|
 |
8ae002 |
From 53f860e80162b09c44b48f207342c1452289072c Mon Sep 17 00:00:00 2001
|
|
|
8ae002 |
From: Stefan Liebler <stli@linux.vnet.ibm.com>
|
|
|
8ae002 |
Date: Mon, 7 Nov 2016 15:37:29 +0100
|
|
|
8ae002 |
Subject: [PATCH 04/17] S390: Optimize builtin iconv-modules.
|
|
|
8ae002 |
|
|
|
8ae002 |
Upstream commit 3b704e26b33e35d99de920f8462d8e438f89be39
|
|
|
8ae002 |
|
|
|
8ae002 |
This patch introduces a s390 specific gconv_simple.c file which provides
|
|
|
8ae002 |
optimized versions for z13 with vector instructions, which will be chosen at
|
|
|
8ae002 |
runtime via ifunc.
|
|
|
8ae002 |
The optimized conversions can convert between internal and ascii, ucs4, ucs4le,
|
|
|
8ae002 |
ucs2, ucs2le.
|
|
|
8ae002 |
If the build-environment lacks vector support, then iconv/gconv_simple.c
|
|
|
8ae002 |
is used wihtout any change. Otherwise iconvdata/gconv_simple.c is used to create
|
|
|
8ae002 |
conversion loop routines without vector instructions as fallback, if vector
|
|
|
8ae002 |
instructions aren't available at runtime.
|
|
|
8ae002 |
|
|
|
8ae002 |
ChangeLog:
|
|
|
8ae002 |
|
|
|
8ae002 |
* sysdeps/s390/multiarch/gconv_simple.c: New File.
|
|
|
8ae002 |
* sysdeps/s390/multiarch/Makefile (sysdep_routines): Add gconv_simple.
|
|
|
8ae002 |
---
|
|
|
8ae002 |
sysdeps/s390/multiarch/Makefile | 4 +
|
|
|
8ae002 |
sysdeps/s390/multiarch/gconv_simple.c | 1266 +++++++++++++++++++++++++++++++++
|
|
|
8ae002 |
2 files changed, 1270 insertions(+)
|
|
|
8ae002 |
create mode 100644 sysdeps/s390/multiarch/gconv_simple.c
|
|
|
8ae002 |
|
|
|
8ae002 |
diff --git a/sysdeps/s390/multiarch/Makefile b/sysdeps/s390/multiarch/Makefile
|
|
|
8ae002 |
index 11ad2b9..24949cd 100644
|
|
|
8ae002 |
--- a/sysdeps/s390/multiarch/Makefile
|
|
|
8ae002 |
+++ b/sysdeps/s390/multiarch/Makefile
|
|
|
8ae002 |
@@ -52,3 +52,7 @@ $(move-if-change) $(@:stmp=T) $(@:stmp=h)
|
|
|
8ae002 |
touch $@
|
|
|
8ae002 |
endef
|
|
|
8ae002 |
endif
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+ifeq ($(subdir),iconv)
|
|
|
8ae002 |
+sysdep_routines += gconv_simple
|
|
|
8ae002 |
+endif
|
|
|
8ae002 |
diff --git a/sysdeps/s390/multiarch/gconv_simple.c b/sysdeps/s390/multiarch/gconv_simple.c
|
|
|
8ae002 |
new file mode 100644
|
|
|
8ae002 |
index 0000000..dc53a48
|
|
|
8ae002 |
--- /dev/null
|
|
|
8ae002 |
+++ b/sysdeps/s390/multiarch/gconv_simple.c
|
|
|
8ae002 |
@@ -0,0 +1,1266 @@
|
|
|
8ae002 |
+/* Simple transformations functions - s390 version.
|
|
|
8ae002 |
+ Copyright (C) 2016 Free Software Foundation, Inc.
|
|
|
8ae002 |
+ This file is part of the GNU C Library.
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+ The GNU C Library is free software; you can redistribute it and/or
|
|
|
8ae002 |
+ modify it under the terms of the GNU Lesser General Public
|
|
|
8ae002 |
+ License as published by the Free Software Foundation; either
|
|
|
8ae002 |
+ version 2.1 of the License, or (at your option) any later version.
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+ The GNU C Library is distributed in the hope that it will be useful,
|
|
|
8ae002 |
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
8ae002 |
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
|
8ae002 |
+ Lesser General Public License for more details.
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+ You should have received a copy of the GNU Lesser General Public
|
|
|
8ae002 |
+ License along with the GNU C Library; if not, see
|
|
|
8ae002 |
+ <http://www.gnu.org/licenses/>. */
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+#if defined HAVE_S390_VX_ASM_SUPPORT
|
|
|
8ae002 |
+# include <ifunc-resolve.h>
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+# if defined HAVE_S390_VX_GCC_SUPPORT
|
|
|
8ae002 |
+# define ASM_CLOBBER_VR(NR) , NR
|
|
|
8ae002 |
+# else
|
|
|
8ae002 |
+# define ASM_CLOBBER_VR(NR)
|
|
|
8ae002 |
+# endif
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+# define ICONV_C_NAME(NAME) __##NAME##_c
|
|
|
8ae002 |
+# define ICONV_VX_NAME(NAME) __##NAME##_vx
|
|
|
8ae002 |
+# define ICONV_VX_IFUNC(FUNC) \
|
|
|
8ae002 |
+ extern __typeof (ICONV_C_NAME (FUNC)) __##FUNC; \
|
|
|
8ae002 |
+ s390_vx_libc_ifunc (__##FUNC) \
|
|
|
8ae002 |
+ int FUNC (struct __gconv_step *step, struct __gconv_step_data *data, \
|
|
|
8ae002 |
+ const unsigned char **inptrp, const unsigned char *inend, \
|
|
|
8ae002 |
+ unsigned char **outbufstart, size_t *irreversible, \
|
|
|
8ae002 |
+ int do_flush, int consume_incomplete) \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ return __##FUNC (step, data, inptrp, inend,outbufstart, \
|
|
|
8ae002 |
+ irreversible, do_flush, consume_incomplete); \
|
|
|
8ae002 |
+ }
|
|
|
8ae002 |
+# define ICONV_VX_SINGLE(NAME) \
|
|
|
8ae002 |
+ static __typeof (NAME##_single) __##NAME##_vx_single __attribute__((alias(#NAME "_single")));
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+/* Generate the transformations which are used, if the target machine does not
|
|
|
8ae002 |
+ support vector instructions. */
|
|
|
8ae002 |
+# define __gconv_transform_ascii_internal \
|
|
|
8ae002 |
+ ICONV_C_NAME (__gconv_transform_ascii_internal)
|
|
|
8ae002 |
+# define __gconv_transform_internal_ascii \
|
|
|
8ae002 |
+ ICONV_C_NAME (__gconv_transform_internal_ascii)
|
|
|
8ae002 |
+# define __gconv_transform_internal_ucs4le \
|
|
|
8ae002 |
+ ICONV_C_NAME (__gconv_transform_internal_ucs4le)
|
|
|
8ae002 |
+# define __gconv_transform_ucs4_internal \
|
|
|
8ae002 |
+ ICONV_C_NAME (__gconv_transform_ucs4_internal)
|
|
|
8ae002 |
+# define __gconv_transform_ucs4le_internal \
|
|
|
8ae002 |
+ ICONV_C_NAME (__gconv_transform_ucs4le_internal)
|
|
|
8ae002 |
+# define __gconv_transform_ucs2_internal \
|
|
|
8ae002 |
+ ICONV_C_NAME (__gconv_transform_ucs2_internal)
|
|
|
8ae002 |
+# define __gconv_transform_ucs2reverse_internal \
|
|
|
8ae002 |
+ ICONV_C_NAME (__gconv_transform_ucs2reverse_internal)
|
|
|
8ae002 |
+# define __gconv_transform_internal_ucs2 \
|
|
|
8ae002 |
+ ICONV_C_NAME (__gconv_transform_internal_ucs2)
|
|
|
8ae002 |
+# define __gconv_transform_internal_ucs2reverse \
|
|
|
8ae002 |
+ ICONV_C_NAME (__gconv_transform_internal_ucs2reverse)
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+# include <iconv/gconv_simple.c>
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+# undef __gconv_transform_ascii_internal
|
|
|
8ae002 |
+# undef __gconv_transform_internal_ascii
|
|
|
8ae002 |
+# undef __gconv_transform_internal_ucs4le
|
|
|
8ae002 |
+# undef __gconv_transform_ucs4_internal
|
|
|
8ae002 |
+# undef __gconv_transform_ucs4le_internal
|
|
|
8ae002 |
+# undef __gconv_transform_ucs2_internal
|
|
|
8ae002 |
+# undef __gconv_transform_ucs2reverse_internal
|
|
|
8ae002 |
+# undef __gconv_transform_internal_ucs2
|
|
|
8ae002 |
+# undef __gconv_transform_internal_ucs2reverse
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+/* Now define the functions with vector support. */
|
|
|
8ae002 |
+# if defined __s390x__
|
|
|
8ae002 |
+# define CONVERT_32BIT_SIZE_T(REG)
|
|
|
8ae002 |
+# else
|
|
|
8ae002 |
+# define CONVERT_32BIT_SIZE_T(REG) "llgfr %" #REG ",%" #REG "\n\t"
|
|
|
8ae002 |
+# endif
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+/* Convert from ISO 646-IRV to the internal (UCS4-like) format. */
|
|
|
8ae002 |
+# define DEFINE_INIT 0
|
|
|
8ae002 |
+# define DEFINE_FINI 0
|
|
|
8ae002 |
+# define MIN_NEEDED_FROM 1
|
|
|
8ae002 |
+# define MIN_NEEDED_TO 4
|
|
|
8ae002 |
+# define FROM_DIRECTION 1
|
|
|
8ae002 |
+# define FROM_LOOP ICONV_VX_NAME (ascii_internal_loop)
|
|
|
8ae002 |
+# define TO_LOOP ICONV_VX_NAME (ascii_internal_loop) /* This is not used. */
|
|
|
8ae002 |
+# define FUNCTION_NAME ICONV_VX_NAME (__gconv_transform_ascii_internal)
|
|
|
8ae002 |
+# define ONE_DIRECTION 1
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+# define MIN_NEEDED_INPUT MIN_NEEDED_FROM
|
|
|
8ae002 |
+# define MIN_NEEDED_OUTPUT MIN_NEEDED_TO
|
|
|
8ae002 |
+# define LOOPFCT FROM_LOOP
|
|
|
8ae002 |
+# define BODY_ORIG_ERROR \
|
|
|
8ae002 |
+ /* The value is too large. We don't try transliteration here since \
|
|
|
8ae002 |
+ this is not an error because of the lack of possibilities to \
|
|
|
8ae002 |
+ represent the result. This is a genuine bug in the input since \
|
|
|
8ae002 |
+ ASCII does not allow such values. */ \
|
|
|
8ae002 |
+ STANDARD_FROM_LOOP_ERR_HANDLER (1);
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+# define BODY_ORIG \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ if (__glibc_unlikely (*inptr > '\x7f')) \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ BODY_ORIG_ERROR \
|
|
|
8ae002 |
+ } \
|
|
|
8ae002 |
+ else \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ /* It's an one byte sequence. */ \
|
|
|
8ae002 |
+ *((uint32_t *) outptr) = *inptr++; \
|
|
|
8ae002 |
+ outptr += sizeof (uint32_t); \
|
|
|
8ae002 |
+ } \
|
|
|
8ae002 |
+ }
|
|
|
8ae002 |
+# define BODY \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ size_t len = inend - inptr; \
|
|
|
8ae002 |
+ if (len > (outend - outptr) / 4) \
|
|
|
8ae002 |
+ len = (outend - outptr) / 4; \
|
|
|
8ae002 |
+ size_t loop_count, tmp; \
|
|
|
8ae002 |
+ __asm__ volatile (".machine push\n\t" \
|
|
|
8ae002 |
+ ".machine \"z13\"\n\t" \
|
|
|
8ae002 |
+ ".machinemode \"zarch_nohighgprs\"\n\t" \
|
|
|
8ae002 |
+ CONVERT_32BIT_SIZE_T ([R_LEN]) \
|
|
|
8ae002 |
+ " vrepib %%v30,0x7f\n\t" /* For compare > 0x7f. */ \
|
|
|
8ae002 |
+ " srlg %[R_LI],%[R_LEN],4\n\t" \
|
|
|
8ae002 |
+ " vrepib %%v31,0x20\n\t" \
|
|
|
8ae002 |
+ " clgije %[R_LI],0,1f\n\t" \
|
|
|
8ae002 |
+ "0: \n\t" /* Handle 16-byte blocks. */ \
|
|
|
8ae002 |
+ " vl %%v16,0(%[R_IN])\n\t" \
|
|
|
8ae002 |
+ /* Checking for values > 0x7f. */ \
|
|
|
8ae002 |
+ " vstrcbs %%v17,%%v16,%%v30,%%v31\n\t" \
|
|
|
8ae002 |
+ " jno 10f\n\t" \
|
|
|
8ae002 |
+ /* Enlarge to UCS4. */ \
|
|
|
8ae002 |
+ " vuplhb %%v17,%%v16\n\t" \
|
|
|
8ae002 |
+ " vupllb %%v18,%%v16\n\t" \
|
|
|
8ae002 |
+ " vuplhh %%v19,%%v17\n\t" \
|
|
|
8ae002 |
+ " vupllh %%v20,%%v17\n\t" \
|
|
|
8ae002 |
+ " vuplhh %%v21,%%v18\n\t" \
|
|
|
8ae002 |
+ " vupllh %%v22,%%v18\n\t" \
|
|
|
8ae002 |
+ /* Store 64bytes to buf_out. */ \
|
|
|
8ae002 |
+ " vstm %%v19,%%v22,0(%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ " la %[R_IN],16(%[R_IN])\n\t" \
|
|
|
8ae002 |
+ " la %[R_OUT],64(%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ " brctg %[R_LI],0b\n\t" \
|
|
|
8ae002 |
+ " lghi %[R_LI],15\n\t" \
|
|
|
8ae002 |
+ " ngr %[R_LEN],%[R_LI]\n\t" \
|
|
|
8ae002 |
+ " je 20f\n\t" /* Jump away if no remaining bytes. */ \
|
|
|
8ae002 |
+ /* Handle remaining bytes. */ \
|
|
|
8ae002 |
+ "1: aghik %[R_LI],%[R_LEN],-1\n\t" \
|
|
|
8ae002 |
+ " jl 20f\n\t" /* Jump away if no remaining bytes. */ \
|
|
|
8ae002 |
+ " vll %%v16,%[R_LI],0(%[R_IN])\n\t" \
|
|
|
8ae002 |
+ /* Checking for values > 0x7f. */ \
|
|
|
8ae002 |
+ " vstrcbs %%v17,%%v16,%%v30,%%v31\n\t" \
|
|
|
8ae002 |
+ " vlgvb %[R_TMP],%%v17,7\n\t" \
|
|
|
8ae002 |
+ " clr %[R_TMP],%[R_LI]\n\t" \
|
|
|
8ae002 |
+ " locrh %[R_TMP],%[R_LEN]\n\t" \
|
|
|
8ae002 |
+ " locghih %[R_LEN],0\n\t" \
|
|
|
8ae002 |
+ " j 12f\n\t" \
|
|
|
8ae002 |
+ "10:\n\t" \
|
|
|
8ae002 |
+ /* Found a value > 0x7f. \
|
|
|
8ae002 |
+ Store the preceding chars. */ \
|
|
|
8ae002 |
+ " vlgvb %[R_TMP],%%v17,7\n\t" \
|
|
|
8ae002 |
+ "12: la %[R_IN],0(%[R_TMP],%[R_IN])\n\t" \
|
|
|
8ae002 |
+ " sllk %[R_TMP],%[R_TMP],2\n\t" \
|
|
|
8ae002 |
+ " ahi %[R_TMP],-1\n\t" \
|
|
|
8ae002 |
+ " jl 20f\n\t" \
|
|
|
8ae002 |
+ " lgr %[R_LI],%[R_TMP]\n\t" \
|
|
|
8ae002 |
+ " vuplhb %%v17,%%v16\n\t" \
|
|
|
8ae002 |
+ " vuplhh %%v19,%%v17\n\t" \
|
|
|
8ae002 |
+ " vstl %%v19,%[R_LI],0(%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ " ahi %[R_LI],-16\n\t" \
|
|
|
8ae002 |
+ " jl 11f\n\t" \
|
|
|
8ae002 |
+ " vupllh %%v20,%%v17\n\t" \
|
|
|
8ae002 |
+ " vstl %%v20,%[R_LI],16(%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ " ahi %[R_LI],-16\n\t" \
|
|
|
8ae002 |
+ " jl 11f\n\t" \
|
|
|
8ae002 |
+ " vupllb %%v18,%%v16\n\t" \
|
|
|
8ae002 |
+ " vuplhh %%v21,%%v18\n\t" \
|
|
|
8ae002 |
+ " vstl %%v21,%[R_LI],32(%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ " ahi %[R_LI],-16\n\t" \
|
|
|
8ae002 |
+ " jl 11f\n\t" \
|
|
|
8ae002 |
+ " vupllh %%v22,%%v18\n\t" \
|
|
|
8ae002 |
+ " vstl %%v22,%[R_LI],48(%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ "11:\n\t" \
|
|
|
8ae002 |
+ " la %[R_OUT],1(%[R_TMP],%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ "20:\n\t" \
|
|
|
8ae002 |
+ ".machine pop" \
|
|
|
8ae002 |
+ : /* outputs */ [R_OUT] "+a" (outptr) \
|
|
|
8ae002 |
+ , [R_IN] "+a" (inptr) \
|
|
|
8ae002 |
+ , [R_LEN] "+d" (len) \
|
|
|
8ae002 |
+ , [R_LI] "=d" (loop_count) \
|
|
|
8ae002 |
+ , [R_TMP] "=a" (tmp) \
|
|
|
8ae002 |
+ : /* inputs */ \
|
|
|
8ae002 |
+ : /* clobber list*/ "memory", "cc" \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v16") ASM_CLOBBER_VR ("v17") \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v18") ASM_CLOBBER_VR ("v19") \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v20") ASM_CLOBBER_VR ("v21") \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v22") ASM_CLOBBER_VR ("v30") \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v31") \
|
|
|
8ae002 |
+ ); \
|
|
|
8ae002 |
+ if (len > 0) \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ /* Found an invalid character at the next input byte. */ \
|
|
|
8ae002 |
+ BODY_ORIG_ERROR \
|
|
|
8ae002 |
+ } \
|
|
|
8ae002 |
+ }
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+# define LOOP_NEED_FLAGS
|
|
|
8ae002 |
+# include <iconv/loop.c>
|
|
|
8ae002 |
+# include <iconv/skeleton.c>
|
|
|
8ae002 |
+# undef BODY_ORIG
|
|
|
8ae002 |
+# undef BODY_ORIG_ERROR
|
|
|
8ae002 |
+ICONV_VX_IFUNC (__gconv_transform_ascii_internal)
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+/* Convert from the internal (UCS4-like) format to ISO 646-IRV. */
|
|
|
8ae002 |
+# define DEFINE_INIT 0
|
|
|
8ae002 |
+# define DEFINE_FINI 0
|
|
|
8ae002 |
+# define MIN_NEEDED_FROM 4
|
|
|
8ae002 |
+# define MIN_NEEDED_TO 1
|
|
|
8ae002 |
+# define FROM_DIRECTION 1
|
|
|
8ae002 |
+# define FROM_LOOP ICONV_VX_NAME (internal_ascii_loop)
|
|
|
8ae002 |
+# define TO_LOOP ICONV_VX_NAME (internal_ascii_loop) /* This is not used. */
|
|
|
8ae002 |
+# define FUNCTION_NAME ICONV_VX_NAME (__gconv_transform_internal_ascii)
|
|
|
8ae002 |
+# define ONE_DIRECTION 1
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+# define MIN_NEEDED_INPUT MIN_NEEDED_FROM
|
|
|
8ae002 |
+# define MIN_NEEDED_OUTPUT MIN_NEEDED_TO
|
|
|
8ae002 |
+# define LOOPFCT FROM_LOOP
|
|
|
8ae002 |
+# define BODY_ORIG_ERROR \
|
|
|
8ae002 |
+ UNICODE_TAG_HANDLER (*((const uint32_t *) inptr), 4); \
|
|
|
8ae002 |
+ STANDARD_TO_LOOP_ERR_HANDLER (4);
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+# define BODY_ORIG \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ if (__glibc_unlikely (*((const uint32_t *) inptr) > 0x7f)) \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ BODY_ORIG_ERROR \
|
|
|
8ae002 |
+ } \
|
|
|
8ae002 |
+ else \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ /* It's an one byte sequence. */ \
|
|
|
8ae002 |
+ *outptr++ = *((const uint32_t *) inptr); \
|
|
|
8ae002 |
+ inptr += sizeof (uint32_t); \
|
|
|
8ae002 |
+ } \
|
|
|
8ae002 |
+ }
|
|
|
8ae002 |
+# define BODY \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ size_t len = (inend - inptr) / 4; \
|
|
|
8ae002 |
+ if (len > outend - outptr) \
|
|
|
8ae002 |
+ len = outend - outptr; \
|
|
|
8ae002 |
+ size_t loop_count, tmp, tmp2; \
|
|
|
8ae002 |
+ __asm__ volatile (".machine push\n\t" \
|
|
|
8ae002 |
+ ".machine \"z13\"\n\t" \
|
|
|
8ae002 |
+ ".machinemode \"zarch_nohighgprs\"\n\t" \
|
|
|
8ae002 |
+ CONVERT_32BIT_SIZE_T ([R_LEN]) \
|
|
|
8ae002 |
+ /* Setup to check for ch > 0x7f. */ \
|
|
|
8ae002 |
+ " vzero %%v21\n\t" \
|
|
|
8ae002 |
+ " srlg %[R_LI],%[R_LEN],4\n\t" \
|
|
|
8ae002 |
+ " vleih %%v21,8192,0\n\t" /* element 0: > */ \
|
|
|
8ae002 |
+ " vleih %%v21,-8192,2\n\t" /* element 1: =<> */ \
|
|
|
8ae002 |
+ " vleif %%v20,127,0\n\t" /* element 0: 127 */ \
|
|
|
8ae002 |
+ " lghi %[R_TMP],0\n\t" \
|
|
|
8ae002 |
+ " clgije %[R_LI],0,1f\n\t" \
|
|
|
8ae002 |
+ "0:\n\t" \
|
|
|
8ae002 |
+ " vlm %%v16,%%v19,0(%[R_IN])\n\t" \
|
|
|
8ae002 |
+ /* Shorten to byte values. */ \
|
|
|
8ae002 |
+ " vpkf %%v23,%%v16,%%v17\n\t" \
|
|
|
8ae002 |
+ " vpkf %%v24,%%v18,%%v19\n\t" \
|
|
|
8ae002 |
+ " vpkh %%v23,%%v23,%%v24\n\t" \
|
|
|
8ae002 |
+ /* Checking for values > 0x7f. */ \
|
|
|
8ae002 |
+ " vstrcfs %%v22,%%v16,%%v20,%%v21\n\t" \
|
|
|
8ae002 |
+ " jno 10f\n\t" \
|
|
|
8ae002 |
+ " vstrcfs %%v22,%%v17,%%v20,%%v21\n\t" \
|
|
|
8ae002 |
+ " jno 11f\n\t" \
|
|
|
8ae002 |
+ " vstrcfs %%v22,%%v18,%%v20,%%v21\n\t" \
|
|
|
8ae002 |
+ " jno 12f\n\t" \
|
|
|
8ae002 |
+ " vstrcfs %%v22,%%v19,%%v20,%%v21\n\t" \
|
|
|
8ae002 |
+ " jno 13f\n\t" \
|
|
|
8ae002 |
+ /* Store 16bytes to outptr. */ \
|
|
|
8ae002 |
+ " vst %%v23,0(%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ " la %[R_IN],64(%[R_IN])\n\t" \
|
|
|
8ae002 |
+ " la %[R_OUT],16(%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ " brctg %[R_LI],0b\n\t" \
|
|
|
8ae002 |
+ " lghi %[R_LI],15\n\t" \
|
|
|
8ae002 |
+ " ngr %[R_LEN],%[R_LI]\n\t" \
|
|
|
8ae002 |
+ " je 20f\n\t" /* Jump away if no remaining bytes. */ \
|
|
|
8ae002 |
+ /* Handle remaining bytes. */ \
|
|
|
8ae002 |
+ "1: sllg %[R_LI],%[R_LEN],2\n\t" \
|
|
|
8ae002 |
+ " aghi %[R_LI],-1\n\t" \
|
|
|
8ae002 |
+ " jl 20f\n\t" /* Jump away if no remaining bytes. */ \
|
|
|
8ae002 |
+ /* Load remaining 1...63 bytes. */ \
|
|
|
8ae002 |
+ " vll %%v16,%[R_LI],0(%[R_IN])\n\t" \
|
|
|
8ae002 |
+ " ahi %[R_LI],-16\n\t" \
|
|
|
8ae002 |
+ " jl 2f\n\t" \
|
|
|
8ae002 |
+ " vll %%v17,%[R_LI],16(%[R_IN])\n\t" \
|
|
|
8ae002 |
+ " ahi %[R_LI],-16\n\t" \
|
|
|
8ae002 |
+ " jl 2f\n\t" \
|
|
|
8ae002 |
+ " vll %%v18,%[R_LI],32(%[R_IN])\n\t" \
|
|
|
8ae002 |
+ " ahi %[R_LI],-16\n\t" \
|
|
|
8ae002 |
+ " jl 2f\n\t" \
|
|
|
8ae002 |
+ " vll %%v19,%[R_LI],48(%[R_IN])\n\t" \
|
|
|
8ae002 |
+ "2:\n\t" \
|
|
|
8ae002 |
+ /* Shorten to byte values. */ \
|
|
|
8ae002 |
+ " vpkf %%v23,%%v16,%%v17\n\t" \
|
|
|
8ae002 |
+ " vpkf %%v24,%%v18,%%v19\n\t" \
|
|
|
8ae002 |
+ " vpkh %%v23,%%v23,%%v24\n\t" \
|
|
|
8ae002 |
+ " sllg %[R_LI],%[R_LEN],2\n\t" \
|
|
|
8ae002 |
+ " aghi %[R_LI],-16\n\t" \
|
|
|
8ae002 |
+ " jl 3f\n\t" /* v16 is not fully loaded. */ \
|
|
|
8ae002 |
+ " vstrcfs %%v22,%%v16,%%v20,%%v21\n\t" \
|
|
|
8ae002 |
+ " jno 10f\n\t" \
|
|
|
8ae002 |
+ " aghi %[R_LI],-16\n\t" \
|
|
|
8ae002 |
+ " jl 4f\n\t" /* v17 is not fully loaded. */ \
|
|
|
8ae002 |
+ " vstrcfs %%v22,%%v17,%%v20,%%v21\n\t" \
|
|
|
8ae002 |
+ " jno 11f\n\t" \
|
|
|
8ae002 |
+ " aghi %[R_LI],-16\n\t" \
|
|
|
8ae002 |
+ " jl 5f\n\t" /* v18 is not fully loaded. */ \
|
|
|
8ae002 |
+ " vstrcfs %%v22,%%v18,%%v20,%%v21\n\t" \
|
|
|
8ae002 |
+ " jno 12f\n\t" \
|
|
|
8ae002 |
+ " aghi %[R_LI],-16\n\t" \
|
|
|
8ae002 |
+ /* v19 is not fully loaded. */ \
|
|
|
8ae002 |
+ " lghi %[R_TMP],12\n\t" \
|
|
|
8ae002 |
+ " vstrcfs %%v22,%%v19,%%v20,%%v21\n\t" \
|
|
|
8ae002 |
+ "6: vlgvb %[R_I],%%v22,7\n\t" \
|
|
|
8ae002 |
+ " aghi %[R_LI],16\n\t" \
|
|
|
8ae002 |
+ " clrjl %[R_I],%[R_LI],14f\n\t" \
|
|
|
8ae002 |
+ " lgr %[R_I],%[R_LEN]\n\t" \
|
|
|
8ae002 |
+ " lghi %[R_LEN],0\n\t" \
|
|
|
8ae002 |
+ " j 15f\n\t" \
|
|
|
8ae002 |
+ "3: vstrcfs %%v22,%%v16,%%v20,%%v21\n\t" \
|
|
|
8ae002 |
+ " j 6b\n\t" \
|
|
|
8ae002 |
+ "4: vstrcfs %%v22,%%v17,%%v20,%%v21\n\t" \
|
|
|
8ae002 |
+ " lghi %[R_TMP],4\n\t" \
|
|
|
8ae002 |
+ " j 6b\n\t" \
|
|
|
8ae002 |
+ "5: vstrcfs %%v22,%%v17,%%v20,%%v21\n\t" \
|
|
|
8ae002 |
+ " lghi %[R_TMP],8\n\t" \
|
|
|
8ae002 |
+ " j 6b\n\t" \
|
|
|
8ae002 |
+ /* Found a value > 0x7f. */ \
|
|
|
8ae002 |
+ "13: ahi %[R_TMP],4\n\t" \
|
|
|
8ae002 |
+ "12: ahi %[R_TMP],4\n\t" \
|
|
|
8ae002 |
+ "11: ahi %[R_TMP],4\n\t" \
|
|
|
8ae002 |
+ "10: vlgvb %[R_I],%%v22,7\n\t" \
|
|
|
8ae002 |
+ "14: srlg %[R_I],%[R_I],2\n\t" \
|
|
|
8ae002 |
+ " agr %[R_I],%[R_TMP]\n\t" \
|
|
|
8ae002 |
+ " je 20f\n\t" \
|
|
|
8ae002 |
+ /* Store characters before invalid one... */ \
|
|
|
8ae002 |
+ "15: aghi %[R_I],-1\n\t" \
|
|
|
8ae002 |
+ " vstl %%v23,%[R_I],0(%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ /* ... and update pointers. */ \
|
|
|
8ae002 |
+ " la %[R_OUT],1(%[R_I],%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ " sllg %[R_I],%[R_I],2\n\t" \
|
|
|
8ae002 |
+ " la %[R_IN],4(%[R_I],%[R_IN])\n\t" \
|
|
|
8ae002 |
+ "20:\n\t" \
|
|
|
8ae002 |
+ ".machine pop" \
|
|
|
8ae002 |
+ : /* outputs */ [R_OUT] "+a" (outptr) \
|
|
|
8ae002 |
+ , [R_IN] "+a" (inptr) \
|
|
|
8ae002 |
+ , [R_LEN] "+d" (len) \
|
|
|
8ae002 |
+ , [R_LI] "=d" (loop_count) \
|
|
|
8ae002 |
+ , [R_I] "=a" (tmp2) \
|
|
|
8ae002 |
+ , [R_TMP] "=d" (tmp) \
|
|
|
8ae002 |
+ : /* inputs */ \
|
|
|
8ae002 |
+ : /* clobber list*/ "memory", "cc" \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v16") ASM_CLOBBER_VR ("v17") \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v18") ASM_CLOBBER_VR ("v19") \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v20") ASM_CLOBBER_VR ("v21") \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v22") ASM_CLOBBER_VR ("v23") \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v24") \
|
|
|
8ae002 |
+ ); \
|
|
|
8ae002 |
+ if (len > 0) \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ /* Found an invalid character > 0x7f at next character. */ \
|
|
|
8ae002 |
+ BODY_ORIG_ERROR \
|
|
|
8ae002 |
+ } \
|
|
|
8ae002 |
+ }
|
|
|
8ae002 |
+# define LOOP_NEED_FLAGS
|
|
|
8ae002 |
+# include <iconv/loop.c>
|
|
|
8ae002 |
+# include <iconv/skeleton.c>
|
|
|
8ae002 |
+# undef BODY_ORIG
|
|
|
8ae002 |
+# undef BODY_ORIG_ERROR
|
|
|
8ae002 |
+ICONV_VX_IFUNC (__gconv_transform_internal_ascii)
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+/* Convert from internal UCS4 to UCS4 little endian form. */
|
|
|
8ae002 |
+# define DEFINE_INIT 0
|
|
|
8ae002 |
+# define DEFINE_FINI 0
|
|
|
8ae002 |
+# define MIN_NEEDED_FROM 4
|
|
|
8ae002 |
+# define MIN_NEEDED_TO 4
|
|
|
8ae002 |
+# define FROM_DIRECTION 1
|
|
|
8ae002 |
+# define FROM_LOOP ICONV_VX_NAME (internal_ucs4le_loop)
|
|
|
8ae002 |
+# define TO_LOOP ICONV_VX_NAME (internal_ucs4le_loop) /* This is not used. */
|
|
|
8ae002 |
+# define FUNCTION_NAME ICONV_VX_NAME (__gconv_transform_internal_ucs4le)
|
|
|
8ae002 |
+# define ONE_DIRECTION 0
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+static inline int
|
|
|
8ae002 |
+__attribute ((always_inline))
|
|
|
8ae002 |
+ICONV_VX_NAME (internal_ucs4le_loop) (struct __gconv_step *step,
|
|
|
8ae002 |
+ struct __gconv_step_data *step_data,
|
|
|
8ae002 |
+ const unsigned char **inptrp,
|
|
|
8ae002 |
+ const unsigned char *inend,
|
|
|
8ae002 |
+ unsigned char **outptrp,
|
|
|
8ae002 |
+ unsigned char *outend,
|
|
|
8ae002 |
+ size_t *irreversible)
|
|
|
8ae002 |
+{
|
|
|
8ae002 |
+ const unsigned char *inptr = *inptrp;
|
|
|
8ae002 |
+ unsigned char *outptr = *outptrp;
|
|
|
8ae002 |
+ int result;
|
|
|
8ae002 |
+ size_t len = MIN (inend - inptr, outend - outptr) / 4;
|
|
|
8ae002 |
+ size_t loop_count;
|
|
|
8ae002 |
+ __asm__ volatile (".machine push\n\t"
|
|
|
8ae002 |
+ ".machine \"z13\"\n\t"
|
|
|
8ae002 |
+ ".machinemode \"zarch_nohighgprs\"\n\t"
|
|
|
8ae002 |
+ CONVERT_32BIT_SIZE_T ([R_LEN])
|
|
|
8ae002 |
+ " bras %[R_LI],1f\n\t"
|
|
|
8ae002 |
+ /* Vector permute mask: */
|
|
|
8ae002 |
+ " .long 0x03020100,0x7060504,0x0B0A0908,0x0F0E0D0C\n\t"
|
|
|
8ae002 |
+ "1: vl %%v20,0(%[R_LI])\n\t"
|
|
|
8ae002 |
+ /* Process 64byte (16char) blocks. */
|
|
|
8ae002 |
+ " srlg %[R_LI],%[R_LEN],4\n\t"
|
|
|
8ae002 |
+ " clgije %[R_LI],0,10f\n\t"
|
|
|
8ae002 |
+ "0: vlm %%v16,%%v19,0(%[R_IN])\n\t"
|
|
|
8ae002 |
+ " vperm %%v16,%%v16,%%v16,%%v20\n\t"
|
|
|
8ae002 |
+ " vperm %%v17,%%v17,%%v17,%%v20\n\t"
|
|
|
8ae002 |
+ " vperm %%v18,%%v18,%%v18,%%v20\n\t"
|
|
|
8ae002 |
+ " vperm %%v19,%%v19,%%v19,%%v20\n\t"
|
|
|
8ae002 |
+ " vstm %%v16,%%v19,0(%[R_OUT])\n\t"
|
|
|
8ae002 |
+ " la %[R_IN],64(%[R_IN])\n\t"
|
|
|
8ae002 |
+ " la %[R_OUT],64(%[R_OUT])\n\t"
|
|
|
8ae002 |
+ " brctg %[R_LI],0b\n\t"
|
|
|
8ae002 |
+ " llgfr %[R_LEN],%[R_LEN]\n\t"
|
|
|
8ae002 |
+ " nilf %[R_LEN],15\n\t"
|
|
|
8ae002 |
+ /* Process 16byte (4char) blocks. */
|
|
|
8ae002 |
+ "10: srlg %[R_LI],%[R_LEN],2\n\t"
|
|
|
8ae002 |
+ " clgije %[R_LI],0,20f\n\t"
|
|
|
8ae002 |
+ "11: vl %%v16,0(%[R_IN])\n\t"
|
|
|
8ae002 |
+ " vperm %%v16,%%v16,%%v16,%%v20\n\t"
|
|
|
8ae002 |
+ " vst %%v16,0(%[R_OUT])\n\t"
|
|
|
8ae002 |
+ " la %[R_IN],16(%[R_IN])\n\t"
|
|
|
8ae002 |
+ " la %[R_OUT],16(%[R_OUT])\n\t"
|
|
|
8ae002 |
+ " brctg %[R_LI],11b\n\t"
|
|
|
8ae002 |
+ " nill %[R_LEN],3\n\t"
|
|
|
8ae002 |
+ /* Process <16bytes. */
|
|
|
8ae002 |
+ "20: sll %[R_LEN],2\n\t"
|
|
|
8ae002 |
+ " ahi %[R_LEN],-1\n\t"
|
|
|
8ae002 |
+ " jl 30f\n\t"
|
|
|
8ae002 |
+ " vll %%v16,%[R_LEN],0(%[R_IN])\n\t"
|
|
|
8ae002 |
+ " vperm %%v16,%%v16,%%v16,%%v20\n\t"
|
|
|
8ae002 |
+ " vstl %%v16,%[R_LEN],0(%[R_OUT])\n\t"
|
|
|
8ae002 |
+ " la %[R_IN],1(%[R_LEN],%[R_IN])\n\t"
|
|
|
8ae002 |
+ " la %[R_OUT],1(%[R_LEN],%[R_OUT])\n\t"
|
|
|
8ae002 |
+ "30: \n\t"
|
|
|
8ae002 |
+ ".machine pop"
|
|
|
8ae002 |
+ : /* outputs */ [R_OUT] "+a" (outptr)
|
|
|
8ae002 |
+ , [R_IN] "+a" (inptr)
|
|
|
8ae002 |
+ , [R_LI] "=a" (loop_count)
|
|
|
8ae002 |
+ , [R_LEN] "+a" (len)
|
|
|
8ae002 |
+ : /* inputs */
|
|
|
8ae002 |
+ : /* clobber list*/ "memory", "cc"
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v16") ASM_CLOBBER_VR ("v17")
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v18") ASM_CLOBBER_VR ("v19")
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v20")
|
|
|
8ae002 |
+ );
|
|
|
8ae002 |
+ *inptrp = inptr;
|
|
|
8ae002 |
+ *outptrp = outptr;
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+ /* Determine the status. */
|
|
|
8ae002 |
+ if (*inptrp == inend)
|
|
|
8ae002 |
+ result = __GCONV_EMPTY_INPUT;
|
|
|
8ae002 |
+ else if (*outptrp + 4 > outend)
|
|
|
8ae002 |
+ result = __GCONV_FULL_OUTPUT;
|
|
|
8ae002 |
+ else
|
|
|
8ae002 |
+ result = __GCONV_INCOMPLETE_INPUT;
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+ return result;
|
|
|
8ae002 |
+}
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+ICONV_VX_SINGLE (internal_ucs4le_loop)
|
|
|
8ae002 |
+# include <iconv/skeleton.c>
|
|
|
8ae002 |
+ICONV_VX_IFUNC (__gconv_transform_internal_ucs4le)
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+/* Transform from UCS4 to the internal, UCS4-like format. Unlike
|
|
|
8ae002 |
+ for the other direction we have to check for correct values here. */
|
|
|
8ae002 |
+# define DEFINE_INIT 0
|
|
|
8ae002 |
+# define DEFINE_FINI 0
|
|
|
8ae002 |
+# define MIN_NEEDED_FROM 4
|
|
|
8ae002 |
+# define MIN_NEEDED_TO 4
|
|
|
8ae002 |
+# define FROM_DIRECTION 1
|
|
|
8ae002 |
+# define FROM_LOOP ICONV_VX_NAME (ucs4_internal_loop)
|
|
|
8ae002 |
+# define TO_LOOP ICONV_VX_NAME (ucs4_internal_loop) /* This is not used. */
|
|
|
8ae002 |
+# define FUNCTION_NAME ICONV_VX_NAME (__gconv_transform_ucs4_internal)
|
|
|
8ae002 |
+# define ONE_DIRECTION 0
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+static inline int
|
|
|
8ae002 |
+__attribute ((always_inline))
|
|
|
8ae002 |
+ICONV_VX_NAME (ucs4_internal_loop) (struct __gconv_step *step,
|
|
|
8ae002 |
+ struct __gconv_step_data *step_data,
|
|
|
8ae002 |
+ const unsigned char **inptrp,
|
|
|
8ae002 |
+ const unsigned char *inend,
|
|
|
8ae002 |
+ unsigned char **outptrp,
|
|
|
8ae002 |
+ unsigned char *outend,
|
|
|
8ae002 |
+ size_t *irreversible)
|
|
|
8ae002 |
+{
|
|
|
8ae002 |
+ int flags = step_data->__flags;
|
|
|
8ae002 |
+ const unsigned char *inptr = *inptrp;
|
|
|
8ae002 |
+ unsigned char *outptr = *outptrp;
|
|
|
8ae002 |
+ int result;
|
|
|
8ae002 |
+ size_t len, loop_count;
|
|
|
8ae002 |
+ do
|
|
|
8ae002 |
+ {
|
|
|
8ae002 |
+ len = MIN (inend - inptr, outend - outptr) / 4;
|
|
|
8ae002 |
+ __asm__ volatile (".machine push\n\t"
|
|
|
8ae002 |
+ ".machine \"z13\"\n\t"
|
|
|
8ae002 |
+ ".machinemode \"zarch_nohighgprs\"\n\t"
|
|
|
8ae002 |
+ CONVERT_32BIT_SIZE_T ([R_LEN])
|
|
|
8ae002 |
+ /* Setup to check for ch > 0x7fffffff. */
|
|
|
8ae002 |
+ " larl %[R_LI],9f\n\t"
|
|
|
8ae002 |
+ " vlm %%v20,%%v21,0(%[R_LI])\n\t"
|
|
|
8ae002 |
+ " srlg %[R_LI],%[R_LEN],2\n\t"
|
|
|
8ae002 |
+ " clgije %[R_LI],0,1f\n\t"
|
|
|
8ae002 |
+ /* Process 16byte (4char) blocks. */
|
|
|
8ae002 |
+ "0: vl %%v16,0(%[R_IN])\n\t"
|
|
|
8ae002 |
+ " vstrcfs %%v22,%%v16,%%v20,%%v21\n\t"
|
|
|
8ae002 |
+ " jno 10f\n\t"
|
|
|
8ae002 |
+ " vst %%v16,0(%[R_OUT])\n\t"
|
|
|
8ae002 |
+ " la %[R_IN],16(%[R_IN])\n\t"
|
|
|
8ae002 |
+ " la %[R_OUT],16(%[R_OUT])\n\t"
|
|
|
8ae002 |
+ " brctg %[R_LI],0b\n\t"
|
|
|
8ae002 |
+ " llgfr %[R_LEN],%[R_LEN]\n\t"
|
|
|
8ae002 |
+ " nilf %[R_LEN],3\n\t"
|
|
|
8ae002 |
+ /* Process <16bytes. */
|
|
|
8ae002 |
+ "1: sll %[R_LEN],2\n\t"
|
|
|
8ae002 |
+ " ahik %[R_LI],%[R_LEN],-1\n\t"
|
|
|
8ae002 |
+ " jl 20f\n\t" /* No further bytes available. */
|
|
|
8ae002 |
+ " vll %%v16,%[R_LI],0(%[R_IN])\n\t"
|
|
|
8ae002 |
+ " vstrcfs %%v22,%%v16,%%v20,%%v21\n\t"
|
|
|
8ae002 |
+ " vlgvb %[R_LI],%%v22,7\n\t"
|
|
|
8ae002 |
+ " clr %[R_LI],%[R_LEN]\n\t"
|
|
|
8ae002 |
+ " locgrhe %[R_LI],%[R_LEN]\n\t"
|
|
|
8ae002 |
+ " locghihe %[R_LEN],0\n\t"
|
|
|
8ae002 |
+ " j 11f\n\t"
|
|
|
8ae002 |
+ /* v20: Vector string range compare values. */
|
|
|
8ae002 |
+ "9: .long 0x7fffffff,0x0,0x0,0x0\n\t"
|
|
|
8ae002 |
+ /* v21: Vector string range compare control-bits.
|
|
|
8ae002 |
+ element 0: >; element 1: =<> (always true) */
|
|
|
8ae002 |
+ " .long 0x20000000,0xE0000000,0x0,0x0\n\t"
|
|
|
8ae002 |
+ /* Found a value > 0x7fffffff. */
|
|
|
8ae002 |
+ "10: vlgvb %[R_LI],%%v22,7\n\t"
|
|
|
8ae002 |
+ /* Store characters before invalid one. */
|
|
|
8ae002 |
+ "11: aghi %[R_LI],-1\n\t"
|
|
|
8ae002 |
+ " jl 20f\n\t"
|
|
|
8ae002 |
+ " vstl %%v16,%[R_LI],0(%[R_OUT])\n\t"
|
|
|
8ae002 |
+ " la %[R_IN],1(%[R_LI],%[R_IN])\n\t"
|
|
|
8ae002 |
+ " la %[R_OUT],1(%[R_LI],%[R_OUT])\n\t"
|
|
|
8ae002 |
+ "20:\n\t"
|
|
|
8ae002 |
+ ".machine pop"
|
|
|
8ae002 |
+ : /* outputs */ [R_OUT] "+a" (outptr)
|
|
|
8ae002 |
+ , [R_IN] "+a" (inptr)
|
|
|
8ae002 |
+ , [R_LI] "=a" (loop_count)
|
|
|
8ae002 |
+ , [R_LEN] "+d" (len)
|
|
|
8ae002 |
+ : /* inputs */
|
|
|
8ae002 |
+ : /* clobber list*/ "memory", "cc"
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v16") ASM_CLOBBER_VR ("v20")
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v21") ASM_CLOBBER_VR ("v22")
|
|
|
8ae002 |
+ );
|
|
|
8ae002 |
+ if (len > 0)
|
|
|
8ae002 |
+ {
|
|
|
8ae002 |
+ /* The value is too large. We don't try transliteration here since
|
|
|
8ae002 |
+ this is not an error because of the lack of possibilities to
|
|
|
8ae002 |
+ represent the result. This is a genuine bug in the input since
|
|
|
8ae002 |
+ UCS4 does not allow such values. */
|
|
|
8ae002 |
+ if (irreversible == NULL)
|
|
|
8ae002 |
+ /* We are transliterating, don't try to correct anything. */
|
|
|
8ae002 |
+ return __GCONV_ILLEGAL_INPUT;
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+ if (flags & __GCONV_IGNORE_ERRORS)
|
|
|
8ae002 |
+ {
|
|
|
8ae002 |
+ /* Just ignore this character. */
|
|
|
8ae002 |
+ ++*irreversible;
|
|
|
8ae002 |
+ inptr += 4;
|
|
|
8ae002 |
+ continue;
|
|
|
8ae002 |
+ }
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+ *inptrp = inptr;
|
|
|
8ae002 |
+ *outptrp = outptr;
|
|
|
8ae002 |
+ return __GCONV_ILLEGAL_INPUT;
|
|
|
8ae002 |
+ }
|
|
|
8ae002 |
+ }
|
|
|
8ae002 |
+ while (len > 0);
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+ *inptrp = inptr;
|
|
|
8ae002 |
+ *outptrp = outptr;
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+ /* Determine the status. */
|
|
|
8ae002 |
+ if (*inptrp == inend)
|
|
|
8ae002 |
+ result = __GCONV_EMPTY_INPUT;
|
|
|
8ae002 |
+ else if (*outptrp + 4 > outend)
|
|
|
8ae002 |
+ result = __GCONV_FULL_OUTPUT;
|
|
|
8ae002 |
+ else
|
|
|
8ae002 |
+ result = __GCONV_INCOMPLETE_INPUT;
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+ return result;
|
|
|
8ae002 |
+}
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+ICONV_VX_SINGLE (ucs4_internal_loop)
|
|
|
8ae002 |
+# include <iconv/skeleton.c>
|
|
|
8ae002 |
+ICONV_VX_IFUNC (__gconv_transform_ucs4_internal)
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+/* Transform from UCS4-LE to the internal encoding. */
|
|
|
8ae002 |
+# define DEFINE_INIT 0
|
|
|
8ae002 |
+# define DEFINE_FINI 0
|
|
|
8ae002 |
+# define MIN_NEEDED_FROM 4
|
|
|
8ae002 |
+# define MIN_NEEDED_TO 4
|
|
|
8ae002 |
+# define FROM_DIRECTION 1
|
|
|
8ae002 |
+# define FROM_LOOP ICONV_VX_NAME (ucs4le_internal_loop)
|
|
|
8ae002 |
+# define TO_LOOP ICONV_VX_NAME (ucs4le_internal_loop) /* This is not used. */
|
|
|
8ae002 |
+# define FUNCTION_NAME ICONV_VX_NAME (__gconv_transform_ucs4le_internal)
|
|
|
8ae002 |
+# define ONE_DIRECTION 0
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+static inline int
|
|
|
8ae002 |
+__attribute ((always_inline))
|
|
|
8ae002 |
+ICONV_VX_NAME (ucs4le_internal_loop) (struct __gconv_step *step,
|
|
|
8ae002 |
+ struct __gconv_step_data *step_data,
|
|
|
8ae002 |
+ const unsigned char **inptrp,
|
|
|
8ae002 |
+ const unsigned char *inend,
|
|
|
8ae002 |
+ unsigned char **outptrp,
|
|
|
8ae002 |
+ unsigned char *outend,
|
|
|
8ae002 |
+ size_t *irreversible)
|
|
|
8ae002 |
+{
|
|
|
8ae002 |
+ int flags = step_data->__flags;
|
|
|
8ae002 |
+ const unsigned char *inptr = *inptrp;
|
|
|
8ae002 |
+ unsigned char *outptr = *outptrp;
|
|
|
8ae002 |
+ int result;
|
|
|
8ae002 |
+ size_t len, loop_count;
|
|
|
8ae002 |
+ do
|
|
|
8ae002 |
+ {
|
|
|
8ae002 |
+ len = MIN (inend - inptr, outend - outptr) / 4;
|
|
|
8ae002 |
+ __asm__ volatile (".machine push\n\t"
|
|
|
8ae002 |
+ ".machine \"z13\"\n\t"
|
|
|
8ae002 |
+ ".machinemode \"zarch_nohighgprs\"\n\t"
|
|
|
8ae002 |
+ CONVERT_32BIT_SIZE_T ([R_LEN])
|
|
|
8ae002 |
+ /* Setup to check for ch > 0x7fffffff. */
|
|
|
8ae002 |
+ " larl %[R_LI],9f\n\t"
|
|
|
8ae002 |
+ " vlm %%v20,%%v22,0(%[R_LI])\n\t"
|
|
|
8ae002 |
+ " srlg %[R_LI],%[R_LEN],2\n\t"
|
|
|
8ae002 |
+ " clgije %[R_LI],0,1f\n\t"
|
|
|
8ae002 |
+ /* Process 16byte (4char) blocks. */
|
|
|
8ae002 |
+ "0: vl %%v16,0(%[R_IN])\n\t"
|
|
|
8ae002 |
+ " vperm %%v16,%%v16,%%v16,%%v22\n\t"
|
|
|
8ae002 |
+ " vstrcfs %%v23,%%v16,%%v20,%%v21\n\t"
|
|
|
8ae002 |
+ " jno 10f\n\t"
|
|
|
8ae002 |
+ " vst %%v16,0(%[R_OUT])\n\t"
|
|
|
8ae002 |
+ " la %[R_IN],16(%[R_IN])\n\t"
|
|
|
8ae002 |
+ " la %[R_OUT],16(%[R_OUT])\n\t"
|
|
|
8ae002 |
+ " brctg %[R_LI],0b\n\t"
|
|
|
8ae002 |
+ " llgfr %[R_LEN],%[R_LEN]\n\t"
|
|
|
8ae002 |
+ " nilf %[R_LEN],3\n\t"
|
|
|
8ae002 |
+ /* Process <16bytes. */
|
|
|
8ae002 |
+ "1: sll %[R_LEN],2\n\t"
|
|
|
8ae002 |
+ " ahik %[R_LI],%[R_LEN],-1\n\t"
|
|
|
8ae002 |
+ " jl 20f\n\t" /* No further bytes available. */
|
|
|
8ae002 |
+ " vll %%v16,%[R_LI],0(%[R_IN])\n\t"
|
|
|
8ae002 |
+ " vperm %%v16,%%v16,%%v16,%%v22\n\t"
|
|
|
8ae002 |
+ " vstrcfs %%v23,%%v16,%%v20,%%v21\n\t"
|
|
|
8ae002 |
+ " vlgvb %[R_LI],%%v23,7\n\t"
|
|
|
8ae002 |
+ " clr %[R_LI],%[R_LEN]\n\t"
|
|
|
8ae002 |
+ " locgrhe %[R_LI],%[R_LEN]\n\t"
|
|
|
8ae002 |
+ " locghihe %[R_LEN],0\n\t"
|
|
|
8ae002 |
+ " j 11f\n\t"
|
|
|
8ae002 |
+ /* v20: Vector string range compare values. */
|
|
|
8ae002 |
+ "9: .long 0x7fffffff,0x0,0x0,0x0\n\t"
|
|
|
8ae002 |
+ /* v21: Vector string range compare control-bits.
|
|
|
8ae002 |
+ element 0: >; element 1: =<> (always true) */
|
|
|
8ae002 |
+ " .long 0x20000000,0xE0000000,0x0,0x0\n\t"
|
|
|
8ae002 |
+ /* v22: Vector permute mask. */
|
|
|
8ae002 |
+ " .long 0x03020100,0x7060504,0x0B0A0908,0x0F0E0D0C\n\t"
|
|
|
8ae002 |
+ /* Found a value > 0x7fffffff. */
|
|
|
8ae002 |
+ "10: vlgvb %[R_LI],%%v23,7\n\t"
|
|
|
8ae002 |
+ /* Store characters before invalid one. */
|
|
|
8ae002 |
+ "11: aghi %[R_LI],-1\n\t"
|
|
|
8ae002 |
+ " jl 20f\n\t"
|
|
|
8ae002 |
+ " vstl %%v16,%[R_LI],0(%[R_OUT])\n\t"
|
|
|
8ae002 |
+ " la %[R_IN],1(%[R_LI],%[R_IN])\n\t"
|
|
|
8ae002 |
+ " la %[R_OUT],1(%[R_LI],%[R_OUT])\n\t"
|
|
|
8ae002 |
+ "20:\n\t"
|
|
|
8ae002 |
+ ".machine pop"
|
|
|
8ae002 |
+ : /* outputs */ [R_OUT] "+a" (outptr)
|
|
|
8ae002 |
+ , [R_IN] "+a" (inptr)
|
|
|
8ae002 |
+ , [R_LI] "=a" (loop_count)
|
|
|
8ae002 |
+ , [R_LEN] "+d" (len)
|
|
|
8ae002 |
+ : /* inputs */
|
|
|
8ae002 |
+ : /* clobber list*/ "memory", "cc"
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v16") ASM_CLOBBER_VR ("v20")
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v21") ASM_CLOBBER_VR ("v22")
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v23")
|
|
|
8ae002 |
+ );
|
|
|
8ae002 |
+ if (len > 0)
|
|
|
8ae002 |
+ {
|
|
|
8ae002 |
+ /* The value is too large. We don't try transliteration here since
|
|
|
8ae002 |
+ this is not an error because of the lack of possibilities to
|
|
|
8ae002 |
+ represent the result. This is a genuine bug in the input since
|
|
|
8ae002 |
+ UCS4 does not allow such values. */
|
|
|
8ae002 |
+ if (irreversible == NULL)
|
|
|
8ae002 |
+ /* We are transliterating, don't try to correct anything. */
|
|
|
8ae002 |
+ return __GCONV_ILLEGAL_INPUT;
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+ if (flags & __GCONV_IGNORE_ERRORS)
|
|
|
8ae002 |
+ {
|
|
|
8ae002 |
+ /* Just ignore this character. */
|
|
|
8ae002 |
+ ++*irreversible;
|
|
|
8ae002 |
+ inptr += 4;
|
|
|
8ae002 |
+ continue;
|
|
|
8ae002 |
+ }
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+ *inptrp = inptr;
|
|
|
8ae002 |
+ *outptrp = outptr;
|
|
|
8ae002 |
+ return __GCONV_ILLEGAL_INPUT;
|
|
|
8ae002 |
+ }
|
|
|
8ae002 |
+ }
|
|
|
8ae002 |
+ while (len > 0);
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+ *inptrp = inptr;
|
|
|
8ae002 |
+ *outptrp = outptr;
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+ /* Determine the status. */
|
|
|
8ae002 |
+ if (*inptrp == inend)
|
|
|
8ae002 |
+ result = __GCONV_EMPTY_INPUT;
|
|
|
8ae002 |
+ else if (*inptrp + 4 > inend)
|
|
|
8ae002 |
+ result = __GCONV_INCOMPLETE_INPUT;
|
|
|
8ae002 |
+ else
|
|
|
8ae002 |
+ {
|
|
|
8ae002 |
+ assert (*outptrp + 4 > outend);
|
|
|
8ae002 |
+ result = __GCONV_FULL_OUTPUT;
|
|
|
8ae002 |
+ }
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+ return result;
|
|
|
8ae002 |
+}
|
|
|
8ae002 |
+ICONV_VX_SINGLE (ucs4le_internal_loop)
|
|
|
8ae002 |
+# include <iconv/skeleton.c>
|
|
|
8ae002 |
+ICONV_VX_IFUNC (__gconv_transform_ucs4le_internal)
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+/* Convert from UCS2 to the internal (UCS4-like) format. */
|
|
|
8ae002 |
+# define DEFINE_INIT 0
|
|
|
8ae002 |
+# define DEFINE_FINI 0
|
|
|
8ae002 |
+# define MIN_NEEDED_FROM 2
|
|
|
8ae002 |
+# define MIN_NEEDED_TO 4
|
|
|
8ae002 |
+# define FROM_DIRECTION 1
|
|
|
8ae002 |
+# define FROM_LOOP ICONV_VX_NAME (ucs2_internal_loop)
|
|
|
8ae002 |
+# define TO_LOOP ICONV_VX_NAME (ucs2_internal_loop) /* This is not used. */
|
|
|
8ae002 |
+# define FUNCTION_NAME ICONV_VX_NAME (__gconv_transform_ucs2_internal)
|
|
|
8ae002 |
+# define ONE_DIRECTION 1
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+# define MIN_NEEDED_INPUT MIN_NEEDED_FROM
|
|
|
8ae002 |
+# define MIN_NEEDED_OUTPUT MIN_NEEDED_TO
|
|
|
8ae002 |
+# define LOOPFCT FROM_LOOP
|
|
|
8ae002 |
+# define BODY_ORIG_ERROR \
|
|
|
8ae002 |
+ /* Surrogate characters in UCS-2 input are not valid. Reject \
|
|
|
8ae002 |
+ them. (Catching this here is not security relevant.) */ \
|
|
|
8ae002 |
+ STANDARD_FROM_LOOP_ERR_HANDLER (2);
|
|
|
8ae002 |
+# define BODY_ORIG \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ uint16_t u1 = get16 (inptr); \
|
|
|
8ae002 |
+ \
|
|
|
8ae002 |
+ if (__glibc_unlikely (u1 >= 0xd800 && u1 < 0xe000)) \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ BODY_ORIG_ERROR \
|
|
|
8ae002 |
+ } \
|
|
|
8ae002 |
+ \
|
|
|
8ae002 |
+ *((uint32_t *) outptr) = u1; \
|
|
|
8ae002 |
+ outptr += sizeof (uint32_t); \
|
|
|
8ae002 |
+ inptr += 2; \
|
|
|
8ae002 |
+ }
|
|
|
8ae002 |
+# define BODY \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ size_t len, tmp, tmp2; \
|
|
|
8ae002 |
+ len = MIN ((inend - inptr) / 2, (outend - outptr) / 4); \
|
|
|
8ae002 |
+ __asm__ volatile (".machine push\n\t" \
|
|
|
8ae002 |
+ ".machine \"z13\"\n\t" \
|
|
|
8ae002 |
+ ".machinemode \"zarch_nohighgprs\"\n\t" \
|
|
|
8ae002 |
+ CONVERT_32BIT_SIZE_T ([R_LEN]) \
|
|
|
8ae002 |
+ /* Setup to check for ch >= 0xd800 && ch < 0xe000. */ \
|
|
|
8ae002 |
+ " larl %[R_TMP],9f\n\t" \
|
|
|
8ae002 |
+ " vlm %%v20,%%v21,0(%[R_TMP])\n\t" \
|
|
|
8ae002 |
+ " srlg %[R_TMP],%[R_LEN],3\n\t" \
|
|
|
8ae002 |
+ " clgije %[R_TMP],0,1f\n\t" \
|
|
|
8ae002 |
+ /* Process 16byte (8char) blocks. */ \
|
|
|
8ae002 |
+ "0: vl %%v16,0(%[R_IN])\n\t" \
|
|
|
8ae002 |
+ " vstrchs %%v19,%%v16,%%v20,%%v21\n\t" \
|
|
|
8ae002 |
+ /* Enlarge UCS2 to UCS4. */ \
|
|
|
8ae002 |
+ " vuplhh %%v17,%%v16\n\t" \
|
|
|
8ae002 |
+ " vupllh %%v18,%%v16\n\t" \
|
|
|
8ae002 |
+ " jno 10f\n\t" \
|
|
|
8ae002 |
+ /* Store 32bytes to buf_out. */ \
|
|
|
8ae002 |
+ " vstm %%v17,%%v18,0(%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ " la %[R_IN],16(%[R_IN])\n\t" \
|
|
|
8ae002 |
+ " la %[R_OUT],32(%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ " brctg %[R_TMP],0b\n\t" \
|
|
|
8ae002 |
+ " llgfr %[R_LEN],%[R_LEN]\n\t" \
|
|
|
8ae002 |
+ " nilf %[R_LEN],7\n\t" \
|
|
|
8ae002 |
+ /* Process <16bytes. */ \
|
|
|
8ae002 |
+ "1: sll %[R_LEN],1\n\t" \
|
|
|
8ae002 |
+ " ahik %[R_TMP],%[R_LEN],-1\n\t" \
|
|
|
8ae002 |
+ " jl 20f\n\t" /* No further bytes available. */ \
|
|
|
8ae002 |
+ " vll %%v16,%[R_TMP],0(%[R_IN])\n\t" \
|
|
|
8ae002 |
+ " vstrchs %%v19,%%v16,%%v20,%%v21\n\t" \
|
|
|
8ae002 |
+ /* Enlarge UCS2 to UCS4. */ \
|
|
|
8ae002 |
+ " vuplhh %%v17,%%v16\n\t" \
|
|
|
8ae002 |
+ " vupllh %%v18,%%v16\n\t" \
|
|
|
8ae002 |
+ " vlgvb %[R_TMP],%%v19,7\n\t" \
|
|
|
8ae002 |
+ " clr %[R_TMP],%[R_LEN]\n\t" \
|
|
|
8ae002 |
+ " locgrhe %[R_TMP],%[R_LEN]\n\t" \
|
|
|
8ae002 |
+ " locghihe %[R_LEN],0\n\t" \
|
|
|
8ae002 |
+ " j 11f\n\t" \
|
|
|
8ae002 |
+ /* v20: Vector string range compare values. */ \
|
|
|
8ae002 |
+ "9: .short 0xd800,0xe000,0x0,0x0,0x0,0x0,0x0,0x0\n\t" \
|
|
|
8ae002 |
+ /* v21: Vector string range compare control-bits. \
|
|
|
8ae002 |
+ element 0: =>; element 1: < */ \
|
|
|
8ae002 |
+ " .short 0xa000,0x4000,0x0,0x0,0x0,0x0,0x0,0x0\n\t" \
|
|
|
8ae002 |
+ /* Found an element: ch >= 0xd800 && ch < 0xe000 */ \
|
|
|
8ae002 |
+ "10: vlgvb %[R_TMP],%%v19,7\n\t" \
|
|
|
8ae002 |
+ "11: la %[R_IN],0(%[R_TMP],%[R_IN])\n\t" \
|
|
|
8ae002 |
+ " sll %[R_TMP],1\n\t" \
|
|
|
8ae002 |
+ " lgr %[R_TMP2],%[R_TMP]\n\t" \
|
|
|
8ae002 |
+ " ahi %[R_TMP],-1\n\t" \
|
|
|
8ae002 |
+ " jl 20f\n\t" \
|
|
|
8ae002 |
+ " vstl %%v17,%[R_TMP],0(%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ " ahi %[R_TMP],-16\n\t" \
|
|
|
8ae002 |
+ " jl 19f\n\t" \
|
|
|
8ae002 |
+ " vstl %%v18,%[R_TMP],16(%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ "19: la %[R_OUT],0(%[R_TMP2],%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ "20: \n\t" \
|
|
|
8ae002 |
+ ".machine pop" \
|
|
|
8ae002 |
+ : /* outputs */ [R_OUT] "+a" (outptr) \
|
|
|
8ae002 |
+ , [R_IN] "+a" (inptr) \
|
|
|
8ae002 |
+ , [R_TMP] "=a" (tmp) \
|
|
|
8ae002 |
+ , [R_TMP2] "=a" (tmp2) \
|
|
|
8ae002 |
+ , [R_LEN] "+d" (len) \
|
|
|
8ae002 |
+ : /* inputs */ \
|
|
|
8ae002 |
+ : /* clobber list*/ "memory", "cc" \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v16") ASM_CLOBBER_VR ("v17") \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v18") ASM_CLOBBER_VR ("v19") \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v20") ASM_CLOBBER_VR ("v21") \
|
|
|
8ae002 |
+ ); \
|
|
|
8ae002 |
+ if (len > 0) \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ /* Found an invalid character at next input-char. */ \
|
|
|
8ae002 |
+ BODY_ORIG_ERROR \
|
|
|
8ae002 |
+ } \
|
|
|
8ae002 |
+ }
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+# define LOOP_NEED_FLAGS
|
|
|
8ae002 |
+# include <iconv/loop.c>
|
|
|
8ae002 |
+# include <iconv/skeleton.c>
|
|
|
8ae002 |
+# undef BODY_ORIG
|
|
|
8ae002 |
+# undef BODY_ORIG_ERROR
|
|
|
8ae002 |
+ICONV_VX_IFUNC (__gconv_transform_ucs2_internal)
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+/* Convert from UCS2 in other endianness to the internal (UCS4-like) format. */
|
|
|
8ae002 |
+# define DEFINE_INIT 0
|
|
|
8ae002 |
+# define DEFINE_FINI 0
|
|
|
8ae002 |
+# define MIN_NEEDED_FROM 2
|
|
|
8ae002 |
+# define MIN_NEEDED_TO 4
|
|
|
8ae002 |
+# define FROM_DIRECTION 1
|
|
|
8ae002 |
+# define FROM_LOOP ICONV_VX_NAME (ucs2reverse_internal_loop)
|
|
|
8ae002 |
+# define TO_LOOP ICONV_VX_NAME (ucs2reverse_internal_loop) /* This is not used.*/
|
|
|
8ae002 |
+# define FUNCTION_NAME ICONV_VX_NAME (__gconv_transform_ucs2reverse_internal)
|
|
|
8ae002 |
+# define ONE_DIRECTION 1
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+# define MIN_NEEDED_INPUT MIN_NEEDED_FROM
|
|
|
8ae002 |
+# define MIN_NEEDED_OUTPUT MIN_NEEDED_TO
|
|
|
8ae002 |
+# define LOOPFCT FROM_LOOP
|
|
|
8ae002 |
+# define BODY_ORIG_ERROR \
|
|
|
8ae002 |
+ /* Surrogate characters in UCS-2 input are not valid. Reject \
|
|
|
8ae002 |
+ them. (Catching this here is not security relevant.) */ \
|
|
|
8ae002 |
+ if (! ignore_errors_p ()) \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ result = __GCONV_ILLEGAL_INPUT; \
|
|
|
8ae002 |
+ break; \
|
|
|
8ae002 |
+ } \
|
|
|
8ae002 |
+ inptr += 2; \
|
|
|
8ae002 |
+ ++*irreversible; \
|
|
|
8ae002 |
+ continue;
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+# define BODY_ORIG \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ uint16_t u1 = bswap_16 (get16 (inptr)); \
|
|
|
8ae002 |
+ \
|
|
|
8ae002 |
+ if (__glibc_unlikely (u1 >= 0xd800 && u1 < 0xe000)) \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ BODY_ORIG_ERROR \
|
|
|
8ae002 |
+ } \
|
|
|
8ae002 |
+ \
|
|
|
8ae002 |
+ *((uint32_t *) outptr) = u1; \
|
|
|
8ae002 |
+ outptr += sizeof (uint32_t); \
|
|
|
8ae002 |
+ inptr += 2; \
|
|
|
8ae002 |
+ }
|
|
|
8ae002 |
+# define BODY \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ size_t len, tmp, tmp2; \
|
|
|
8ae002 |
+ len = MIN ((inend - inptr) / 2, (outend - outptr) / 4); \
|
|
|
8ae002 |
+ __asm__ volatile (".machine push\n\t" \
|
|
|
8ae002 |
+ ".machine \"z13\"\n\t" \
|
|
|
8ae002 |
+ ".machinemode \"zarch_nohighgprs\"\n\t" \
|
|
|
8ae002 |
+ CONVERT_32BIT_SIZE_T ([R_LEN]) \
|
|
|
8ae002 |
+ /* Setup to check for ch >= 0xd800 && ch < 0xe000. */ \
|
|
|
8ae002 |
+ " larl %[R_TMP],9f\n\t" \
|
|
|
8ae002 |
+ " vlm %%v20,%%v22,0(%[R_TMP])\n\t" \
|
|
|
8ae002 |
+ " srlg %[R_TMP],%[R_LEN],3\n\t" \
|
|
|
8ae002 |
+ " clgije %[R_TMP],0,1f\n\t" \
|
|
|
8ae002 |
+ /* Process 16byte (8char) blocks. */ \
|
|
|
8ae002 |
+ "0: vl %%v16,0(%[R_IN])\n\t" \
|
|
|
8ae002 |
+ " vperm %%v16,%%v16,%%v16,%%v22\n\t" \
|
|
|
8ae002 |
+ " vstrchs %%v19,%%v16,%%v20,%%v21\n\t" \
|
|
|
8ae002 |
+ /* Enlarge UCS2 to UCS4. */ \
|
|
|
8ae002 |
+ " vuplhh %%v17,%%v16\n\t" \
|
|
|
8ae002 |
+ " vupllh %%v18,%%v16\n\t" \
|
|
|
8ae002 |
+ " jno 10f\n\t" \
|
|
|
8ae002 |
+ /* Store 32bytes to buf_out. */ \
|
|
|
8ae002 |
+ " vstm %%v17,%%v18,0(%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ " la %[R_IN],16(%[R_IN])\n\t" \
|
|
|
8ae002 |
+ " la %[R_OUT],32(%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ " brctg %[R_TMP],0b\n\t" \
|
|
|
8ae002 |
+ " llgfr %[R_LEN],%[R_LEN]\n\t" \
|
|
|
8ae002 |
+ " nilf %[R_LEN],7\n\t" \
|
|
|
8ae002 |
+ /* Process <16bytes. */ \
|
|
|
8ae002 |
+ "1: sll %[R_LEN],1\n\t" \
|
|
|
8ae002 |
+ " ahik %[R_TMP],%[R_LEN],-1\n\t" \
|
|
|
8ae002 |
+ " jl 20f\n\t" /* No further bytes available. */ \
|
|
|
8ae002 |
+ " vll %%v16,%[R_TMP],0(%[R_IN])\n\t" \
|
|
|
8ae002 |
+ " vperm %%v16,%%v16,%%v16,%%v22\n\t" \
|
|
|
8ae002 |
+ " vstrchs %%v19,%%v16,%%v20,%%v21\n\t" \
|
|
|
8ae002 |
+ /* Enlarge UCS2 to UCS4. */ \
|
|
|
8ae002 |
+ " vuplhh %%v17,%%v16\n\t" \
|
|
|
8ae002 |
+ " vupllh %%v18,%%v16\n\t" \
|
|
|
8ae002 |
+ " vlgvb %[R_TMP],%%v19,7\n\t" \
|
|
|
8ae002 |
+ " clr %[R_TMP],%[R_LEN]\n\t" \
|
|
|
8ae002 |
+ " locgrhe %[R_TMP],%[R_LEN]\n\t" \
|
|
|
8ae002 |
+ " locghihe %[R_LEN],0\n\t" \
|
|
|
8ae002 |
+ " j 11f\n\t" \
|
|
|
8ae002 |
+ /* v20: Vector string range compare values. */ \
|
|
|
8ae002 |
+ "9: .short 0xd800,0xe000,0x0,0x0,0x0,0x0,0x0,0x0\n\t" \
|
|
|
8ae002 |
+ /* v21: Vector string range compare control-bits. \
|
|
|
8ae002 |
+ element 0: =>; element 1: < */ \
|
|
|
8ae002 |
+ " .short 0xa000,0x4000,0x0,0x0,0x0,0x0,0x0,0x0\n\t" \
|
|
|
8ae002 |
+ /* v22: Vector permute mask. */ \
|
|
|
8ae002 |
+ " .short 0x0100,0x0302,0x0504,0x0706\n\t" \
|
|
|
8ae002 |
+ " .short 0x0908,0x0b0a,0x0d0c,0x0f0e\n\t" \
|
|
|
8ae002 |
+ /* Found an element: ch >= 0xd800 && ch < 0xe000 */ \
|
|
|
8ae002 |
+ "10: vlgvb %[R_TMP],%%v19,7\n\t" \
|
|
|
8ae002 |
+ "11: la %[R_IN],0(%[R_TMP],%[R_IN])\n\t" \
|
|
|
8ae002 |
+ " sll %[R_TMP],1\n\t" \
|
|
|
8ae002 |
+ " lgr %[R_TMP2],%[R_TMP]\n\t" \
|
|
|
8ae002 |
+ " ahi %[R_TMP],-1\n\t" \
|
|
|
8ae002 |
+ " jl 20f\n\t" \
|
|
|
8ae002 |
+ " vstl %%v17,%[R_TMP],0(%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ " ahi %[R_TMP],-16\n\t" \
|
|
|
8ae002 |
+ " jl 19f\n\t" \
|
|
|
8ae002 |
+ " vstl %%v18,%[R_TMP],16(%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ "19: la %[R_OUT],0(%[R_TMP2],%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ "20: \n\t" \
|
|
|
8ae002 |
+ ".machine pop" \
|
|
|
8ae002 |
+ : /* outputs */ [R_OUT] "+a" (outptr) \
|
|
|
8ae002 |
+ , [R_IN] "+a" (inptr) \
|
|
|
8ae002 |
+ , [R_TMP] "=a" (tmp) \
|
|
|
8ae002 |
+ , [R_TMP2] "=a" (tmp2) \
|
|
|
8ae002 |
+ , [R_LEN] "+d" (len) \
|
|
|
8ae002 |
+ : /* inputs */ \
|
|
|
8ae002 |
+ : /* clobber list*/ "memory", "cc" \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v16") ASM_CLOBBER_VR ("v17") \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v18") ASM_CLOBBER_VR ("v19") \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v20") ASM_CLOBBER_VR ("v21") \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v22") \
|
|
|
8ae002 |
+ ); \
|
|
|
8ae002 |
+ if (len > 0) \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ /* Found an invalid character at next input-char. */ \
|
|
|
8ae002 |
+ BODY_ORIG_ERROR \
|
|
|
8ae002 |
+ } \
|
|
|
8ae002 |
+ }
|
|
|
8ae002 |
+# define LOOP_NEED_FLAGS
|
|
|
8ae002 |
+# include <iconv/loop.c>
|
|
|
8ae002 |
+# include <iconv/skeleton.c>
|
|
|
8ae002 |
+# undef BODY_ORIG
|
|
|
8ae002 |
+# undef BODY_ORIG_ERROR
|
|
|
8ae002 |
+ICONV_VX_IFUNC (__gconv_transform_ucs2reverse_internal)
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+/* Convert from the internal (UCS4-like) format to UCS2. */
|
|
|
8ae002 |
+#define DEFINE_INIT 0
|
|
|
8ae002 |
+#define DEFINE_FINI 0
|
|
|
8ae002 |
+#define MIN_NEEDED_FROM 4
|
|
|
8ae002 |
+#define MIN_NEEDED_TO 2
|
|
|
8ae002 |
+#define FROM_DIRECTION 1
|
|
|
8ae002 |
+#define FROM_LOOP ICONV_VX_NAME (internal_ucs2_loop)
|
|
|
8ae002 |
+#define TO_LOOP ICONV_VX_NAME (internal_ucs2_loop) /* This is not used. */
|
|
|
8ae002 |
+#define FUNCTION_NAME ICONV_VX_NAME (__gconv_transform_internal_ucs2)
|
|
|
8ae002 |
+#define ONE_DIRECTION 1
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+#define MIN_NEEDED_INPUT MIN_NEEDED_FROM
|
|
|
8ae002 |
+#define MIN_NEEDED_OUTPUT MIN_NEEDED_TO
|
|
|
8ae002 |
+#define LOOPFCT FROM_LOOP
|
|
|
8ae002 |
+#define BODY_ORIG \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ uint32_t val = *((const uint32_t *) inptr); \
|
|
|
8ae002 |
+ \
|
|
|
8ae002 |
+ if (__glibc_unlikely (val >= 0x10000)) \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ UNICODE_TAG_HANDLER (val, 4); \
|
|
|
8ae002 |
+ STANDARD_TO_LOOP_ERR_HANDLER (4); \
|
|
|
8ae002 |
+ } \
|
|
|
8ae002 |
+ else if (__glibc_unlikely (val >= 0xd800 && val < 0xe000)) \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ /* Surrogate characters in UCS-4 input are not valid. \
|
|
|
8ae002 |
+ We must catch this, because the UCS-2 output might be \
|
|
|
8ae002 |
+ interpreted as UTF-16 by other programs. If we let \
|
|
|
8ae002 |
+ surrogates pass through, attackers could make a security \
|
|
|
8ae002 |
+ hole exploit by synthesizing any desired plane 1-16 \
|
|
|
8ae002 |
+ character. */ \
|
|
|
8ae002 |
+ result = __GCONV_ILLEGAL_INPUT; \
|
|
|
8ae002 |
+ if (! ignore_errors_p ()) \
|
|
|
8ae002 |
+ break; \
|
|
|
8ae002 |
+ inptr += 4; \
|
|
|
8ae002 |
+ ++*irreversible; \
|
|
|
8ae002 |
+ continue; \
|
|
|
8ae002 |
+ } \
|
|
|
8ae002 |
+ else \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ put16 (outptr, val); \
|
|
|
8ae002 |
+ outptr += sizeof (uint16_t); \
|
|
|
8ae002 |
+ inptr += 4; \
|
|
|
8ae002 |
+ } \
|
|
|
8ae002 |
+ }
|
|
|
8ae002 |
+# define BODY \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ if (__builtin_expect (inend - inptr < 32, 1) \
|
|
|
8ae002 |
+ || outend - outptr < 16) \
|
|
|
8ae002 |
+ /* Convert remaining bytes with c code. */ \
|
|
|
8ae002 |
+ BODY_ORIG \
|
|
|
8ae002 |
+ else \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ /* Convert in 32 byte blocks. */ \
|
|
|
8ae002 |
+ size_t loop_count = (inend - inptr) / 32; \
|
|
|
8ae002 |
+ size_t tmp, tmp2; \
|
|
|
8ae002 |
+ if (loop_count > (outend - outptr) / 16) \
|
|
|
8ae002 |
+ loop_count = (outend - outptr) / 16; \
|
|
|
8ae002 |
+ __asm__ volatile (".machine push\n\t" \
|
|
|
8ae002 |
+ ".machine \"z13\"\n\t" \
|
|
|
8ae002 |
+ ".machinemode \"zarch_nohighgprs\"\n\t" \
|
|
|
8ae002 |
+ CONVERT_32BIT_SIZE_T ([R_LI]) \
|
|
|
8ae002 |
+ " larl %[R_I],3f\n\t" \
|
|
|
8ae002 |
+ " vlm %%v20,%%v23,0(%[R_I])\n\t" \
|
|
|
8ae002 |
+ "0: \n\t" \
|
|
|
8ae002 |
+ " vlm %%v16,%%v17,0(%[R_IN])\n\t" \
|
|
|
8ae002 |
+ /* Shorten UCS4 to UCS2. */ \
|
|
|
8ae002 |
+ " vpkf %%v18,%%v16,%%v17\n\t" \
|
|
|
8ae002 |
+ " vstrcfs %%v19,%%v16,%%v20,%%v21\n\t" \
|
|
|
8ae002 |
+ " jno 11f\n\t" \
|
|
|
8ae002 |
+ "1: vstrcfs %%v19,%%v17,%%v20,%%v21\n\t" \
|
|
|
8ae002 |
+ " jno 10f\n\t" \
|
|
|
8ae002 |
+ /* Store 16bytes to buf_out. */ \
|
|
|
8ae002 |
+ "2: vst %%v18,0(%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ " la %[R_IN],32(%[R_IN])\n\t" \
|
|
|
8ae002 |
+ " la %[R_OUT],16(%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ " brctg %[R_LI],0b\n\t" \
|
|
|
8ae002 |
+ " j 20f\n\t" \
|
|
|
8ae002 |
+ /* Setup to check for ch >= 0xd800. (v20, v21) */ \
|
|
|
8ae002 |
+ "3: .long 0xd800,0xd800,0x0,0x0\n\t" \
|
|
|
8ae002 |
+ " .long 0xa0000000,0xa0000000,0x0,0x0\n\t" \
|
|
|
8ae002 |
+ /* Setup to check for ch >= 0xe000 \
|
|
|
8ae002 |
+ && ch < 0x10000. (v22,v23) */ \
|
|
|
8ae002 |
+ " .long 0xe000,0x10000,0x0,0x0\n\t" \
|
|
|
8ae002 |
+ " .long 0xa0000000,0x40000000,0x0,0x0\n\t" \
|
|
|
8ae002 |
+ /* v16 contains only valid chars. Check in v17: \
|
|
|
8ae002 |
+ ch >= 0xe000 && ch <= 0xffff. */ \
|
|
|
8ae002 |
+ "10: vstrcfs %%v19,%%v17,%%v22,%%v23,8\n\t" \
|
|
|
8ae002 |
+ " jo 2b\n\t" /* All ch's in this range, proceed. */ \
|
|
|
8ae002 |
+ " lghi %[R_TMP],16\n\t" \
|
|
|
8ae002 |
+ " j 12f\n\t" \
|
|
|
8ae002 |
+ /* Maybe v16 contains invalid chars. \
|
|
|
8ae002 |
+ Check ch >= 0xe000 && ch <= 0xffff. */ \
|
|
|
8ae002 |
+ "11: vstrcfs %%v19,%%v16,%%v22,%%v23,8\n\t" \
|
|
|
8ae002 |
+ " jo 1b\n\t" /* All ch's in this range, proceed. */ \
|
|
|
8ae002 |
+ " lghi %[R_TMP],0\n\t" \
|
|
|
8ae002 |
+ "12: vlgvb %[R_I],%%v19,7\n\t" \
|
|
|
8ae002 |
+ " agr %[R_I],%[R_TMP]\n\t" \
|
|
|
8ae002 |
+ " la %[R_IN],0(%[R_I],%[R_IN])\n\t" \
|
|
|
8ae002 |
+ " srl %[R_I],1\n\t" \
|
|
|
8ae002 |
+ " ahi %[R_I],-1\n\t" \
|
|
|
8ae002 |
+ " jl 20f\n\t" \
|
|
|
8ae002 |
+ " vstl %%v18,%[R_I],0(%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ " la %[R_OUT],1(%[R_I],%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ "20:\n\t" \
|
|
|
8ae002 |
+ ".machine pop" \
|
|
|
8ae002 |
+ : /* outputs */ [R_OUT] "+a" (outptr) \
|
|
|
8ae002 |
+ , [R_IN] "+a" (inptr) \
|
|
|
8ae002 |
+ , [R_LI] "+d" (loop_count) \
|
|
|
8ae002 |
+ , [R_I] "=a" (tmp2) \
|
|
|
8ae002 |
+ , [R_TMP] "=d" (tmp) \
|
|
|
8ae002 |
+ : /* inputs */ \
|
|
|
8ae002 |
+ : /* clobber list*/ "memory", "cc" \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v16") ASM_CLOBBER_VR ("v17") \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v18") ASM_CLOBBER_VR ("v19") \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v20") ASM_CLOBBER_VR ("v21") \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v22") ASM_CLOBBER_VR ("v23") \
|
|
|
8ae002 |
+ ); \
|
|
|
8ae002 |
+ if (loop_count > 0) \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ /* Found an invalid character at next character. */ \
|
|
|
8ae002 |
+ BODY_ORIG \
|
|
|
8ae002 |
+ } \
|
|
|
8ae002 |
+ } \
|
|
|
8ae002 |
+ }
|
|
|
8ae002 |
+#define LOOP_NEED_FLAGS
|
|
|
8ae002 |
+#include <iconv/loop.c>
|
|
|
8ae002 |
+#include <iconv/skeleton.c>
|
|
|
8ae002 |
+# undef BODY_ORIG
|
|
|
8ae002 |
+ICONV_VX_IFUNC (__gconv_transform_internal_ucs2)
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+/* Convert from the internal (UCS4-like) format to UCS2 in other endianness. */
|
|
|
8ae002 |
+#define DEFINE_INIT 0
|
|
|
8ae002 |
+#define DEFINE_FINI 0
|
|
|
8ae002 |
+#define MIN_NEEDED_FROM 4
|
|
|
8ae002 |
+#define MIN_NEEDED_TO 2
|
|
|
8ae002 |
+#define FROM_DIRECTION 1
|
|
|
8ae002 |
+#define FROM_LOOP ICONV_VX_NAME (internal_ucs2reverse_loop)
|
|
|
8ae002 |
+#define TO_LOOP ICONV_VX_NAME (internal_ucs2reverse_loop)/* This is not used.*/
|
|
|
8ae002 |
+#define FUNCTION_NAME ICONV_VX_NAME (__gconv_transform_internal_ucs2reverse)
|
|
|
8ae002 |
+#define ONE_DIRECTION 1
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+#define MIN_NEEDED_INPUT MIN_NEEDED_FROM
|
|
|
8ae002 |
+#define MIN_NEEDED_OUTPUT MIN_NEEDED_TO
|
|
|
8ae002 |
+#define LOOPFCT FROM_LOOP
|
|
|
8ae002 |
+#define BODY_ORIG \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ uint32_t val = *((const uint32_t *) inptr); \
|
|
|
8ae002 |
+ if (__glibc_unlikely (val >= 0x10000)) \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ UNICODE_TAG_HANDLER (val, 4); \
|
|
|
8ae002 |
+ STANDARD_TO_LOOP_ERR_HANDLER (4); \
|
|
|
8ae002 |
+ } \
|
|
|
8ae002 |
+ else if (__glibc_unlikely (val >= 0xd800 && val < 0xe000)) \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ /* Surrogate characters in UCS-4 input are not valid. \
|
|
|
8ae002 |
+ We must catch this, because the UCS-2 output might be \
|
|
|
8ae002 |
+ interpreted as UTF-16 by other programs. If we let \
|
|
|
8ae002 |
+ surrogates pass through, attackers could make a security \
|
|
|
8ae002 |
+ hole exploit by synthesizing any desired plane 1-16 \
|
|
|
8ae002 |
+ character. */ \
|
|
|
8ae002 |
+ if (! ignore_errors_p ()) \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ result = __GCONV_ILLEGAL_INPUT; \
|
|
|
8ae002 |
+ break; \
|
|
|
8ae002 |
+ } \
|
|
|
8ae002 |
+ inptr += 4; \
|
|
|
8ae002 |
+ ++*irreversible; \
|
|
|
8ae002 |
+ continue; \
|
|
|
8ae002 |
+ } \
|
|
|
8ae002 |
+ else \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ put16 (outptr, bswap_16 (val)); \
|
|
|
8ae002 |
+ outptr += sizeof (uint16_t); \
|
|
|
8ae002 |
+ inptr += 4; \
|
|
|
8ae002 |
+ } \
|
|
|
8ae002 |
+ }
|
|
|
8ae002 |
+# define BODY \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ if (__builtin_expect (inend - inptr < 32, 1) \
|
|
|
8ae002 |
+ || outend - outptr < 16) \
|
|
|
8ae002 |
+ /* Convert remaining bytes with c code. */ \
|
|
|
8ae002 |
+ BODY_ORIG \
|
|
|
8ae002 |
+ else \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ /* Convert in 32 byte blocks. */ \
|
|
|
8ae002 |
+ size_t loop_count = (inend - inptr) / 32; \
|
|
|
8ae002 |
+ size_t tmp, tmp2; \
|
|
|
8ae002 |
+ if (loop_count > (outend - outptr) / 16) \
|
|
|
8ae002 |
+ loop_count = (outend - outptr) / 16; \
|
|
|
8ae002 |
+ __asm__ volatile (".machine push\n\t" \
|
|
|
8ae002 |
+ ".machine \"z13\"\n\t" \
|
|
|
8ae002 |
+ ".machinemode \"zarch_nohighgprs\"\n\t" \
|
|
|
8ae002 |
+ CONVERT_32BIT_SIZE_T ([R_LI]) \
|
|
|
8ae002 |
+ " larl %[R_I],3f\n\t" \
|
|
|
8ae002 |
+ " vlm %%v20,%%v24,0(%[R_I])\n\t" \
|
|
|
8ae002 |
+ "0: \n\t" \
|
|
|
8ae002 |
+ " vlm %%v16,%%v17,0(%[R_IN])\n\t" \
|
|
|
8ae002 |
+ /* Shorten UCS4 to UCS2 and byteswap. */ \
|
|
|
8ae002 |
+ " vpkf %%v18,%%v16,%%v17\n\t" \
|
|
|
8ae002 |
+ " vperm %%v18,%%v18,%%v18,%%v24\n\t" \
|
|
|
8ae002 |
+ " vstrcfs %%v19,%%v16,%%v20,%%v21\n\t" \
|
|
|
8ae002 |
+ " jno 11f\n\t" \
|
|
|
8ae002 |
+ "1: vstrcfs %%v19,%%v17,%%v20,%%v21\n\t" \
|
|
|
8ae002 |
+ " jno 10f\n\t" \
|
|
|
8ae002 |
+ /* Store 16bytes to buf_out. */ \
|
|
|
8ae002 |
+ "2: vst %%v18,0(%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ " la %[R_IN],32(%[R_IN])\n\t" \
|
|
|
8ae002 |
+ " la %[R_OUT],16(%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ " brctg %[R_LI],0b\n\t" \
|
|
|
8ae002 |
+ " j 20f\n\t" \
|
|
|
8ae002 |
+ /* Setup to check for ch >= 0xd800. (v20, v21) */ \
|
|
|
8ae002 |
+ "3: .long 0xd800,0xd800,0x0,0x0\n\t" \
|
|
|
8ae002 |
+ " .long 0xa0000000,0xa0000000,0x0,0x0\n\t" \
|
|
|
8ae002 |
+ /* Setup to check for ch >= 0xe000 \
|
|
|
8ae002 |
+ && ch < 0x10000. (v22,v23) */ \
|
|
|
8ae002 |
+ " .long 0xe000,0x10000,0x0,0x0\n\t" \
|
|
|
8ae002 |
+ " .long 0xa0000000,0x40000000,0x0,0x0\n\t" \
|
|
|
8ae002 |
+ /* Vector permute mask (v24) */ \
|
|
|
8ae002 |
+ " .short 0x0100,0x0302,0x0504,0x0706\n\t" \
|
|
|
8ae002 |
+ " .short 0x0908,0x0b0a,0x0d0c,0x0f0e\n\t" \
|
|
|
8ae002 |
+ /* v16 contains only valid chars. Check in v17: \
|
|
|
8ae002 |
+ ch >= 0xe000 && ch <= 0xffff. */ \
|
|
|
8ae002 |
+ "10: vstrcfs %%v19,%%v17,%%v22,%%v23,8\n\t" \
|
|
|
8ae002 |
+ " jo 2b\n\t" /* All ch's in this range, proceed. */ \
|
|
|
8ae002 |
+ " lghi %[R_TMP],16\n\t" \
|
|
|
8ae002 |
+ " j 12f\n\t" \
|
|
|
8ae002 |
+ /* Maybe v16 contains invalid chars. \
|
|
|
8ae002 |
+ Check ch >= 0xe000 && ch <= 0xffff. */ \
|
|
|
8ae002 |
+ "11: vstrcfs %%v19,%%v16,%%v22,%%v23,8\n\t" \
|
|
|
8ae002 |
+ " jo 1b\n\t" /* All ch's in this range, proceed. */ \
|
|
|
8ae002 |
+ " lghi %[R_TMP],0\n\t" \
|
|
|
8ae002 |
+ "12: vlgvb %[R_I],%%v19,7\n\t" \
|
|
|
8ae002 |
+ " agr %[R_I],%[R_TMP]\n\t" \
|
|
|
8ae002 |
+ " la %[R_IN],0(%[R_I],%[R_IN])\n\t" \
|
|
|
8ae002 |
+ " srl %[R_I],1\n\t" \
|
|
|
8ae002 |
+ " ahi %[R_I],-1\n\t" \
|
|
|
8ae002 |
+ " jl 20f\n\t" \
|
|
|
8ae002 |
+ " vstl %%v18,%[R_I],0(%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ " la %[R_OUT],1(%[R_I],%[R_OUT])\n\t" \
|
|
|
8ae002 |
+ "20:\n\t" \
|
|
|
8ae002 |
+ ".machine pop" \
|
|
|
8ae002 |
+ : /* outputs */ [R_OUT] "+a" (outptr) \
|
|
|
8ae002 |
+ , [R_IN] "+a" (inptr) \
|
|
|
8ae002 |
+ , [R_LI] "+d" (loop_count) \
|
|
|
8ae002 |
+ , [R_I] "=a" (tmp2) \
|
|
|
8ae002 |
+ , [R_TMP] "=d" (tmp) \
|
|
|
8ae002 |
+ : /* inputs */ \
|
|
|
8ae002 |
+ : /* clobber list*/ "memory", "cc" \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v16") ASM_CLOBBER_VR ("v17") \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v18") ASM_CLOBBER_VR ("v19") \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v20") ASM_CLOBBER_VR ("v21") \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v22") ASM_CLOBBER_VR ("v23") \
|
|
|
8ae002 |
+ ASM_CLOBBER_VR ("v24") \
|
|
|
8ae002 |
+ ); \
|
|
|
8ae002 |
+ if (loop_count > 0) \
|
|
|
8ae002 |
+ { \
|
|
|
8ae002 |
+ /* Found an invalid character at next character. */ \
|
|
|
8ae002 |
+ BODY_ORIG \
|
|
|
8ae002 |
+ } \
|
|
|
8ae002 |
+ } \
|
|
|
8ae002 |
+ }
|
|
|
8ae002 |
+#define LOOP_NEED_FLAGS
|
|
|
8ae002 |
+#include <iconv/loop.c>
|
|
|
8ae002 |
+#include <iconv/skeleton.c>
|
|
|
8ae002 |
+# undef BODY_ORIG
|
|
|
8ae002 |
+ICONV_VX_IFUNC (__gconv_transform_internal_ucs2reverse)
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+
|
|
|
8ae002 |
+#else
|
|
|
8ae002 |
+/* Generate the internal transformations without ifunc if build environment
|
|
|
8ae002 |
+ lacks vector support. Instead simply include the common version. */
|
|
|
8ae002 |
+# include <iconv/gconv_simple.c>
|
|
|
8ae002 |
+#endif /* !defined HAVE_S390_VX_ASM_SUPPORT */
|
|
|
8ae002 |
--
|
|
|
8ae002 |
1.8.3.1
|
|
|
8ae002 |
|