (DImode, neg_optab, operands[2], NULL_RTX, 0);
|
825 |
|
|
|
826 |
|
|
emit_insn (gen_madddi4 (operands[0], div, op2_neg, operands[1]));
|
827 |
|
|
DONE;
|
828 |
|
|
})
|
829 |
|
|
|
830 |
|
|
(define_expand "divdi3_internal_lat"
|
831 |
|
|
[(set (match_operand:RF 0 "fr_register_operand" "")
|
832 |
|
|
(float:RF (div:DI (match_operand:RF 1 "fr_register_operand" "")
|
833 |
|
|
(match_operand:RF 2 "fr_register_operand" ""))))]
|
834 |
|
|
"TARGET_INLINE_INT_DIV"
|
835 |
|
|
{
|
836 |
|
|
rtx a = operands[1];
|
837 |
|
|
rtx b = operands[2];
|
838 |
|
|
rtx y = gen_reg_rtx (RFmode);
|
839 |
|
|
rtx y1 = gen_reg_rtx (RFmode);
|
840 |
|
|
rtx y2 = gen_reg_rtx (RFmode);
|
841 |
|
|
rtx e = gen_reg_rtx (RFmode);
|
842 |
|
|
rtx e1 = gen_reg_rtx (RFmode);
|
843 |
|
|
rtx q = gen_reg_rtx (RFmode);
|
844 |
|
|
rtx q1 = gen_reg_rtx (RFmode);
|
845 |
|
|
rtx q2 = gen_reg_rtx (RFmode);
|
846 |
|
|
rtx r = gen_reg_rtx (RFmode);
|
847 |
|
|
rtx cond = gen_reg_rtx (BImode);
|
848 |
|
|
rtx zero = CONST0_RTX (RFmode);
|
849 |
|
|
rtx one = CONST1_RTX (RFmode);
|
850 |
|
|
rtx status1 = CONST1_RTX (SImode);
|
851 |
|
|
rtx trunc_off = CONST2_RTX (SImode);
|
852 |
|
|
|
853 |
|
|
/* y = 1 / b */
|
854 |
|
|
emit_insn (gen_recip_approx_rf (y, a, b, cond, status1));
|
855 |
|
|
/* e = 1 - (b * y) */
|
856 |
|
|
emit_insn (gen_m2subrf4_cond (e, cond, one, b, y, zero, status1, trunc_off));
|
857 |
|
|
/* q = a * y */
|
858 |
|
|
emit_insn (gen_mulrf3_cond (q, cond, a, y, zero, status1, trunc_off));
|
859 |
|
|
/* q1 = q + (q * e) */
|
860 |
|
|
emit_insn (gen_m2addrf4_cond (q1, cond, q, q, e, zero, status1, trunc_off));
|
861 |
|
|
/* e1 = e * e */
|
862 |
|
|
emit_insn (gen_mulrf3_cond (e1, cond, e, e, zero, status1, trunc_off));
|
863 |
|
|
/* q2 = q1 + (e1 * q1) */
|
864 |
|
|
emit_insn (gen_m2addrf4_cond (q2, cond, q1, e1, q1, zero, status1, trunc_off));
|
865 |
|
|
/* y1 = y + (y * e) */
|
866 |
|
|
emit_insn (gen_m2addrf4_cond (y1, cond, y, y, e, zero, status1, trunc_off));
|
867 |
|
|
/* r = a - (b * q2) */
|
868 |
|
|
emit_insn (gen_m2subrf4_cond (r, cond, a, b, q2, zero, status1, trunc_off));
|
869 |
|
|
/* y2 = y1 + (y1 * e1) */
|
870 |
|
|
emit_insn (gen_m2addrf4_cond (y2, cond, y1, y1, e1, zero, status1, trunc_off));
|
871 |
|
|
/* q3 = q2 + (r * y2) */
|
872 |
|
|
emit_insn (gen_m2addrf4_cond (operands[0], cond, q2, r, y2, y, status1, trunc_off));
|
873 |
|
|
DONE;
|
874 |
|
|
})
|
875 |
|
|
|
876 |
|
|
(define_expand "divdi3_internal_thr"
|
877 |
|
|
[(set (match_operand:RF 0 "fr_register_operand" "")
|
878 |
|
|
(float:RF (div:DI (match_operand:RF 1 "fr_register_operand" "")
|
879 |
|
|
(match_operand:RF 2 "fr_register_operand" ""))))]
|
880 |
|
|
"TARGET_INLINE_INT_DIV"
|
881 |
|
|
{
|
882 |
|
|
rtx a = operands[1];
|
883 |
|
|
rtx b = operands[2];
|
884 |
|
|
rtx y = gen_reg_rtx (RFmode);
|
885 |
|
|
rtx y1 = gen_reg_rtx (RFmode);
|
886 |
|
|
rtx y2 = gen_reg_rtx (RFmode);
|
887 |
|
|
rtx e = gen_reg_rtx (RFmode);
|
888 |
|
|
rtx e1 = gen_reg_rtx (RFmode);
|
889 |
|
|
rtx q2 = gen_reg_rtx (RFmode);
|
890 |
|
|
rtx r = gen_reg_rtx (RFmode);
|
891 |
|
|
rtx cond = gen_reg_rtx (BImode);
|
892 |
|
|
rtx zero = CONST0_RTX (RFmode);
|
893 |
|
|
rtx one = CONST1_RTX (RFmode);
|
894 |
|
|
rtx status1 = CONST1_RTX (SImode);
|
895 |
|
|
rtx trunc_off = CONST2_RTX (SImode);
|
896 |
|
|
|
897 |
|
|
/* y = 1 / b */
|
898 |
|
|
emit_insn (gen_recip_approx_rf (y, a, b, cond, status1));
|
899 |
|
|
/* e = 1 - (b * y) */
|
900 |
|
|
emit_insn (gen_m2subrf4_cond (e, cond, one, b, y, zero, status1, trunc_off));
|
901 |
|
|
/* y1 = y + (y * e) */
|
902 |
|
|
emit_insn (gen_m2addrf4_cond (y1, cond, y, y, e, zero, status1, trunc_off));
|
903 |
|
|
/* e1 = e * e */
|
904 |
|
|
emit_insn (gen_mulrf3_cond (e1, cond, e, e, zero, status1, trunc_off));
|
905 |
|
|
/* y2 = y1 + (y1 * e1) */
|
906 |
|
|
emit_insn (gen_m2addrf4_cond (y2, cond, y1, y1, e1, zero, status1, trunc_off));
|
907 |
|
|
/* q2 = y2 * a */
|
908 |
|
|
emit_insn (gen_mulrf3_cond (q2, cond, y2, a, zero, status1, trunc_off));
|
909 |
|
|
/* r = a - (b * q2) */
|
910 |
|
|
emit_insn (gen_m2subrf4_cond (r, cond, a, b, q2, zero, status1, trunc_off));
|
911 |
|
|
/* q3 = q2 + (r * y2) */
|
912 |
|
|
emit_insn (gen_m2addrf4_cond (operands[0], cond, q2, r, y2, y, status1, trunc_off));
|
913 |
|
|
DONE;
|
914 |
|
|
})
|
915 |
|
|
|
916 |
|
|
;; SQRT operations
|
917 |
|
|
|
918 |
|
|
|
919 |
|
|
(define_insn "sqrt_approx_rf"
|
920 |
|
|
[(set (match_operand:RF 0 "fr_register_operand" "=f")
|
921 |
|
|
(unspec:RF [(match_operand:RF 1 "fr_reg_or_fp01_operand" "fG")]
|
922 |
|
|
UNSPEC_FR_SQRT_RECIP_APPROX_RES))
|
923 |
|
|
(set (match_operand:BI 2 "register_operand" "=c")
|
924 |
|
|
(unspec:BI [(match_dup 1)] UNSPEC_FR_SQRT_RECIP_APPROX))
|
925 |
|
|
(use (match_operand:SI 3 "const_int_operand" ""))]
|
926 |
|
|
""
|
927 |
|
|
"frsqrta.s%3 %0, %2 = %F1"
|
928 |
|
|
[(set_attr "itanium_class" "fmisc")
|
929 |
|
|
(set_attr "predicable" "no")])
|
930 |
|
|
|
931 |
|
|
(define_expand "sqrtsf2"
|
932 |
|
|
[(set (match_operand:SF 0 "fr_register_operand" "=&f")
|
933 |
|
|
(sqrt:SF (match_operand:SF 1 "fr_reg_or_fp01_operand" "fG")))]
|
934 |
|
|
"TARGET_INLINE_SQRT"
|
935 |
|
|
{
|
936 |
|
|
rtx insn;
|
937 |
|
|
if (TARGET_INLINE_SQRT == INL_MIN_LAT)
|
938 |
|
|
insn = gen_sqrtsf2_internal_lat (operands[0], operands[1]);
|
939 |
|
|
else
|
940 |
|
|
insn = gen_sqrtsf2_internal_thr (operands[0], operands[1]);
|
941 |
|
|
emit_insn (insn);
|
942 |
|
|
DONE;
|
943 |
|
|
})
|
944 |
|
|
|
945 |
|
|
(define_expand "sqrtsf2_internal_thr"
|
946 |
|
|
[(set (match_operand:SF 0 "fr_register_operand" "")
|
947 |
|
|
(sqrt:SF (match_operand:SF 1 "fr_register_operand" "")))]
|
948 |
|
|
"TARGET_INLINE_SQRT"
|
949 |
|
|
{
|
950 |
|
|
rtx y = gen_reg_rtx (RFmode);
|
951 |
|
|
rtx b = gen_reg_rtx (RFmode);
|
952 |
|
|
rtx g = gen_reg_rtx (RFmode);
|
953 |
|
|
rtx e = gen_reg_rtx (RFmode);
|
954 |
|
|
rtx s = gen_reg_rtx (RFmode);
|
955 |
|
|
rtx f = gen_reg_rtx (RFmode);
|
956 |
|
|
rtx y1 = gen_reg_rtx (RFmode);
|
957 |
|
|
rtx g1 = gen_reg_rtx (RFmode);
|
958 |
|
|
rtx h = gen_reg_rtx (RFmode);
|
959 |
|
|
rtx d = gen_reg_rtx (RFmode);
|
960 |
|
|
rtx g2 = gen_reg_rtx (RFmode);
|
961 |
|
|
rtx cond = gen_reg_rtx (BImode);
|
962 |
|
|
rtx zero = CONST0_RTX (RFmode);
|
963 |
|
|
rtx one = CONST1_RTX (RFmode);
|
964 |
|
|
rtx c1 = ia64_dconst_0_5();
|
965 |
|
|
rtx c2 = ia64_dconst_0_375();
|
966 |
|
|
rtx reg_df_c1 = gen_reg_rtx (DFmode);
|
967 |
|
|
rtx reg_df_c2 = gen_reg_rtx (DFmode);
|
968 |
|
|
rtx reg_rf_c1 = gen_reg_rtx (RFmode);
|
969 |
|
|
rtx reg_rf_c2 = gen_reg_rtx (RFmode);
|
970 |
|
|
rtx status0 = CONST0_RTX (SImode);
|
971 |
|
|
rtx status1 = CONST1_RTX (SImode);
|
972 |
|
|
rtx trunc_sgl = CONST0_RTX (SImode);
|
973 |
|
|
rtx trunc_off = CONST2_RTX (SImode);
|
974 |
|
|
|
975 |
|
|
/* Put needed constants into registers. */
|
976 |
|
|
emit_insn (gen_movdf (reg_df_c1, c1));
|
977 |
|
|
emit_insn (gen_movdf (reg_df_c2, c2));
|
978 |
|
|
emit_insn (gen_extenddfrf2 (reg_rf_c1, reg_df_c1));
|
979 |
|
|
emit_insn (gen_extenddfrf2 (reg_rf_c2, reg_df_c2));
|
980 |
|
|
/* Empty conversion to put input into RFmode. */
|
981 |
|
|
emit_insn (gen_extendsfrf2 (b, operands[1]));
|
982 |
|
|
/* y = sqrt (1 / b) */
|
983 |
|
|
emit_insn (gen_sqrt_approx_rf (y, b, cond, status0));
|
984 |
|
|
/* g = b * y */
|
985 |
|
|
emit_insn (gen_mulrf3_cond (g, cond, b, y, zero, status1, trunc_off));
|
986 |
|
|
/* e = 1 - (g * y) */
|
987 |
|
|
emit_insn (gen_m2subrf4_cond (e, cond, one, g, y, zero, status1, trunc_off));
|
988 |
|
|
/* s = 0.5 + (0.375 * e) */
|
989 |
|
|
emit_insn (gen_m2addrf4_cond (s, cond, reg_rf_c1, reg_rf_c2, e, zero, status1, trunc_off));
|
990 |
|
|
/* f = y * e */
|
991 |
|
|
emit_insn (gen_mulrf3_cond (f, cond, y, e, zero, status1, trunc_off));
|
992 |
|
|
/* y1 = y + (f * s) */
|
993 |
|
|
emit_insn (gen_m2addrf4_cond (y1, cond, y, f, s, zero, status1, trunc_off));
|
994 |
|
|
/* g1 = single (b * y1) */
|
995 |
|
|
emit_insn (gen_mulrf3_cond (g1, cond, b, y1, zero, status1, trunc_sgl));
|
996 |
|
|
/* h = 0.5 * y1 */
|
997 |
|
|
emit_insn (gen_mulrf3_cond (h, cond, reg_rf_c1, y1, zero, status1, trunc_off));
|
998 |
|
|
/* d = b - g1 * g1 */
|
999 |
|
|
emit_insn (gen_m2subrf4_cond (d, cond, b, g1, g1, zero, status1, trunc_off));
|
1000 |
|
|
/* g2 = single(g1 + (d * h)) */
|
1001 |
|
|
emit_insn (gen_m2addrf4_cond (g2, cond, g1, d, h, y, status0, trunc_sgl));
|
1002 |
|
|
/* Conversion back into SFmode. */
|
1003 |
|
|
emit_insn (gen_truncrfsf2 (operands[0], g2));
|
1004 |
|
|
DONE;
|
1005 |
|
|
})
|
1006 |
|
|
|
1007 |
|
|
(define_expand "sqrtsf2_internal_lat"
|
1008 |
|
|
[(set (match_operand:SF 0 "fr_register_operand" "")
|
1009 |
|
|
(sqrt:SF (match_operand:SF 1 "fr_register_operand" "")))]
|
1010 |
|
|
"TARGET_INLINE_SQRT"
|
1011 |
|
|
{
|
1012 |
|
|
rtx y = gen_reg_rtx (RFmode);
|
1013 |
|
|
rtx b = gen_reg_rtx (RFmode);
|
1014 |
|
|
rtx g = gen_reg_rtx (RFmode);
|
1015 |
|
|
rtx g1 = gen_reg_rtx (RFmode);
|
1016 |
|
|
rtx g2 = gen_reg_rtx (RFmode);
|
1017 |
|
|
rtx e = gen_reg_rtx (RFmode);
|
1018 |
|
|
rtx s = gen_reg_rtx (RFmode);
|
1019 |
|
|
rtx f = gen_reg_rtx (RFmode);
|
1020 |
|
|
rtx f1 = gen_reg_rtx (RFmode);
|
1021 |
|
|
rtx h = gen_reg_rtx (RFmode);
|
1022 |
|
|
rtx h1 = gen_reg_rtx (RFmode);
|
1023 |
|
|
rtx d = gen_reg_rtx (RFmode);
|
1024 |
|
|
rtx cond = gen_reg_rtx (BImode);
|
1025 |
|
|
rtx zero = CONST0_RTX (RFmode);
|
1026 |
|
|
rtx one = CONST1_RTX (RFmode);
|
1027 |
|
|
rtx c1 = ia64_dconst_0_5();
|
1028 |
|
|
rtx c2 = ia64_dconst_0_375();
|
1029 |
|
|
rtx reg_df_c1 = gen_reg_rtx (DFmode);
|
1030 |
|
|
rtx reg_df_c2 = gen_reg_rtx (DFmode);
|
1031 |
|
|
rtx reg_rf_c1 = gen_reg_rtx (RFmode);
|
1032 |
|
|
rtx reg_rf_c2 = gen_reg_rtx (RFmode);
|
1033 |
|
|
rtx status0 = CONST0_RTX (SImode);
|
1034 |
|
|
rtx status1 = CONST1_RTX (SImode);
|
1035 |
|
|
rtx trunc_sgl = CONST0_RTX (SImode);
|
1036 |
|
|
rtx trunc_off = CONST2_RTX (SImode);
|
1037 |
|
|
|
1038 |
|
|
/* Put needed constants into registers. */
|
1039 |
|
|
emit_insn (gen_movdf (reg_df_c1, c1));
|
1040 |
|
|
emit_insn (gen_movdf (reg_df_c2, c2));
|
1041 |
|
|
emit_insn (gen_extenddfrf2 (reg_rf_c1, reg_df_c1));
|
1042 |
|
|
emit_insn (gen_extenddfrf2 (reg_rf_c2, reg_df_c2));
|
1043 |
|
|
/* Empty conversion to put input into RFmode. */
|
1044 |
|
|
emit_insn (gen_extendsfrf2 (b, operands[1]));
|
1045 |
|
|
/* y = sqrt (1 / b) */
|
1046 |
|
|
emit_insn (gen_sqrt_approx_rf (y, b, cond, status0));
|
1047 |
|
|
/* g = b * y */
|
1048 |
|
|
emit_insn (gen_mulrf3_cond (g, cond, b, y, zero, status1, trunc_off));
|
1049 |
|
|
/* e = 1 - (g * y) */
|
1050 |
|
|
emit_insn (gen_m2subrf4_cond (e, cond, one, g, y, zero, status1, trunc_off));
|
1051 |
|
|
/* h = 0.5 * y */
|
1052 |
|
|
emit_insn (gen_mulrf3_cond (h, cond, reg_rf_c1, y, zero, status1, trunc_off));
|
1053 |
|
|
/* s = 0.5 + (0.375 * e) */
|
1054 |
|
|
emit_insn (gen_m2addrf4_cond (s, cond, reg_rf_c1, reg_rf_c2, e, zero, status1, trunc_off));
|
1055 |
|
|
/* f = e * g */
|
1056 |
|
|
emit_insn (gen_mulrf3_cond (f, cond, e, g, zero, status1, trunc_off));
|
1057 |
|
|
/* g1 = single (g + (f * s)) */
|
1058 |
|
|
emit_insn (gen_m2addrf4_cond (g1, cond, g, f, s, zero, status1, trunc_sgl));
|
1059 |
|
|
/* f1 = e * h */
|
1060 |
|
|
emit_insn (gen_mulrf3_cond (f1, cond, e, h, zero, status1, trunc_off));
|
1061 |
|
|
/* d = b - g1 * g1 */
|
1062 |
|
|
emit_insn (gen_m2subrf4_cond (d, cond, b, g1, g1, zero, status1, trunc_off));
|
1063 |
|
|
/* h1 = h + (f1 * s) */
|
1064 |
|
|
emit_insn (gen_m2addrf4_cond (h1, cond, h, f1, s, zero, status1, trunc_off));
|
1065 |
|
|
/* g2 = single(g1 + (d * h1)) */
|
1066 |
|
|
emit_insn (gen_m2addrf4_cond (g2, cond, g1, d, h1, y, status0, trunc_sgl));
|
1067 |
|
|
/* Conversion back into SFmode. */
|
1068 |
|
|
emit_insn (gen_truncrfsf2 (operands[0], g2));
|
1069 |
|
|
DONE;
|
1070 |
|
|
})
|
1071 |
|
|
|
1072 |
|
|
(define_expand "sqrtdf2"
|
1073 |
|
|
[(set (match_operand:DF 0 "fr_register_operand" "=&f")
|
1074 |
|
|
(sqrt:DF (match_operand:DF 1 "fr_reg_or_fp01_operand" "fG")))]
|
1075 |
|
|
"TARGET_INLINE_SQRT"
|
1076 |
|
|
{
|
1077 |
|
|
rtx insn;
|
1078 |
|
|
#if 0
|
1079 |
|
|
if (TARGET_INLINE_SQRT == INL_MIN_LAT)
|
1080 |
|
|
insn = gen_sqrtdf2_internal_lat (operands[0], operands[1]);
|
1081 |
|
|
else
|
1082 |
|
|
#endif
|
1083 |
|
|
insn = gen_sqrtdf2_internal_thr (operands[0], operands[1]);
|
1084 |
|
|
emit_insn (insn);
|
1085 |
|
|
DONE;
|
1086 |
|
|
})
|
1087 |
|
|
|
1088 |
|
|
(define_expand "sqrtdf2_internal_thr"
|
1089 |
|
|
[(set (match_operand:DF 0 "fr_register_operand" "")
|
1090 |
|
|
(sqrt:DF (match_operand:DF 1 "fr_register_operand" "")))]
|
1091 |
|
|
"TARGET_INLINE_SQRT"
|
1092 |
|
|
{
|
1093 |
|
|
rtx y = gen_reg_rtx (RFmode);
|
1094 |
|
|
rtx b = gen_reg_rtx (RFmode);
|
1095 |
|
|
rtx g = gen_reg_rtx (RFmode);
|
1096 |
|
|
rtx g1 = gen_reg_rtx (RFmode);
|
1097 |
|
|
rtx g2 = gen_reg_rtx (RFmode);
|
1098 |
|
|
rtx g3 = gen_reg_rtx (RFmode);
|
1099 |
|
|
rtx g4 = gen_reg_rtx (RFmode);
|
1100 |
|
|
rtx r = gen_reg_rtx (RFmode);
|
1101 |
|
|
rtx r1 = gen_reg_rtx (RFmode);
|
1102 |
|
|
rtx h = gen_reg_rtx (RFmode);
|
1103 |
|
|
rtx h1 = gen_reg_rtx (RFmode);
|
1104 |
|
|
rtx h2 = gen_reg_rtx (RFmode);
|
1105 |
|
|
rtx d = gen_reg_rtx (RFmode);
|
1106 |
|
|
rtx d1 = gen_reg_rtx (RFmode);
|
1107 |
|
|
rtx cond = gen_reg_rtx (BImode);
|
1108 |
|
|
rtx zero = CONST0_RTX (RFmode);
|
1109 |
|
|
rtx c1 = ia64_dconst_0_5();
|
1110 |
|
|
rtx reg_df_c1 = gen_reg_rtx (DFmode);
|
1111 |
|
|
rtx reg_rf_c1 = gen_reg_rtx (RFmode);
|
1112 |
|
|
rtx status0 = CONST0_RTX (SImode);
|
1113 |
|
|
rtx status1 = CONST1_RTX (SImode);
|
1114 |
|
|
rtx trunc_dbl = CONST1_RTX (SImode);
|
1115 |
|
|
rtx trunc_off = CONST2_RTX (SImode);
|
1116 |
|
|
|
1117 |
|
|
/* Put needed constants into registers. */
|
1118 |
|
|
emit_insn (gen_movdf (reg_df_c1, c1));
|
1119 |
|
|
emit_insn (gen_extenddfrf2 (reg_rf_c1, reg_df_c1));
|
1120 |
|
|
/* Empty conversion to put input into RFmode. */
|
1121 |
|
|
emit_insn (gen_extenddfrf2 (b, operands[1]));
|
1122 |
|
|
/* y = sqrt (1 / b) */
|
1123 |
|
|
emit_insn (gen_sqrt_approx_rf (y, b, cond, status0));
|
1124 |
|
|
/* g = b * y */
|
1125 |
|
|
emit_insn (gen_mulrf3_cond (g, cond, b, y, zero, status1, trunc_off));
|
1126 |
|
|
/* h = 0.5 * y */
|
1127 |
|
|
emit_insn (gen_mulrf3_cond (h, cond, reg_rf_c1, y, zero, status1, trunc_off));
|
1128 |
|
|
/* r = 0.5 - (g * h) */
|
1129 |
|
|
emit_insn (gen_m2subrf4_cond (r, cond, reg_rf_c1, g, h, zero, status1, trunc_off));
|
1130 |
|
|
/* g1 = g + (g * r) */
|
1131 |
|
|
emit_insn (gen_m2addrf4_cond (g1, cond, g, g, r, zero, status1, trunc_off));
|
1132 |
|
|
/* h1 = h + (h * r) */
|
1133 |
|
|
emit_insn (gen_m2addrf4_cond (h1, cond, h, h, r, zero, status1, trunc_off));
|
1134 |
|
|
/* r1 = 0.5 - (g1 * h1) */
|
1135 |
|
|
emit_insn (gen_m2subrf4_cond (r1, cond, reg_rf_c1, g1, h1, zero, status1, trunc_off));
|
1136 |
|
|
/* g2 = g1 + (g1 * r1) */
|
1137 |
|
|
emit_insn (gen_m2addrf4_cond (g2, cond, g1, g1, r1, zero, status1, trunc_off));
|
1138 |
|
|
/* h2 = h1 + (h1 * r1) */
|
1139 |
|
|
emit_insn (gen_m2addrf4_cond (h2, cond, h1, h1, r1, zero, status1, trunc_off));
|
1140 |
|
|
/* d = b - (g2 * g2) */
|
1141 |
|
|
emit_insn (gen_m2subrf4_cond (d, cond, b, g2, g2, zero, status1, trunc_off));
|
1142 |
|
|
/* g3 = g2 + (d * h2) */
|
1143 |
|
|
emit_insn (gen_m2addrf4_cond (g3, cond, g2, d, h2, zero, status1, trunc_off));
|
1144 |
|
|
/* d1 = b - (g3 * g3) */
|
1145 |
|
|
emit_insn (gen_m2subrf4_cond (d1, cond, b, g3, g3, zero, status1, trunc_off));
|
1146 |
|
|
/* g4 = g3 + (d1 * h2) */
|
1147 |
|
|
emit_insn (gen_m2addrf4_cond (g4, cond, g3, d1, h2, y, status1, trunc_dbl));
|
1148 |
|
|
/* Conversion back into SFmode. */
|
1149 |
|
|
emit_insn (gen_truncrfdf2 (operands[0], g4));
|
1150 |
|
|
DONE;
|
1151 |
|
|
})
|
1152 |
|
|
|
1153 |
|
|
(define_expand "sqrtxf2"
|
1154 |
|
|
[(set (match_operand:XF 0 "fr_register_operand" "")
|
1155 |
|
|
(sqrt:XF (match_operand:XF 1 "fr_register_operand" "")))]
|
1156 |
|
|
"TARGET_INLINE_SQRT"
|
1157 |
|
|
{
|
1158 |
|
|
rtx y = gen_reg_rtx (RFmode);
|
1159 |
|
|
rtx b = gen_reg_rtx (RFmode);
|
1160 |
|
|
rtx g = gen_reg_rtx (RFmode);
|
1161 |
|
|
rtx g1 = gen_reg_rtx (RFmode);
|
1162 |
|
|
rtx g2 = gen_reg_rtx (RFmode);
|
1163 |
|
|
rtx g3 = gen_reg_rtx (RFmode);
|
1164 |
|
|
rtx g4 = gen_reg_rtx (RFmode);
|
1165 |
|
|
rtx e = gen_reg_rtx (RFmode);
|
1166 |
|
|
rtx e1 = gen_reg_rtx (RFmode);
|
1167 |
|
|
rtx e2 = gen_reg_rtx (RFmode);
|
1168 |
|
|
rtx h = gen_reg_rtx (RFmode);
|
1169 |
|
|
rtx h1 = gen_reg_rtx (RFmode);
|
1170 |
|
|
rtx h2 = gen_reg_rtx (RFmode);
|
1171 |
|
|
rtx h3 = gen_reg_rtx (RFmode);
|
1172 |
|
|
rtx d = gen_reg_rtx (RFmode);
|
1173 |
|
|
rtx d1 = gen_reg_rtx (RFmode);
|
1174 |
|
|
rtx cond = gen_reg_rtx (BImode);
|
1175 |
|
|
rtx zero = CONST0_RTX (RFmode);
|
1176 |
|
|
rtx c1 = ia64_dconst_0_5();
|
1177 |
|
|
rtx reg_df_c1 = gen_reg_rtx (DFmode);
|
1178 |
|
|
rtx reg_rf_c1 = gen_reg_rtx (RFmode);
|
1179 |
|
|
rtx status0 = CONST0_RTX (SImode);
|
1180 |
|
|
rtx status1 = CONST1_RTX (SImode);
|
1181 |
|
|
rtx trunc_off = CONST2_RTX (SImode);
|
1182 |
|
|
|
1183 |
|
|
/* Put needed constants into registers. */
|
1184 |
|
|
emit_insn (gen_movdf (reg_df_c1, c1));
|
1185 |
|
|
emit_insn (gen_extenddfrf2 (reg_rf_c1, reg_df_c1));
|
1186 |
|
|
/* Empty conversion to put input into RFmode. */
|
1187 |
|
|
emit_insn (gen_extendxfrf2 (b, operands[1]));
|
1188 |
|
|
/* y = sqrt (1 / b) */
|
1189 |
|
|
emit_insn (gen_sqrt_approx_rf (y, b, cond, status0));
|
1190 |
|
|
/* g = b * y */
|
1191 |
|
|
emit_insn (gen_mulrf3_cond (g, cond, b, y, zero, status1, trunc_off));
|
1192 |
|
|
/* h = 0.5 * y */
|
1193 |
|
|
emit_insn (gen_mulrf3_cond (h, cond, reg_rf_c1, y, zero, status1, trunc_off));
|
1194 |
|
|
/* e = 0.5 - (g * h) */
|
1195 |
|
|
emit_insn (gen_m2subrf4_cond (e, cond, reg_rf_c1, g, h, zero, status1, trunc_off));
|
1196 |
|
|
/* g1 = g + (g * e) */
|
1197 |
|
|
emit_insn (gen_m2addrf4_cond (g1, cond, g, g, e, zero, status1, trunc_off));
|
1198 |
|
|
/* h1 = h + (h * e) */
|
1199 |
|
|
emit_insn (gen_m2addrf4_cond (h1, cond, h, h, e, zero, status1, trunc_off));
|
1200 |
|
|
/* e1 = 0.5 - (g1 * h1) */
|
1201 |
|
|
emit_insn (gen_m2subrf4_cond (e1, cond, reg_rf_c1, g1, h1, zero, status1, trunc_off));
|
1202 |
|
|
/* g2 = g1 + (g1 * e1) */
|
1203 |
|
|
emit_insn (gen_m2addrf4_cond (g2, cond, g1, g1, e1, zero, status1, trunc_off));
|
1204 |
|
|
/* h2 = h1 + (h1 * e1) */
|
1205 |
|
|
emit_insn (gen_m2addrf4_cond (h2, cond, h1, h1, e1, zero, status1, trunc_off));
|
1206 |
|
|
/* d = b - (g2 * g2) */
|
1207 |
|
|
emit_insn (gen_m2subrf4_cond (d, cond, b, g2, g2, zero, status1, trunc_off));
|
1208 |
|
|
/* e2 = 0.5 - (g2 * h2) */
|
1209 |
|
|
emit_insn (gen_m2subrf4_cond (e2, cond, reg_rf_c1, g2, h2, zero, status1, trunc_off));
|
1210 |
|
|
/* g3 = g2 + (d * h2) */
|
1211 |
|
|
emit_insn (gen_m2addrf4_cond (g3, cond, g2, d, h2, zero, status1, trunc_off));
|
1212 |
|
|
/* h3 = h2 + (e2 * h2) */
|
1213 |
|
|
emit_insn (gen_m2addrf4_cond (h3, cond, h2, e2, h2, zero, status1, trunc_off));
|
1214 |
|
|
/* d1 = b - (g3 * g3) */
|
1215 |
|
|
emit_insn (gen_m2subrf4_cond (d1, cond, b, g3, g3, zero, status1, trunc_off));
|
1216 |
|
|
/* g4 = g3 + (d1 * h3) */
|
1217 |
|
|
emit_insn (gen_m2addrf4_cond (g4, cond, g3, d1, h3, y, status1, trunc_off));
|
1218 |
|
|
/* Conversion back into SFmode. */
|
1219 |
|
|
emit_insn (gen_truncrfxf2 (operands[0], g4));
|
1220 |
|
|
DONE;
|
1221 |
|
|
})
|