Line data    Source code

       1             : /*      $OpenBSD: fpu.h,v 1.15 2018/06/24 00:49:25 guenther Exp $       */
       2             : /*      $NetBSD: fpu.h,v 1.1 2003/04/26 18:39:40 fvdl Exp $     */
       3             : 
       4             : #ifndef _MACHINE_FPU_H_
       5             : #define _MACHINE_FPU_H_
       6             : 
       7             : #include <sys/types.h>
       8             : 
       9             : /*
      10             :  * If the CPU supports xsave/xrstor then we use them so that we can provide
      11             :  * AVX support.  Otherwise we require fxsave/fxrstor, as the SSE registers
      12             :  * are part of the ABI for passing floating point values.
      13             :  * While fxsave/fxrstor only required 16-byte alignment for the save area,
      14             :  * xsave/xrstor requires the save area to have 64-byte alignment.
      15             :  */
      16             : 
      17             : struct fxsave64 {
      18             :         u_int16_t  fx_fcw;
      19             :         u_int16_t  fx_fsw;
      20             :         u_int8_t   fx_ftw;
      21             :         u_int8_t   fx_unused1;
      22             :         u_int16_t  fx_fop;
      23             :         u_int64_t  fx_rip;
      24             :         u_int64_t  fx_rdp;
      25             :         u_int32_t  fx_mxcsr;
      26             :         u_int32_t  fx_mxcsr_mask;
      27             :         u_int64_t  fx_st[8][2];   /* 8 normal FP regs */
      28             :         u_int64_t  fx_xmm[16][2]; /* 16 SSE2 registers */
      29             :         u_int8_t   fx_unused3[96];
      30             : } __packed;
      31             : 
      32             : struct xstate_hdr {
      33             :         uint64_t        xstate_bv;
      34             :         uint64_t        xstate_xcomp_bv;
      35             :         uint8_t         xstate_rsrv0[0];
      36             :         uint8_t         xstate_rsrv[40];
      37             : } __packed;
      38             : 
      39             : struct savefpu {
      40             :         struct fxsave64 fp_fxsave;      /* see above */
      41             :         struct xstate_hdr fp_xstate;
      42             :         u_int64_t fp_ymm[16][2];
      43             :         u_int16_t fp_ex_sw;             /* saved status from last exception */
      44             :         u_int16_t fp_ex_tw;             /* saved tag from last exception */
      45             : };
      46             : 
      47             : /*
      48             :  * The i387 defaults to Intel extended precision mode and round to nearest,
      49             :  * with all exceptions masked.
      50             :  */
      51             : #define __INITIAL_NPXCW__       0x037f
      52             : #define __INITIAL_MXCSR__       0x1f80
      53             : #define __INITIAL_MXCSR_MASK__  0xffbf
      54             : 
      55             : #ifdef _KERNEL
      56             : /*
      57             :  * XXX
      58             :  */
      59             : struct trapframe;
      60             : struct cpu_info;
      61             : 
      62             : extern size_t   fpu_save_len;
      63             : extern uint32_t fpu_mxcsr_mask;
      64             : extern uint64_t xsave_mask;
      65             : 
      66             : void fpuinit(struct cpu_info *);
      67             : int fputrap(int _type);
      68             : void fpusave(struct savefpu *);
      69             : void fpusavereset(struct savefpu *);
      70             : void fpu_kernel_enter(void);
      71             : void fpu_kernel_exit(void);
      72             : 
      73             : int     xrstor_user(struct savefpu *_addr, uint64_t _mask);
      74             : #define fpureset() \
      75             :         xrstor_user(&proc0.p_addr->u_pcb.pcb_savefpu, xsave_mask)
      76             : 
      77             : #define fninit()                __asm("fninit")
      78             : #define fwait()                 __asm("fwait")
      79             : /* should be fxsave64, but where we use this it doesn't matter */
      80             : #define fxsave(addr)            __asm("fxsave %0" : "=m" (*addr))
      81             : #define ldmxcsr(addr)           __asm("ldmxcsr %0" : : "m" (*addr))
      82             : #define fldcw(addr)             __asm("fldcw %0" : : "m" (*addr))
      83             : 
      84             : static inline void
      85           0 : xsave(struct savefpu *addr, uint64_t mask)
      86             : {
      87             :         uint32_t lo, hi;
      88             : 
      89           0 :         lo = mask;
      90           0 :         hi = mask >> 32;
      91             :         /* should be xsave64, but where we use this it doesn't matter */
      92           0 :         __asm volatile("xsave %0" : "=m" (*addr) : "a" (lo), "d" (hi) :
      93             :             "memory");
      94           0 : }
      95             : 
      96             : #endif
      97             : 
      98             : #endif /* _MACHINE_FPU_H_ */