[PATCH 07/14] S390: Optimize utf8-utf32 module.

Stefan Liebler stli@linux.vnet.ibm.com
Thu Apr 21 15:15:00 GMT 2016


Here is an updated patch, where the labels in inline assemblies are 
out-dented as suggested by Florian.

On 02/23/2016 10:21 AM, Stefan Liebler wrote:
> This patch reworks the s390 specific module to convert between utf8 and utf32.
> Now ifunc is used to choose either the c or etf3eh (with convert utf
> instruction) variants at runtime.
> Furthermore a new vector variant for z13 is introduced which will be build
> and chosen if vector support is available at build / runtime.
> The vector variants optimize input of 1byte utf8 characters. The convert utf
> instruction is used if a multibyte utf8 character is found.
>
> This patch also fixes some whitespace errors. The c variants are rejecting
> UTF-16 surrogates and values above 0x10ffff now.
> Furthermore, the etf3eh variants are handling the "UTF-xx//IGNORE" case now.
> Before they ignored the ignore-case and always stopped at an error.
>
> ChangeLog:
>
> 	* sysdeps/s390/s390-64/utf8-utf32-z9.c: Use ifunc to select c, etf3eh
> 	or new vector loop-variant.
> ---
>   sysdeps/s390/s390-64/utf8-utf32-z9.c | 664 +++++++++++++++++++++++++----------
>   1 file changed, 480 insertions(+), 184 deletions(-)
>
> diff --git a/sysdeps/s390/s390-64/utf8-utf32-z9.c b/sysdeps/s390/s390-64/utf8-utf32-z9.c
> index defd47d..e89dc70 100644
> --- a/sysdeps/s390/s390-64/utf8-utf32-z9.c
> +++ b/sysdeps/s390/s390-64/utf8-utf32-z9.c
> @@ -30,35 +30,25 @@
>   #include <dl-procinfo.h>
>   #include <gconv.h>
>
> -/* UTF-32 big endian byte order mark.  */
> -#define BOM	                0x0000feffu
> +#if defined HAVE_S390_VX_GCC_SUPPORT
> +# define ASM_CLOBBER_VR(NR) , NR
> +#else
> +# define ASM_CLOBBER_VR(NR)
> +#endif
>
> +/* Defines for skeleton.c.  */
>   #define DEFINE_INIT		0
>   #define DEFINE_FINI		0
> -/* These definitions apply to the UTF-8 to UTF-32 direction.  The
> -   software implementation for UTF-8 still supports multibyte
> -   characters up to 6 bytes whereas the hardware variant does not.  */
>   #define MIN_NEEDED_FROM		1
>   #define MAX_NEEDED_FROM		6
>   #define MIN_NEEDED_TO		4
> -#define FROM_LOOP		from_utf8_loop
> -#define TO_LOOP			to_utf8_loop
> +#define FROM_LOOP		__from_utf8_loop
> +#define TO_LOOP			__to_utf8_loop
>   #define FROM_DIRECTION		(dir == from_utf8)
>   #define ONE_DIRECTION           0
> -#define PREPARE_LOOP							\
> -  enum direction dir = ((struct utf8_data *) step->__data)->dir;	\
> -  int emit_bom = ((struct utf8_data *) step->__data)->emit_bom;		\
> -									\
> -  if (emit_bom && !data->__internal_use					\
> -      && data->__invocation_counter == 0)				\
> -    {									\
> -      /* Emit the Byte Order Mark.  */					\
> -      if (__glibc_unlikely (outbuf + 4 > outend))			      \
> -	return __GCONV_FULL_OUTPUT;					\
> -									\
> -      put32u (outbuf, BOM);						\
> -      outbuf += 4;							\
> -    }
> +
> +/* UTF-32 big endian byte order mark.  */
> +#define BOM			0x0000feffu
>
>   /* Direction of the transformation.  */
>   enum direction
> @@ -155,16 +145,16 @@ gconv_end (struct __gconv_step *data)
>       register unsigned long long outlen __asm__("11") = outend - outptr;	\
>       uint64_t cc = 0;							\
>   									\
> -    __asm__ volatile (".machine push       \n\t"			\
> -		      ".machine \"z9-109\" \n\t"			\
> -		      "0: " INSTRUCTION "  \n\t"			\
> -		      ".machine pop        \n\t"			\
> -		      "   jo     0b        \n\t"			\
> -		      "   ipm    %2        \n"				\
> -		      : "+a" (pOutput), "+a" (pInput), "+d" (cc),	\
> -		      "+d" (outlen), "+d" (inlen)			\
> -		      :							\
> -		      : "cc", "memory");				\
> +    __asm__ __volatile__ (".machine push       \n\t"			\
> +			  ".machine \"z9-109\" \n\t"			\
> +			  "0: " INSTRUCTION "  \n\t"			\
> +			  ".machine pop        \n\t"			\
> +			  "   jo     0b        \n\t"			\
> +			  "   ipm    %2        \n"			\
> +			  : "+a" (pOutput), "+a" (pInput), "+d" (cc),	\
> +			    "+d" (outlen), "+d" (inlen)			\
> +			  :						\
> +			  : "cc", "memory");				\
>   									\
>       inptr = pInput;							\
>       outptr = pOutput;							\
> @@ -173,49 +163,150 @@ gconv_end (struct __gconv_step *data)
>       if (cc == 1)							\
>         {									\
>   	result = __GCONV_FULL_OUTPUT;					\
> -	break;								\
>         }									\
>       else if (cc == 2)							\
>         {									\
>   	result = __GCONV_ILLEGAL_INPUT;					\
> -	break;								\
>         }									\
>     }
>
> +#define PREPARE_LOOP							\
> +  enum direction dir = ((struct utf8_data *) step->__data)->dir;	\
> +  int emit_bom = ((struct utf8_data *) step->__data)->emit_bom;		\
> +									\
> +  if (emit_bom && !data->__internal_use					\
> +      && data->__invocation_counter == 0)				\
> +    {									\
> +      /* Emit the Byte Order Mark.  */					\
> +      if (__glibc_unlikely (outbuf + 4 > outend))			\
> +	return __GCONV_FULL_OUTPUT;					\
> +									\
> +      put32u (outbuf, BOM);						\
> +      outbuf += 4;							\
> +    }
> +
>   /* Conversion function from UTF-8 to UTF-32 internal/BE.  */
>
> -#define MIN_NEEDED_INPUT	MIN_NEEDED_FROM
> -#define MAX_NEEDED_INPUT	MAX_NEEDED_FROM
> -#define MIN_NEEDED_OUTPUT	MIN_NEEDED_TO
> -#define LOOPFCT			FROM_LOOP
> -/* The software routine is copied from gconv_simple.c.  */
> -#define BODY								\
> +#define STORE_REST_COMMON						      \
> +  {									      \
> +    /* We store the remaining bytes while converting them into the UCS4	      \
> +       format.  We can assume that the first byte in the buffer is	      \
> +       correct and that it requires a larger number of bytes than there	      \
> +       are in the input buffer.  */					      \
> +    wint_t ch = **inptrp;						      \
> +    size_t cnt, r;							      \
> +									      \
> +    state->__count = inend - *inptrp;					      \
> +									      \
> +    assert (ch != 0xc0 && ch != 0xc1);					      \
> +    if (ch >= 0xc2 && ch < 0xe0)					      \
> +      {									      \
> +	/* We expect two bytes.  The first byte cannot be 0xc0 or	      \
> +	   0xc1, otherwise the wide character could have been		      \
> +	   represented using a single byte.  */				      \
> +	cnt = 2;							      \
> +	ch &= 0x1f;							      \
> +      }									      \
> +    else if (__glibc_likely ((ch & 0xf0) == 0xe0))			      \
> +      {									      \
> +	/* We expect three bytes.  */					      \
> +	cnt = 3;							      \
> +	ch &= 0x0f;							      \
> +      }									      \
> +    else if (__glibc_likely ((ch & 0xf8) == 0xf0))			      \
> +      {									      \
> +	/* We expect four bytes.  */					      \
> +	cnt = 4;							      \
> +	ch &= 0x07;							      \
> +      }									      \
> +    else if (__glibc_likely ((ch & 0xfc) == 0xf8))			      \
> +      {									      \
> +	/* We expect five bytes.  */					      \
> +	cnt = 5;							      \
> +	ch &= 0x03;							      \
> +      }									      \
> +    else								      \
> +      {									      \
> +	/* We expect six bytes.  */					      \
> +	cnt = 6;							      \
> +	ch &= 0x01;							      \
> +      }									      \
> +									      \
> +    /* The first byte is already consumed.  */				      \
> +    r = cnt - 1;							      \
> +    while (++(*inptrp) < inend)						      \
> +      {									      \
> +	ch <<= 6;							      \
> +	ch |= **inptrp & 0x3f;						      \
> +	--r;								      \
> +      }									      \
> +									      \
> +    /* Shift for the so far missing bytes.  */				      \
> +    ch <<= r * 6;							      \
> +									      \
> +    /* Store the number of bytes expected for the entire sequence.  */	      \
> +    state->__count |= cnt << 8;						      \
> +									      \
> +    /* Store the value.  */						      \
> +    state->__value.__wch = ch;						      \
> +  }
> +
> +#define UNPACK_BYTES_COMMON \
> +  {									      \
> +    static const unsigned char inmask[5] = { 0xc0, 0xe0, 0xf0, 0xf8, 0xfc };  \
> +    wint_t wch = state->__value.__wch;					      \
> +    size_t ntotal = state->__count >> 8;				      \
> +									      \
> +    inlen = state->__count & 255;					      \
> +									      \
> +    bytebuf[0] = inmask[ntotal - 2];					      \
> +									      \
> +    do									      \
> +      {									      \
> +	if (--ntotal < inlen)						      \
> +	  bytebuf[ntotal] = 0x80 | (wch & 0x3f);			      \
> +	wch >>= 6;							      \
> +      }									      \
> +    while (ntotal > 1);							      \
> +									      \
> +    bytebuf[0] |= wch;							      \
> +  }
> +
> +#define CLEAR_STATE_COMMON \
> +  state->__count = 0
> +
> +#define BODY_FROM_HW(ASM)						\
>     {									\
> -    if (GLRO (dl_hwcap) & HWCAP_S390_ETF3EH)				\
> -      {									\
> -	HARDWARE_CONVERT ("cu14 %0, %1, 1");				\
> +    ASM;								\
> +    if (__glibc_likely (inptr == inend)					\
> +	|| result == __GCONV_FULL_OUTPUT)				\
> +      break;								\
>   									\
> -	if (inptr != inend)						\
> -	  {								\
> -	    int i;							\
> -	    for (i = 1; inptr + i < inend; ++i)				\
> -	      if ((inptr[i] & 0xc0) != 0x80)				\
> -		break;							\
> +    int i;								\
> +    for (i = 1; inptr + i < inend && i < 5; ++i)			\
> +      if ((inptr[i] & 0xc0) != 0x80)					\
> +	break;								\
>   									\
> -	    if (__glibc_likely (inptr + i == inend))			      \
> -	      {								\
> -		result = __GCONV_INCOMPLETE_INPUT;			\
> -		break;							\
> -	      }								\
> -	    STANDARD_FROM_LOOP_ERR_HANDLER (i);				\
> -	  }								\
> -	continue;							\
> +    if (__glibc_likely (inptr + i == inend				\
> +			&& result == __GCONV_EMPTY_INPUT))		\
> +      {									\
> +	result = __GCONV_INCOMPLETE_INPUT;				\
> +	break;								\
>         }									\
> -									\
> +    STANDARD_FROM_LOOP_ERR_HANDLER (i);					\
> +  }
> +
> +/* This hardware routine uses the Convert UTF8 to UTF32 (cu14) instruction.  */
> +#define BODY_FROM_ETF3EH BODY_FROM_HW (HARDWARE_CONVERT ("cu14 %0, %1, 1"))
> +
> +
> +/* The software routine is copied from gconv_simple.c.  */
> +#define BODY_FROM_C							\
> +  {									\
>       /* Next input byte.  */						\
>       uint32_t ch = *inptr;						\
>   									\
> -    if (__glibc_likely (ch < 0x80))					      \
> +    if (__glibc_likely (ch < 0x80))					\
>         {									\
>   	/* One byte sequence.  */					\
>   	++inptr;							\
> @@ -233,30 +324,18 @@ gconv_end (struct __gconv_step *data)
>   	    cnt = 2;							\
>   	    ch &= 0x1f;							\
>   	  }								\
> -        else if (__glibc_likely ((ch & 0xf0) == 0xe0))			      \
> +	else if (__glibc_likely ((ch & 0xf0) == 0xe0))			\
>   	  {								\
>   	    /* We expect three bytes.  */				\
>   	    cnt = 3;							\
>   	    ch &= 0x0f;							\
>   	  }								\
> -	else if (__glibc_likely ((ch & 0xf8) == 0xf0))			      \
> +	else if (__glibc_likely ((ch & 0xf8) == 0xf0))			\
>   	  {								\
>   	    /* We expect four bytes.  */				\
>   	    cnt = 4;							\
>   	    ch &= 0x07;							\
>   	  }								\
> -	else if (__glibc_likely ((ch & 0xfc) == 0xf8))			      \
> -	  {								\
> -	    /* We expect five bytes.  */				\
> -	    cnt = 5;							\
> -	    ch &= 0x03;							\
> -	  }								\
> -	else if (__glibc_likely ((ch & 0xfe) == 0xfc))			      \
> -	  {								\
> -	    /* We expect six bytes.  */					\
> -	    cnt = 6;							\
> -	    ch &= 0x01;							\
> -	  }								\
>   	else								\
>   	  {								\
>   	    /* Search the end of this ill-formed UTF-8 character.  This	\
> @@ -272,7 +351,7 @@ gconv_end (struct __gconv_step *data)
>   	    STANDARD_FROM_LOOP_ERR_HANDLER (i);				\
>   	  }								\
>   									\
> -	if (__glibc_unlikely (inptr + cnt > inend))			      \
> +	if (__glibc_unlikely (inptr + cnt > inend))			\
>   	  {								\
>   	    /* We don't have enough input.  But before we report	\
>   	       that check that all the bytes are correct.  */		\
> @@ -280,7 +359,7 @@ gconv_end (struct __gconv_step *data)
>   	      if ((inptr[i] & 0xc0) != 0x80)				\
>   		break;							\
>   									\
> -	    if (__glibc_likely (inptr + i == inend))			      \
> +	    if (__glibc_likely (inptr + i == inend))			\
>   	      {								\
>   		result = __GCONV_INCOMPLETE_INPUT;			\
>   		break;							\
> @@ -305,7 +384,10 @@ gconv_end (struct __gconv_step *data)
>   	/* If i < cnt, some trail byte was not >= 0x80, < 0xc0.		\
>   	   If cnt > 2 and ch < 2^(5*cnt-4), the wide character ch could	\
>   	   have been represented with fewer than cnt bytes.  */		\
> -	if (i < cnt || (cnt > 2 && (ch >> (5 * cnt - 4)) == 0))		\
> +	if (i < cnt || (cnt > 2 && (ch >> (5 * cnt - 4)) == 0)		\
> +	    /* Do not accept UTF-16 surrogates.  */			\
> +	    || (ch >= 0xd800 && ch <= 0xdfff)				\
> +	    || (ch > 0x10ffff))						\
>   	  {								\
>   	    /* This is an illegal encoding.  */				\
>   	    goto errout;						\
> @@ -318,137 +400,212 @@ gconv_end (struct __gconv_step *data)
>       *((uint32_t *) outptr) = ch;					\
>       outptr += sizeof (uint32_t);					\
>     }
> -#define LOOP_NEED_FLAGS
>
> -#define STORE_REST							\
> -  {									      \
> -    /* We store the remaining bytes while converting them into the UCS4	      \
> -       format.  We can assume that the first byte in the buffer is	      \
> -       correct and that it requires a larger number of bytes than there	      \
> -       are in the input buffer.  */					      \
> -    wint_t ch = **inptrp;						      \
> -    size_t cnt, r;							      \
> -									      \
> -    state->__count = inend - *inptrp;					      \
> -									      \
> -    if (ch >= 0xc2 && ch < 0xe0)					      \
> -      {									      \
> -	/* We expect two bytes.  The first byte cannot be 0xc0 or	      \
> -	   0xc1, otherwise the wide character could have been		      \
> -	   represented using a single byte.  */				      \
> -	cnt = 2;							      \
> -	ch &= 0x1f;							      \
> -      }									      \
> -    else if (__glibc_likely ((ch & 0xf0) == 0xe0))			      \
> -      {									      \
> -	/* We expect three bytes.  */					      \
> -	cnt = 3;							      \
> -	ch &= 0x0f;							      \
> -      }									      \
> -    else if (__glibc_likely ((ch & 0xf8) == 0xf0))			      \
> -      {									      \
> -	/* We expect four bytes.  */					      \
> -	cnt = 4;							      \
> -	ch &= 0x07;							      \
> -      }									      \
> -    else if (__glibc_likely ((ch & 0xfc) == 0xf8))			      \
> -      {									      \
> -	/* We expect five bytes.  */					      \
> -	cnt = 5;							      \
> -	ch &= 0x03;							      \
> -      }									      \
> -    else								      \
> -      {									      \
> -	/* We expect six bytes.  */					      \
> -	cnt = 6;							      \
> -	ch &= 0x01;							      \
> -      }									      \
> -									      \
> -    /* The first byte is already consumed.  */				      \
> -    r = cnt - 1;							      \
> -    while (++(*inptrp) < inend)						      \
> -      {									      \
> -	ch <<= 6;							      \
> -	ch |= **inptrp & 0x3f;						      \
> -	--r;								      \
> -      }									      \
> -									      \
> -    /* Shift for the so far missing bytes.  */				      \
> -    ch <<= r * 6;							      \
> -									      \
> -    /* Store the number of bytes expected for the entire sequence.  */	      \
> -    state->__count |= cnt << 8;						      \
> -									      \
> -    /* Store the value.  */						      \
> -    state->__value.__wch = ch;						      \
> +#define HW_FROM_VX							\
> +  {									\
> +    register const unsigned char* pInput asm ("8") = inptr;		\
> +    register size_t inlen asm ("9") = inend - inptr;			\
> +    register unsigned char* pOutput asm ("10") = outptr;		\
> +    register size_t outlen asm("11") = outend - outptr;			\
> +    unsigned long tmp, tmp2, tmp3;					\
> +    asm volatile (".machine push\n\t"					\
> +		  ".machine \"z13\"\n\t"				\
> +		  ".machinemode \"zarch_nohighgprs\"\n\t"		\
> +		  "vrepib %%v30,0x7f\n\t" /* For compare > 0x7f.  */	\
> +		  "vrepib %%v31,0x20\n\t"				\
> +		  /* Loop which handles UTF-8 chars <=0x7f.  */		\
> +		  "0: clgijl %[R_INLEN],16,20f\n\t"			\
> +		  "clgijl %[R_OUTLEN],64,20f\n\t"			\
> +		  "1: vl %%v16,0(%[R_IN])\n\t"				\
> +		  "vstrcbs %%v17,%%v16,%%v30,%%v31\n\t"			\
> +		  "jno 10f\n\t" /* Jump away if not all bytes are 1byte	\
> +				   UTF8 chars.  */			\
> +		  /* Enlarge to UCS4.  */				\
> +		  "vuplhb %%v18,%%v16\n\t"				\
> +		  "vupllb %%v19,%%v16\n\t"				\
> +		  "la %[R_IN],16(%[R_IN])\n\t"				\
> +		  "vuplhh %%v20,%%v18\n\t"				\
> +		  "aghi %[R_INLEN],-16\n\t"				\
> +		  "vupllh %%v21,%%v18\n\t"				\
> +		  "aghi %[R_OUTLEN],-64\n\t"				\
> +		  "vuplhh %%v22,%%v19\n\t"				\
> +		  "vupllh %%v23,%%v19\n\t"				\
> +		  /* Store 64 bytes to buf_out.  */			\
> +		  "vstm %%v20,%%v23,0(%[R_OUT])\n\t"			\
> +		  "la %[R_OUT],64(%[R_OUT])\n\t"			\
> +		  "clgijl %[R_INLEN],16,20f\n\t"			\
> +		  "clgijl %[R_OUTLEN],64,20f\n\t"			\
> +		  "j 1b\n\t"						\
> +		  "10:\n\t"						\
> +		  /* At least one byte is > 0x7f.			\
> +		     Store the preceding 1-byte chars.  */		\
> +		  "vlgvb %[R_TMP],%%v17,7\n\t"				\
> +		  "sllk %[R_TMP2],%[R_TMP],2\n\t" /* Compute highest	\
> +						     index to store. */ \
> +		  "llgfr %[R_TMP3],%[R_TMP2]\n\t"			\
> +		  "ahi %[R_TMP2],-1\n\t"				\
> +		  "jl 20f\n\t"						\
> +		  "vuplhb %%v18,%%v16\n\t"				\
> +		  "vuplhh %%v20,%%v18\n\t"				\
> +		  "vstl %%v20,%[R_TMP2],0(%[R_OUT])\n\t"		\
> +		  "ahi %[R_TMP2],-16\n\t"				\
> +		  "jl 11f\n\t"						\
> +		  "vupllh %%v21,%%v18\n\t"				\
> +		  "vstl %%v21,%[R_TMP2],16(%[R_OUT])\n\t"		\
> +		  "ahi %[R_TMP2],-16\n\t"				\
> +		  "jl 11f\n\t"						\
> +		  "vupllb %%v19,%%v16\n\t"				\
> +		  "vuplhh %%v22,%%v19\n\t"				\
> +		  "vstl %%v22,%[R_TMP2],32(%[R_OUT])\n\t"		\
> +		  "ahi %[R_TMP2],-16\n\t"				\
> +		  "jl 11f\n\t"						\
> +		  "vupllh %%v23,%%v19\n\t"				\
> +		  "vstl %%v23,%[R_TMP2],48(%[R_OUT])\n\t"		\
> +		  "11:\n\t"						\
> +		  /* Update pointers.  */				\
> +		  "la %[R_IN],0(%[R_TMP],%[R_IN])\n\t"			\
> +		  "slgr %[R_INLEN],%[R_TMP]\n\t"			\
> +		  "la %[R_OUT],0(%[R_TMP3],%[R_OUT])\n\t"		\
> +		  "slgr %[R_OUTLEN],%[R_TMP3]\n\t"			\
> +		  /* Handle multibyte utf8-char with convert instruction. */ \
> +		  "20: cu14 %[R_OUT],%[R_IN],1\n\t"			\
> +		  "jo 0b\n\t" /* Try vector implemenation again.  */	\
> +		  "lochil %[R_RES],%[RES_OUT_FULL]\n\t" /* cc == 1.  */	\
> +		  "lochih %[R_RES],%[RES_IN_ILL]\n\t" /* cc == 2.  */	\
> +		  ".machine pop"					\
> +		  : /* outputs */ [R_IN] "+a" (pInput)			\
> +		    , [R_INLEN] "+d" (inlen), [R_OUT] "+a" (pOutput)	\
> +		    , [R_OUTLEN] "+d" (outlen), [R_TMP] "=a" (tmp)	\
> +		    , [R_TMP2] "=d" (tmp2), [R_TMP3] "=a" (tmp3)	\
> +		    , [R_RES] "+d" (result)				\
> +		  : /* inputs */					\
> +		    [RES_OUT_FULL] "i" (__GCONV_FULL_OUTPUT)		\
> +		    , [RES_IN_ILL] "i" (__GCONV_ILLEGAL_INPUT)		\
> +		  : /* clobber list */ "memory", "cc"			\
> +		    ASM_CLOBBER_VR ("v16") ASM_CLOBBER_VR ("v17")	\
> +		    ASM_CLOBBER_VR ("v18") ASM_CLOBBER_VR ("v19")	\
> +		    ASM_CLOBBER_VR ("v20") ASM_CLOBBER_VR ("v21")	\
> +		    ASM_CLOBBER_VR ("v22") ASM_CLOBBER_VR ("v30")	\
> +		    ASM_CLOBBER_VR ("v31")				\
> +		  );							\
> +    inptr = pInput;							\
> +    outptr = pOutput;							\
>     }
> +#define BODY_FROM_VX BODY_FROM_HW (HW_FROM_VX)
>
> -#define UNPACK_BYTES \
> -  {									      \
> -    static const unsigned char inmask[5] = { 0xc0, 0xe0, 0xf0, 0xf8, 0xfc };  \
> -    wint_t wch = state->__value.__wch;					      \
> -    size_t ntotal = state->__count >> 8;				      \
> -									      \
> -    inlen = state->__count & 255;					      \
> -									      \
> -    bytebuf[0] = inmask[ntotal - 2];					      \
> -									      \
> -    do									      \
> -      {									      \
> -	if (--ntotal < inlen)						      \
> -	  bytebuf[ntotal] = 0x80 | (wch & 0x3f);			      \
> -	wch >>= 6;							      \
> -      }									      \
> -    while (ntotal > 1);							      \
> -									      \
> -    bytebuf[0] |= wch;							      \
> -  }
> +/* These definitions apply to the UTF-8 to UTF-32 direction.  The
> +   software implementation for UTF-8 still supports multibyte
> +   characters up to 6 bytes whereas the hardware variant does not.  */
> +#define MIN_NEEDED_INPUT	MIN_NEEDED_FROM
> +#define MAX_NEEDED_INPUT	MAX_NEEDED_FROM
> +#define MIN_NEEDED_OUTPUT	MIN_NEEDED_TO
> +#define LOOPFCT			__from_utf8_loop_c
>
> -#define CLEAR_STATE \
> -  state->__count = 0
> +#define LOOP_NEED_FLAGS
>
> +#define STORE_REST		STORE_REST_COMMON
> +#define UNPACK_BYTES		UNPACK_BYTES_COMMON
> +#define CLEAR_STATE		CLEAR_STATE_COMMON
> +#define BODY			BODY_FROM_C
>   #include <iconv/loop.c>
>
> +
> +/* Generate loop-function with hardware utf-convert instruction.  */
> +#define MIN_NEEDED_INPUT	MIN_NEEDED_FROM
> +#define MAX_NEEDED_INPUT	MAX_NEEDED_FROM
> +#define MIN_NEEDED_OUTPUT	MIN_NEEDED_TO
> +#define LOOPFCT			__from_utf8_loop_etf3eh
> +
> +#define LOOP_NEED_FLAGS
> +
> +#define STORE_REST		STORE_REST_COMMON
> +#define UNPACK_BYTES		UNPACK_BYTES_COMMON
> +#define CLEAR_STATE		CLEAR_STATE_COMMON
> +#define BODY			BODY_FROM_ETF3EH
> +#include <iconv/loop.c>
> +
> +#if defined HAVE_S390_VX_ASM_SUPPORT
> +/* Generate loop-function with hardware vector instructions.  */
> +# define MIN_NEEDED_INPUT	MIN_NEEDED_FROM
> +# define MAX_NEEDED_INPUT	MAX_NEEDED_FROM
> +# define MIN_NEEDED_OUTPUT	MIN_NEEDED_TO
> +# define LOOPFCT		__from_utf8_loop_vx
> +
> +# define LOOP_NEED_FLAGS
> +
> +# define STORE_REST		STORE_REST_COMMON
> +# define UNPACK_BYTES		UNPACK_BYTES_COMMON
> +# define CLEAR_STATE		CLEAR_STATE_COMMON
> +# define BODY			BODY_FROM_VX
> +# include <iconv/loop.c>
> +#endif
> +
> +
> +/* Generate ifunc'ed loop function.  */
> +__typeof(__from_utf8_loop_c)
> +__attribute__ ((ifunc ("__from_utf8_loop_resolver")))
> +__from_utf8_loop;
> +
> +static void *
> +__from_utf8_loop_resolver (unsigned long int dl_hwcap)
> +{
> +#if defined HAVE_S390_VX_ASM_SUPPORT
> +  if (dl_hwcap & HWCAP_S390_VX)
> +    return __from_utf8_loop_vx;
> +  else
> +#endif
> +  if (dl_hwcap & HWCAP_S390_ETF3EH)
> +    return __from_utf8_loop_etf3eh;
> +  else
> +    return __from_utf8_loop_c;
> +}
> +
> +strong_alias (__from_utf8_loop_c_single, __from_utf8_loop_single)
> +
> +
>   /* Conversion from UTF-32 internal/BE to UTF-8.  */
> +#define BODY_TO_HW(ASM)							\
> +  {									\
> +    ASM;								\
> +    if (__glibc_likely (inptr == inend)					\
> +	|| result == __GCONV_FULL_OUTPUT)				\
> +      break;								\
> +    if (inptr + 4 > inend)						\
> +      {									\
> +	result = __GCONV_INCOMPLETE_INPUT;				\
> +	break;								\
> +      }									\
> +    STANDARD_TO_LOOP_ERR_HANDLER (4);					\
> +  }
> +
> +/* The hardware routine uses the S/390 cu41 instruction.  */
> +#define BODY_TO_ETF3EH BODY_TO_HW (HARDWARE_CONVERT ("cu41 %0, %1"))
> +
> +/* The hardware routine uses the S/390 vector and cu41 instructions.  */
> +#define BODY_TO_VX BODY_TO_HW (HW_TO_VX)
>
> -#define MIN_NEEDED_INPUT	MIN_NEEDED_TO
> -#define MIN_NEEDED_OUTPUT	MIN_NEEDED_FROM
> -#define MAX_NEEDED_OUTPUT	MAX_NEEDED_FROM
> -#define LOOPFCT			TO_LOOP
>   /* The software routine mimics the S/390 cu41 instruction.  */
> -#define BODY							\
> +#define BODY_TO_C						\
>     {								\
> -    if (GLRO (dl_hwcap) & HWCAP_S390_ETF3EH)			\
> -      {								\
> -	HARDWARE_CONVERT ("cu41 %0, %1");			\
> -								\
> -	if (inptr != inend)					\
> -	  {							\
> -	    result = __GCONV_INCOMPLETE_INPUT;			\
> -	    break;						\
> -	  }							\
> -	continue;						\
> -      }								\
> -								\
>       uint32_t wc = *((const uint32_t *) inptr);			\
>   								\
> -    if (__glibc_likely (wc <= 0x7f))					      \
> +    if (__glibc_likely (wc <= 0x7f))				\
>         {								\
> -        /* Single UTF-8 char.  */				\
> -        *outptr = (uint8_t)wc;					\
> +	/* Single UTF-8 char.  */				\
> +	*outptr = (uint8_t)wc;					\
>   	outptr++;						\
>         }								\
>       else if (wc <= 0x7ff)					\
>         {								\
> -        /* Two UTF-8 chars.  */					\
> -        if (__glibc_unlikely (outptr + 2 > outend))			      \
> +	/* Two UTF-8 chars.  */					\
> +	if (__glibc_unlikely (outptr + 2 > outend))		\
>   	  {							\
>   	    /* Overflow in the output buffer.  */		\
>   	    result = __GCONV_FULL_OUTPUT;			\
>   	    break;						\
>   	  }							\
>   								\
> -        outptr[0] = 0xc0;					\
> +	outptr[0] = 0xc0;					\
>   	outptr[0] |= wc >> 6;					\
>   								\
>   	outptr[1] = 0x80;					\
> @@ -459,12 +616,18 @@ gconv_end (struct __gconv_step *data)
>       else if (wc <= 0xffff)					\
>         {								\
>   	/* Three UTF-8 chars.  */				\
> -	if (__glibc_unlikely (outptr + 3 > outend))			      \
> +	if (__glibc_unlikely (outptr + 3 > outend))		\
>   	  {							\
>   	    /* Overflow in the output buffer.  */		\
>   	    result = __GCONV_FULL_OUTPUT;			\
>   	    break;						\
>   	  }							\
> +	if (wc >= 0xd800 && wc < 0xdc00)			\
> +	  {							\
> +	    /* Do not accept UTF-16 surrogates.   */		\
> +	    result = __GCONV_ILLEGAL_INPUT;			\
> +	    STANDARD_TO_LOOP_ERR_HANDLER (4);			\
> +	  }							\
>   	outptr[0] = 0xe0;					\
>   	outptr[0] |= wc >> 12;					\
>   								\
> @@ -479,7 +642,7 @@ gconv_end (struct __gconv_step *data)
>         else if (wc <= 0x10ffff)					\
>   	{							\
>   	  /* Four UTF-8 chars.  */				\
> -	  if (__glibc_unlikely (outptr + 4 > outend))			      \
> +	  if (__glibc_unlikely (outptr + 4 > outend))		\
>   	    {							\
>   	      /* Overflow in the output buffer.  */		\
>   	      result = __GCONV_FULL_OUTPUT;			\
> @@ -505,7 +668,140 @@ gconv_end (struct __gconv_step *data)
>   	}							\
>       inptr += 4;							\
>     }
> +
> +#define HW_TO_VX							\
> +  {									\
> +    register const unsigned char* pInput asm ("8") = inptr;		\
> +    register size_t inlen asm ("9") = inend - inptr;			\
> +    register unsigned char* pOutput asm ("10") = outptr;		\
> +    register size_t outlen asm("11") = outend - outptr;			\
> +    unsigned long tmp, tmp2;						\
> +    asm volatile (".machine push\n\t"					\
> +		  ".machine \"z13\"\n\t"				\
> +		  ".machinemode \"zarch_nohighgprs\"\n\t"		\
> +		  "vleif %%v20,127,0\n\t"   /* element 0: 127  */	\
> +		  "vzero %%v21\n\t"					\
> +		  "vleih %%v21,8192,0\n\t"  /* element 0:   >  */	\
> +		  "vleih %%v21,-8192,2\n\t" /* element 1: =<>  */	\
> +		  /* Loop which handles UTF-32 chars <=0x7f.  */	\
> +		  "0: clgijl %[R_INLEN],64,20f\n\t"			\
> +		  "clgijl %[R_OUTLEN],16,20f\n\t"			\
> +		  "1: vlm %%v16,%%v19,0(%[R_IN])\n\t"			\
> +		  "lghi %[R_TMP],0\n\t"					\
> +		  /* Shorten to byte values.  */			\
> +		  "vpkf %%v23,%%v16,%%v17\n\t"				\
> +		  "vpkf %%v24,%%v18,%%v19\n\t"				\
> +		  "vpkh %%v23,%%v23,%%v24\n\t"				\
> +		  /* Checking for values > 0x7f.  */			\
> +		  "vstrcfs %%v22,%%v16,%%v20,%%v21\n\t"			\
> +		  "jno 10f\n\t"						\
> +		  "vstrcfs %%v22,%%v17,%%v20,%%v21\n\t"			\
> +		  "jno 11f\n\t"						\
> +		  "vstrcfs %%v22,%%v18,%%v20,%%v21\n\t"			\
> +		  "jno 12f\n\t"						\
> +		  "vstrcfs %%v22,%%v19,%%v20,%%v21\n\t"			\
> +		  "jno 13f\n\t"						\
> +		  /* Store 16bytes to outptr.  */			\
> +		  "vst %%v23,0(%[R_OUT])\n\t"				\
> +		  "aghi %[R_INLEN],-64\n\t"				\
> +		  "aghi %[R_OUTLEN],-16\n\t"				\
> +		  "la %[R_IN],64(%[R_IN])\n\t"				\
> +		  "la %[R_OUT],16(%[R_OUT])\n\t"			\
> +		  "clgijl %[R_INLEN],64,20f\n\t"			\
> +		  "clgijl %[R_OUTLEN],16,20f\n\t"			\
> +		  "j 1b\n\t"						\
> +		  /* Found a value > 0x7f.  */				\
> +		  "13: ahi %[R_TMP],4\n\t"				\
> +		  "12: ahi %[R_TMP],4\n\t"				\
> +		  "11: ahi %[R_TMP],4\n\t"				\
> +		  "10: vlgvb %[R_I],%%v22,7\n\t"			\
> +		  "srlg %[R_I],%[R_I],2\n\t"				\
> +		  "agr %[R_I],%[R_TMP]\n\t"				\
> +		  "je 20f\n\t"						\
> +		  /* Store characters before invalid one...  */		\
> +		  "slgr %[R_OUTLEN],%[R_I]\n\t"				\
> +		  "15: aghi %[R_I],-1\n\t"				\
> +		  "vstl %%v23,%[R_I],0(%[R_OUT])\n\t"			\
> +		  /* ... and update pointers.  */			\
> +		  "aghi %[R_I],1\n\t"					\
> +		  "la %[R_OUT],0(%[R_I],%[R_OUT])\n\t"			\
> +		  "sllg %[R_I],%[R_I],2\n\t"				\
> +		  "la %[R_IN],0(%[R_I],%[R_IN])\n\t"			\
> +		  "slgr %[R_INLEN],%[R_I]\n\t"				\
> +		  /* Handle multibyte utf8-char with convert instruction. */ \
> +		  "20: cu41 %[R_OUT],%[R_IN]\n\t"			\
> +		  "jo 0b\n\t" /* Try vector implemenation again.  */	\
> +		  "lochil %[R_RES],%[RES_OUT_FULL]\n\t" /* cc == 1.  */	\
> +		  "lochih %[R_RES],%[RES_IN_ILL]\n\t" /* cc == 2.  */	\
> +		  ".machine pop"					\
> +		  : /* outputs */ [R_IN] "+a" (pInput)			\
> +		    , [R_INLEN] "+d" (inlen), [R_OUT] "+a" (pOutput)	\
> +		    , [R_OUTLEN] "+d" (outlen), [R_TMP] "=d" (tmp)	\
> +		    , [R_I] "=a" (tmp2)					\
> +		    , [R_RES] "+d" (result)				\
> +		  : /* inputs */					\
> +		    [RES_OUT_FULL] "i" (__GCONV_FULL_OUTPUT)		\
> +		    , [RES_IN_ILL] "i" (__GCONV_ILLEGAL_INPUT)		\
> +		  : /* clobber list */ "memory", "cc"			\
> +		    ASM_CLOBBER_VR ("v16") ASM_CLOBBER_VR ("v17")	\
> +		    ASM_CLOBBER_VR ("v18") ASM_CLOBBER_VR ("v19")	\
> +		    ASM_CLOBBER_VR ("v20") ASM_CLOBBER_VR ("v21")	\
> +		    ASM_CLOBBER_VR ("v22") ASM_CLOBBER_VR ("v23")	\
> +		    ASM_CLOBBER_VR ("v24")				\
> +		  );							\
> +    inptr = pInput;							\
> +    outptr = pOutput;							\
> +  }
> +
> +/* Generate loop-function with software routing.  */
> +#define MIN_NEEDED_INPUT	MIN_NEEDED_TO
> +#define MIN_NEEDED_OUTPUT	MIN_NEEDED_FROM
> +#define MAX_NEEDED_OUTPUT	MAX_NEEDED_FROM
> +#define LOOPFCT			__to_utf8_loop_c
> +#define BODY			BODY_TO_C
> +#define LOOP_NEED_FLAGS
> +#include <iconv/loop.c>
> +
> +/* Generate loop-function with hardware utf-convert instruction.  */
> +#define MIN_NEEDED_INPUT	MIN_NEEDED_TO
> +#define MIN_NEEDED_OUTPUT	MIN_NEEDED_FROM
> +#define MAX_NEEDED_OUTPUT	MAX_NEEDED_FROM
> +#define LOOPFCT			__to_utf8_loop_etf3eh
>   #define LOOP_NEED_FLAGS
> +#define BODY			BODY_TO_ETF3EH
>   #include <iconv/loop.c>
>
> +#if defined HAVE_S390_VX_ASM_SUPPORT
> +/* Generate loop-function with hardware vector and utf-convert instructions.  */
> +# define MIN_NEEDED_INPUT	MIN_NEEDED_TO
> +# define MIN_NEEDED_OUTPUT	MIN_NEEDED_FROM
> +# define MAX_NEEDED_OUTPUT	MAX_NEEDED_FROM
> +# define LOOPFCT		__to_utf8_loop_vx
> +# define BODY			BODY_TO_VX
> +# define LOOP_NEED_FLAGS
> +# include <iconv/loop.c>
> +#endif
> +
> +/* Generate ifunc'ed loop function.  */
> +__typeof(__to_utf8_loop_c)
> +__attribute__ ((ifunc ("__to_utf8_loop_resolver")))
> +__to_utf8_loop;
> +
> +static void *
> +__to_utf8_loop_resolver (unsigned long int dl_hwcap)
> +{
> +#if defined HAVE_S390_VX_ASM_SUPPORT
> +  if (dl_hwcap & HWCAP_S390_VX)
> +    return __to_utf8_loop_vx;
> +  else
> +#endif
> +  if (dl_hwcap & HWCAP_S390_ETF3EH)
> +    return __to_utf8_loop_etf3eh;
> +  else
> +    return __to_utf8_loop_c;
> +}
> +
> +strong_alias (__to_utf8_loop_c_single, __to_utf8_loop_single)
> +
> +
>   #include <iconv/skeleton.c>
>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0007-S390-Optimize-utf8-utf32-module.patch
Type: text/x-patch
Size: 29532 bytes
Desc: not available
URL: <http://sourceware.org/pipermail/libc-alpha/attachments/20160421/4078a505/attachment.bin>


More information about the Libc-alpha mailing list