From b35b582ef3f6575447097585174302fde1761078 Mon Sep 17 00:00:00 2001
From: Nagaraju <nmekala@xilinx.com>
Date: Wed, 24 Apr 2019 23:29:21 +0530
Subject: [PATCH 11/11] Removing the Assembly implementation of 64bit string
 function. Revisit in next release and fix it

---
 newlib/libc/machine/microblaze/mb_endian.h |  4 ++
 newlib/libc/machine/microblaze/strcmp.c    | 93 ++++++++++--------------------
 newlib/libc/machine/microblaze/strcpy.c    | 82 ++++++++------------------
 newlib/libc/machine/microblaze/strlen.c    | 59 +++++++------------
 4 files changed, 81 insertions(+), 157 deletions(-)

diff --git a/newlib/libc/machine/microblaze/mb_endian.h b/newlib/libc/machine/microblaze/mb_endian.h
index fb217ec..17772c8 100644
--- a/newlib/libc/machine/microblaze/mb_endian.h
+++ b/newlib/libc/machine/microblaze/mb_endian.h
@@ -8,8 +8,12 @@
 #ifdef __LITTLE_ENDIAN__
 #define LOAD4BYTES(rD,rA,rB)   "\tlwr\t" rD ", " rA ", " rB "\n"
 #define STORE4BYTES(rD,rA,rB)  "\tswr\t" rD ", " rA ", " rB "\n"
+#define LOAD8BYTES(rD,rA,rB)   "\tllr\t" rD ", " rA ", " rB "\n"
+#define STORE8BYTES(rD,rA,rB)  "\tslr\t" rD ", " rA ", " rB "\n"
 #else
 #define LOAD4BYTES(rD,rA,rB)   "\tlw\t" rD ", " rA ", " rB "\n"
 #define STORE4BYTES(rD,rA,rB)  "\tsw\t" rD ", " rA ", " rB "\n"
+#define LOAD8BYTES(rD,rA,rB)   "\tll\t" rD ", " rA ", " rB "\n"
+#define STORE8BYTES(rD,rA,rB)  "\tsl\t" rD ", " rA ", " rB "\n"
 #endif
 #endif
diff --git a/newlib/libc/machine/microblaze/strcmp.c b/newlib/libc/machine/microblaze/strcmp.c
index acfe4cd..e34c64a 100644
--- a/newlib/libc/machine/microblaze/strcmp.c
+++ b/newlib/libc/machine/microblaze/strcmp.c
@@ -129,70 +129,42 @@ strcmp (const char *s1,
   return (*(unsigned char *) s1) - (*(unsigned char *) s2);
 #endif /* not PREFER_SIZE_OVER_SPEED */
 
+#elif __arch64__
+  unsigned int *a1;
+  unsigned int *a2;
+
+  /* If s1 or s2 are unaligned, then compare bytes. */
+  if (!UNALIGNED (s1, s2))
+    {  
+      /* If s1 and s2 are word-aligned, compare them a word at a time. */
+      a1 = (unsigned int*)s1;
+      a2 = (unsigned int*)s2;
+      while (*a1 == *a2)
+        {
+          /* To get here, *a1 == *a2, thus if we find a null in *a1,
+	     then the strings must be equal, so return zero.  */
+          if (DETECTNULL (*a1))
+	    return 0;
+
+          a1++;
+          a2++;
+        }
+
+      /* A difference was detected in last few bytes of s1, so search bytewise */
+      s1 = (char*)a1;
+      s2 = (char*)a2;
+    }
+
+  while (*s1 != '\0' && *s1 == *s2)
+    {
+      s1++;
+      s2++;
+    }
+  return (*(unsigned char *) s1) - (*(unsigned char *) s2);
 #else
 
 #include "mb_endian.h"
 
-#ifdef __arch64__
-    asm volatile ("                                          \n\
-        orl      r9, r0, r0               /* Index register */ \n\
-check_alignment:                                             \n\
-        andli    r3, r5, 3                                    \n\
-        andli    r4, r6, 3                                    \n\
-        beanei    r3, try_align_args                           \n\
-        beanei    r4, regular_strcmp     /* At this point we don't have a choice */ \n\
-cmp_loop:                                                                       \n"
-        LOAD4BYTES("r3", "r5", "r9")
-        LOAD4BYTES("r4", "r6", "r9")
-"                                                                                      \n\
-        pcmplbf  r7, r3, r0              /* See if there is Null byte */                         \n\
-        beanei    r7, end_cmp_loop        /* IF yes (r7 > 0) use byte compares in end_cmp_loop */ \n\
-        cmplu    r7, r4, r3              /* ELSE compare whole word */                   \n\
-        beanei    r7, end_cmp                                                             \n\
-        addlik   r9, r9, 4               /* delay slot */                                \n\
-        breaid    cmp_loop                                                                \n\
-        nop		               /* delay slot */                                \n\
-end_cmp_loop:                                                                           \n\
-        lbu     r3, r5, r9              /* byte compare loop */                         \n\
-        lbu     r4, r6, r9                                                              \n\
-        cmplu    r7, r4, r3              /* Compare bytes */                             \n\
-        beanei    r7, end_cmp_early                                                       \n\
-        addlik   r9, r9, 1               /* delay slot */                        \n\
-        beaneid   r3, end_cmp_loop        /* If reached null on one string, terminate */  \n\
-        nop					                              \n\
-end_cmp_early:                                                                  \n\
-        orl   r3, r0, r7               /* delay slot */                        \n\
-        rtsd    r15, 8                                                          \n\
-        nop		                                                         \n\
-try_align_args:                                                                 \n\
-        xorl     r7, r4, r3                                                      \n\
-        beanei    r7, regular_strcmp      /* cannot align args */                 \n\
-        rsublik  r10, r3, 4              /* Number of initial bytes to align */  \n\
-align_loop:                                                                     \n\
-        lbu     r3, r5, r9                                                      \n\
-        lbu     r4, r6, r9                                                      \n\
-        cmplu    r7, r4, r3                                                      \n\
-        beanei    r7, end_cmp                                                     \n\
-        beaeqi    r3, end_cmp                                                     \n\
-        addlik   r10, r10, -1                                                    \n\
-        addlik   r9, r9, 1                                                       \n\
-        beaeqid   r10, cmp_loop                                                   \n\
-        nop		                                                           \n\
-        breai     align_loop                                                      \n\
-regular_strcmp:                                                                 \n\
-        lbu     r3, r5, r9                                                      \n\
-        lbu     r4, r6, r9                                                      \n\
-        cmplu    r7, r4, r3                                                      \n\
-        beanei    r7, end_cmp                                                     \n\
-        beaeqi    r3, end_cmp                                                     \n\
-        addlik   r9, r9, 1                                                       \n\
-        breaid    regular_strcmp                                                  \n\
-        nop                                                  \n\
-end_cmp:                                                                        \n\
-        orl       r3, r0, r7                                                       \n\
-        rtsd    r15, 8                                                          \n\
-        nop              /* Return strcmp result */");
-#else
     asm volatile ("                                          \n\
         or      r9, r0, r0               /* Index register */\n\
 check_alignment:                                             \n\
@@ -246,7 +218,6 @@ end_cmp:
         rtsd    r15, 8                                                          \n\
         or      r3, r0, r7              /* Return strcmp result */");
 
-#endif
 #endif /* ! HAVE_HW_PCMP */
 }
 
diff --git a/newlib/libc/machine/microblaze/strcpy.c b/newlib/libc/machine/microblaze/strcpy.c
index 6dbc60d..ddb6922 100644
--- a/newlib/libc/machine/microblaze/strcpy.c
+++ b/newlib/libc/machine/microblaze/strcpy.c
@@ -121,67 +121,36 @@ strcpy (char *__restrict dst0,
     ;
   return dst0;
 #endif /* not PREFER_SIZE_OVER_SPEED */
+#elif __arch64__
+  char *dst = dst0;
+  const char *src = src0;
+  long *aligned_dst;
+  const long *aligned_src;
 
-#else    
+  /* If SRC or DEST is unaligned, then copy bytes.  */
+  if (!UNALIGNED (src, dst))
+    {
+      aligned_dst = (long*)dst;
+      aligned_src = (long*)src;
 
-#include "mb_endian.h"
-#ifdef __arch64__
+      /* SRC and DEST are both "long int" aligned, try to do "long int"
+         sized copies.  */
+      while (!DETECTNULL(*aligned_src))
+        {
+          *aligned_dst++ = *aligned_src++;
+        }
 
-  asm volatile ("                                                   \n\
-        orl      r9, r0, r0              /* Index register */        \n\
-check_alignment:                                                    \n\
-        andli    r3, r5, 3                                           \n\
-        andli    r4, r6, 3                                           \n\
-        beanei    r3, try_align_args                                  \n\
-        beanei    r4, regular_strcpy      /* At this point we dont have a choice */       \n\
-cpy_loop:                                   \n"
-        LOAD4BYTES("r3", "r6", "r9")
-"                                           \n\
-        pcmplbf  r4, r0, r3                  \n\
-        beanei    r4, cpy_bytes           /* If r4 != 0, then null present within string */\n"
-        STORE4BYTES("r3", "r5", "r9")
-"                                           \n\
-        addlik   r9, r9, 4                   \n\
-        breaid    cpy_loop                    \n\
-	nop					\n\
-cpy_bytes:                                  \n\
-        lbu     r3, r6, r9                  \n\
-        sb      r3, r5, r9                  \n\
-        addlik   r4, r4, -1                  \n\
-        addlik   r9, r9, 1               /* delay slot */\n\
-        beaneid   r4, cpy_bytes               \n\
-	nop					\n\
-cpy_null:                                   \n\
-        orl      r3, r0, r5              /* Return strcpy result */\n\
-        rtsd    r15, 8                      \n\
-	nop					\n\
-try_align_args:                             \n\
-        xorl     r7, r4, r3                  \n\
-        beanei    r7, regular_strcpy      /* cannot align args */\n\
-        rsublik  r10, r3, 4              /* Number of initial bytes to align */\n\
-align_loop:                                 \n\
-        lbu     r3, r6, r9                  \n\
-        sb      r3, r5, r9                  \n\
-        addlik   r10, r10, -1                \n\
-        beaeqid   r3, end_cpy             /* Break if we have seen null character */\n\
-	nop					\n\
-        addlik   r9, r9, 1                   \n\
-        beaneid   r10, align_loop             \n\
-	nop					\n\
-        breai     cpy_loop                    \n\
-regular_strcpy:                             \n\
-        lbu     r3, r6, r9                  \n\
-        sb      r3, r5, r9                  \n\
-        addlik   r9, r9, 1                   \n\
-        beaneid   r3, regular_strcpy          \n\
-	nop					\n\
-end_cpy:                                    \n\
-        orl      r3, r0, r5  			\n\
-        rtsd    r15, 8                      \n\
-	nop			 /* Return strcpy result */");
+      dst = (char*)aligned_dst;
+      src = (char*)aligned_src;
+    }
 
-#else
+  while (*dst++ = *src++)
+    ;
+  return dst0;
+
+#else    
 
+#include "mb_endian.h"
   asm volatile ("                                                   \n\
         or      r9, r0, r0              /* Index register */        \n\
 check_alignment:                                                    \n\
@@ -227,7 +196,6 @@ regular_strcpy:                             \n\
 end_cpy:                                    \n\
         rtsd    r15, 8                      \n\
         or      r3, r0, r5              /* Return strcpy result */");
-#endif
 #endif /* ! HAVE_HW_PCMP */
 }
 
diff --git a/newlib/libc/machine/microblaze/strlen.c b/newlib/libc/machine/microblaze/strlen.c
index b6f2d3c..9407539 100644
--- a/newlib/libc/machine/microblaze/strlen.c
+++ b/newlib/libc/machine/microblaze/strlen.c
@@ -112,47 +112,29 @@ strlen (const char *str)
   return str - start;
 #endif /* not PREFER_SIZE_OVER_SPEED */
 
-#else
-
-#include "mb_endian.h"
+#elif __arch64__
+  const char *start = str;
+  unsigned long *aligned_addr;
 
-#ifdef __arch64__
-  asm volatile ("                                               \n\
-        orl      r9, r0, r0              /* Index register */    \n\
-check_alignment:                                                \n\
-        andli    r3, r5, 3                                       \n\
-        beanei    r3, align_arg                                   \n\
-len_loop:                                                       \n"
-        LOAD4BYTES("r3", "r5", "r9")
-"                                                               \n\
-        pcmplbf  r4, r3, r0                                      \n\
-        beanei    r4, end_len                                     \n\
-        addlik   r9, r9, 4                                       \n\
-        breaid    len_loop                                        \n\
-        nop		                                        \n\
-end_len:                                                        \n\
-        lbu     r3, r5, r9                                      \n\
-        beaeqi    r3, done_len                                    \n\
-        addlik   r9, r9, 1                                       \n\
-        breaid    end_len                                         \n\
-        nop		                                       \n\
-done_len:                                                       \n\
-        orl      r3, r0, r9              /* Return len */        \n\
-        rtsd    r15, 8                                          \n\
-        nop						        \n\
-align_arg:                                                      \n\
-        rsublik  r10, r3, 4                                      \n\
-align_loop:                                                     \n\
-        lbu     r3, r5, r9                                      \n\
-        addlik   r10, r10, -1                                    \n\
-        beaeqid   r3, done_len                                    \n\
-        nop		                                       \n\
-        addlik   r9, r9, 1                                       \n\
-        beaneid   r10, align_loop                                 \n\
-        nop		                                        \n\
-        breai     len_loop");
+  if (!UNALIGNED (str))
+    {
+      /* If the string is word-aligned, we can check for the presence of 
+         a null in each word-sized block.  */
+      aligned_addr = (unsigned long*)str;
+      while (!DETECTNULL (*aligned_addr))
+        aligned_addr++;
 
+      /* Once a null is detected, we check each byte in that block for a
+         precise position of the null.  */
+      str = (char*)aligned_addr;
+    }
+ 
+  while (*str)
+    str++;
+  return str - start;
 #else
+
+#include "mb_endian.h"
   asm volatile ("                                               \n\
         or      r9, r0, r0              /* Index register */    \n\
 check_alignment:                                                \n\
@@ -183,6 +165,5 @@ align_loop:                                                     \n\
         addik   r9, r9, 1                                       \n\
         bri     len_loop");
 
-#endif
 #endif  /* ! HAVE_HW_PCMP */
 }
-- 
2.7.4