00db10
From 832572eac8a661d25efe0f2bcc6a861e2c29c3b8 Mon Sep 17 00:00:00 2001
00db10
From: Stefan Liebler <stli@linux.vnet.ibm.com>
00db10
Date: Mon, 7 Nov 2016 15:50:46 +0100
00db10
Subject: [PATCH 05/17] S390: Optimize iso-8859-1 to ibm037 iconv-module.
00db10
00db10
Upstream commit 81c6380887c6d62c56e5f0f85a241f759f58b2fd
00db10
00db10
This patch reworks the s390 specific module which used the z900
00db10
translate one to one instruction. Now the g5 translate instruction is used,
00db10
because it outperforms the troo instruction.
00db10
00db10
ChangeLog:
00db10
00db10
	* sysdeps/s390/s390-64/iso-8859-1_cp037_z900.c (TROO_LOOP):
00db10
	Rename to TR_LOOP and usage of tr instead of troo instruction.
00db10
---
00db10
 sysdeps/s390/s390-64/iso-8859-1_cp037_z900.c | 95 +++++++++++++++++-----------
00db10
 1 file changed, 57 insertions(+), 38 deletions(-)
00db10
00db10
diff --git a/sysdeps/s390/s390-64/iso-8859-1_cp037_z900.c b/sysdeps/s390/s390-64/iso-8859-1_cp037_z900.c
00db10
index 58641f5..3b63e6a 100644
00db10
--- a/sysdeps/s390/s390-64/iso-8859-1_cp037_z900.c
00db10
+++ b/sysdeps/s390/s390-64/iso-8859-1_cp037_z900.c
00db10
@@ -1,8 +1,7 @@
00db10
 /* Conversion between ISO 8859-1 and IBM037.
00db10
 
00db10
-   This module uses the Z900 variant of the Translate One To One
00db10
-   instruction.
00db10
-   Copyright (C) 1997-2009 Free Software Foundation, Inc.
00db10
+   This module uses the translate instruction.
00db10
+   Copyright (C) 1997-2016 Free Software Foundation, Inc.
00db10
 
00db10
    Author: Andreas Krebbel  <Andreas.Krebbel@de.ibm.com>
00db10
    Based on the work by Ulrich Drepper  <drepper@cygnus.com>, 1997.
00db10
@@ -176,50 +175,70 @@ __attribute__ ((aligned (8))) =
00db10
 #define MIN_NEEDED_FROM		1
00db10
 #define MIN_NEEDED_TO		1
00db10
 
00db10
-/* The Z900 variant of troo forces us to always specify a test
00db10
-   character which ends the translation.  So if we run into the
00db10
-   situation where the translation has been interrupted due to the
00db10
-   test character we translate the character by hand and jump back
00db10
-   into the instruction.  */
00db10
-
00db10
-#define TROO_LOOP(TABLE)						\
00db10
+#define TR_LOOP(TABLE)							\
00db10
   {									\
00db10
-    register const unsigned char test asm ("0") = 0;			\
00db10
-    register const unsigned char *pTable asm ("1") = TABLE;		\
00db10
-    register unsigned char *pOutput asm ("2") = outptr;			\
00db10
-    register uint64_t length asm ("3");					\
00db10
-    const unsigned char* pInput = inptr;				\
00db10
-    uint64_t tmp;							\
00db10
-									\
00db10
-    length = (inend - inptr < outend - outptr				\
00db10
-	      ? inend - inptr : outend - outptr);			\
00db10
+    size_t length = (inend - inptr < outend - outptr			\
00db10
+		     ? inend - inptr : outend - outptr);		\
00db10
 									\
00db10
-    asm volatile ("0:                        \n\t"			\
00db10
-                  "  troo    %0,%1           \n\t"			\
00db10
-                  "  jz      1f              \n\t"			\
00db10
-                  "  jo      0b              \n\t"			\
00db10
-                  "  llgc    %3,0(%1)        \n\t"			\
00db10
-                  "  la      %3,0(%3,%4)     \n\t"			\
00db10
-                  "  mvc     0(1,%0),0(%3)   \n\t"			\
00db10
-                  "  aghi    %1,1            \n\t"			\
00db10
-                  "  aghi    %0,1            \n\t"			\
00db10
-                  "  aghi    %2,-1           \n\t"			\
00db10
-                  "  j       0b              \n\t"			\
00db10
-                  "1:                        \n"			\
00db10
+    /* Process in 256 byte blocks.  */					\
00db10
+    if (__builtin_expect (length >= 256, 0))				\
00db10
+      {									\
00db10
+	size_t blocks = length / 256;					\
00db10
+	__asm__ __volatile__("0: mvc 0(256,%[R_OUT]),0(%[R_IN])\n\t"	\
00db10
+			     "   tr 0(256,%[R_OUT]),0(%[R_TBL])\n\t"	\
00db10
+			     "   la %[R_IN],256(%[R_IN])\n\t"		\
00db10
+			     "   la %[R_OUT],256(%[R_OUT])\n\t"		\
00db10
+			     "   brctg %[R_LI],0b\n\t"			\
00db10
+			     : /* outputs */ [R_IN] "+a" (inptr)	\
00db10
+			       , [R_OUT] "+a" (outptr), [R_LI] "+d" (blocks) \
00db10
+			     : /* inputs */ [R_TBL] "a" (TABLE)		\
00db10
+			     : /* clobber list */ "memory"		\
00db10
+			     );						\
00db10
+	length = length % 256;						\
00db10
+      }									\
00db10
 									\
00db10
-     : "+a" (pOutput), "+a" (pInput), "+d" (length), "=&a" (tmp)        \
00db10
-     : "a" (pTable), "d" (test)						\
00db10
-     : "cc");								\
00db10
+    /* Process remaining 0...248 bytes in 8byte blocks.  */		\
00db10
+    if (length >= 8)							\
00db10
+      {									\
00db10
+	size_t blocks = length / 8;					\
00db10
+	for (int i = 0; i < blocks; i++)				\
00db10
+	  {								\
00db10
+	    outptr[0] = TABLE[inptr[0]];				\
00db10
+	    outptr[1] = TABLE[inptr[1]];				\
00db10
+	    outptr[2] = TABLE[inptr[2]];				\
00db10
+	    outptr[3] = TABLE[inptr[3]];				\
00db10
+	    outptr[4] = TABLE[inptr[4]];				\
00db10
+	    outptr[5] = TABLE[inptr[5]];				\
00db10
+	    outptr[6] = TABLE[inptr[6]];				\
00db10
+	    outptr[7] = TABLE[inptr[7]];				\
00db10
+	    inptr += 8;							\
00db10
+	    outptr += 8;						\
00db10
+	  }								\
00db10
+	length = length % 8;						\
00db10
+      }									\
00db10
 									\
00db10
-    inptr = pInput;							\
00db10
-    outptr = pOutput;							\
00db10
+    /* Process remaining 0...7 bytes.  */				\
00db10
+    switch (length)							\
00db10
+      {									\
00db10
+      case 7: outptr[6] = TABLE[inptr[6]];				\
00db10
+      case 6: outptr[5] = TABLE[inptr[5]];				\
00db10
+      case 5: outptr[4] = TABLE[inptr[4]];				\
00db10
+      case 4: outptr[3] = TABLE[inptr[3]];				\
00db10
+      case 3: outptr[2] = TABLE[inptr[2]];				\
00db10
+      case 2: outptr[1] = TABLE[inptr[1]];				\
00db10
+      case 1: outptr[0] = TABLE[inptr[0]];				\
00db10
+      case 0: break;							\
00db10
+      }									\
00db10
+    inptr += length;							\
00db10
+    outptr += length;							\
00db10
   }
00db10
 
00db10
+
00db10
 /* First define the conversion function from ISO 8859-1 to CP037.  */
00db10
 #define MIN_NEEDED_INPUT	MIN_NEEDED_FROM
00db10
 #define MIN_NEEDED_OUTPUT	MIN_NEEDED_TO
00db10
 #define LOOPFCT			FROM_LOOP
00db10
-#define BODY TROO_LOOP (table_iso8859_1_to_cp037)
00db10
+#define BODY			TR_LOOP (table_iso8859_1_to_cp037)
00db10
 
00db10
 #include <iconv/loop.c>
00db10
 
00db10
@@ -228,7 +247,7 @@ __attribute__ ((aligned (8))) =
00db10
 #define MIN_NEEDED_INPUT	MIN_NEEDED_TO
00db10
 #define MIN_NEEDED_OUTPUT	MIN_NEEDED_FROM
00db10
 #define LOOPFCT			TO_LOOP
00db10
-#define BODY TROO_LOOP (table_cp037_iso8859_1);
00db10
+#define BODY			TR_LOOP (table_cp037_iso8859_1);
00db10
 
00db10
 #include <iconv/loop.c>
00db10
 
00db10
-- 
00db10
1.8.3.1
00db10