1 /* $NetBSD: altivec.c,v 1.1 2003/02/03 17:10:09 matt Exp $ */ 2 3 /* 4 * Copyright (C) 1996 Wolfgang Solfrank. 5 * Copyright (C) 1996 TooLs GmbH. 6 * All rights reserved. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. All advertising materials mentioning features or use of this software 17 * must display the following acknowledgement: 18 * This product includes software developed by TooLs GmbH. 19 * 4. The name of TooLs GmbH may not be used to endorse or promote products 20 * derived from this software without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY TOOLS GMBH ``AS IS'' AND ANY EXPRESS OR 23 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 24 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 25 * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, 27 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; 28 * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, 29 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR 30 * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF 31 * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 32 */ 33 #include <sys/param.h> 34 #include <sys/proc.h> 35 #include <sys/sa.h> 36 #include <sys/systm.h> 37 #include <sys/user.h> 38 #include <sys/malloc.h> 39 #include <sys/pool.h> 40 41 #include <powerpc/altivec.h> 42 #include <powerpc/spr.h> 43 #include <powerpc/psl.h> 44 45 struct pool vecpool; 46 47 void 48 enable_vec() 49 { 50 struct cpu_info *ci = curcpu(); 51 struct lwp *l = curlwp; 52 struct pcb *pcb = &l->l_addr->u_pcb; 53 struct trapframe *tf = trapframe(l); 54 struct vreg *vr = pcb->pcb_vr; 55 int msr, scratch; 56 57 KASSERT(pcb->pcb_veccpu == NULL); 58 59 /* 60 * Allocate a vreg structure if we haven't done so. 61 */ 62 if (!(pcb->pcb_flags & PCB_ALTIVEC)) { 63 vr = pcb->pcb_vr = pool_get(&vecpool, PR_WAITOK); 64 pcb->pcb_flags |= PCB_ALTIVEC; 65 66 /* 67 * Initialize the vectors with NaNs 68 */ 69 70 for (scratch = 0; scratch < 32; scratch++) { 71 vr->vreg[scratch][0] = 0x7FFFDEAD; 72 vr->vreg[scratch][1] = 0x7FFFDEAD; 73 vr->vreg[scratch][2] = 0x7FFFDEAD; 74 vr->vreg[scratch][3] = 0x7FFFDEAD; 75 } 76 vr->vscr = 0; 77 vr->vrsave = tf->tf_xtra[TF_VRSAVE]; 78 } 79 80 /* 81 * Enable AltiVec temporarily (and disable interrupts). 82 */ 83 msr = mfmsr(); 84 mtmsr((msr & ~PSL_EE) | PSL_VEC); 85 __asm __volatile ("isync"); 86 if (ci->ci_veclwp) { 87 save_vec_cpu(); 88 } 89 KASSERT(curcpu()->ci_veclwp == NULL); 90 91 /* 92 * Restore VSCR by first loading it into a vector and then into VSCR. 93 * (this needs to done before loading the user's vector registers 94 * since we need to use a scratch vector register) 95 */ 96 __asm __volatile("vxor %2,%2,%2; lvewx %2,%0,%1; mtvscr %2" \ 97 :: "r"(vr), "r"(offsetof(struct vreg, vscr)), "n"(0)); 98 99 /* 100 * VRSAVE will be restored when trap frame returns 101 */ 102 tf->tf_xtra[TF_VRSAVE] = vr->vrsave; 103 104 #define LVX(n,vr) __asm /*__volatile*/("lvx %2,%0,%1" \ 105 :: "r"(vr), "r"(offsetof(struct vreg, vreg[n])), "n"(n)); 106 107 /* 108 * Load all 32 vector registers 109 */ 110 LVX( 0,vr); LVX( 1,vr); LVX( 2,vr); LVX( 3,vr); 111 LVX( 4,vr); LVX( 5,vr); LVX( 6,vr); LVX( 7,vr); 112 LVX( 8,vr); LVX( 9,vr); LVX(10,vr); LVX(11,vr); 113 LVX(12,vr); LVX(13,vr); LVX(14,vr); LVX(15,vr); 114 115 LVX(16,vr); LVX(17,vr); LVX(18,vr); LVX(19,vr); 116 LVX(20,vr); LVX(21,vr); LVX(22,vr); LVX(23,vr); 117 LVX(24,vr); LVX(25,vr); LVX(26,vr); LVX(27,vr); 118 LVX(28,vr); LVX(29,vr); LVX(30,vr); LVX(31,vr); 119 __asm __volatile ("isync"); 120 121 /* 122 * Enable AltiVec when we return to user-mode. 123 * Record the new ownership of the AltiVec unit. 124 */ 125 tf->srr1 |= PSL_VEC; 126 curcpu()->ci_veclwp = l; 127 pcb->pcb_veccpu = curcpu(); 128 __asm __volatile ("sync"); 129 130 /* 131 * Restore MSR (turn off AltiVec) 132 */ 133 mtmsr(msr); 134 } 135 136 void 137 save_vec_cpu(void) 138 { 139 struct cpu_info *ci = curcpu(); 140 struct lwp *l; 141 struct pcb *pcb; 142 struct vreg *vr; 143 struct trapframe *tf; 144 int msr; 145 146 /* 147 * Turn on AltiVEC, turn off interrupts. 148 */ 149 msr = mfmsr(); 150 mtmsr((msr & ~PSL_EE) | PSL_VEC); 151 __asm __volatile ("isync"); 152 l = ci->ci_veclwp; 153 if (l == NULL) { 154 goto out; 155 } 156 pcb = &l->l_addr->u_pcb; 157 vr = pcb->pcb_vr; 158 tf = trapframe(l); 159 160 #define STVX(n,vr) __asm /*__volatile*/("stvx %2,%0,%1" \ 161 :: "r"(vr), "r"(offsetof(struct vreg, vreg[n])), "n"(n)); 162 163 /* 164 * Save the vector registers. 165 */ 166 STVX( 0,vr); STVX( 1,vr); STVX( 2,vr); STVX( 3,vr); 167 STVX( 4,vr); STVX( 5,vr); STVX( 6,vr); STVX( 7,vr); 168 STVX( 8,vr); STVX( 9,vr); STVX(10,vr); STVX(11,vr); 169 STVX(12,vr); STVX(13,vr); STVX(14,vr); STVX(15,vr); 170 171 STVX(16,vr); STVX(17,vr); STVX(18,vr); STVX(19,vr); 172 STVX(20,vr); STVX(21,vr); STVX(22,vr); STVX(23,vr); 173 STVX(24,vr); STVX(25,vr); STVX(26,vr); STVX(27,vr); 174 STVX(28,vr); STVX(29,vr); STVX(30,vr); STVX(31,vr); 175 176 /* 177 * Save VSCR (this needs to be done after save the vector registers 178 * since we need to use one as scratch). 179 */ 180 __asm __volatile("mfvscr %2; stvewx %2,%0,%1" \ 181 :: "r"(vr), "r"(offsetof(struct vreg, vscr)), "n"(0)); 182 183 /* 184 * Save VRSAVE 185 */ 186 vr->vrsave = tf->tf_xtra[TF_VRSAVE]; 187 188 /* 189 * Note that we aren't using any CPU resources and stop any 190 * data streams. 191 */ 192 tf->srr1 &= ~PSL_VEC; 193 pcb->pcb_veccpu = NULL; 194 ci->ci_veclwp = NULL; 195 __asm __volatile ("dssall; sync"); 196 197 out: 198 199 /* 200 * Restore MSR (turn off AltiVec) 201 */ 202 mtmsr(msr); 203 } 204 205 /* 206 * Save a process's AltiVEC state to its PCB. The state may be in any CPU. 207 * The process must either be curproc or traced by curproc (and stopped). 208 * (The point being that the process must not run on another CPU during 209 * this function). 210 */ 211 void 212 save_vec_lwp(l) 213 struct lwp *l; 214 { 215 struct pcb *pcb = &l->l_addr->u_pcb; 216 struct cpu_info *ci = curcpu(); 217 218 /* 219 * If it's already in the PCB, there's nothing to do. 220 */ 221 222 if (pcb->pcb_veccpu == NULL) { 223 return; 224 } 225 226 /* 227 * If the state is in the current CPU, just flush the current CPU's 228 * state. 229 */ 230 231 if (l == ci->ci_veclwp) { 232 save_vec_cpu(); 233 return; 234 } 235 236 #ifdef MULTIPROCESSOR 237 238 /* 239 * It must be on another CPU, flush it from there. 240 */ 241 242 mp_save_vec_lwp(l); 243 #endif 244 } 245 246 #define ZERO_VEC 19 247 248 void 249 vzeropage(paddr_t pa) 250 { 251 const paddr_t ea = pa + NBPG; 252 uint32_t vec[7], *vp = (void *) roundup((uintptr_t) vec, 16); 253 uint32_t omsr, msr; 254 255 __asm __volatile("mfmsr %0" : "=r"(omsr) :); 256 257 /* 258 * Turn on AltiVec, turn off interrupts. 259 */ 260 msr = (omsr & ~PSL_EE) | PSL_VEC; 261 __asm __volatile("sync; mtmsr %0; isync" :: "r"(msr)); 262 263 /* 264 * Save the VEC register we are going to use before we disable 265 * relocation. 266 */ 267 __asm("stvx %1,0,%0" :: "r"(vp), "n"(ZERO_VEC)); 268 __asm("vxor %0,%0,%0" :: "n"(ZERO_VEC)); 269 270 /* 271 * Turn off data relocation (DMMU off). 272 */ 273 msr &= ~PSL_DR; 274 __asm __volatile("sync; mtmsr %0; isync" :: "r"(msr)); 275 276 /* 277 * Zero the page using a single cache line. 278 */ 279 do { 280 __asm("stvx %2,%0,%1" :: "r"(pa), "r"( 0), "n"(ZERO_VEC)); 281 __asm("stvxl %2,%0,%1" :: "r"(pa), "r"(16), "n"(ZERO_VEC)); 282 __asm("stvx %2,%0,%1" :: "r"(pa), "r"(32), "n"(ZERO_VEC)); 283 __asm("stvxl %2,%0,%1" :: "r"(pa), "r"(48), "n"(ZERO_VEC)); 284 pa += 64; 285 } while (pa < ea); 286 287 /* 288 * Restore data relocation (DMMU on); 289 */ 290 msr |= PSL_DR; 291 __asm __volatile("sync; mtmsr %0; isync" :: "r"(msr)); 292 293 /* 294 * Restore VEC register (now that we can access the stack again). 295 */ 296 __asm("lvx %1,0,%0" :: "r"(vp), "n"(ZERO_VEC)); 297 298 /* 299 * Restore old MSR (AltiVec OFF). 300 */ 301 __asm __volatile("sync; mtmsr %0; isync" :: "r"(omsr)); 302 } 303 304 #define LO_VEC 16 305 #define HI_VEC 17 306 307 void 308 vcopypage(paddr_t dst, paddr_t src) 309 { 310 const paddr_t edst = dst + NBPG; 311 uint32_t vec[11], *vp = (void *) roundup((uintptr_t) vec, 16); 312 uint32_t omsr, msr; 313 314 __asm __volatile("mfmsr %0" : "=r"(omsr) :); 315 316 /* 317 * Turn on AltiVec, turn off interrupts. 318 */ 319 msr = (omsr & ~PSL_EE) | PSL_VEC; 320 __asm __volatile("sync; mtmsr %0; isync" :: "r"(msr)); 321 322 /* 323 * Save the VEC registers we will be using before we disable 324 * relocation. 325 */ 326 __asm("stvx %2,%1,%0" :: "r"(vp), "r"( 0), "n"(LO_VEC)); 327 __asm("stvx %2,%1,%0" :: "r"(vp), "r"(16), "n"(HI_VEC)); 328 329 /* 330 * Turn off data relocation (DMMU off). 331 */ 332 msr &= ~PSL_DR; 333 __asm __volatile("sync; mtmsr %0; isync" :: "r"(msr)); 334 335 /* 336 * Copy the page using a single cache line. On most PPCs, two 337 * vector registers occupy one cache line. 338 */ 339 do { 340 __asm("lvx %2,%0,%1" :: "r"(src), "r"( 0), "n"(LO_VEC)); 341 __asm("stvx %2,%0,%1" :: "r"(dst), "r"( 0), "n"(LO_VEC)); 342 __asm("lvxl %2,%0,%1" :: "r"(src), "r"(16), "n"(HI_VEC)); 343 __asm("stvxl %2,%0,%1" :: "r"(dst), "r"(16), "n"(HI_VEC)); 344 src += 32; 345 dst += 32; 346 } while (dst < edst); 347 348 /* 349 * Restore data relocation (DMMU on); 350 */ 351 msr |= PSL_DR; 352 __asm __volatile("sync; mtmsr %0; isync" :: "r"(msr)); 353 354 /* 355 * Restore VEC registers (now that we can access the stack again). 356 */ 357 __asm("lvx %2,%1,%0" :: "r"(vp), "r"( 0), "n"(LO_VEC)); 358 __asm("lvx %2,%1,%0" :: "r"(vp), "r"(16), "n"(HI_VEC)); 359 360 /* 361 * Restore old MSR (AltiVec OFF). 362 */ 363 __asm __volatile("sync; mtmsr %0; isync" :: "r"(omsr)); 364 } 365 366 void 367 init_vec(void) 368 { 369 pool_init(&vecpool, sizeof(struct vreg), 16, 0, 0, "vecpl", NULL); 370 } 371