>
> this is a new version of the Katmai patch. Changes:
>
> - disables PIII serial# on all CPUs (Doug Ledford)
> [we might make this configurable in the future for Wine
> compatibility]
>
> - more correct tagword conversion (Gabriel Paubert)
Nitpicking again :-)
[trimmed]
> --- linux/include/asm-i386/processor.h.orig Tue Jan 26 10:47:10 1999
> +++ linux/include/asm-i386/processor.h Tue Mar 2 21:47:41 1999
> @@ -164,6 +214,27 @@
> */
> #define IO_BITMAP_SIZE 32
>
> +/*
> + * We have two (incompatible) floating-point state 'hard' layouts:
> + *
> + * - the new MMX2-ready layout used by newer PII/Xeon/PIII(Katmai) CPUs
> + * - the 'old' i387 FPU layout used by everything else
> + *
> + * we use the one apropriate for the CPU picked at kernel compilation time.
> + *
> + * to keep the FPU code a bit more transparent, we define the same
> + * structure in both cases, the parameter names are identical too.
> + * for cases where we cannot make this transparent, we define access
> + * functions.
> + *
> + * we also have two FPU formats visible in APIs, 'hard format' and
> + * 'user-format'. The user-format is decoupled from the actual hard
> + * format (which can be i387 or KNI) and conversion routines are
> + * provided. user-format is currently i387, which we cannot change
> + * (only extend) due to external APIs (iBCS2, ELF coredumps, ptrace API).
> + */
> +
> +#ifndef CONFIG_X86_FX
> struct i387_hard_struct {
> long cwd;
> long swd;
> @@ -176,6 +247,84 @@
> long status; /* software status information */
> };
>
> +/*
> + * Fill out the reserved bits
> + */
> +#define i387_set_cwd(x,v) do { (x).cwd = ((long)(v) | 0xffff0000); } while (0)
> +#define i387_set_swd(x,v) do { (x).swd = ((long)(v) | 0xffff0000); } while (0)
> +#define i387_set_twd(x,v) do { (x).twd = ((long)(v) | 0xffff0000); } while (0)
> +
> +#define i387_save_hard(x) \
> + do { __asm__("fnsave %0; fwait;": "=m" (x)); } while (0)
> +
> +#define i387_restore_hard(x) \
> + do { __asm__("frstor %0": :"m" (x)); } while (0)
> +
> +#else
> +
> +/*
> + * has to be 128-bit aligned
> + */
> +struct i387_hard_struct {
> + unsigned short cwd;
> + unsigned short swd;
> + unsigned char twd;
> + unsigned char __reserved_01;
> + unsigned short fopcode;
> + long fip;
> + long fcs;
> + long foo;
> + long fos;
> + long __reserved_02, __reserved_03;
> + long st_space[32]; /* 8*16 bytes for each FP/MMX-reg = 128 bytes */
> + long __reserved_04 [22*16]; /* 22 cachelines for MMX2 registers */
> +} __attribute__ ((aligned (16)));
Actually __reserved_02 is now defined as mxcsr. And unless I'm mistaken
sizeof(i387_hard_struct) here is around 1568 bytes while intel only
requires 512. The last element should be split into long xmmregs[32] and
a reserved area of 224 bytes (56 long).
> +
> +/*
> + * tag word conversion (thanks to Gabriel Paubert for noticing the
> + * subtle format difference and implementing these functions)
> + *
> + * there are several erratas wrt. the tag word in the i387, thus
> + * any software relying on it's value is questionable, but we
> + * definitely want to be as close as possible.
> + */
> +static inline unsigned short fputag_KNIto387(unsigned char tb) {
> + unsigned short tw = tb;
> + tw = ((tw<<4) | tw) &0x0f0f; /* zzzz7654zzzz3210 */
> + tw = ((tw<<2) | tw) &0x3333; /* zz76zz54zz32zz10 */
> + tw = ((tw<<1) | tw) &0x5555; /* z7z6z5z4z3z2z1z0 */
> + return ~(tw*3);
> +}
> +
> +static inline unsigned char fputag_387toKNI(unsigned short tw) {
> + tw = ~tw;
> + tw = (tw | (tw>>1)) & 0x5555; /* z7z6z5z4z3z2z1z0 */
> + tw = (tw | (tw>>1)) & 0x3333; /* zz76zz54zz32zz10 */
> + tw = (tw | (tw>>3)) & 0x0f0f; /* zzzz7654zzzz3210 */
^
typo here: should be 2, my fault :-(
> + return (tw|(tw>>4)) & 0x00ff; /* zzzzzzzz76543210 */
> +}
> +
> +#define i387_set_cwd(x,v) do { (x).cwd = (short)(v); } while (0)
> +#define i387_set_swd(x,v) do { (x).swd = (short)(v); } while (0)
> +#define i387_set_twd(x,v) do { (x).twd = fputag_387toKNI(v); } while (0)
> +
> +/*
> + * GAS is not ready yet, so we encode the FXSAVE/FXRSTOR
> + * opcodes directly:
> + */
> +#define i387_save_hard(x) \
> +do { __asm__ __volatile__(".byte 0x0f, 0xae, 0x06": :"S" (&(x))); } while (0)
> +
> +#define i387_restore_hard(x) \
> +do { __asm__ __volatile__(".byte 0x0f, 0xae, 0x4f, 0x00": :"D" (&(x))); } while (0)
> +
> +#endif
> +
> +extern int i387_hard_to_user ( struct user_i387_struct * user,
> + struct i387_hard_struct * hard);
> +extern int i387_user_to_hard (struct i387_hard_struct * hard,
> + struct user_i387_struct * user);
> +
> struct i387_soft_struct {
> long cwd;
> long swd;
Gabriel.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/