2 * arch/powerpc/math-emu/math_efp.c
4 * Copyright (C) 2006-2008 Freescale Semiconductor, Inc. All rights reserved.
6 * Author: Ebony Zhu, <ebony.zhu@freescale.com>
7 * Yu Liu, <yu.liu@freescale.com>
9 * Derived from arch/alpha/math-emu/math.c
10 * arch/powerpc/math-emu/math.c
13 * This file is the exception handler to make E500 SPE instructions
14 * fully comply with IEEE-754 floating point standard.
16 * This program is free software; you can redistribute it and/or
17 * modify it under the terms of the GNU General Public License
18 * as published by the Free Software Foundation; either version
19 * 2 of the License, or (at your option) any later version.
22 #include <linux/types.h>
24 #include <asm/uaccess.h>
27 #define FP_EX_BOOKE_E500_SPE
28 #include <asm/sfp-machine.h>
30 #include <math-emu/soft-fp.h>
31 #include <math-emu/single.h>
32 #include <math-emu/double.h>
47 #define EFSCMPGT 0x2cc
48 #define EFSCMPLT 0x2cd
49 #define EFSCMPEQ 0x2ce
56 #define EFSCTUIZ 0x2d8
57 #define EFSCTSIZ 0x2da
62 #define EVFSNABS 0x285
66 #define EVFSCMPGT 0x28c
67 #define EVFSCMPLT 0x28d
68 #define EVFSCMPEQ 0x28e
69 #define EVFSCTUI 0x294
70 #define EVFSCTSI 0x295
71 #define EVFSCTUF 0x296
72 #define EVFSCTSF 0x297
73 #define EVFSCTUIZ 0x298
74 #define EVFSCTSIZ 0x29a
83 #define EFDCTUIDZ 0x2ea
84 #define EFDCTSIDZ 0x2eb
85 #define EFDCMPGT 0x2ec
86 #define EFDCMPLT 0x2ed
87 #define EFDCMPEQ 0x2ee
93 #define EFDCTUIZ 0x2f8
94 #define EFDCTSIZ 0x2fa
102 #define SIGN_BIT_S (1UL << 31)
103 #define SIGN_BIT_D (1ULL << 63)
104 #define FP_EX_MASK (FP_EX_INEXACT | FP_EX_INVALID | FP_EX_DIVZERO | \
105 FP_EX_UNDERFLOW | FP_EX_OVERFLOW)
112 static unsigned long insn_type(unsigned long speinsn)
114 unsigned long ret = NOTYPE;
116 switch (speinsn & 0x7ff) {
117 case EFSABS: ret = XA; break;
118 case EFSADD: ret = AB; break;
119 case EFSCFD: ret = XB; break;
120 case EFSCMPEQ: ret = XCR; break;
121 case EFSCMPGT: ret = XCR; break;
122 case EFSCMPLT: ret = XCR; break;
123 case EFSCTSF: ret = XB; break;
124 case EFSCTSI: ret = XB; break;
125 case EFSCTSIZ: ret = XB; break;
126 case EFSCTUF: ret = XB; break;
127 case EFSCTUI: ret = XB; break;
128 case EFSCTUIZ: ret = XB; break;
129 case EFSDIV: ret = AB; break;
130 case EFSMUL: ret = AB; break;
131 case EFSNABS: ret = XA; break;
132 case EFSNEG: ret = XA; break;
133 case EFSSUB: ret = AB; break;
134 case EFSCFSI: ret = XB; break;
136 case EVFSABS: ret = XA; break;
137 case EVFSADD: ret = AB; break;
138 case EVFSCMPEQ: ret = XCR; break;
139 case EVFSCMPGT: ret = XCR; break;
140 case EVFSCMPLT: ret = XCR; break;
141 case EVFSCTSF: ret = XB; break;
142 case EVFSCTSI: ret = XB; break;
143 case EVFSCTSIZ: ret = XB; break;
144 case EVFSCTUF: ret = XB; break;
145 case EVFSCTUI: ret = XB; break;
146 case EVFSCTUIZ: ret = XB; break;
147 case EVFSDIV: ret = AB; break;
148 case EVFSMUL: ret = AB; break;
149 case EVFSNABS: ret = XA; break;
150 case EVFSNEG: ret = XA; break;
151 case EVFSSUB: ret = AB; break;
153 case EFDABS: ret = XA; break;
154 case EFDADD: ret = AB; break;
155 case EFDCFS: ret = XB; break;
156 case EFDCMPEQ: ret = XCR; break;
157 case EFDCMPGT: ret = XCR; break;
158 case EFDCMPLT: ret = XCR; break;
159 case EFDCTSF: ret = XB; break;
160 case EFDCTSI: ret = XB; break;
161 case EFDCTSIDZ: ret = XB; break;
162 case EFDCTSIZ: ret = XB; break;
163 case EFDCTUF: ret = XB; break;
164 case EFDCTUI: ret = XB; break;
165 case EFDCTUIDZ: ret = XB; break;
166 case EFDCTUIZ: ret = XB; break;
167 case EFDDIV: ret = AB; break;
168 case EFDMUL: ret = AB; break;
169 case EFDNABS: ret = XA; break;
170 case EFDNEG: ret = XA; break;
171 case EFDSUB: ret = AB; break;
174 printk(KERN_ERR "\nOoops! SPE instruction no type found.");
175 printk(KERN_ERR "\ninst code: %08lx\n", speinsn);
181 int do_spe_mathemu(struct pt_regs *regs)
186 unsigned long type, func, fc, fa, fb, src, speinsn;
187 union dw_union vc, va, vb;
189 if (get_user(speinsn, (unsigned int __user *) regs->nip))
191 if ((speinsn >> 26) != EFAPU)
192 return -EINVAL; /* not an spe instruction */
194 type = insn_type(speinsn);
198 func = speinsn & 0x7ff;
199 fc = (speinsn >> 21) & 0x1f;
200 fa = (speinsn >> 16) & 0x1f;
201 fb = (speinsn >> 11) & 0x1f;
202 src = (speinsn >> 5) & 0x7;
204 vc.wp[0] = current->thread.evr[fc];
205 vc.wp[1] = regs->gpr[fc];
206 va.wp[0] = current->thread.evr[fa];
207 va.wp[1] = regs->gpr[fa];
208 vb.wp[0] = current->thread.evr[fb];
209 vb.wp[1] = regs->gpr[fb];
211 __FPU_FPSCR = mfspr(SPRN_SPEFSCR);
214 printk("speinsn:%08lx spefscr:%08lx\n", speinsn, __FPU_FPSCR);
215 printk("vc: %08x %08x\n", vc.wp[0], vc.wp[1]);
216 printk("va: %08x %08x\n", va.wp[0], va.wp[1]);
217 printk("vb: %08x %08x\n", vb.wp[0], vb.wp[1]);
222 FP_DECL_S(SA); FP_DECL_S(SB); FP_DECL_S(SR);
227 FP_UNPACK_SP(SA, va.wp + 1);
229 FP_UNPACK_SP(SB, vb.wp + 1);
232 FP_UNPACK_SP(SA, va.wp + 1);
237 printk("SA: %ld %08lx %ld (%ld)\n", SA_s, SA_f, SA_e, SA_c);
238 printk("SB: %ld %08lx %ld (%ld)\n", SB_s, SB_f, SB_e, SB_c);
243 vc.wp[1] = va.wp[1] & ~SIGN_BIT_S;
247 vc.wp[1] = va.wp[1] | SIGN_BIT_S;
251 vc.wp[1] = va.wp[1] ^ SIGN_BIT_S;
255 FP_ADD_S(SR, SA, SB);
259 FP_SUB_S(SR, SA, SB);
263 FP_MUL_S(SR, SA, SB);
267 FP_DIV_S(SR, SA, SB);
284 if (!((vb.wp[1] >> 23) == 0xff && ((vb.wp[1] & 0x7fffff) > 0))) {
286 if (((vb.wp[1] >> 23) & 0xff) == 0) {
289 } else if ((vb.wp[1] >> 31) == 0) {
290 /* positive normal */
291 vc.wp[1] = (func == EFSCTSF) ?
292 0x7fffffff : 0xffffffff;
293 } else { /* negative normal */
294 vc.wp[1] = (func == EFSCTSF) ?
297 } else { /* rB is NaN */
305 FP_UNPACK_DP(DB, vb.dp);
307 printk("DB: %ld %08lx %08lx %ld (%ld)\n",
308 DB_s, DB_f1, DB_f0, DB_e, DB_c);
310 FP_CONV(S, D, 1, 2, SR, DB);
321 _FP_ROUND_ZERO(1, SB);
323 FP_TO_INT_S(vc.wp[1], SB, 32, ((func & 0x3) != 0));
333 printk("SR: %ld %08lx %ld (%ld)\n", SR_s, SR_f, SR_e, SR_c);
335 FP_PACK_SP(vc.wp + 1, SR);
339 FP_CMP_S(IR, SA, SB, 3);
340 if (IR == 3 && (FP_ISSIGNAN_S(SA) || FP_ISSIGNAN_S(SB)))
341 FP_SET_EXCEPTION(FP_EX_INVALID);
351 FP_DECL_D(DA); FP_DECL_D(DB); FP_DECL_D(DR);
356 FP_UNPACK_DP(DA, va.dp);
358 FP_UNPACK_DP(DB, vb.dp);
361 FP_UNPACK_DP(DA, va.dp);
366 printk("DA: %ld %08lx %08lx %ld (%ld)\n",
367 DA_s, DA_f1, DA_f0, DA_e, DA_c);
368 printk("DB: %ld %08lx %08lx %ld (%ld)\n",
369 DB_s, DB_f1, DB_f0, DB_e, DB_c);
374 vc.dp[0] = va.dp[0] & ~SIGN_BIT_D;
378 vc.dp[0] = va.dp[0] | SIGN_BIT_D;
382 vc.dp[0] = va.dp[0] ^ SIGN_BIT_D;
386 FP_ADD_D(DR, DA, DB);
390 FP_SUB_D(DR, DA, DB);
394 FP_MUL_D(DR, DA, DB);
398 FP_DIV_D(DR, DA, DB);
415 if (!((vb.wp[0] >> 20) == 0x7ff &&
416 ((vb.wp[0] & 0xfffff) > 0 || (vb.wp[1] > 0)))) {
418 if (((vb.wp[0] >> 20) & 0x7ff) == 0) {
421 } else if ((vb.wp[0] >> 31) == 0) {
422 /* positive normal */
423 vc.wp[1] = (func == EFDCTSF) ?
424 0x7fffffff : 0xffffffff;
425 } else { /* negative normal */
426 vc.wp[1] = (func == EFDCTSF) ?
437 FP_UNPACK_SP(SB, vb.wp + 1);
439 printk("SB: %ld %08lx %ld (%ld)\n",
440 SB_s, SB_f, SB_e, SB_c);
442 FP_CONV(D, S, 2, 1, DR, SB);
448 _FP_ROUND_ZERO(2, DB);
449 FP_TO_INT_D(vc.dp[0], DB, 64, ((func & 0x1) == 0));
459 _FP_ROUND_ZERO(2, DB);
461 FP_TO_INT_D(vc.wp[1], DB, 32, ((func & 0x3) != 0));
471 printk("DR: %ld %08lx %08lx %ld (%ld)\n",
472 DR_s, DR_f1, DR_f0, DR_e, DR_c);
474 FP_PACK_DP(vc.dp, DR);
478 FP_CMP_D(IR, DA, DB, 3);
479 if (IR == 3 && (FP_ISSIGNAN_D(DA) || FP_ISSIGNAN_D(DB)))
480 FP_SET_EXCEPTION(FP_EX_INVALID);
491 FP_DECL_S(SA0); FP_DECL_S(SB0); FP_DECL_S(SR0);
492 FP_DECL_S(SA1); FP_DECL_S(SB1); FP_DECL_S(SR1);
498 FP_UNPACK_SP(SA0, va.wp);
499 FP_UNPACK_SP(SA1, va.wp + 1);
501 FP_UNPACK_SP(SB0, vb.wp);
502 FP_UNPACK_SP(SB1, vb.wp + 1);
505 FP_UNPACK_SP(SA0, va.wp);
506 FP_UNPACK_SP(SA1, va.wp + 1);
511 printk("SA0: %ld %08lx %ld (%ld)\n", SA0_s, SA0_f, SA0_e, SA0_c);
512 printk("SA1: %ld %08lx %ld (%ld)\n", SA1_s, SA1_f, SA1_e, SA1_c);
513 printk("SB0: %ld %08lx %ld (%ld)\n", SB0_s, SB0_f, SB0_e, SB0_c);
514 printk("SB1: %ld %08lx %ld (%ld)\n", SB1_s, SB1_f, SB1_e, SB1_c);
519 vc.wp[0] = va.wp[0] & ~SIGN_BIT_S;
520 vc.wp[1] = va.wp[1] & ~SIGN_BIT_S;
524 vc.wp[0] = va.wp[0] | SIGN_BIT_S;
525 vc.wp[1] = va.wp[1] | SIGN_BIT_S;
529 vc.wp[0] = va.wp[0] ^ SIGN_BIT_S;
530 vc.wp[1] = va.wp[1] ^ SIGN_BIT_S;
534 FP_ADD_S(SR0, SA0, SB0);
535 FP_ADD_S(SR1, SA1, SB1);
539 FP_SUB_S(SR0, SA0, SB0);
540 FP_SUB_S(SR1, SA1, SB1);
544 FP_MUL_S(SR0, SA0, SB0);
545 FP_MUL_S(SR1, SA1, SB1);
549 FP_DIV_S(SR0, SA0, SB0);
550 FP_DIV_S(SR1, SA1, SB1);
566 __asm__ __volatile__ ("mtspr 512, %4\n"
569 : "=r" (vc.wp[0]), "=r" (vc.wp[1])
570 : "r" (vb.wp[0]), "r" (vb.wp[1]), "r" (0));
574 __asm__ __volatile__ ("mtspr 512, %4\n"
577 : "=r" (vc.wp[0]), "=r" (vc.wp[1])
578 : "r" (vb.wp[0]), "r" (vb.wp[1]), "r" (0));
589 _FP_ROUND_ZERO(1, SB0);
590 _FP_ROUND_ZERO(1, SB1);
592 FP_TO_INT_S(vc.wp[0], SB0, 32, ((func & 0x3) != 0));
593 FP_TO_INT_S(vc.wp[1], SB1, 32, ((func & 0x3) != 0));
603 printk("SR0: %ld %08lx %ld (%ld)\n", SR0_s, SR0_f, SR0_e, SR0_c);
604 printk("SR1: %ld %08lx %ld (%ld)\n", SR1_s, SR1_f, SR1_e, SR1_c);
606 FP_PACK_SP(vc.wp, SR0);
607 FP_PACK_SP(vc.wp + 1, SR1);
614 FP_CMP_S(IR0, SA0, SB0, 3);
615 FP_CMP_S(IR1, SA1, SB1, 3);
616 if (IR0 == 3 && (FP_ISSIGNAN_S(SA0) || FP_ISSIGNAN_S(SB0)))
617 FP_SET_EXCEPTION(FP_EX_INVALID);
618 if (IR1 == 3 && (FP_ISSIGNAN_S(SA1) || FP_ISSIGNAN_S(SB1)))
619 FP_SET_EXCEPTION(FP_EX_INVALID);
620 ch = (IR0 == cmp) ? 1 : 0;
621 cl = (IR1 == cmp) ? 1 : 0;
622 IR = (ch << 3) | (cl << 2) | ((ch | cl) << 1) |
632 regs->ccr &= ~(15 << ((7 - ((speinsn >> 23) & 0x7)) << 2));
633 regs->ccr |= (IR << ((7 - ((speinsn >> 23) & 0x7)) << 2));
636 __FPU_FPSCR &= ~FP_EX_MASK;
637 __FPU_FPSCR |= (FP_CUR_EXCEPTIONS & FP_EX_MASK);
638 mtspr(SPRN_SPEFSCR, __FPU_FPSCR);
640 current->thread.evr[fc] = vc.wp[0];
641 regs->gpr[fc] = vc.wp[1];
644 printk("ccr = %08lx\n", regs->ccr);
645 printk("cur exceptions = %08x spefscr = %08lx\n",
646 FP_CUR_EXCEPTIONS, __FPU_FPSCR);
647 printk("vc: %08x %08x\n", vc.wp[0], vc.wp[1]);
648 printk("va: %08x %08x\n", va.wp[0], va.wp[1]);
649 printk("vb: %08x %08x\n", vb.wp[0], vb.wp[1]);
655 printk(KERN_ERR "\nOoops! IEEE-754 compliance handler encountered un-supported instruction.\ninst code: %08lx\n", speinsn);
659 int speround_handler(struct pt_regs *regs)
663 unsigned long speinsn, type, fc;
665 if (get_user(speinsn, (unsigned int __user *) regs->nip))
667 if ((speinsn >> 26) != 4)
668 return -EINVAL; /* not an spe instruction */
670 type = insn_type(speinsn & 0x7ff);
671 if (type == XCR) return -ENOSYS;
673 fc = (speinsn >> 21) & 0x1f;
674 s_lo = regs->gpr[fc] & SIGN_BIT_S;
675 s_hi = current->thread.evr[fc] & SIGN_BIT_S;
676 fgpr.wp[0] = current->thread.evr[fc];
677 fgpr.wp[1] = regs->gpr[fc];
679 __FPU_FPSCR = mfspr(SPRN_SPEFSCR);
681 switch ((speinsn >> 5) & 0x7) {
682 /* Since SPE instructions on E500 core can handle round to nearest
683 * and round toward zero with IEEE-754 complied, we just need
684 * to handle round toward +Inf and round toward -Inf by software.
687 if ((FP_ROUNDMODE) == FP_RND_PINF) {
688 if (!s_lo) fgpr.wp[1]++; /* Z > 0, choose Z1 */
689 } else { /* round to -Inf */
690 if (s_lo) fgpr.wp[1]++; /* Z < 0, choose Z2 */
695 if (FP_ROUNDMODE == FP_RND_PINF) {
696 if (!s_hi) fgpr.dp[0]++; /* Z > 0, choose Z1 */
697 } else { /* round to -Inf */
698 if (s_hi) fgpr.dp[0]++; /* Z < 0, choose Z2 */
703 if (FP_ROUNDMODE == FP_RND_PINF) {
704 if (!s_lo) fgpr.wp[1]++; /* Z_low > 0, choose Z1 */
705 if (!s_hi) fgpr.wp[0]++; /* Z_high word > 0, choose Z1 */
706 } else { /* round to -Inf */
707 if (s_lo) fgpr.wp[1]++; /* Z_low < 0, choose Z2 */
708 if (s_hi) fgpr.wp[0]++; /* Z_high < 0, choose Z2 */
716 current->thread.evr[fc] = fgpr.wp[0];
717 regs->gpr[fc] = fgpr.wp[1];