From patchwork Thu Apr 21 14:49:58 2016 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Stefan Liebler X-Patchwork-Id: 11839 Received: (qmail 37853 invoked by alias); 21 Apr 2016 15:15:23 -0000 Mailing-List: contact libc-alpha-help@sourceware.org; run by ezmlm Precedence: bulk List-Id: List-Unsubscribe: List-Subscribe: List-Archive: List-Post: List-Help: , Sender: libc-alpha-owner@sourceware.org Delivered-To: mailing list libc-alpha@sourceware.org Received: (qmail 37722 invoked by uid 89); 21 Apr 2016 15:15:14 -0000 Authentication-Results: sourceware.org; auth=none X-Virus-Found: No X-Spam-SWARE-Status: No, score=-3.2 required=5.0 tests=AWL, BAYES_00, RCVD_IN_DNSWL_LOW, RP_MATCHES_RCVD, SPF_HELO_PASS, SPF_PASS autolearn=ham version=3.3.2 spammy=2727, bom, Order X-HELO: plane.gmane.org To: libc-alpha@sourceware.org From: Stefan Liebler Subject: Re: [PATCH 07/14] S390: Optimize utf8-utf32 module. Date: Thu, 21 Apr 2016 16:49:58 +0200 Lines: 1688 Message-ID: References: <1456219278-5258-1-git-send-email-stli@linux.vnet.ibm.com> <1456219278-5258-8-git-send-email-stli@linux.vnet.ibm.com> Mime-Version: 1.0 User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:38.0) Gecko/20100101 Thunderbird/38.7.0 In-Reply-To: <1456219278-5258-8-git-send-email-stli@linux.vnet.ibm.com> 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 > #include > > -/* 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 > > + > +/* 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 > + > +#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 > +#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 > + > +/* 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 > > +#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 > +#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 > From 12647b88906a3e8bfff9deae09c54fa2c933304a Mon Sep 17 00:00:00 2001 From: Stefan Liebler Date: Thu, 21 Apr 2016 12:42:49 +0200 Subject: [PATCH 07/14] S390: Optimize utf8-utf32 module. 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..f9c9199 100644 --- a/sysdeps/s390/s390-64/utf8-utf32-z9.c +++ b/sysdeps/s390/s390-64/utf8-utf32-z9.c @@ -30,35 +30,25 @@ #include #include -/* 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 + +/* 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 + +#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 +#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 + +/* 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 +#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 +#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 -- 2.5.5