Home | History | Annotate | Line # | Download | only in vfp
vfp_init.c revision 1.64
      1  1.64     joerg /*      $NetBSD: vfp_init.c,v 1.64 2019/10/29 16:18:23 joerg Exp $ */
      2   1.1  rearnsha 
      3   1.1  rearnsha /*
      4   1.1  rearnsha  * Copyright (c) 2008 ARM Ltd
      5   1.1  rearnsha  * All rights reserved.
      6   1.1  rearnsha  *
      7   1.1  rearnsha  * Redistribution and use in source and binary forms, with or without
      8   1.1  rearnsha  * modification, are permitted provided that the following conditions
      9   1.1  rearnsha  * are met:
     10   1.1  rearnsha  * 1. Redistributions of source code must retain the above copyright
     11   1.1  rearnsha  *    notice, this list of conditions and the following disclaimer.
     12   1.1  rearnsha  * 2. Redistributions in binary form must reproduce the above copyright
     13   1.1  rearnsha  *    notice, this list of conditions and the following disclaimer in the
     14   1.1  rearnsha  *    documentation and/or other materials provided with the distribution.
     15   1.1  rearnsha  * 3. The name of the company may not be used to endorse or promote
     16   1.1  rearnsha  *    products derived from this software without specific prior written
     17   1.1  rearnsha  *    permission.
     18   1.1  rearnsha  *
     19   1.1  rearnsha  * THIS SOFTWARE IS PROVIDED BY ARM LTD ``AS IS'' AND ANY EXPRESS OR
     20   1.1  rearnsha  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
     21   1.1  rearnsha  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
     22   1.1  rearnsha  * ARE DISCLAIMED.  IN NO EVENT SHALL ARM LTD BE LIABLE FOR ANY
     23   1.1  rearnsha  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     24   1.1  rearnsha  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE
     25   1.1  rearnsha  * GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     26   1.1  rearnsha  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
     27   1.1  rearnsha  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
     28   1.1  rearnsha  * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
     29   1.1  rearnsha  * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
     30   1.1  rearnsha  */
     31   1.1  rearnsha 
     32  1.59     skrll #include "opt_cputypes.h"
     33  1.59     skrll 
     34  1.58     skrll #include <sys/cdefs.h>
     35  1.64     joerg __KERNEL_RCSID(0, "$NetBSD: vfp_init.c,v 1.64 2019/10/29 16:18:23 joerg Exp $");
     36  1.58     skrll 
     37   1.1  rearnsha #include <sys/param.h>
     38   1.1  rearnsha #include <sys/types.h>
     39   1.1  rearnsha #include <sys/systm.h>
     40   1.1  rearnsha #include <sys/device.h>
     41   1.1  rearnsha #include <sys/proc.h>
     42   1.4      matt #include <sys/cpu.h>
     43   1.1  rearnsha 
     44  1.23      matt #include <arm/locore.h>
     45   1.5      matt #include <arm/pcb.h>
     46   1.1  rearnsha #include <arm/undefined.h>
     47   1.1  rearnsha #include <arm/vfpreg.h>
     48   1.8      matt #include <arm/mcontext.h>
     49   1.1  rearnsha 
     50  1.12      matt #include <uvm/uvm_extern.h>		/* for pmap.h */
     51  1.12      matt 
     52  1.11      matt #ifdef FPU_VFP
     53  1.11      matt 
     54  1.29      matt #ifdef CPU_CORTEX
     55  1.56  christos #define SETFPU __asm(".fpu\tvfpv4")
     56  1.29      matt #else
     57  1.56  christos #define SETFPU __asm(".fpu\tvfp")
     58  1.29      matt #endif
     59  1.56  christos SETFPU;
     60  1.29      matt 
     61   1.1  rearnsha /* FLDMD <X>, {d0-d15} */
     62  1.11      matt static inline void
     63  1.13      matt load_vfpregs_lo(const uint64_t *p)
     64  1.10      matt {
     65  1.56  christos 	SETFPU;
     66  1.64     joerg 	__asm __volatile(".fpu vfp\n vldmia\t%0, {d0-d15}" :: "r" (p) : "memory");
     67  1.10      matt }
     68  1.10      matt 
     69  1.10      matt /* FSTMD <X>, {d0-d15} */
     70  1.11      matt static inline void
     71  1.10      matt save_vfpregs_lo(uint64_t *p)
     72  1.10      matt {
     73  1.56  christos 	SETFPU;
     74  1.64     joerg 	__asm __volatile(".fpu vfp\n vstmia\t%0, {d0-d15}" :: "r" (p) : "memory");
     75  1.10      matt }
     76  1.10      matt 
     77  1.10      matt #ifdef CPU_CORTEX
     78  1.10      matt /* FLDMD <X>, {d16-d31} */
     79  1.11      matt static inline void
     80  1.13      matt load_vfpregs_hi(const uint64_t *p)
     81  1.10      matt {
     82  1.56  christos 	SETFPU;
     83  1.64     joerg 	__asm __volatile(".fpu neon-vfpv4\n vldmia\t%0, {d16-d31}" :: "r" (&p[16]) : "memory");
     84  1.10      matt }
     85  1.10      matt 
     86  1.10      matt /* FLDMD <X>, {d16-d31} */
     87  1.11      matt static inline void
     88  1.10      matt save_vfpregs_hi(uint64_t *p)
     89  1.10      matt {
     90  1.56  christos 	SETFPU;
     91  1.64     joerg 	__asm __volatile(".fpu neon-vfpv4\nvstmia\t%0, {d16-d31}" :: "r" (&p[16]) : "memory");
     92  1.10      matt }
     93  1.10      matt #endif
     94   1.1  rearnsha 
     95  1.13      matt static inline void
     96  1.13      matt load_vfpregs(const struct vfpreg *fregs)
     97  1.13      matt {
     98  1.13      matt 	load_vfpregs_lo(fregs->vfp_regs);
     99  1.13      matt #ifdef CPU_CORTEX
    100  1.13      matt #ifdef CPU_ARM11
    101  1.13      matt 	switch (curcpu()->ci_vfp_id) {
    102  1.13      matt 	case FPU_VFP_CORTEXA5:
    103  1.13      matt 	case FPU_VFP_CORTEXA7:
    104  1.13      matt 	case FPU_VFP_CORTEXA8:
    105  1.13      matt 	case FPU_VFP_CORTEXA9:
    106  1.20      matt 	case FPU_VFP_CORTEXA15:
    107  1.42       slp 	case FPU_VFP_CORTEXA15_QEMU:
    108  1.50     skrll 	case FPU_VFP_CORTEXA53:
    109  1.53  jmcneill 	case FPU_VFP_CORTEXA57:
    110  1.13      matt #endif
    111  1.13      matt 		load_vfpregs_hi(fregs->vfp_regs);
    112  1.13      matt #ifdef CPU_ARM11
    113  1.13      matt 		break;
    114  1.13      matt 	}
    115  1.13      matt #endif
    116  1.13      matt #endif
    117  1.13      matt }
    118  1.13      matt 
    119  1.13      matt static inline void
    120  1.13      matt save_vfpregs(struct vfpreg *fregs)
    121  1.13      matt {
    122  1.13      matt 	save_vfpregs_lo(fregs->vfp_regs);
    123  1.13      matt #ifdef CPU_CORTEX
    124  1.13      matt #ifdef CPU_ARM11
    125  1.13      matt 	switch (curcpu()->ci_vfp_id) {
    126  1.13      matt 	case FPU_VFP_CORTEXA5:
    127  1.13      matt 	case FPU_VFP_CORTEXA7:
    128  1.13      matt 	case FPU_VFP_CORTEXA8:
    129  1.13      matt 	case FPU_VFP_CORTEXA9:
    130  1.20      matt 	case FPU_VFP_CORTEXA15:
    131  1.42       slp 	case FPU_VFP_CORTEXA15_QEMU:
    132  1.50     skrll 	case FPU_VFP_CORTEXA53:
    133  1.53  jmcneill 	case FPU_VFP_CORTEXA57:
    134  1.13      matt #endif
    135  1.13      matt 		save_vfpregs_hi(fregs->vfp_regs);
    136  1.13      matt #ifdef CPU_ARM11
    137  1.13      matt 		break;
    138  1.13      matt 	}
    139  1.13      matt #endif
    140  1.13      matt #endif
    141  1.13      matt }
    142  1.13      matt 
    143   1.1  rearnsha /* The real handler for VFP bounces.  */
    144   1.1  rearnsha static int vfp_handler(u_int, u_int, trapframe_t *, int);
    145  1.13      matt #ifdef CPU_CORTEX
    146  1.13      matt static int neon_handler(u_int, u_int, trapframe_t *, int);
    147  1.13      matt #endif
    148   1.1  rearnsha 
    149  1.13      matt static void vfp_state_load(lwp_t *, u_int);
    150  1.39     rmind static void vfp_state_save(lwp_t *);
    151  1.39     rmind static void vfp_state_release(lwp_t *);
    152   1.4      matt 
    153   1.4      matt const pcu_ops_t arm_vfp_ops = {
    154   1.4      matt 	.pcu_id = PCU_FPU,
    155  1.13      matt 	.pcu_state_save = vfp_state_save,
    156   1.4      matt 	.pcu_state_load = vfp_state_load,
    157   1.4      matt 	.pcu_state_release = vfp_state_release,
    158   1.4      matt };
    159   1.1  rearnsha 
    160  1.34      matt /* determine what bits can be changed */
    161  1.34      matt uint32_t vfp_fpscr_changable = VFP_FPSCR_CSUM;
    162  1.34      matt /* default to run fast */
    163  1.34      matt uint32_t vfp_fpscr_default = (VFP_FPSCR_DN | VFP_FPSCR_FZ | VFP_FPSCR_RN);
    164  1.34      matt 
    165   1.1  rearnsha /*
    166   1.1  rearnsha  * Used to test for a VFP. The following function is installed as a coproc10
    167   1.1  rearnsha  * handler on the undefined instruction vector and then we issue a VFP
    168   1.1  rearnsha  * instruction. If undefined_test is non zero then the VFP did not handle
    169   1.1  rearnsha  * the instruction so must be absent, or disabled.
    170   1.1  rearnsha  */
    171   1.1  rearnsha 
    172   1.1  rearnsha static int undefined_test;
    173   1.1  rearnsha 
    174   1.1  rearnsha static int
    175   1.4      matt vfp_test(u_int address, u_int insn, trapframe_t *frame, int fault_code)
    176   1.1  rearnsha {
    177   1.1  rearnsha 
    178   1.1  rearnsha 	frame->tf_pc += INSN_SIZE;
    179   1.1  rearnsha 	++undefined_test;
    180   1.4      matt 	return 0;
    181   1.4      matt }
    182   1.4      matt 
    183  1.35      matt #else
    184  1.35      matt /* determine what bits can be changed */
    185  1.35      matt uint32_t vfp_fpscr_changable = VFP_FPSCR_CSUM|VFP_FPSCR_ESUM|VFP_FPSCR_RMODE;
    186   1.4      matt #endif /* FPU_VFP */
    187   1.4      matt 
    188   1.4      matt static int
    189   1.4      matt vfp_fpscr_handler(u_int address, u_int insn, trapframe_t *frame, int fault_code)
    190   1.4      matt {
    191   1.4      matt 	struct lwp * const l = curlwp;
    192   1.4      matt 	const u_int regno = (insn >> 12) & 0xf;
    193   1.4      matt 	/*
    194   1.4      matt 	 * Only match move to/from the FPSCR register and we
    195   1.4      matt 	 * can't be using the SP,LR,PC as a source.
    196   1.4      matt 	 */
    197   1.4      matt 	if ((insn & 0xffef0fff) != 0xeee10a10 || regno > 12)
    198   1.4      matt 		return 1;
    199   1.4      matt 
    200   1.4      matt 	struct pcb * const pcb = lwp_getpcb(l);
    201   1.4      matt 
    202   1.4      matt #ifdef FPU_VFP
    203   1.4      matt 	/*
    204   1.4      matt 	 * If FPU is valid somewhere, let's just reenable VFP and
    205   1.4      matt 	 * retry the instruction (only safe thing to do since the
    206   1.4      matt 	 * pcb has a stale copy).
    207   1.4      matt 	 */
    208   1.4      matt 	if (pcb->pcb_vfp.vfp_fpexc & VFP_FPEXC_EN)
    209   1.4      matt 		return 1;
    210   1.4      matt 
    211  1.51       chs 	if (__predict_false(!vfp_used_p(l))) {
    212  1.35      matt 		pcb->pcb_vfp.vfp_fpscr = vfp_fpscr_default;
    213   1.4      matt 	}
    214  1.26      matt #endif
    215   1.4      matt 
    216   1.4      matt 	/*
    217  1.30     skrll 	 * We now know the pcb has the saved copy.
    218   1.4      matt 	 */
    219   1.4      matt 	register_t * const regp = &frame->tf_r0 + regno;
    220   1.4      matt 	if (insn & 0x00100000) {
    221   1.4      matt 		*regp = pcb->pcb_vfp.vfp_fpscr;
    222   1.4      matt 	} else {
    223  1.34      matt 		pcb->pcb_vfp.vfp_fpscr &= ~vfp_fpscr_changable;
    224  1.34      matt 		pcb->pcb_vfp.vfp_fpscr |= *regp & vfp_fpscr_changable;
    225   1.4      matt 	}
    226   1.4      matt 
    227  1.37      matt 	curcpu()->ci_vfp_evs[0].ev_count++;
    228  1.61     skrll 
    229   1.4      matt 	frame->tf_pc += INSN_SIZE;
    230   1.4      matt 	return 0;
    231   1.1  rearnsha }
    232   1.1  rearnsha 
    233   1.4      matt #ifndef FPU_VFP
    234   1.4      matt /*
    235   1.4      matt  * If we don't want VFP support, we still need to handle emulating VFP FPSCR
    236   1.4      matt  * instructions.
    237   1.4      matt  */
    238   1.4      matt void
    239  1.37      matt vfp_attach(struct cpu_info *ci)
    240   1.4      matt {
    241  1.37      matt 	if (CPU_IS_PRIMARY(ci)) {
    242  1.37      matt 		install_coproc_handler(VFP_COPROC, vfp_fpscr_handler);
    243  1.37      matt 	}
    244  1.37      matt 	evcnt_attach_dynamic(&ci->ci_vfp_evs[0], EVCNT_TYPE_TRAP, NULL,
    245  1.37      matt 	    ci->ci_cpuname, "vfp fpscr traps");
    246   1.4      matt }
    247   1.4      matt 
    248   1.4      matt #else
    249   1.1  rearnsha void
    250  1.37      matt vfp_attach(struct cpu_info *ci)
    251   1.1  rearnsha {
    252   1.4      matt 	const char *model = NULL;
    253   1.1  rearnsha 
    254  1.37      matt 	if (CPU_ID_ARM11_P(ci->ci_arm_cpuid)
    255  1.37      matt 	    || CPU_ID_MV88SV58XX_P(ci->ci_arm_cpuid)
    256  1.37      matt 	    || CPU_ID_CORTEX_P(ci->ci_arm_cpuid)) {
    257  1.37      matt #if 0
    258  1.37      matt 		const uint32_t nsacr = armreg_nsacr_read();
    259  1.37      matt 		const uint32_t nsacr_vfp = __BITS(VFP_COPROC,VFP_COPROC2);
    260  1.37      matt 		if ((nsacr & nsacr_vfp) != nsacr_vfp) {
    261  1.40      matt 			aprint_normal_dev(ci->ci_dev,
    262  1.40      matt 			    "VFP access denied (NSACR=%#x)\n", nsacr);
    263  1.62     skrll 			if (CPU_IS_PRIMARY(ci))
    264  1.62     skrll 				install_coproc_handler(VFP_COPROC, vfp_fpscr_handler);
    265  1.37      matt 			ci->ci_vfp_id = 0;
    266  1.37      matt 			evcnt_attach_dynamic(&ci->ci_vfp_evs[0],
    267  1.37      matt 			    EVCNT_TYPE_TRAP, NULL, ci->ci_cpuname,
    268  1.37      matt 			    "vfp fpscr traps");
    269  1.37      matt 			return;
    270  1.37      matt 		}
    271  1.37      matt #endif
    272   1.7      matt 		const uint32_t cpacr_vfp = CPACR_CPn(VFP_COPROC);
    273   1.7      matt 		const uint32_t cpacr_vfp2 = CPACR_CPn(VFP_COPROC2);
    274   1.1  rearnsha 
    275   1.7      matt 		/*
    276   1.7      matt 		 * We first need to enable access to the coprocessors.
    277   1.7      matt 		 */
    278   1.7      matt 		uint32_t cpacr = armreg_cpacr_read();
    279   1.7      matt 		cpacr |= __SHIFTIN(CPACR_ALL, cpacr_vfp);
    280   1.7      matt 		cpacr |= __SHIFTIN(CPACR_ALL, cpacr_vfp2);
    281   1.7      matt 		armreg_cpacr_write(cpacr);
    282   1.1  rearnsha 
    283  1.48  jmcneill 		arm_isb();
    284  1.48  jmcneill 
    285   1.7      matt 		/*
    286   1.7      matt 		 * If we could enable them, then they exist.
    287   1.7      matt 		 */
    288   1.7      matt 		cpacr = armreg_cpacr_read();
    289  1.40      matt 		bool vfp_p = __SHIFTOUT(cpacr, cpacr_vfp2) == CPACR_ALL
    290  1.40      matt 		    && __SHIFTOUT(cpacr, cpacr_vfp) == CPACR_ALL;
    291  1.28      matt 		if (!vfp_p) {
    292  1.40      matt 			aprint_normal_dev(ci->ci_dev,
    293  1.40      matt 			    "VFP access denied (CPACR=%#x)\n", cpacr);
    294  1.62     skrll 			if (CPU_IS_PRIMARY(ci))
    295  1.62     skrll 				install_coproc_handler(VFP_COPROC, vfp_fpscr_handler);
    296  1.28      matt 			ci->ci_vfp_id = 0;
    297  1.37      matt 			evcnt_attach_dynamic(&ci->ci_vfp_evs[0],
    298  1.37      matt 			    EVCNT_TYPE_TRAP, NULL, ci->ci_cpuname,
    299  1.37      matt 			    "vfp fpscr traps");
    300  1.28      matt 			return;
    301  1.28      matt 		}
    302   1.6      matt 	}
    303   1.6      matt 
    304   1.7      matt 	void *uh = install_coproc_handler(VFP_COPROC, vfp_test);
    305   1.7      matt 
    306   1.7      matt 	undefined_test = 0;
    307   1.7      matt 
    308  1.21      matt 	const uint32_t fpsid = armreg_fpsid_read();
    309   1.1  rearnsha 
    310   1.1  rearnsha 	remove_coproc_handler(uh);
    311   1.1  rearnsha 
    312   1.1  rearnsha 	if (undefined_test != 0) {
    313   1.4      matt 		aprint_normal_dev(ci->ci_dev, "No VFP detected\n");
    314  1.62     skrll 		if (CPU_IS_PRIMARY(ci))
    315  1.62     skrll 			install_coproc_handler(VFP_COPROC, vfp_fpscr_handler);
    316   1.4      matt 		ci->ci_vfp_id = 0;
    317   1.1  rearnsha 		return;
    318   1.1  rearnsha 	}
    319   1.1  rearnsha 
    320   1.4      matt 	ci->ci_vfp_id = fpsid;
    321   1.4      matt 	switch (fpsid & ~ VFP_FPSID_REV_MSK) {
    322   1.4      matt 	case FPU_VFP10_ARM10E:
    323   1.4      matt 		model = "VFP10 R1";
    324   1.4      matt 		break;
    325   1.4      matt 	case FPU_VFP11_ARM11:
    326   1.4      matt 		model = "VFP11";
    327   1.4      matt 		break;
    328  1.36      matt 	case FPU_VFP_MV88SV58XX:
    329  1.36      matt 		model = "VFP3";
    330  1.36      matt 		break;
    331   1.7      matt 	case FPU_VFP_CORTEXA5:
    332   1.7      matt 	case FPU_VFP_CORTEXA7:
    333   1.7      matt 	case FPU_VFP_CORTEXA8:
    334   1.7      matt 	case FPU_VFP_CORTEXA9:
    335  1.63       tnn 	case FPU_VFP_CORTEXA12:
    336  1.20      matt 	case FPU_VFP_CORTEXA15:
    337  1.42       slp 	case FPU_VFP_CORTEXA15_QEMU:
    338  1.63       tnn 	case FPU_VFP_CORTEXA17:
    339  1.50     skrll 	case FPU_VFP_CORTEXA53:
    340  1.53  jmcneill 	case FPU_VFP_CORTEXA57:
    341  1.37      matt 		if (armreg_cpacr_read() & CPACR_V7_ASEDIS) {
    342  1.37      matt 			model = "VFP 4.0+";
    343  1.37      matt 		} else {
    344  1.37      matt 			model = "NEON MPE (VFP 3.0+)";
    345  1.37      matt 			cpu_neon_present = 1;
    346  1.37      matt 		}
    347   1.6      matt 		break;
    348   1.4      matt 	default:
    349  1.36      matt 		aprint_normal_dev(ci->ci_dev, "unrecognized VFP version %#x\n",
    350   1.4      matt 		    fpsid);
    351  1.62     skrll 		if (CPU_IS_PRIMARY(ci))
    352  1.62     skrll 			install_coproc_handler(VFP_COPROC, vfp_fpscr_handler);
    353  1.35      matt 		vfp_fpscr_changable = VFP_FPSCR_CSUM|VFP_FPSCR_ESUM
    354  1.35      matt 		    |VFP_FPSCR_RMODE;
    355  1.35      matt 		vfp_fpscr_default = 0;
    356   1.4      matt 		return;
    357   1.4      matt 	}
    358   1.1  rearnsha 
    359  1.17      matt 	cpu_fpu_present = 1;
    360  1.21      matt 	cpu_media_and_vfp_features[0] = armreg_mvfr0_read();
    361  1.21      matt 	cpu_media_and_vfp_features[1] = armreg_mvfr1_read();
    362   1.1  rearnsha 	if (fpsid != 0) {
    363  1.34      matt 		uint32_t f0 = armreg_mvfr0_read();
    364  1.41      matt 		uint32_t f1 = armreg_mvfr1_read();
    365  1.34      matt 		aprint_normal("vfp%d at %s: %s%s%s%s%s\n",
    366  1.37      matt 		    device_unit(ci->ci_dev),
    367  1.37      matt 		    device_xname(ci->ci_dev),
    368  1.34      matt 		    model,
    369  1.34      matt 		    ((f0 & ARM_MVFR0_ROUNDING_MASK) ? ", rounding" : ""),
    370  1.34      matt 		    ((f0 & ARM_MVFR0_EXCEPT_MASK) ? ", exceptions" : ""),
    371  1.38      matt 		    ((f1 & ARM_MVFR1_D_NAN_MASK) ? ", NaN propagation" : ""),
    372  1.34      matt 		    ((f1 & ARM_MVFR1_FTZ_MASK) ? ", denormals" : ""));
    373  1.49  jmcneill 		aprint_debug("vfp%d: mvfr: [0]=%#x [1]=%#x\n",
    374  1.37      matt 		    device_unit(ci->ci_dev), f0, f1);
    375  1.37      matt 		if (CPU_IS_PRIMARY(ci)) {
    376  1.37      matt 			if (f0 & ARM_MVFR0_ROUNDING_MASK) {
    377  1.37      matt 				vfp_fpscr_changable |= VFP_FPSCR_RMODE;
    378  1.37      matt 			}
    379  1.37      matt 			if (f1 & ARM_MVFR0_EXCEPT_MASK) {
    380  1.37      matt 				vfp_fpscr_changable |= VFP_FPSCR_ESUM;
    381  1.37      matt 			}
    382  1.38      matt 			// If hardware supports propagation of NaNs, select it.
    383  1.37      matt 			if (f1 & ARM_MVFR1_D_NAN_MASK) {
    384  1.37      matt 				vfp_fpscr_default &= ~VFP_FPSCR_DN;
    385  1.37      matt 				vfp_fpscr_changable |= VFP_FPSCR_DN;
    386  1.37      matt 			}
    387  1.37      matt 			// If hardware supports denormalized numbers, use it.
    388  1.37      matt 			if (cpu_media_and_vfp_features[1] & ARM_MVFR1_FTZ_MASK) {
    389  1.37      matt 				vfp_fpscr_default &= ~VFP_FPSCR_FZ;
    390  1.37      matt 				vfp_fpscr_changable |= VFP_FPSCR_FZ;
    391  1.37      matt 			}
    392  1.37      matt 		}
    393  1.37      matt 	}
    394  1.37      matt 	evcnt_attach_dynamic(&ci->ci_vfp_evs[0], EVCNT_TYPE_MISC, NULL,
    395  1.37      matt 	    ci->ci_cpuname, "vfp coproc use");
    396  1.37      matt 	evcnt_attach_dynamic(&ci->ci_vfp_evs[1], EVCNT_TYPE_MISC, NULL,
    397  1.37      matt 	    ci->ci_cpuname, "vfp coproc re-use");
    398  1.37      matt 	evcnt_attach_dynamic(&ci->ci_vfp_evs[2], EVCNT_TYPE_TRAP, NULL,
    399  1.37      matt 	    ci->ci_cpuname, "vfp coproc fault");
    400  1.62     skrll 	if (CPU_IS_PRIMARY(ci)) {
    401  1.62     skrll 		install_coproc_handler(VFP_COPROC, vfp_handler);
    402  1.62     skrll 		install_coproc_handler(VFP_COPROC2, vfp_handler);
    403  1.13      matt #ifdef CPU_CORTEX
    404  1.62     skrll 		if (cpu_neon_present)
    405  1.62     skrll 			install_coproc_handler(CORE_UNKNOWN_HANDLER, neon_handler);
    406  1.13      matt #endif
    407  1.62     skrll 	}
    408   1.1  rearnsha }
    409   1.1  rearnsha 
    410   1.1  rearnsha /* The real handler for VFP bounces.  */
    411   1.4      matt static int
    412  1.21      matt vfp_handler(u_int address, u_int insn, trapframe_t *frame, int fault_code)
    413   1.1  rearnsha {
    414   1.4      matt 	struct cpu_info * const ci = curcpu();
    415   1.1  rearnsha 
    416   1.1  rearnsha 	/* This shouldn't ever happen.  */
    417   1.1  rearnsha 	if (fault_code != FAULT_USER)
    418  1.14      matt 		panic("VFP fault at %#x in non-user mode", frame->tf_pc);
    419   1.1  rearnsha 
    420  1.27      matt 	if (ci->ci_vfp_id == 0) {
    421   1.1  rearnsha 		/* No VFP detected, just fault.  */
    422   1.1  rearnsha 		return 1;
    423  1.27      matt 	}
    424  1.27      matt 
    425  1.47      matt 	/*
    426  1.47      matt 	 * If we already own the FPU and it's enabled (and no exception), raise
    427  1.47      matt 	 * SIGILL.  If there is an exception, drop through to raise a SIGFPE.
    428  1.47      matt 	 */
    429  1.46      matt 	if (curcpu()->ci_pcu_curlwp[PCU_FPU] == curlwp
    430  1.47      matt 	    && (armreg_fpexc_read() & (VFP_FPEXC_EX|VFP_FPEXC_EN)) == VFP_FPEXC_EN)
    431  1.47      matt 		return 1;
    432  1.44      matt 
    433  1.27      matt 	/*
    434  1.27      matt 	 * Make sure we own the FP.
    435  1.27      matt 	 */
    436  1.27      matt 	pcu_load(&arm_vfp_ops);
    437   1.1  rearnsha 
    438  1.21      matt 	uint32_t fpexc = armreg_fpexc_read();
    439  1.21      matt 	if (fpexc & VFP_FPEXC_EX) {
    440  1.21      matt 		ksiginfo_t ksi;
    441  1.21      matt 		KASSERT(fpexc & VFP_FPEXC_EN);
    442  1.21      matt 
    443  1.37      matt 		curcpu()->ci_vfp_evs[2].ev_count++;
    444  1.21      matt 
    445  1.21      matt 		/*
    446  1.21      matt 		 * Need the clear the exception condition so any signal
    447  1.33     skrll 		 * and future use can proceed.
    448  1.21      matt 		 */
    449  1.31     skrll 		armreg_fpexc_write(fpexc & ~(VFP_FPEXC_EX|VFP_FPEXC_FSUM));
    450  1.21      matt 
    451  1.51       chs 		pcu_save(&arm_vfp_ops, curlwp);
    452  1.33     skrll 
    453  1.33     skrll 		/*
    454  1.33     skrll 		 * XXX Need to emulate bounce instructions here to get correct
    455  1.33     skrll 		 * XXX exception codes, etc.
    456  1.33     skrll 		 */
    457  1.21      matt 		KSI_INIT_TRAP(&ksi);
    458  1.21      matt 		ksi.ksi_signo = SIGFPE;
    459  1.21      matt 		if (fpexc & VFP_FPEXC_IXF)
    460  1.21      matt 			ksi.ksi_code = FPE_FLTRES;
    461  1.21      matt 		else if (fpexc & VFP_FPEXC_UFF)
    462  1.21      matt 			ksi.ksi_code = FPE_FLTUND;
    463  1.21      matt 		else if (fpexc & VFP_FPEXC_OFF)
    464  1.21      matt 			ksi.ksi_code = FPE_FLTOVF;
    465  1.21      matt 		else if (fpexc & VFP_FPEXC_DZF)
    466  1.21      matt 			ksi.ksi_code = FPE_FLTDIV;
    467  1.21      matt 		else if (fpexc & VFP_FPEXC_IOF)
    468  1.21      matt 			ksi.ksi_code = FPE_FLTINV;
    469  1.21      matt 		ksi.ksi_addr = (uint32_t *)address;
    470  1.21      matt 		ksi.ksi_trap = 0;
    471  1.21      matt 		trapsignal(curlwp, &ksi);
    472  1.21      matt 		return 0;
    473  1.21      matt 	}
    474  1.21      matt 
    475   1.4      matt 	/* Need to restart the faulted instruction.  */
    476   1.4      matt //	frame->tf_pc -= INSN_SIZE;
    477   1.4      matt 	return 0;
    478   1.4      matt }
    479   1.1  rearnsha 
    480  1.13      matt #ifdef CPU_CORTEX
    481  1.13      matt /* The real handler for NEON bounces.  */
    482  1.13      matt static int
    483  1.21      matt neon_handler(u_int address, u_int insn, trapframe_t *frame, int fault_code)
    484  1.13      matt {
    485  1.13      matt 	struct cpu_info * const ci = curcpu();
    486  1.13      matt 
    487  1.13      matt 	if (ci->ci_vfp_id == 0)
    488  1.13      matt 		/* No VFP detected, just fault.  */
    489  1.13      matt 		return 1;
    490  1.13      matt 
    491  1.13      matt 	if ((insn & 0xfe000000) != 0xf2000000
    492  1.13      matt 	    && (insn & 0xfe000000) != 0xf4000000)
    493  1.13      matt 		/* Not NEON instruction, just fault.  */
    494  1.13      matt 		return 1;
    495  1.13      matt 
    496  1.13      matt 	/* This shouldn't ever happen.  */
    497  1.13      matt 	if (fault_code != FAULT_USER)
    498  1.13      matt 		panic("NEON fault in non-user mode");
    499  1.13      matt 
    500  1.45      matt 	/* if we already own the FPU and it's enabled, raise SIGILL */
    501  1.45      matt 	if (curcpu()->ci_pcu_curlwp[PCU_FPU] == curlwp
    502  1.45      matt 	    && (armreg_fpexc_read() & VFP_FPEXC_EN) != 0)
    503  1.47      matt 		return 1;
    504  1.43      matt 
    505  1.13      matt 	pcu_load(&arm_vfp_ops);
    506  1.13      matt 
    507  1.13      matt 	/* Need to restart the faulted instruction.  */
    508  1.13      matt //	frame->tf_pc -= INSN_SIZE;
    509  1.13      matt 	return 0;
    510  1.13      matt }
    511  1.13      matt #endif
    512  1.13      matt 
    513   1.4      matt static void
    514  1.13      matt vfp_state_load(lwp_t *l, u_int flags)
    515   1.4      matt {
    516   1.4      matt 	struct pcb * const pcb = lwp_getpcb(l);
    517   1.4      matt 	struct vfpreg * const fregs = &pcb->pcb_vfp;
    518   1.1  rearnsha 
    519   1.1  rearnsha 	/*
    520   1.1  rearnsha 	 * Instrument VFP usage -- if a process has not previously
    521   1.1  rearnsha 	 * used the VFP, mark it as having used VFP for the first time,
    522   1.1  rearnsha 	 * and count this event.
    523   1.1  rearnsha 	 *
    524   1.1  rearnsha 	 * If a process has used the VFP, count a "used VFP, and took
    525   1.1  rearnsha 	 * a trap to use it again" event.
    526   1.1  rearnsha 	 */
    527  1.39     rmind 	if (__predict_false((flags & PCU_VALID) == 0)) {
    528  1.37      matt 		curcpu()->ci_vfp_evs[0].ev_count++;
    529  1.34      matt 		pcb->pcb_vfp.vfp_fpscr = vfp_fpscr_default;
    530   1.4      matt 	} else {
    531  1.37      matt 		curcpu()->ci_vfp_evs[1].ev_count++;
    532   1.4      matt 	}
    533   1.1  rearnsha 
    534  1.54    bouyer 	KASSERT((armreg_fpexc_read() & VFP_FPEXC_EN) == 0);
    535  1.39     rmind 	/*
    536  1.39     rmind 	 * If the VFP is already enabled we must be bouncing an instruction.
    537  1.39     rmind 	 */
    538  1.39     rmind 	if (flags & PCU_REENABLE) {
    539  1.39     rmind 		uint32_t fpexc = armreg_fpexc_read();
    540  1.39     rmind 		armreg_fpexc_write(fpexc | VFP_FPEXC_EN);
    541  1.54    bouyer 		fregs->vfp_fpexc |= VFP_FPEXC_EN;
    542  1.39     rmind 		return;
    543  1.39     rmind 	}
    544  1.54    bouyer 	KASSERT((fregs->vfp_fpexc & VFP_FPEXC_EN) == 0);
    545  1.33     skrll 
    546  1.39     rmind 	/*
    547  1.39     rmind 	 * Load and Enable the VFP (so that we can write the registers).
    548  1.39     rmind 	 */
    549  1.39     rmind 	fregs->vfp_fpexc |= VFP_FPEXC_EN;
    550  1.39     rmind 	armreg_fpexc_write(fregs->vfp_fpexc);
    551  1.54    bouyer 	KASSERT(curcpu()->ci_pcu_curlwp[PCU_FPU] == NULL);
    552  1.54    bouyer 	KASSERT(l->l_pcu_cpu[PCU_FPU] == NULL);
    553  1.13      matt 
    554  1.39     rmind 	load_vfpregs(fregs);
    555  1.39     rmind 	armreg_fpscr_write(fregs->vfp_fpscr);
    556  1.13      matt 
    557  1.39     rmind 	if (fregs->vfp_fpexc & VFP_FPEXC_EX) {
    558  1.39     rmind 		/* Need to restore the exception handling state.  */
    559  1.52       chs 		armreg_fpinst_write(fregs->vfp_fpinst);
    560  1.39     rmind 		if (fregs->vfp_fpexc & VFP_FPEXC_FP2V)
    561  1.52       chs 			armreg_fpinst2_write(fregs->vfp_fpinst2);
    562   1.1  rearnsha 	}
    563   1.1  rearnsha }
    564   1.1  rearnsha 
    565   1.1  rearnsha void
    566  1.39     rmind vfp_state_save(lwp_t *l)
    567   1.1  rearnsha {
    568   1.4      matt 	struct pcb * const pcb = lwp_getpcb(l);
    569  1.39     rmind 	struct vfpreg * const fregs = &pcb->pcb_vfp;
    570  1.21      matt 	uint32_t fpexc = armreg_fpexc_read();
    571  1.33     skrll 
    572  1.54    bouyer 	KASSERT(curcpu()->ci_pcu_curlwp[PCU_FPU] == l);
    573  1.54    bouyer 	KASSERT(curcpu() == l->l_pcu_cpu[PCU_FPU]);
    574  1.54    bouyer 	KASSERT(curlwp == l || curlwp->l_pcu_cpu[PCU_FPU] != curcpu());
    575  1.33     skrll 	/*
    576  1.33     skrll 	 * Enable the VFP (so we can read the registers).
    577  1.33     skrll 	 * Make sure the exception bit is cleared so that we can
    578  1.33     skrll 	 * safely dump the registers.
    579  1.33     skrll 	 */
    580  1.21      matt 	armreg_fpexc_write((fpexc | VFP_FPEXC_EN) & ~VFP_FPEXC_EX);
    581   1.1  rearnsha 
    582   1.4      matt 	fregs->vfp_fpexc = fpexc;
    583   1.4      matt 	if (fpexc & VFP_FPEXC_EX) {
    584   1.4      matt 		/* Need to save the exception handling state */
    585  1.21      matt 		fregs->vfp_fpinst = armreg_fpinst_read();
    586  1.21      matt 		if (fpexc & VFP_FPEXC_FP2V)
    587  1.21      matt 			fregs->vfp_fpinst2 = armreg_fpinst2_read();
    588   1.1  rearnsha 	}
    589  1.21      matt 	fregs->vfp_fpscr = armreg_fpscr_read();
    590  1.13      matt 	save_vfpregs(fregs);
    591   1.4      matt 
    592   1.1  rearnsha 	/* Disable the VFP.  */
    593  1.33     skrll 	armreg_fpexc_write(fpexc & ~VFP_FPEXC_EN);
    594   1.1  rearnsha }
    595   1.1  rearnsha 
    596   1.1  rearnsha void
    597  1.39     rmind vfp_state_release(lwp_t *l)
    598   1.1  rearnsha {
    599   1.4      matt 	struct pcb * const pcb = lwp_getpcb(l);
    600   1.1  rearnsha 
    601  1.39     rmind 	/*
    602  1.39     rmind 	 * Now mark the VFP as disabled (and our state
    603  1.39     rmind 	 * has been already saved or is being discarded).
    604  1.39     rmind 	 */
    605  1.39     rmind 	pcb->pcb_vfp.vfp_fpexc &= ~VFP_FPEXC_EN;
    606   1.1  rearnsha 
    607   1.1  rearnsha 	/*
    608   1.4      matt 	 * Turn off the FPU so the next time a VFP instruction is issued
    609   1.4      matt 	 * an exception happens.  We don't know if this LWP's state was
    610   1.4      matt 	 * loaded but if we turned off the FPU for some other LWP, when
    611   1.4      matt 	 * pcu_load invokes vfp_state_load it will see that VFP_FPEXC_EN
    612  1.13      matt 	 * is still set so it just restore fpexc and return since its
    613   1.4      matt 	 * contents are still sitting in the VFP.
    614   1.1  rearnsha 	 */
    615  1.21      matt 	armreg_fpexc_write(armreg_fpexc_read() & ~VFP_FPEXC_EN);
    616   1.1  rearnsha }
    617   1.1  rearnsha 
    618   1.1  rearnsha void
    619  1.51       chs vfp_savecontext(lwp_t *l)
    620   1.1  rearnsha {
    621  1.51       chs 	pcu_save(&arm_vfp_ops, l);
    622   1.1  rearnsha }
    623   1.1  rearnsha 
    624   1.1  rearnsha void
    625  1.51       chs vfp_discardcontext(lwp_t *l, bool used_p)
    626   1.1  rearnsha {
    627  1.51       chs 	pcu_discard(&arm_vfp_ops, l, used_p);
    628  1.25      matt }
    629  1.25      matt 
    630  1.25      matt bool
    631  1.51       chs vfp_used_p(const lwp_t *l)
    632  1.25      matt {
    633  1.51       chs 	return pcu_valid_p(&arm_vfp_ops, l);
    634  1.13      matt }
    635  1.13      matt 
    636  1.13      matt void
    637   1.8      matt vfp_getcontext(struct lwp *l, mcontext_t *mcp, int *flagsp)
    638   1.8      matt {
    639  1.51       chs 	if (vfp_used_p(l)) {
    640   1.8      matt 		const struct pcb * const pcb = lwp_getpcb(l);
    641  1.51       chs 
    642  1.51       chs 		pcu_save(&arm_vfp_ops, l);
    643   1.8      matt 		mcp->__fpu.__vfpregs.__vfp_fpscr = pcb->pcb_vfp.vfp_fpscr;
    644   1.8      matt 		memcpy(mcp->__fpu.__vfpregs.__vfp_fstmx, pcb->pcb_vfp.vfp_regs,
    645   1.8      matt 		    sizeof(mcp->__fpu.__vfpregs.__vfp_fstmx));
    646  1.10      matt 		*flagsp |= _UC_FPU|_UC_ARM_VFP;
    647   1.8      matt 	}
    648   1.8      matt }
    649   1.8      matt 
    650   1.8      matt void
    651   1.8      matt vfp_setcontext(struct lwp *l, const mcontext_t *mcp)
    652   1.8      matt {
    653   1.8      matt 	struct pcb * const pcb = lwp_getpcb(l);
    654  1.51       chs 
    655  1.51       chs 	pcu_discard(&arm_vfp_ops, l, true);
    656   1.8      matt 	pcb->pcb_vfp.vfp_fpscr = mcp->__fpu.__vfpregs.__vfp_fpscr;
    657   1.8      matt 	memcpy(pcb->pcb_vfp.vfp_regs, mcp->__fpu.__vfpregs.__vfp_fstmx,
    658   1.8      matt 	    sizeof(mcp->__fpu.__vfpregs.__vfp_fstmx));
    659   1.8      matt }
    660   1.8      matt 
    661   1.4      matt #endif /* FPU_VFP */
    662