57871462 |
1 | /* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * |
7e605697 |
2 | * linkage_arm.s for PCSX * |
0bbd1454 |
3 | * Copyright (C) 2009-2011 Ari64 * |
b1f89e6f |
4 | * Copyright (C) 2010-2013 GraÅžvydas "notaz" Ignotas * |
57871462 |
5 | * * |
6 | * This program is free software; you can redistribute it and/or modify * |
7 | * it under the terms of the GNU General Public License as published by * |
8 | * the Free Software Foundation; either version 2 of the License, or * |
9 | * (at your option) any later version. * |
10 | * * |
11 | * This program is distributed in the hope that it will be useful, * |
12 | * but WITHOUT ANY WARRANTY; without even the implied warranty of * |
13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the * |
14 | * GNU General Public License for more details. * |
15 | * * |
16 | * You should have received a copy of the GNU General Public License * |
17 | * along with this program; if not, write to the * |
18 | * Free Software Foundation, Inc., * |
19 | * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA. * |
20 | * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */ |
b021ee75 |
21 | |
665f33e1 |
22 | #include "arm_features.h" |
d148d265 |
23 | #include "new_dynarec_config.h" |
b1f89e6f |
24 | #include "linkage_offsets.h" |
25 | |
26 | |
27 | #ifdef __MACH__ |
28 | #define dynarec_local ESYM(dynarec_local) |
3d680478 |
29 | #define add_jump_out ESYM(add_jump_out) |
b1f89e6f |
30 | #define new_recompile_block ESYM(new_recompile_block) |
398d6924 |
31 | #define ndrc_try_restore_block ESYM(ndrc_try_restore_block) |
b1f89e6f |
32 | #define get_addr ESYM(get_addr) |
33 | #define get_addr_ht ESYM(get_addr_ht) |
b1f89e6f |
34 | #define gen_interupt ESYM(gen_interupt) |
b1f89e6f |
35 | #define invalidate_addr ESYM(invalidate_addr) |
81dbbf4c |
36 | #define gteCheckStallRaw ESYM(gteCheckStallRaw) |
d1150cd6 |
37 | #define psxException ESYM(psxException) |
b1f89e6f |
38 | #endif |
f95a77f7 |
39 | |
57871462 |
40 | .bss |
41 | .align 4 |
b1f89e6f |
42 | .global dynarec_local |
57871462 |
43 | .type dynarec_local, %object |
b1f89e6f |
44 | .size dynarec_local, LO_dynarec_local_size |
57871462 |
45 | dynarec_local: |
b1f89e6f |
46 | .space LO_dynarec_local_size |
47 | |
48 | #define DRC_VAR_(name, vname, size_) \ |
49 | vname = dynarec_local + LO_##name; \ |
50 | .global vname; \ |
51 | .type vname, %object; \ |
52 | .size vname, size_ |
53 | |
54 | #define DRC_VAR(name, size_) \ |
55 | DRC_VAR_(name, ESYM(name), size_) |
56 | |
57 | DRC_VAR(next_interupt, 4) |
58 | DRC_VAR(cycle_count, 4) |
59 | DRC_VAR(last_count, 4) |
60 | DRC_VAR(pending_exception, 4) |
61 | DRC_VAR(stop, 4) |
687b4580 |
62 | DRC_VAR(branch_target, 4) |
b1f89e6f |
63 | DRC_VAR(address, 4) |
7f94b097 |
64 | DRC_VAR(hack_addr, 4) |
b1f89e6f |
65 | DRC_VAR(psxRegs, LO_psxRegs_end - LO_psxRegs) |
f95a77f7 |
66 | |
67 | /* psxRegs */ |
7c3a5182 |
68 | @DRC_VAR(reg, 128) |
b1f89e6f |
69 | DRC_VAR(lo, 4) |
70 | DRC_VAR(hi, 4) |
71 | DRC_VAR(reg_cop0, 128) |
72 | DRC_VAR(reg_cop2d, 128) |
73 | DRC_VAR(reg_cop2c, 128) |
74 | DRC_VAR(pcaddr, 4) |
75 | @DRC_VAR(code, 4) |
76 | @DRC_VAR(cycle, 4) |
77 | @DRC_VAR(interrupt, 4) |
78 | @DRC_VAR(intCycle, 256) |
79 | |
80 | DRC_VAR(rcnts, 7*4*4) |
687b4580 |
81 | DRC_VAR(inv_code_start, 4) |
82 | DRC_VAR(inv_code_end, 4) |
b1f89e6f |
83 | DRC_VAR(mem_rtab, 4) |
84 | DRC_VAR(mem_wtab, 4) |
85 | DRC_VAR(psxH_ptr, 4) |
86 | DRC_VAR(zeromem_ptr, 4) |
687b4580 |
87 | DRC_VAR(invc_ptr, 4) |
c6d5790c |
88 | DRC_VAR(scratch_buf_ptr, 4) |
37387d8b |
89 | DRC_VAR(ram_offset, 4) |
b1f89e6f |
90 | DRC_VAR(mini_ht, 256) |
63cb0298 |
91 | |
57871462 |
92 | |
0e4ad319 |
93 | #ifdef TEXRELS_FORBIDDEN |
b861c0a9 |
94 | .data |
95 | .align 2 |
96 | ptr_jump_in: |
97 | .word ESYM(jump_in) |
b861c0a9 |
98 | ptr_hash_table: |
99 | .word ESYM(hash_table) |
100 | #endif |
101 | |
102 | |
103 | .syntax unified |
104 | .text |
105 | .align 2 |
106 | |
665f33e1 |
107 | #ifndef HAVE_ARMV5 |
108 | .macro blx rd |
109 | mov lr, pc |
110 | bx \rd |
111 | .endm |
112 | #endif |
113 | |
c67af2ac |
114 | .macro load_varadr reg var |
0e4ad319 |
115 | #if defined(HAVE_ARMV7) && defined(TEXRELS_FORBIDDEN) |
1f4e070a |
116 | movw \reg, #:lower16:(\var-(1678f+8)) |
117 | movt \reg, #:upper16:(\var-(1678f+8)) |
b861c0a9 |
118 | 1678: |
119 | add \reg, pc |
0e4ad319 |
120 | #elif defined(HAVE_ARMV7) && !defined(__PIC__) |
121 | movw \reg, #:lower16:\var |
122 | movt \reg, #:upper16:\var |
c67af2ac |
123 | #else |
274c4243 |
124 | ldr \reg, =\var |
c67af2ac |
125 | #endif |
274c4243 |
126 | .endm |
127 | |
b861c0a9 |
128 | .macro load_varadr_ext reg var |
0e4ad319 |
129 | #if defined(HAVE_ARMV7) && defined(TEXRELS_FORBIDDEN) |
1f4e070a |
130 | movw \reg, #:lower16:(ptr_\var-(1678f+8)) |
131 | movt \reg, #:upper16:(ptr_\var-(1678f+8)) |
b861c0a9 |
132 | 1678: |
133 | ldr \reg, [pc, \reg] |
134 | #else |
135 | load_varadr \reg \var |
136 | #endif |
137 | .endm |
138 | |
b1be1eee |
139 | .macro mov_16 reg imm |
8f2bb0cb |
140 | #ifdef HAVE_ARMV7 |
b1be1eee |
141 | movw \reg, #\imm |
c67af2ac |
142 | #else |
b1be1eee |
143 | mov \reg, #(\imm & 0x00ff) |
144 | orr \reg, #(\imm & 0xff00) |
c67af2ac |
145 | #endif |
b1be1eee |
146 | .endm |
147 | |
148 | .macro mov_24 reg imm |
8f2bb0cb |
149 | #ifdef HAVE_ARMV7 |
b1be1eee |
150 | movw \reg, #(\imm & 0xffff) |
151 | movt \reg, #(\imm >> 16) |
c67af2ac |
152 | #else |
b1be1eee |
153 | mov \reg, #(\imm & 0x0000ff) |
154 | orr \reg, #(\imm & 0x00ff00) |
155 | orr \reg, #(\imm & 0xff0000) |
c67af2ac |
156 | #endif |
b1be1eee |
157 | .endm |
158 | |
398d6924 |
159 | /* r4 = virtual target address */ |
160 | /* r5 = instruction to patch */ |
76f71c27 |
161 | .macro dyna_linker_main |
d148d265 |
162 | #ifndef NO_WRITE_EXEC |
b861c0a9 |
163 | load_varadr_ext r3, jump_in |
f968d35d |
164 | /* get_page */ |
398d6924 |
165 | lsr r2, r4, #12 |
f968d35d |
166 | mov r6, #4096 |
167 | bic r2, r2, #0xe0000 |
57871462 |
168 | sub r6, r6, #1 |
f968d35d |
169 | cmp r2, #0x1000 |
398d6924 |
170 | ldr r7, [r5] |
f968d35d |
171 | biclt r2, #0x0e00 |
172 | and r6, r6, r2 |
57871462 |
173 | cmp r2, #2048 |
174 | add r12, r7, #2 |
175 | orrcs r2, r6, #2048 |
398d6924 |
176 | ldr r1, [r3, r2, lsl #2] |
57871462 |
177 | lsl r12, r12, #8 |
398d6924 |
178 | add r6, r5, r12, asr #6 /* old target */ |
76f71c27 |
179 | mov r8, #0 |
57871462 |
180 | /* jump_in lookup */ |
76f71c27 |
181 | 1: |
398d6924 |
182 | movs r0, r1 |
76f71c27 |
183 | beq 2f |
398d6924 |
184 | ldr r3, [r1] /* ll_entry .vaddr */ |
185 | ldrd r0, r1, [r0, #8] /* ll_entry .addr, .next */ |
186 | teq r3, r4 |
76f71c27 |
187 | bne 1b |
398d6924 |
188 | teq r0, r6 |
189 | moveq pc, r0 /* Stale i-cache */ |
190 | mov r8, r0 |
76f71c27 |
191 | b 1b /* jump_in may have dupes, continue search */ |
192 | 2: |
193 | tst r8, r8 |
398d6924 |
194 | beq 3f /* r4 not in jump_in */ |
76f71c27 |
195 | |
398d6924 |
196 | mov r0, r4 |
76f71c27 |
197 | mov r1, r6 |
3d680478 |
198 | bl add_jump_out |
76f71c27 |
199 | sub r2, r8, r5 |
57871462 |
200 | and r1, r7, #0xff000000 |
201 | lsl r2, r2, #6 |
202 | sub r1, r1, #2 |
203 | add r1, r1, r2, lsr #8 |
204 | str r1, [r5] |
76f71c27 |
205 | mov pc, r8 |
206 | 3: |
398d6924 |
207 | mov r0, r4 |
208 | bl ndrc_try_restore_block |
209 | tst r0, r0 |
210 | movne pc, r0 |
d148d265 |
211 | #else |
212 | /* XXX: should be able to do better than this... */ |
398d6924 |
213 | mov r0, r4 |
d148d265 |
214 | bl get_addr_ht |
215 | mov pc, r0 |
216 | #endif |
76f71c27 |
217 | .endm |
218 | |
5c6457c3 |
219 | |
220 | FUNCTION(dyna_linker): |
76f71c27 |
221 | /* r0 = virtual target address */ |
222 | /* r1 = instruction to patch */ |
57871462 |
223 | mov r4, r0 |
224 | mov r5, r1 |
398d6924 |
225 | 10: |
226 | dyna_linker_main |
227 | |
228 | mov r0, r4 |
57871462 |
229 | bl new_recompile_block |
230 | tst r0, r0 |
398d6924 |
231 | beq 10b |
232 | |
57871462 |
233 | /* pagefault */ |
398d6924 |
234 | mov r0, r4 |
57871462 |
235 | mov r1, r0 |
b4ab351d |
236 | mov r2, #(4<<2) /* Address error (fetch) */ |
57871462 |
237 | .size dyna_linker, .-dyna_linker |
5c6457c3 |
238 | |
239 | FUNCTION(exec_pagefault): |
57871462 |
240 | /* r0 = instruction pointer */ |
241 | /* r1 = fault address */ |
242 | /* r2 = cause */ |
b1f89e6f |
243 | ldr r3, [fp, #LO_reg_cop0+48] /* Status */ |
b1f89e6f |
244 | str r0, [fp, #LO_reg_cop0+56] /* EPC */ |
57871462 |
245 | orr r3, r3, #2 |
b1f89e6f |
246 | str r1, [fp, #LO_reg_cop0+32] /* BadVAddr */ |
b1f89e6f |
247 | str r3, [fp, #LO_reg_cop0+48] /* Status */ |
b1f89e6f |
248 | str r2, [fp, #LO_reg_cop0+52] /* Cause */ |
57871462 |
249 | mov r0, #0x80000000 |
b4ab351d |
250 | orr r0, r0, #0x80 |
57871462 |
251 | bl get_addr_ht |
252 | mov pc, r0 |
253 | .size exec_pagefault, .-exec_pagefault |
7139f3c8 |
254 | |
57871462 |
255 | /* Special dynamic linker for the case where a page fault |
256 | may occur in a branch delay slot */ |
5c6457c3 |
257 | FUNCTION(dyna_linker_ds): |
57871462 |
258 | /* r0 = virtual target address */ |
259 | /* r1 = instruction to patch */ |
57871462 |
260 | mov r4, r0 |
57871462 |
261 | mov r5, r1 |
398d6924 |
262 | 10: |
263 | dyna_linker_main |
264 | |
265 | bic r0, r4, #7 |
57871462 |
266 | orr r0, r0, #1 |
267 | bl new_recompile_block |
268 | tst r0, r0 |
398d6924 |
269 | beq 10b |
270 | |
57871462 |
271 | /* pagefault */ |
398d6924 |
272 | mov r0, r4 |
57871462 |
273 | bic r1, r0, #7 |
274 | mov r2, #0x80000008 /* High bit set indicates pagefault in delay slot */ |
275 | sub r0, r1, #4 |
276 | b exec_pagefault |
277 | .size dyna_linker_ds, .-dyna_linker_ds |
7139f3c8 |
278 | |
57871462 |
279 | .align 2 |
5c6457c3 |
280 | |
281 | FUNCTION(jump_vaddr_r0): |
57871462 |
282 | eor r2, r0, r0, lsl #16 |
283 | b jump_vaddr |
284 | .size jump_vaddr_r0, .-jump_vaddr_r0 |
5c6457c3 |
285 | FUNCTION(jump_vaddr_r1): |
57871462 |
286 | eor r2, r1, r1, lsl #16 |
287 | mov r0, r1 |
288 | b jump_vaddr |
289 | .size jump_vaddr_r1, .-jump_vaddr_r1 |
5c6457c3 |
290 | FUNCTION(jump_vaddr_r2): |
57871462 |
291 | mov r0, r2 |
292 | eor r2, r2, r2, lsl #16 |
293 | b jump_vaddr |
294 | .size jump_vaddr_r2, .-jump_vaddr_r2 |
5c6457c3 |
295 | FUNCTION(jump_vaddr_r3): |
57871462 |
296 | eor r2, r3, r3, lsl #16 |
297 | mov r0, r3 |
298 | b jump_vaddr |
299 | .size jump_vaddr_r3, .-jump_vaddr_r3 |
5c6457c3 |
300 | FUNCTION(jump_vaddr_r4): |
57871462 |
301 | eor r2, r4, r4, lsl #16 |
302 | mov r0, r4 |
303 | b jump_vaddr |
304 | .size jump_vaddr_r4, .-jump_vaddr_r4 |
5c6457c3 |
305 | FUNCTION(jump_vaddr_r5): |
57871462 |
306 | eor r2, r5, r5, lsl #16 |
307 | mov r0, r5 |
308 | b jump_vaddr |
309 | .size jump_vaddr_r5, .-jump_vaddr_r5 |
5c6457c3 |
310 | FUNCTION(jump_vaddr_r6): |
57871462 |
311 | eor r2, r6, r6, lsl #16 |
312 | mov r0, r6 |
313 | b jump_vaddr |
314 | .size jump_vaddr_r6, .-jump_vaddr_r6 |
5c6457c3 |
315 | FUNCTION(jump_vaddr_r8): |
57871462 |
316 | eor r2, r8, r8, lsl #16 |
317 | mov r0, r8 |
318 | b jump_vaddr |
319 | .size jump_vaddr_r8, .-jump_vaddr_r8 |
5c6457c3 |
320 | FUNCTION(jump_vaddr_r9): |
57871462 |
321 | eor r2, r9, r9, lsl #16 |
322 | mov r0, r9 |
323 | b jump_vaddr |
324 | .size jump_vaddr_r9, .-jump_vaddr_r9 |
5c6457c3 |
325 | FUNCTION(jump_vaddr_r10): |
57871462 |
326 | eor r2, r10, r10, lsl #16 |
327 | mov r0, r10 |
328 | b jump_vaddr |
329 | .size jump_vaddr_r10, .-jump_vaddr_r10 |
5c6457c3 |
330 | FUNCTION(jump_vaddr_r12): |
57871462 |
331 | eor r2, r12, r12, lsl #16 |
332 | mov r0, r12 |
333 | b jump_vaddr |
334 | .size jump_vaddr_r12, .-jump_vaddr_r12 |
5c6457c3 |
335 | FUNCTION(jump_vaddr_r7): |
57871462 |
336 | eor r2, r7, r7, lsl #16 |
337 | add r0, r7, #0 |
338 | .size jump_vaddr_r7, .-jump_vaddr_r7 |
5c6457c3 |
339 | FUNCTION(jump_vaddr): |
b861c0a9 |
340 | load_varadr_ext r1, hash_table |
57871462 |
341 | mvn r3, #15 |
342 | and r2, r3, r2, lsr #12 |
343 | ldr r2, [r1, r2]! |
344 | teq r2, r0 |
df4dc2b1 |
345 | ldreq pc, [r1, #8] |
346 | ldr r2, [r1, #4] |
57871462 |
347 | teq r2, r0 |
348 | ldreq pc, [r1, #12] |
b1f89e6f |
349 | str r10, [fp, #LO_cycle_count] |
57871462 |
350 | bl get_addr |
b1f89e6f |
351 | ldr r10, [fp, #LO_cycle_count] |
57871462 |
352 | mov pc, r0 |
353 | .size jump_vaddr, .-jump_vaddr |
7139f3c8 |
354 | |
57871462 |
355 | .align 2 |
5c6457c3 |
356 | |
357 | FUNCTION(verify_code_ds): |
3968e69e |
358 | str r8, [fp, #LO_branch_target] @ preserve HOST_BTREG? |
5c6457c3 |
359 | FUNCTION(verify_code): |
57871462 |
360 | /* r1 = source */ |
361 | /* r2 = target */ |
362 | /* r3 = length */ |
363 | tst r3, #4 |
364 | mov r4, #0 |
365 | add r3, r1, r3 |
366 | mov r5, #0 |
367 | ldrne r4, [r1], #4 |
368 | mov r12, #0 |
369 | ldrne r5, [r2], #4 |
370 | teq r1, r3 |
371 | beq .D3 |
372 | .D2: |
373 | ldr r7, [r1], #4 |
374 | eor r9, r4, r5 |
375 | ldr r8, [r2], #4 |
376 | orrs r9, r9, r12 |
377 | bne .D4 |
378 | ldr r4, [r1], #4 |
379 | eor r12, r7, r8 |
380 | ldr r5, [r2], #4 |
381 | cmp r1, r3 |
382 | bcc .D2 |
383 | teq r7, r8 |
384 | .D3: |
385 | teqeq r4, r5 |
386 | .D4: |
b1f89e6f |
387 | ldr r8, [fp, #LO_branch_target] |
57871462 |
388 | moveq pc, lr |
389 | .D5: |
390 | bl get_addr |
391 | mov pc, r0 |
392 | .size verify_code, .-verify_code |
7c3a5182 |
393 | .size verify_code_ds, .-verify_code_ds |
7139f3c8 |
394 | |
57871462 |
395 | .align 2 |
5c6457c3 |
396 | FUNCTION(cc_interrupt): |
b1f89e6f |
397 | ldr r0, [fp, #LO_last_count] |
57871462 |
398 | mov r1, #0 |
57871462 |
399 | add r10, r0, r10 |
b1f89e6f |
400 | str r1, [fp, #LO_pending_exception] |
b1f89e6f |
401 | str r10, [fp, #LO_cycle] /* PCSX cycles */ |
b4ab351d |
402 | @@ str r10, [fp, #LO_reg_cop0+36] /* Count - not on PSX */ |
57871462 |
403 | mov r10, lr |
398d6924 |
404 | |
57871462 |
405 | bl gen_interupt |
406 | mov lr, r10 |
b1f89e6f |
407 | ldr r10, [fp, #LO_cycle] |
408 | ldr r0, [fp, #LO_next_interupt] |
409 | ldr r1, [fp, #LO_pending_exception] |
410 | ldr r2, [fp, #LO_stop] |
411 | str r0, [fp, #LO_last_count] |
57871462 |
412 | sub r10, r10, r0 |
413 | tst r2, r2 |
b861c0a9 |
414 | ldmfdne sp!, {r4, r5, r6, r7, r8, r9, sl, fp, ip, pc} |
57871462 |
415 | tst r1, r1 |
416 | moveq pc, lr |
b1f89e6f |
417 | ldr r0, [fp, #LO_pcaddr] |
57871462 |
418 | bl get_addr_ht |
419 | mov pc, r0 |
57871462 |
420 | .size cc_interrupt, .-cc_interrupt |
7139f3c8 |
421 | |
57871462 |
422 | .align 2 |
5c6457c3 |
423 | FUNCTION(fp_exception): |
57871462 |
424 | mov r2, #0x10000000 |
425 | .E7: |
b1f89e6f |
426 | ldr r1, [fp, #LO_reg_cop0+48] /* Status */ |
57871462 |
427 | mov r3, #0x80000000 |
b1f89e6f |
428 | str r0, [fp, #LO_reg_cop0+56] /* EPC */ |
57871462 |
429 | orr r1, #2 |
430 | add r2, r2, #0x2c |
b1f89e6f |
431 | str r1, [fp, #LO_reg_cop0+48] /* Status */ |
432 | str r2, [fp, #LO_reg_cop0+52] /* Cause */ |
7139f3c8 |
433 | add r0, r3, #0x80 |
57871462 |
434 | bl get_addr_ht |
435 | mov pc, r0 |
436 | .size fp_exception, .-fp_exception |
437 | .align 2 |
5c6457c3 |
438 | FUNCTION(fp_exception_ds): |
57871462 |
439 | mov r2, #0x90000000 /* Set high bit if delay slot */ |
440 | b .E7 |
441 | .size fp_exception_ds, .-fp_exception_ds |
7139f3c8 |
442 | |
57871462 |
443 | .align 2 |
d1150cd6 |
444 | FUNCTION(jump_break_ds): |
445 | mov r0, #0x24 |
446 | mov r1, #1 |
447 | b call_psxException |
448 | FUNCTION(jump_break): |
449 | mov r0, #0x24 |
450 | mov r1, #0 |
451 | b call_psxException |
452 | FUNCTION(jump_syscall_ds): |
453 | mov r0, #0x20 |
454 | mov r1, #1 |
455 | b call_psxException |
5c6457c3 |
456 | FUNCTION(jump_syscall): |
d1150cd6 |
457 | mov r0, #0x20 |
458 | mov r1, #0 |
459 | |
460 | call_psxException: |
461 | ldr r3, [fp, #LO_last_count] |
462 | str r2, [fp, #LO_pcaddr] |
463 | add r10, r3, r10 |
464 | str r10, [fp, #LO_cycle] /* PCSX cycles */ |
465 | bl psxException |
7139f3c8 |
466 | |
b1f89e6f |
467 | /* note: psxException might do recursive recompiler call from it's HLE code, |
7139f3c8 |
468 | * so be ready for this */ |
3968e69e |
469 | FUNCTION(jump_to_new_pc): |
b1f89e6f |
470 | ldr r1, [fp, #LO_next_interupt] |
471 | ldr r10, [fp, #LO_cycle] |
472 | ldr r0, [fp, #LO_pcaddr] |
822b27d1 |
473 | sub r10, r10, r1 |
b1f89e6f |
474 | str r1, [fp, #LO_last_count] |
7139f3c8 |
475 | bl get_addr_ht |
476 | mov pc, r0 |
3968e69e |
477 | .size jump_to_new_pc, .-jump_to_new_pc |
0d16cda2 |
478 | |
7139f3c8 |
479 | .align 2 |
5c6457c3 |
480 | FUNCTION(new_dyna_leave): |
b1f89e6f |
481 | ldr r0, [fp, #LO_last_count] |
7139f3c8 |
482 | add r12, fp, #28 |
483 | add r10, r0, r10 |
b1f89e6f |
484 | str r10, [fp, #LO_cycle] |
b021ee75 |
485 | ldmfd sp!, {r4, r5, r6, r7, r8, r9, sl, fp, ip, pc} |
7139f3c8 |
486 | .size new_dyna_leave, .-new_dyna_leave |
487 | |
0bbd1454 |
488 | .align 2 |
5c6457c3 |
489 | FUNCTION(invalidate_addr_r0): |
5df0e313 |
490 | stmia fp, {r0, r1, r2, r3, EXTRA_UNSAVED_REGS r12, lr} |
0bbd1454 |
491 | b invalidate_addr_call |
492 | .size invalidate_addr_r0, .-invalidate_addr_r0 |
493 | .align 2 |
5c6457c3 |
494 | FUNCTION(invalidate_addr_r1): |
5df0e313 |
495 | stmia fp, {r0, r1, r2, r3, EXTRA_UNSAVED_REGS r12, lr} |
9be4ba64 |
496 | mov r0, r1 |
0bbd1454 |
497 | b invalidate_addr_call |
498 | .size invalidate_addr_r1, .-invalidate_addr_r1 |
499 | .align 2 |
5c6457c3 |
500 | FUNCTION(invalidate_addr_r2): |
5df0e313 |
501 | stmia fp, {r0, r1, r2, r3, EXTRA_UNSAVED_REGS r12, lr} |
9be4ba64 |
502 | mov r0, r2 |
0bbd1454 |
503 | b invalidate_addr_call |
504 | .size invalidate_addr_r2, .-invalidate_addr_r2 |
505 | .align 2 |
5c6457c3 |
506 | FUNCTION(invalidate_addr_r3): |
5df0e313 |
507 | stmia fp, {r0, r1, r2, r3, EXTRA_UNSAVED_REGS r12, lr} |
9be4ba64 |
508 | mov r0, r3 |
0bbd1454 |
509 | b invalidate_addr_call |
510 | .size invalidate_addr_r3, .-invalidate_addr_r3 |
511 | .align 2 |
5c6457c3 |
512 | FUNCTION(invalidate_addr_r4): |
5df0e313 |
513 | stmia fp, {r0, r1, r2, r3, EXTRA_UNSAVED_REGS r12, lr} |
9be4ba64 |
514 | mov r0, r4 |
0bbd1454 |
515 | b invalidate_addr_call |
516 | .size invalidate_addr_r4, .-invalidate_addr_r4 |
517 | .align 2 |
5c6457c3 |
518 | FUNCTION(invalidate_addr_r5): |
5df0e313 |
519 | stmia fp, {r0, r1, r2, r3, EXTRA_UNSAVED_REGS r12, lr} |
9be4ba64 |
520 | mov r0, r5 |
0bbd1454 |
521 | b invalidate_addr_call |
522 | .size invalidate_addr_r5, .-invalidate_addr_r5 |
523 | .align 2 |
5c6457c3 |
524 | FUNCTION(invalidate_addr_r6): |
5df0e313 |
525 | stmia fp, {r0, r1, r2, r3, EXTRA_UNSAVED_REGS r12, lr} |
9be4ba64 |
526 | mov r0, r6 |
0bbd1454 |
527 | b invalidate_addr_call |
528 | .size invalidate_addr_r6, .-invalidate_addr_r6 |
529 | .align 2 |
5c6457c3 |
530 | FUNCTION(invalidate_addr_r7): |
5df0e313 |
531 | stmia fp, {r0, r1, r2, r3, EXTRA_UNSAVED_REGS r12, lr} |
9be4ba64 |
532 | mov r0, r7 |
0bbd1454 |
533 | b invalidate_addr_call |
534 | .size invalidate_addr_r7, .-invalidate_addr_r7 |
535 | .align 2 |
5c6457c3 |
536 | FUNCTION(invalidate_addr_r8): |
5df0e313 |
537 | stmia fp, {r0, r1, r2, r3, EXTRA_UNSAVED_REGS r12, lr} |
9be4ba64 |
538 | mov r0, r8 |
0bbd1454 |
539 | b invalidate_addr_call |
540 | .size invalidate_addr_r8, .-invalidate_addr_r8 |
541 | .align 2 |
5c6457c3 |
542 | FUNCTION(invalidate_addr_r9): |
5df0e313 |
543 | stmia fp, {r0, r1, r2, r3, EXTRA_UNSAVED_REGS r12, lr} |
9be4ba64 |
544 | mov r0, r9 |
0bbd1454 |
545 | b invalidate_addr_call |
546 | .size invalidate_addr_r9, .-invalidate_addr_r9 |
547 | .align 2 |
5c6457c3 |
548 | FUNCTION(invalidate_addr_r10): |
5df0e313 |
549 | stmia fp, {r0, r1, r2, r3, EXTRA_UNSAVED_REGS r12, lr} |
9be4ba64 |
550 | mov r0, r10 |
0bbd1454 |
551 | b invalidate_addr_call |
552 | .size invalidate_addr_r10, .-invalidate_addr_r10 |
553 | .align 2 |
5c6457c3 |
554 | FUNCTION(invalidate_addr_r12): |
5df0e313 |
555 | stmia fp, {r0, r1, r2, r3, EXTRA_UNSAVED_REGS r12, lr} |
9be4ba64 |
556 | mov r0, r12 |
0bbd1454 |
557 | .size invalidate_addr_r12, .-invalidate_addr_r12 |
558 | .align 2 |
b1f89e6f |
559 | invalidate_addr_call: |
560 | ldr r12, [fp, #LO_inv_code_start] |
561 | ldr lr, [fp, #LO_inv_code_end] |
9be4ba64 |
562 | cmp r0, r12 |
563 | cmpcs lr, r0 |
564 | blcc invalidate_addr |
5df0e313 |
565 | ldmia fp, {r0, r1, r2, r3, EXTRA_UNSAVED_REGS r12, pc} |
0bbd1454 |
566 | .size invalidate_addr_call, .-invalidate_addr_call |
567 | |
57871462 |
568 | .align 2 |
5c6457c3 |
569 | FUNCTION(new_dyna_start): |
b021ee75 |
570 | /* ip is stored to conform EABI alignment */ |
571 | stmfd sp!, {r4, r5, r6, r7, r8, r9, sl, fp, ip, lr} |
be516ebe |
572 | mov fp, r0 /* dynarec_local */ |
b1f89e6f |
573 | ldr r0, [fp, #LO_pcaddr] |
7139f3c8 |
574 | bl get_addr_ht |
b1f89e6f |
575 | ldr r1, [fp, #LO_next_interupt] |
576 | ldr r10, [fp, #LO_cycle] |
577 | str r1, [fp, #LO_last_count] |
7139f3c8 |
578 | sub r10, r10, r1 |
579 | mov pc, r0 |
57871462 |
580 | .size new_dyna_start, .-new_dyna_start |
7139f3c8 |
581 | |
7e605697 |
582 | /* --------------------------------------- */ |
7139f3c8 |
583 | |
7e605697 |
584 | .align 2 |
c6c3b1b3 |
585 | |
586 | .macro pcsx_read_mem readop tab_shift |
587 | /* r0 = address, r1 = handler_tab, r2 = cycles */ |
588 | lsl r3, r0, #20 |
589 | lsr r3, #(20+\tab_shift) |
b1f89e6f |
590 | ldr r12, [fp, #LO_last_count] |
c6c3b1b3 |
591 | ldr r1, [r1, r3, lsl #2] |
592 | add r2, r2, r12 |
593 | lsls r1, #1 |
594 | .if \tab_shift == 1 |
595 | lsl r3, #1 |
596 | \readop r0, [r1, r3] |
597 | .else |
598 | \readop r0, [r1, r3, lsl #\tab_shift] |
599 | .endif |
600 | movcc pc, lr |
b1f89e6f |
601 | str r2, [fp, #LO_cycle] |
c6c3b1b3 |
602 | bx r1 |
603 | .endm |
604 | |
5c6457c3 |
605 | FUNCTION(jump_handler_read8): |
c6c3b1b3 |
606 | add r1, #0x1000/4*4 + 0x1000/2*4 @ shift to r8 part |
b861c0a9 |
607 | pcsx_read_mem ldrbcc, 0 |
c6c3b1b3 |
608 | |
5c6457c3 |
609 | FUNCTION(jump_handler_read16): |
c6c3b1b3 |
610 | add r1, #0x1000/4*4 @ shift to r16 part |
10858959 |
611 | pcsx_read_mem ldrhcc, 1 |
c6c3b1b3 |
612 | |
5c6457c3 |
613 | FUNCTION(jump_handler_read32): |
c6c3b1b3 |
614 | pcsx_read_mem ldrcc, 2 |
615 | |
b96d3df7 |
616 | |
9b9af0d1 |
617 | .macro memhandler_post |
618 | ldr r0, [fp, #LO_next_interupt] |
619 | ldr r2, [fp, #LO_cycle] @ memhandlers can modify cc, like dma |
620 | str r0, [fp, #LO_last_count] |
621 | sub r0, r2, r0 |
622 | .endm |
623 | |
b96d3df7 |
624 | .macro pcsx_write_mem wrtop tab_shift |
625 | /* r0 = address, r1 = data, r2 = cycles, r3 = handler_tab */ |
626 | lsl r12,r0, #20 |
627 | lsr r12, #(20+\tab_shift) |
628 | ldr r3, [r3, r12, lsl #2] |
b1f89e6f |
629 | str r0, [fp, #LO_address] @ some handlers still need it.. |
b96d3df7 |
630 | lsls r3, #1 |
9b9af0d1 |
631 | mov r0, r2 @ cycle return in case of direct store |
b96d3df7 |
632 | .if \tab_shift == 1 |
633 | lsl r12, #1 |
634 | \wrtop r1, [r3, r12] |
635 | .else |
636 | \wrtop r1, [r3, r12, lsl #\tab_shift] |
637 | .endif |
638 | movcc pc, lr |
b1f89e6f |
639 | ldr r12, [fp, #LO_last_count] |
b96d3df7 |
640 | mov r0, r1 |
641 | add r2, r2, r12 |
b1f89e6f |
642 | str r2, [fp, #LO_cycle] |
9b9af0d1 |
643 | |
644 | str lr, [fp, #LO_saved_lr] |
b96d3df7 |
645 | blx r3 |
9b9af0d1 |
646 | ldr lr, [fp, #LO_saved_lr] |
b96d3df7 |
647 | |
9b9af0d1 |
648 | memhandler_post |
687b4580 |
649 | bx lr |
b96d3df7 |
650 | .endm |
651 | |
5c6457c3 |
652 | FUNCTION(jump_handler_write8): |
b96d3df7 |
653 | add r3, #0x1000/4*4 + 0x1000/2*4 @ shift to r8 part |
b861c0a9 |
654 | pcsx_write_mem strbcc, 0 |
b96d3df7 |
655 | |
5c6457c3 |
656 | FUNCTION(jump_handler_write16): |
b96d3df7 |
657 | add r3, #0x1000/4*4 @ shift to r16 part |
b861c0a9 |
658 | pcsx_write_mem strhcc, 1 |
b96d3df7 |
659 | |
5c6457c3 |
660 | FUNCTION(jump_handler_write32): |
b96d3df7 |
661 | pcsx_write_mem strcc, 2 |
662 | |
5c6457c3 |
663 | FUNCTION(jump_handler_write_h): |
b96d3df7 |
664 | /* r0 = address, r1 = data, r2 = cycles, r3 = handler */ |
b1f89e6f |
665 | ldr r12, [fp, #LO_last_count] |
666 | str r0, [fp, #LO_address] @ some handlers still need it.. |
b96d3df7 |
667 | add r2, r2, r12 |
668 | mov r0, r1 |
b1f89e6f |
669 | str r2, [fp, #LO_cycle] |
9b9af0d1 |
670 | |
671 | str lr, [fp, #LO_saved_lr] |
b96d3df7 |
672 | blx r3 |
9b9af0d1 |
673 | ldr lr, [fp, #LO_saved_lr] |
b96d3df7 |
674 | |
9b9af0d1 |
675 | memhandler_post |
687b4580 |
676 | bx lr |
b96d3df7 |
677 | |
5c6457c3 |
678 | FUNCTION(jump_handle_swl): |
b96d3df7 |
679 | /* r0 = address, r1 = data, r2 = cycles */ |
b1f89e6f |
680 | ldr r3, [fp, #LO_mem_wtab] |
b96d3df7 |
681 | mov r12,r0,lsr #12 |
682 | ldr r3, [r3, r12, lsl #2] |
683 | lsls r3, #1 |
684 | bcs 4f |
685 | add r3, r0, r3 |
686 | mov r0, r2 |
687 | tst r3, #2 |
688 | beq 101f |
689 | tst r3, #1 |
690 | beq 2f |
691 | 3: |
692 | str r1, [r3, #-3] |
693 | bx lr |
694 | 2: |
695 | lsr r2, r1, #8 |
696 | lsr r1, #24 |
697 | strh r2, [r3, #-2] |
698 | strb r1, [r3] |
699 | bx lr |
700 | 101: |
701 | tst r3, #1 |
702 | lsrne r1, #16 @ 1 |
703 | lsreq r12, r1, #24 @ 0 |
b861c0a9 |
704 | strhne r1, [r3, #-1] |
705 | strbeq r12, [r3] |
b96d3df7 |
706 | bx lr |
707 | 4: |
708 | mov r0, r2 |
63cb0298 |
709 | @ b abort |
b96d3df7 |
710 | bx lr @ TODO? |
711 | |
712 | |
5c6457c3 |
713 | FUNCTION(jump_handle_swr): |
b96d3df7 |
714 | /* r0 = address, r1 = data, r2 = cycles */ |
b1f89e6f |
715 | ldr r3, [fp, #LO_mem_wtab] |
b96d3df7 |
716 | mov r12,r0,lsr #12 |
717 | ldr r3, [r3, r12, lsl #2] |
718 | lsls r3, #1 |
719 | bcs 4f |
720 | add r3, r0, r3 |
721 | and r12,r3, #3 |
722 | mov r0, r2 |
723 | cmp r12,#2 |
b861c0a9 |
724 | strbgt r1, [r3] @ 3 |
725 | strheq r1, [r3] @ 2 |
b96d3df7 |
726 | cmp r12,#1 |
727 | strlt r1, [r3] @ 0 |
728 | bxne lr |
729 | lsr r2, r1, #8 @ 1 |
730 | strb r1, [r3] |
731 | strh r2, [r3, #1] |
732 | bx lr |
733 | 4: |
734 | mov r0, r2 |
63cb0298 |
735 | @ b abort |
b96d3df7 |
736 | bx lr @ TODO? |
737 | |
738 | |
b1be1eee |
739 | .macro rcntx_read_mode0 num |
740 | /* r0 = address, r2 = cycles */ |
b1f89e6f |
741 | ldr r3, [fp, #LO_rcnts+6*4+7*4*\num] @ cycleStart |
b1be1eee |
742 | mov r0, r2, lsl #16 |
b861c0a9 |
743 | sub r0, r0, r3, lsl #16 |
b1be1eee |
744 | lsr r0, #16 |
745 | bx lr |
746 | .endm |
747 | |
5c6457c3 |
748 | FUNCTION(rcnt0_read_count_m0): |
b1be1eee |
749 | rcntx_read_mode0 0 |
750 | |
5c6457c3 |
751 | FUNCTION(rcnt1_read_count_m0): |
b1be1eee |
752 | rcntx_read_mode0 1 |
753 | |
5c6457c3 |
754 | FUNCTION(rcnt2_read_count_m0): |
b1be1eee |
755 | rcntx_read_mode0 2 |
756 | |
5c6457c3 |
757 | FUNCTION(rcnt0_read_count_m1): |
b1be1eee |
758 | /* r0 = address, r2 = cycles */ |
b1f89e6f |
759 | ldr r3, [fp, #LO_rcnts+6*4+7*4*0] @ cycleStart |
b1be1eee |
760 | mov_16 r1, 0x3334 |
761 | sub r2, r2, r3 |
762 | mul r0, r1, r2 @ /= 5 |
763 | lsr r0, #16 |
764 | bx lr |
765 | |
5c6457c3 |
766 | FUNCTION(rcnt1_read_count_m1): |
b1be1eee |
767 | /* r0 = address, r2 = cycles */ |
b1f89e6f |
768 | ldr r3, [fp, #LO_rcnts+6*4+7*4*1] |
b1be1eee |
769 | mov_24 r1, 0x1e6cde |
770 | sub r2, r2, r3 |
771 | umull r3, r0, r1, r2 @ ~ /= hsync_cycles, max ~0x1e6cdd |
772 | bx lr |
773 | |
5c6457c3 |
774 | FUNCTION(rcnt2_read_count_m1): |
b1be1eee |
775 | /* r0 = address, r2 = cycles */ |
b1f89e6f |
776 | ldr r3, [fp, #LO_rcnts+6*4+7*4*2] |
b1be1eee |
777 | mov r0, r2, lsl #16-3 |
b861c0a9 |
778 | sub r0, r0, r3, lsl #16-3 |
b1be1eee |
779 | lsr r0, #16 @ /= 8 |
780 | bx lr |
781 | |
81dbbf4c |
782 | FUNCTION(call_gteStall): |
783 | /* r0 = op_cycles, r1 = cycles */ |
784 | ldr r2, [fp, #LO_last_count] |
785 | str lr, [fp, #LO_saved_lr] |
786 | add r1, r1, r2 |
787 | str r1, [fp, #LO_cycle] |
788 | add r1, fp, #LO_psxRegs |
789 | bl gteCheckStallRaw |
790 | ldr lr, [fp, #LO_saved_lr] |
791 | add r10, r10, r0 |
792 | bx lr |
793 | |
cdc2da64 |
794 | #ifdef HAVE_ARMV6 |
795 | |
796 | FUNCTION(get_reg): |
797 | ldr r12, [r0] |
798 | and r1, r1, #0xff |
799 | ldr r2, [r0, #4] |
800 | orr r1, r1, r1, lsl #8 |
801 | ldr r3, [r0, #8] |
802 | orr r1, r1, r1, lsl #16 @ searched char in every byte |
803 | ldrb r0, [r0, #12] @ last byte |
804 | eor r12, r12, r1 |
805 | eor r2, r2, r1 |
806 | eor r3, r3, r1 |
807 | cmp r0, r1, lsr #24 |
808 | mov r0, #12 |
809 | mvn r1, #0 @ r1=~0 |
810 | bxeq lr |
811 | orr r3, r3, #0xff000000 @ EXCLUDE_REG |
812 | uadd8 r0, r12, r1 @ add and set GE bits when not 0 (match) |
813 | mov r12, #0 |
814 | sel r0, r12, r1 @ 0 if no match, else ff in some byte |
815 | uadd8 r2, r2, r1 |
816 | sel r2, r12, r1 |
817 | uadd8 r3, r3, r1 |
818 | sel r3, r12, r1 |
819 | mov r12, #3 |
820 | clz r0, r0 @ 0, 8, 16, 24 or 32 |
821 | clz r2, r2 |
822 | clz r3, r3 |
823 | sub r0, r12, r0, lsr #3 @ 3, 2, 1, 0 or -1 |
824 | sub r2, r12, r2, lsr #3 |
825 | sub r3, r12, r3, lsr #3 |
826 | orr r2, r2, #4 |
827 | orr r3, r3, #8 |
828 | and r0, r0, r2 |
829 | and r0, r0, r3 |
830 | bx lr |
831 | |
832 | #endif /* HAVE_ARMV6 */ |
833 | |
7e605697 |
834 | @ vim:filetype=armasm |