0
|
1
|
|
2 ;; For the internal conditional math routines:
|
|
3
|
|
4 ;; operand 0 is always the result
|
|
5 ;; operand 1 is always the predicate
|
|
6 ;; operand 2, 3, and sometimes 4 are the input values.
|
|
7 ;; operand 4 or 5 is the floating point status register to use.
|
|
8 ;; operand 5 or 6 is the rounding to do. (0 = single, 1 = double, 2 = none)
|
|
9 ;;
|
|
10 ;; addrf3_cond - F0 = F2 + F3
|
|
11 ;; subrf3_cond - F0 = F2 - F3
|
|
12 ;; mulrf3_cond - F0 = F2 * F3
|
|
13 ;; nmulrf3_cond - F0 = - (F2 * F3)
|
|
14 ;; m1addrf4_cond - F0 = (F2 * F3) + F4
|
|
15 ;; m1subrf4_cond - F0 = (F2 * F3) - F4
|
|
16 ;; m2addrf4_cond - F0 = F2 + (F3 * F4)
|
|
17 ;; m2subrf4_cond - F0 = F2 - (F3 * F4)
|
|
18
|
|
19 ;; Basic plus/minus/mult operations
|
|
20
|
|
21 (define_insn "addrf3_cond"
|
|
22 [(set (match_operand:RF 0 "fr_register_operand" "=f,f")
|
|
23 (if_then_else:RF (ne:RF (match_operand:BI 1 "register_operand" "c,c")
|
|
24 (const_int 0))
|
|
25 (plus:RF
|
|
26 (match_operand:RF 2 "fr_reg_or_fp01_operand" "fG,fG")
|
|
27 (match_operand:RF 3 "fr_reg_or_fp01_operand" "fG,fG"))
|
|
28 (match_operand:RF 4 "fr_reg_or_0_operand" "0,H")))
|
|
29 (use (match_operand:SI 5 "const_int_operand" ""))
|
|
30 (use (match_operand:SI 6 "const_int_operand" ""))]
|
|
31 ""
|
|
32 "(%1) fadd%R6.s%5 %0 = %F2, %F3"
|
|
33 [(set_attr "itanium_class" "fmac")
|
|
34 (set_attr "predicable" "no")])
|
|
35
|
|
36 (define_insn "subrf3_cond"
|
|
37 [(set (match_operand:RF 0 "fr_register_operand" "=f,f")
|
|
38 (if_then_else:RF (ne:RF (match_operand:BI 1 "register_operand" "c,c")
|
|
39 (const_int 0))
|
|
40 (minus:RF
|
|
41 (match_operand:RF 2 "fr_reg_or_fp01_operand" "fG,fG")
|
|
42 (match_operand:RF 3 "fr_reg_or_fp01_operand" "fG,fG"))
|
|
43 (match_operand:RF 4 "fr_reg_or_0_operand" "0,H")))
|
|
44 (use (match_operand:SI 5 "const_int_operand" ""))
|
|
45 (use (match_operand:SI 6 "const_int_operand" ""))]
|
|
46 ""
|
|
47 "(%1) fsub%R6.s%5 %0 = %F2, %F3"
|
|
48 [(set_attr "itanium_class" "fmac")
|
|
49 (set_attr "predicable" "no")])
|
|
50
|
|
51 (define_insn "mulrf3_cond"
|
|
52 [(set (match_operand:RF 0 "fr_register_operand" "=f,f")
|
|
53 (if_then_else:RF (ne:RF (match_operand:BI 1 "register_operand" "c,c")
|
|
54 (const_int 0))
|
|
55 (mult:RF
|
|
56 (match_operand:RF 2 "fr_reg_or_fp01_operand" "fG,fG")
|
|
57 (match_operand:RF 3 "fr_reg_or_fp01_operand" "fG,fG"))
|
|
58 (match_operand:RF 4 "fr_reg_or_0_operand" "0,H")))
|
|
59 (use (match_operand:SI 5 "const_int_operand" ""))
|
|
60 (use (match_operand:SI 6 "const_int_operand" ""))]
|
|
61 ""
|
|
62 "(%1) fmpy%R6.s%5 %0 = %F2, %F3"
|
|
63 [(set_attr "itanium_class" "fmac")
|
|
64 (set_attr "predicable" "no")])
|
|
65
|
|
66 ;; neg-mult operation
|
|
67
|
|
68 (define_insn "nmulrf3_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 (neg:RF (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) fnmpy%R6.s%5 %0 = %F2, %F3"
|
|
80 [(set_attr "itanium_class" "fmac")
|
|
81 (set_attr "predicable" "no")])
|
|
82
|
|
83 ;; add-mult/sub-mult operations (mult as op1)
|
|
84
|
|
85 (define_insn "m1addrf4_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 (plus:RF
|
|
90 (mult:RF
|
|
91 (match_operand:RF 2 "fr_reg_or_fp01_operand" "fG,fG")
|
|
92 (match_operand:RF 3 "fr_reg_or_fp01_operand" "fG,fG"))
|
|
93 (match_operand:RF 4 "fr_reg_or_fp01_operand" "fG,fG"))
|
|
94 (match_operand:RF 5 "fr_reg_or_0_operand" "0,H")))
|
|
95 (use (match_operand:SI 6 "const_int_operand" ""))
|
|
96 (use (match_operand:SI 7 "const_int_operand" ""))]
|
|
97 ""
|
|
98 "(%1) fma%R7.s%6 %0 = %F2, %F3, %F4"
|
|
99 [(set_attr "itanium_class" "fmac")
|
|
100 (set_attr "predicable" "no")])
|
|
101
|
|
102 (define_insn "m1subrf4_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 (minus: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) fms%R7.s%6 %0 = %F2, %F3, %F4"
|
|
116 [(set_attr "itanium_class" "fmac")
|
|
117 (set_attr "predicable" "no")])
|
|
118
|
|
119 ;; add-mult/sub-mult operations (mult as op2)
|
|
120
|
|
121 (define_insn "m2addrf4_cond"
|
|
122 [(set (match_operand:RF 0 "fr_register_operand" "=f,f")
|
|
123 (if_then_else:RF (ne:RF (match_operand:BI 1 "register_operand" "c,c")
|
|
124 (const_int 0))
|
|
125 (plus:RF
|
|
126 (match_operand:RF 2 "fr_reg_or_fp01_operand" "fG,fG")
|
|
127 (mult:RF
|
|
128 (match_operand:RF 3 "fr_reg_or_fp01_operand" "fG,fG")
|
|
129 (match_operand:RF 4 "fr_reg_or_fp01_operand" "fG,fG")))
|
|
130 (match_operand:RF 5 "fr_reg_or_0_operand" "0,H")))
|
|
131 (use (match_operand:SI 6 "const_int_operand" ""))
|
|
132 (use (match_operand:SI 7 "const_int_operand" ""))]
|
|
133 ""
|
|
134 "(%1) fma%R7.s%6 %0 = %F3, %F4, %F2"
|
|
135 [(set_attr "itanium_class" "fmac")
|
|
136 (set_attr "predicable" "no")])
|
|
137
|
|
138 (define_insn "m2subrf4_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 (minus: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) fnma%R7.s%6 %0 = %F3, %F4, %F2"
|
|
152 [(set_attr "itanium_class" "fmac")
|
|
153 (set_attr "predicable" "no")])
|
|
154
|
|
155 ;; Conversions to/from RF and SF/DF/XF
|
|
156 ;; These conversions should not generate any code but make it possible
|
|
157 ;; for all the instructions used to implement floating point division
|
|
158 ;; to be written for RFmode only and to not have to handle multiple
|
|
159 ;; modes or to have to handle a register in more than one mode.
|
|
160
|
|
161 (define_mode_iterator SDX_F [SF DF XF])
|
|
162
|
|
163 (define_insn "extend<mode>rf2"
|
|
164 [(set (match_operand:RF 0 "fr_register_operand" "=f")
|
|
165 (float_extend:RF (match_operand:SDX_F 1 "fr_register_operand" "f")))]
|
|
166 ""
|
|
167 "#"
|
|
168 [(set_attr "itanium_class" "fmisc")
|
|
169 (set_attr "predicable" "yes")])
|
|
170
|
|
171 (define_split
|
|
172 [(set (match_operand:RF 0 "fr_register_operand" "")
|
|
173 (float_extend:RF (match_operand:SDX_F 1 "fr_register_operand" "")))]
|
|
174 "reload_completed"
|
|
175 [(set (match_dup 0) (match_dup 2))]
|
|
176 {
|
|
177 operands[2] = gen_rtx_REG (RFmode, REGNO (operands[1]));
|
|
178 })
|
|
179
|
|
180
|
|
181 (define_insn "truncrf<mode>2"
|
|
182 [(set (match_operand:SDX_F 0 "fr_register_operand" "=f")
|
|
183 (float_truncate:SDX_F (match_operand:RF 1 "fr_register_operand" "f")))]
|
|
184 ""
|
|
185 "#"
|
|
186 [(set_attr "itanium_class" "fmisc")
|
|
187 (set_attr "predicable" "yes")])
|
|
188
|
|
189 (define_split
|
|
190 [(set (match_operand:SDX_F 0 "fr_register_operand" "")
|
|
191 (float_truncate:SDX_F (match_operand:RF 1 "fr_register_operand" "")))]
|
|
192 "reload_completed"
|
|
193 [(set (match_dup 0) (match_dup 2))]
|
|
194 {
|
|
195 operands[2] = gen_rtx_REG (<MODE>mode, REGNO (operands[1]));
|
|
196 })
|
|
197
|
|
198 ;; Reciprocal approximation
|
|
199
|
|
200 (define_insn "recip_approx_rf"
|
|
201 [(set (match_operand:RF 0 "fr_register_operand" "=f")
|
|
202 (unspec:RF [(match_operand:RF 1 "fr_register_operand" "f")
|
|
203 (match_operand:RF 2 "fr_register_operand" "f")]
|
|
204 UNSPEC_FR_RECIP_APPROX_RES))
|
|
205 (set (match_operand:BI 3 "register_operand" "=c")
|
|
206 (unspec:BI [(match_dup 1) (match_dup 2)] UNSPEC_FR_RECIP_APPROX))
|
|
207 (use (match_operand:SI 4 "const_int_operand" ""))]
|
|
208 ""
|
|
209 "frcpa.s%4 %0, %3 = %1, %2"
|
|
210 [(set_attr "itanium_class" "fmisc")
|
|
211 (set_attr "predicable" "no")])
|
|
212
|
|
213 ;; Single precision floating point division (maximum throughput algorithm).
|
|
214
|
|
215 (define_expand "divsf3_internal_thr"
|
|
216 [(set (match_operand:SF 0 "fr_register_operand" "")
|
|
217 (div:SF (match_operand:SF 1 "fr_register_operand" "")
|
|
218 (match_operand:SF 2 "fr_register_operand" "")))]
|
|
219 "TARGET_INLINE_FLOAT_DIV"
|
|
220 {
|
|
221 rtx y = gen_reg_rtx (RFmode);
|
|
222 rtx a = gen_reg_rtx (RFmode);
|
|
223 rtx b = gen_reg_rtx (RFmode);
|
|
224 rtx e = gen_reg_rtx (RFmode);
|
|
225 rtx y1 = gen_reg_rtx (RFmode);
|
|
226 rtx y2 = gen_reg_rtx (RFmode);
|
|
227 rtx q = gen_reg_rtx (RFmode);
|
|
228 rtx r = gen_reg_rtx (RFmode);
|
|
229 rtx q_res = gen_reg_rtx (RFmode);
|
|
230 rtx cond = gen_reg_rtx (BImode);
|
|
231 rtx zero = CONST0_RTX (RFmode);
|
|
232 rtx one = CONST1_RTX (RFmode);
|
|
233 rtx status0 = CONST0_RTX (SImode);
|
|
234 rtx status1 = CONST1_RTX (SImode);
|
|
235 rtx trunc_sgl = CONST0_RTX (SImode);
|
|
236 rtx trunc_off = CONST2_RTX (SImode);
|
|
237
|
|
238 /* Empty conversions to put inputs into RFmode. */
|
|
239 emit_insn (gen_extendsfrf2 (a, operands[1]));
|
|
240 emit_insn (gen_extendsfrf2 (b, operands[2]));
|
|
241 /* y = 1 / b */
|
|
242 emit_insn (gen_recip_approx_rf (y, a, b, cond, status0));
|
|
243 /* e = 1 - (b * y) */
|
|
244 emit_insn (gen_m2subrf4_cond (e, cond, one, b, y, zero, status1, trunc_off));
|
|
245 /* y1 = y + (y * e) */
|
|
246 emit_insn (gen_m2addrf4_cond (y1, cond, y, y, e, zero, status1, trunc_off));
|
|
247 /* y2 = y + (y1 * e) */
|
|
248 emit_insn (gen_m2addrf4_cond (y2, cond, y, y1, e, zero, status1, trunc_off));
|
|
249 /* q = single(a * y2) */
|
|
250 emit_insn (gen_mulrf3_cond (q, cond, a, y2, zero, status1, trunc_sgl));
|
|
251 /* r = a - (q * b) */
|
|
252 emit_insn (gen_m2subrf4_cond (r, cond, a, q, b, zero, status1, trunc_off));
|
|
253 /* Q = single (q + (r * y2)) */
|
|
254 emit_insn (gen_m2addrf4_cond (q_res, cond, q, r, y2, y, status0, trunc_sgl));
|
|
255 /* Conversion back into SFmode. */
|
|
256 emit_insn (gen_truncrfsf2 (operands[0], q_res));
|
|
257 DONE;
|
|
258 })
|
|
259
|
|
260
|
|
261 ;; Double precision floating point division (maximum throughput algorithm).
|
|
262
|
|
263 (define_expand "divdf3_internal_thr"
|
|
264 [(set (match_operand:DF 0 "fr_register_operand" "")
|
|
265 (div:DF (match_operand:DF 1 "fr_register_operand" "")
|
|
266 (match_operand:DF 2 "fr_register_operand" "")))]
|
|
267 "TARGET_INLINE_FLOAT_DIV"
|
|
268 {
|
|
269 rtx q_res = gen_reg_rtx (RFmode);
|
|
270 rtx a = gen_reg_rtx (RFmode);
|
|
271 rtx b = gen_reg_rtx (RFmode);
|
|
272 rtx y = gen_reg_rtx (RFmode);
|
|
273 rtx e = gen_reg_rtx (RFmode);
|
|
274 rtx y1 = gen_reg_rtx (RFmode);
|
|
275 rtx e1 = gen_reg_rtx (RFmode);
|
|
276 rtx y2 = gen_reg_rtx (RFmode);
|
|
277 rtx e2 = gen_reg_rtx (RFmode);
|
|
278 rtx y3 = gen_reg_rtx (RFmode);
|
|
279 rtx q = gen_reg_rtx (RFmode);
|
|
280 rtx r = gen_reg_rtx (RFmode);
|
|
281 rtx cond = gen_reg_rtx (BImode);
|
|
282 rtx zero = CONST0_RTX (RFmode);
|
|
283 rtx one = CONST1_RTX (RFmode);
|
|
284 rtx status0 = CONST0_RTX (SImode);
|
|
285 rtx status1 = CONST1_RTX (SImode);
|
|
286 rtx trunc_dbl = CONST1_RTX (SImode);
|
|
287 rtx trunc_off = CONST2_RTX (SImode);
|
|
288 /* Empty conversions to put inputs into RFmode */
|
|
289 emit_insn (gen_extenddfrf2 (a, operands[1]));
|
|
290 emit_insn (gen_extenddfrf2 (b, operands[2]));
|
|
291 /* y = 1 / b */
|
|
292 emit_insn (gen_recip_approx_rf (y, a, b, cond, status0));
|
|
293 /* e = 1 - (b * y) */
|
|
294 emit_insn (gen_m2subrf4_cond (e, cond, one, b, y, zero, status1, trunc_off));
|
|
295 /* y1 = y + (y * e) */
|
|
296 emit_insn (gen_m2addrf4_cond (y1, cond, y, y, e, zero, status1, trunc_off));
|
|
297 /* e1 = e * e */
|
|
298 emit_insn (gen_mulrf3_cond (e1, cond, e, e, zero, status1, trunc_off));
|
|
299 /* y2 = y1 + (y1 * e1) */
|
|
300 emit_insn (gen_m2addrf4_cond (y2, cond, y1, y1, e1, zero, status1, trunc_off));
|
|
301 /* e2 = e1 * e1 */
|
|
302 emit_insn (gen_mulrf3_cond (e2, cond, e1, e1, zero, status1, trunc_off));
|
|
303 /* y3 = y2 + (y2 * e2) */
|
|
304 emit_insn (gen_m2addrf4_cond (y3, cond, y2, y2, e2, zero, status1, trunc_off));
|
|
305 /* q = double (a * y3) */
|
|
306 emit_insn (gen_mulrf3_cond (q, cond, a, y3, zero, status1, trunc_dbl));
|
|
307 /* r = a - (b * q) */
|
|
308 emit_insn (gen_m2subrf4_cond (r, cond, a, b, q, zero, status1, trunc_off));
|
|
309 /* Q = double (q + (r * y3)) */
|
|
310 emit_insn (gen_m2addrf4_cond (q_res, cond, q, r, y3, y, status0, trunc_dbl));
|
|
311 /* Conversion back into DFmode */
|
|
312 emit_insn (gen_truncrfdf2 (operands[0], q_res));
|
|
313 DONE;
|
|
314 })
|