xref: /src/sys/arch/x86/include/cpu_extended_state.h

/*	$NetBSD: cpu_extended_state.h,v 1.7 2014/02/15 10:11:15 dsl Exp $	*/

#ifndef _X86_CPU_EXTENDED_STATE_H_
#define _X86_CPU_EXTENDED_STATE_H_

/*
 * This file contains definitions of structures that match the memory
 * layouts used x86 processors to save floating point registers and other
 * extended cpu state.
 * This includes registers (etc) used by SSE/SSE2/SSE3/SSSE3/SSE4 and
 * the later AVX instructions.
 * The definitions are such that any future 'extended state' should
 * be handled (provided the kernel doesn't need to know the actual contents.
 *
 * The actual structures the cpu accesses must be aligned to 16 for
 * FXSAVE and 64 for XSAVE. The types aren't aligned because copies
 * do not need extra alignment.
 *
 * The slightly different layout saved by the i387 fsave in also defined.
 * This is only normally written by pre Pentium II type cpus that don't
 * support the fxsave instruction.
 *
 * Associated save instructions:
 * FNSAVE:  Saves x87 state in 108 bytes (original i387 layout).
 *          Then reinitialies the fpu.
 * FSAVE:   Encodes to FWAIT followed by FNSAVE.
 * FXSAVE:  Saves the x87 state and XMM (aka SSE) registers to the
 *          first 448 (max) bytes of a 512 byte area.
 *          This layout does not match that written by FNSAVE.
 * XSAVE:   Uses the same layout for the x87 and XMM registers,
 *          followed by a 64byte header and separate save areas
 *          for additional extended cpu state.
 *          The x87 state is always saved, the others conditionally.
 * XSAVEOPT: As XSAVE but (IIRC) only writes the registers blocks
 *          that have been modified.
 */

#ifdef __lint__
/* Lint has different packing rules and doesn't understand __aligned() */
#define __CTASSERT_NOLINT(x) __CTASSERT(1)
#else
#define __CTASSERT_NOLINT(x) __CTASSERT(x)
#endif

/*
 * Layout for code/data pointers relating to FP exceptions.
 * Marked 'packed' because they aren't always 64bit aligned.
 * Since the x86 cpu supports misaligned accesses it isn't
 * worth avoiding the 'packed' attribute.
 */
union fp_addr {
	uint64_t fa_64;	/* Linear address for 64bit systems */
	struct {
		uint32_t fa_off;	/* linear address for 32 bit */
		uint16_t fa_seg;	/* code/data (etc) segment */
		uint16_t fa_opcode;	/* last opcode (sometimes) */
	} fa_32;
} __packed __aligned(4);

/* The x87 registers are 80 bits */
struct fpacc87 {
	uint64_t	f87_mantissa;	/* mantissa */
	uint16_t	f87_exp_sign;	/* exponent and sign */
} __packed __aligned(2);

/* The x87 registers padded out to 16 bytes for fxsave */
struct fpaccfx {
	struct fpacc87 r __aligned(16);
};

/* The SSE/SSE2 registers are 128 bits */
struct xmmreg {
	uint8_t xmm_bytes[16];
};

/* The AVX registers are 256 bits, but the low bits are the xmmregs */
struct ymmreg {
	uint8_t ymm_bytes[16];
};

/*
 * Floating point unit registers (fsave instruction).
 * The s87_ac[] and fx_87_ac[] are relative to the stack top.
 * The 'tag word' contains 2 bits per register and refers to
 * absolute register numbers.
 * The cpu sets the tag values 0b01 (zero) and 0b10 (special) when a value
 * is loaded. The software need only set 0b00 (used) and 0xb11 (unused).
 * The fxsave 'Abridged tag word' in inverted.
 */
struct save87 {
	uint16_t	s87_cw __aligned(4);	/* control word (16bits) */
	uint16_t	s87_sw __aligned(4);	/* status word (16bits) */
	uint16_t	s87_tw __aligned(4);	/* tag word (16bits) */
	union fp_addr	s87_ip;		/* floating point instruction pointer */
#define s87_opcode s87_ip.fa_32.fa_opcode	/* opcode last executed (11bits) */
	union fp_addr	s87_dp;		/* floating operand offset */
	struct fpacc87	s87_ac[8];	/* accumulator contents, 0-7 */
};
__CTASSERT_NOLINT(sizeof (struct save87) == 108);

/* FPU/MMX/SSE/SSE2 context */
struct fxsave {
/*0*/	uint16_t	fx_cw;		/* FPU Control Word */
	uint16_t	fx_sw;		/* FPU Status Word */
	uint8_t		fx_tw;		/* FPU Tag Word (abridged) */
	uint16_t	fx_opcode;	/* FPU Opcode */
	union fp_addr	fx_ip;		/* FPU Instruction Pointer */
/*16*/	union fp_addr	fx_dp;		/* FPU Data pointer */
	uint32_t	fx_mxcsr;	/* MXCSR Register State */
	uint32_t	fx_mxcsr_mask;
	struct fpaccfx	fx_87_ac[8];	/* 8 x87 registers */
	struct xmmreg	fx_xmm[16];	/* XMM regs (8 in 32bit modes) */
	uint8_t		fx_rsvd[48];
	uint8_t		fx_kernel[48];	/* Not written by the hardware */
} __aligned(16);
__CTASSERT_NOLINT(sizeof (struct fxsave) == 512);

#ifndef _KERNEL
/* Backwards compatibility for firefox (looks at fx_xmm) */
#define fxsave64 fxsave
#endif

/* The end of the fsave buffer can be used by the operating system */
struct fxsave_os {
	uint8_t		fxo_fxsave[512 - 48];
	/* 48 bytes available, NB copied to/from userspace */
	uint16_t	fxo_dflt_cw;	/* Control word for signal handlers */
};

union savefpu {
	struct save87		sv_87;
	struct fxsave		sv_xmm;
	struct fxsave_os	sv_os;
};

/*
 * For XSAVE a 64byte header follows the above.
 * Currently it only contains one field of which only 3 bits are defined.
 * Some other parts must be zero - zero it all.
 *
 * The xsh_xstate_bv bits match those of XCR0:
 *   XCR0_X87        0x00000001      x87 FPU/MMX state (always set)
 *   XCR0_SSE        0x00000002      SSE state
 *   XCR0_AVX        0x00000004      AVX state (ymmn registers)
 *
 * The offsets and sizes of any save areas can be found by reading
 * the correct control registers.
 */

struct xsave_header {
	uint64_t	xsh_xstate_bv;	/* bitmap of saved sub structures */
	uint64_t	xsh_rsrvd[2];	/* must be zero */
	uint64_t	xsh_reserved[5];/* best if zero */
};
__CTASSERT(sizeof (struct xsave_header) == 64);

/*
 * The ymm save area actually follows the xsave_header.
 */
struct xsave_ymm {
	struct ymmreg	xs_ymm[16];	/* High bits of YMM registers */
};
__CTASSERT(sizeof (struct xsave_ymm) == 256);


/*
 * 80387 control and status word bits
 *
 * The only reference I can find to bits 0x40 and 0x80 in the control word
 * is for the Weitek 1167/3167.
 * I (dsl) can't find why the default word has 0x40 set.
 *
 * A stack error is signalled as an INVOP that also sets STACK_FAULT
 * (other INVOP do not clear STACK_FAULT).
 */
/* Interrupt masks (set masks interrupt) and status bits */
#define EN_SW_INVOP		0x0001  /* Invalid operation */
#define EN_SW_DENORM		0x0002  /* Denormalized operand */
#define EN_SW_ZERODIV		0x0004  /* Divide by zero */
#define EN_SW_OVERFLOW		0x0008  /* Overflow */
#define EN_SW_UNDERFLOW		0x0010  /* Underflow */
#define EN_SW_PRECLOSS		0x0020  /* Loss of precision */
/* Status word bits (reserved in control word) */
#define EN_SW_STACK_FAULT	0x0040	/* Stack under/overflow */
#define EN_SW_ERROR_SUMMARY	0x0080	/* Unmasked error has ocurred */
/* Control bits (badly named) */
#define EN_SW_CTL_PREC		0x0300	/* Precision control */
#define EN_SW_PREC_24		0x0000	/* Single precision */
#define EN_SW_PREC_53		0x0200	/* Double precision */
#define EN_SW_PREC_64		0x0300	/* Extended precision */
#define EN_SW_CTL_ROUND		0x0c00	/* Rounding control */
#define EN_SW_ROUND_EVEN	0x0000	/* Round to nearest even */
#define EN_SW_ROUND_DOWN	0x0400	/* Round towards minus infinity */
#define EN_SW_ROUND_UP		0x0800	/* Round towards plus infinity */
#define EN_SW_ROUND_ZERO	0x0c00	/* Round towards zero (truncates) */
#define EN_SW_CTL_INF		0x1000	/* Infinity control, not used  */

/*
 * The standard 0x87 control word from finit is 0x37F, giving:
 *	round to nearest
 *	64-bit precision
 *	all exceptions masked.
 *
 * NetBSD used to select:
 *	round to nearest
 *	53-bit precision
 *	all exceptions masked.
 * Stating: 64-bit precision often gives bad results with high level
 * languages because it makes the results of calculations depend on whether
 * intermediate values are stored in memory or in FPU registers.
 * Also some 'pathological divisions' give an error in the LSB because
 * the value is first rounded up when the 64bit mantissa is generated,
 * and then again when it is truncated to 53 bits.
 *
 * However the C language explicitly allows the extra precision.
 *
 * The iBCS control word has underflow, overflow, zero divide, and invalid
 * operation exceptions unmasked.  But that causes an unexpected exception
 * in the test program 'paranoia' and makes denormals useless (DBL_MIN / 2
 * underflows).  It doesn't make a lot of sense to trap underflow without
 * trapping denormals.
 */
#define	__INITIAL_NPXCW__	0x037f
/* Modern NetBSD uses the default control word.. */
#define	__NetBSD_NPXCW__	__INITIAL_NPXCW__
/* NetBSD before 6.99.26 forced IEEE double precision. */
#define	__NetBSD_COMPAT_NPXCW__	0x127f
/* FreeBSD leaves some exceptions unmasked as well. */
#define	__FreeBSD_NPXCW__	0x1272
/* iBCS2 goes a bit further and leaves the underflow exception unmasked. */
#define	__iBCS2_NPXCW__		0x0262
/* Linux just uses the default control word. */
#define	__Linux_NPXCW__		__INITIAL_NPXCW__
/* SVR4 uses the same control word as iBCS2. */
#define	__SVR4_NPXCW__		0x0262

/*
 * The default MXCSR value at reset is 0x1f80, IA-32 Instruction
 * Set Reference, pg. 3-369.
 *
 * The low 6 bits of the mxcsr are the fp status bits (same order as x87).
 * Bit 6 is 'denormals are zero' (speeds up calculations).
 * Bits 7-16 are the interrupt mask bits (same order, 1 to mask).
 * Bits 13 and 14 are rounding control.
 * Bit 15 is 'flush to zero' - affects underflow.
 * Bits 16-31 must be zero.
 */
#define	__INITIAL_MXCSR__	0x1f80
#define	__INITIAL_MXCSR_MASK__	0xffbf

#endif /* _X86_CPU_EXTENDED_STATE_H_ */