* libc/machine/arm/strcmp-arm-tiny.S: New file.

	* libc/machine/arm/strcmp-armv4.S: New file.
	* libc/machine/arm/strcmp-armv4t.S: New file.
	* libc/machine/arm/strcmp-armv6.S: New file.
	* libc/machine/arm/strcmp-armv7.S: New file.
	* libc/machine/arm/strcmp-armv7m.S: New file.
	* libc/machine/arm/strcmp.S: Replace with wrapper for various
	implementations.
	* libc/machine/arm/Makefile.am (strcmp.o, strcmp.obj): Add
	dependencies.
	* libc/machine/arm/Makefile.in: Regenerated.

1 files changed, 36 insertions, 741 deletions

diff --git a/newlib/libc/machine/arm/strcmp.S b/newlib/libc/machine/arm/strcmp.S
index f3e738776..1742322ee 100644
--- a/newlib/libc/machine/arm/strcmp.S
+++ b/newlib/libc/machine/arm/strcmp.S
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2012 ARM Ltd
+ * Copyright (c) 2012-2014 ARM Ltd
  * All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
@@ -26,769 +26,64 @@
  * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  */
 
+/* Wrapper for the various implementations of strcmp.  */
+
 #include "arm_asm.h"
 
-#ifdef __ARMEB__
-#define S2LOMEM lsl
-#define S2LOMEMEQ lsleq
-#define S2HIMEM lsr
+#ifdef __ARM_BIG_ENDIAN
+#define S2LO lsl
+#define S2LOEQ lsleq
+#define S2HI lsr
 #define MSB 0x000000ff
 #define LSB 0xff000000
 #define BYTE0_OFFSET 24
 #define BYTE1_OFFSET 16
 #define BYTE2_OFFSET 8
 #define BYTE3_OFFSET 0
-#else /* not  __ARMEB__ */
-#define S2LOMEM lsr
-#define S2LOMEMEQ lsreq
-#define S2HIMEM lsl
+#else /* not  __ARM_BIG_ENDIAN */
+#define S2LO lsr
+#define S2LOEQ lsreq
+#define S2HI lsl
 #define BYTE0_OFFSET 0
 #define BYTE1_OFFSET 8
 #define BYTE2_OFFSET 16
 #define BYTE3_OFFSET 24
 #define MSB 0xff000000
 #define LSB 0x000000ff
-#endif /* not  __ARMEB__ */
-
-.syntax         unified
-
-#if defined (__thumb__)
-        .thumb
-        .thumb_func
-#if !defined (__thumb2__)
-	/* If we have thumb1 only, we need to explictly mark the
-	   compatibility.  */
-	.arch armv4t
-	.eabi_attribute Tag_also_compatible_with, "\006\013"  /* v6-M.  */
-	.eabi_attribute Tag_ARM_ISA_use, 0
-#endif
-#endif
-        .global strcmp
-        .type   strcmp, %function
-strcmp:
-
-#if (defined (__thumb__) && !defined (__thumb2__))
-1:
-        ldrb    r2, [r0]
-        ldrb    r3, [r1]
-        adds    r0, r0, #1
-        adds    r1, r1, #1
-        cmp     r2, #0
-        beq     2f
-        cmp     r2, r3
-        beq     1b
-2:
-        subs    r0, r2, r3
-        bx      lr
-#elif (defined (__OPTIMIZE_SIZE__) || defined (PREFER_SIZE_OVER_SPEED))
-1:
-        ldrb    r2, [r0], #1
-        ldrb    r3, [r1], #1
-        cmp     r2, #1
-        it      cs
-        cmpcs   r2, r3
-        beq     1b
-        subs    r0, r2, r3
-        RETURN
-
-
-#elif (defined (_ISA_THUMB_2) || defined (_ISA_ARM_6))
-      /* Use LDRD whenever possible.  */
-
-/* The main thing to look out for when comparing large blocks is that
-   the loads do not cross a page boundary when loading past the index
-   of the byte with the first difference or the first string-terminator.
-
-   For example, if the strings are identical and the string-terminator
-   is at index k, byte by byte comparison will not load beyond address
-   s1+k and s2+k; word by word comparison may load up to 3 bytes beyond
-   k; double word - up to 7 bytes.  If the load of these bytes crosses
-   a page boundary, it might cause a memory fault (if the page is not mapped)
-   that would not have happened in byte by byte comparison.
-
-   If an address is (double) word aligned, then a load of a (double) word
-   from that address will not cross a page boundary.
-   Therefore, the algorithm below considers word and double-word alignment
-   of strings separately.  */
-
-/* High-level description of the algorithm.
-
-   * The fast path: if both strings are double-word aligned,
-     use LDRD to load two words from each string in every loop iteration.
-   * If the strings have the same offset from a word boundary,
-     use LDRB to load and compare byte by byte until
-     the first string is aligned to a word boundary (at most 3 bytes).
-     This is optimized for quick return on short unaligned strings.
-   * If the strings have the same offset from a double-word boundary,
-     use LDRD to load two words from each string in every loop iteration, as in the fast path.
-   * If the strings do not have the same offset from a double-word boundary,
-     load a word from the second string before the loop to initialize the queue.
-     Use LDRD to load two words from every string in every loop iteration.
-     Inside the loop, load the second word from the second string only after comparing
-     the first word, using the queued value, to guarantee safety across page boundaries.
-   * If the strings do not have the same offset from a word boundary,
-     use LDR and a shift queue. Order of loads and comparisons matters,
-     similarly to the previous case.
-
-   * Use UADD8 and SEL to compare words, and use REV and CLZ to compute the return value.
-   * The only difference between ARM and Thumb modes is the use of CBZ instruction.
-   * The only difference between big and little endian is the use of REV in little endian
-     to compute the return value, instead of MOV.
-   * No preload. [TODO.]
-*/
-
-        .macro m_cbz reg label
-#ifdef __thumb2__
-        cbz     \reg, \label
-#else   /* not defined __thumb2__ */
-        cmp     \reg, #0
-        beq     \label
-#endif /* not defined __thumb2__ */
-        .endm /* m_cbz */
-
-        .macro m_cbnz reg label
-#ifdef __thumb2__
-        cbnz    \reg, \label
-#else   /* not defined __thumb2__ */
-        cmp     \reg, #0
-        bne     \label
-#endif /* not defined __thumb2__ */
-        .endm /* m_cbnz */
-
-        .macro  init
-        /* Macro to save temporary registers and prepare magic values.  */
-        subs    sp, sp, #16
-        strd    r4, r5, [sp, #8]
-        strd    r6, r7, [sp]
-        mvn     r6, #0  /* all F */
-        mov     r7, #0  /* all 0 */
-        .endm   /* init */
-
-        .macro  magic_compare_and_branch w1 w2 label
-        /* Macro to compare registers w1 and w2 and conditionally branch to label.  */
-        cmp     \w1, \w2        /* Are w1 and w2 the same?  */
-        magic_find_zero_bytes \w1
-        it      eq
-        cmpeq   ip, #0          /* Is there a zero byte in w1?  */
-        bne     \label
-        .endm /* magic_compare_and_branch */
-
-        .macro  magic_find_zero_bytes w1
-        /* Macro to find all-zero bytes in w1, result is in ip.  */
-#if (defined (__ARM_FEATURE_DSP))
-        uadd8   ip, \w1, r6
-        sel     ip, r7, r6
-#else /* not defined (__ARM_FEATURE_DSP) */
-        /* __ARM_FEATURE_DSP is not defined for some Cortex-M processors.
-        Coincidently, these processors only have Thumb-2 mode, where we can use the
-        the (large) magic constant available directly as an immediate in instructions.
-        Note that we cannot use the magic constant in ARM mode, where we need
-        to create the constant in a register.  */
-        sub     ip, \w1, #0x01010101
-        bic     ip, ip, \w1
-        and     ip, ip, #0x80808080
-#endif /* not defined (__ARM_FEATURE_DSP) */
-        .endm /* magic_find_zero_bytes */
-
-        .macro  setup_return w1 w2
-#ifdef __ARMEB__
-        mov     r1, \w1
-        mov     r2, \w2
-#else /* not  __ARMEB__ */
-        rev     r1, \w1
-        rev     r2, \w2
-#endif /* not  __ARMEB__ */
-        .endm /* setup_return */
-
-        /*
-        optpld r0, #0
-        optpld r1, #0
-        */
-
-        /* Are both strings double-word aligned?  */
-        orr     ip, r0, r1
-        tst     ip, #7
-        bne     .Ldo_align
-
-        /* Fast path.  */
-        init
-
-.Ldoubleword_aligned:
-
-        /* Get here when the strings to compare are double-word aligned.  */
-        /* Compare two words in every iteration.  */
-        .p2align        2
-2:
-        /*
-        optpld r0, #16
-        optpld r1, #16
-        */
-
-        /* Load the next double-word from each string.  */
-        ldrd    r2, r3, [r0], #8
-        ldrd    r4, r5, [r1], #8
-
-        magic_compare_and_branch w1=r2, w2=r4, label=.Lreturn_24
-        magic_compare_and_branch w1=r3, w2=r5, label=.Lreturn_35
-        b       2b
-
-.Ldo_align:
-        /* Is the first string word-aligned?  */
-        ands    ip, r0, #3
-        beq     .Lword_aligned_r0
-
-        /* Fast compare byte by byte until the first string is word-aligned.  */
-        /* The offset of r0 from a word boundary is in ip. Thus, the number of bytes
-        to read until the next word boudnary is 4-ip.  */
-        bic     r0, r0, #3
-        ldr     r2, [r0], #4
-        lsls    ip, ip, #31
-        beq     .Lbyte2
-        bcs     .Lbyte3
-
-.Lbyte1:
-        ldrb    ip, [r1], #1
-        uxtb    r3, r2, ror #BYTE1_OFFSET
-        subs    ip, r3, ip
-        bne     .Lfast_return
-        m_cbz   reg=r3, label=.Lfast_return
-
-.Lbyte2:
-        ldrb    ip, [r1], #1
-        uxtb    r3, r2, ror #BYTE2_OFFSET
-        subs    ip, r3, ip
-        bne     .Lfast_return
-        m_cbz   reg=r3, label=.Lfast_return
-
-.Lbyte3:
-        ldrb    ip, [r1], #1
-        uxtb    r3, r2, ror #BYTE3_OFFSET
-        subs    ip, r3, ip
-        bne     .Lfast_return
-        m_cbnz  reg=r3, label=.Lword_aligned_r0
-
-.Lfast_return:
-        mov     r0, ip
-        bx      lr
-
-.Lword_aligned_r0:
-        init
-        /* The first string is word-aligned.  */
-        /* Is the second string word-aligned?  */
-        ands    ip, r1, #3
-        bne     .Lstrcmp_unaligned
-
-.Lword_aligned:
-        /* The strings are word-aligned. */
-        /* Is the first string double-word aligned?  */
-        tst     r0, #4
-        beq     .Ldoubleword_aligned_r0
-
-        /* If r0 is not double-word aligned yet, align it by loading
-        and comparing the next word from each string.  */
-        ldr     r2, [r0], #4
-        ldr     r4, [r1], #4
-        magic_compare_and_branch w1=r2 w2=r4 label=.Lreturn_24
-
-.Ldoubleword_aligned_r0:
-        /* Get here when r0 is double-word aligned.  */
-        /* Is r1 doubleword_aligned?  */
-        tst     r1, #4
-        beq     .Ldoubleword_aligned
-
-        /* Get here when the strings to compare are word-aligned,
-        r0 is double-word aligned, but r1 is not double-word aligned.  */
-
-        /* Initialize the queue.  */
-        ldr     r5, [r1], #4
-
-        /* Compare two words in every iteration.  */
-        .p2align        2
-3:
-        /*
-        optpld r0, #16
-        optpld r1, #16
-        */
-
-        /* Load the next double-word from each string and compare.  */
-        ldrd    r2, r3, [r0], #8
-        magic_compare_and_branch w1=r2 w2=r5 label=.Lreturn_25
-        ldrd    r4, r5, [r1], #8
-        magic_compare_and_branch w1=r3 w2=r4 label=.Lreturn_34
-        b       3b
+#endif /* not  __ARM_BIG_ENDIAN */
 
-        .macro miscmp_word offsetlo offsethi
-        /* Macro to compare misaligned strings.  */
-        /* r0, r1 are word-aligned, and at least one of the strings
-        is not double-word aligned.  */
-        /* Compare one word in every loop iteration.  */
-        /* OFFSETLO is the original bit-offset of r1 from a word-boundary,
-        OFFSETHI is 32 - OFFSETLO (i.e., offset from the next word).  */
+	.macro def_fn f p2align=0
+	.text
+	.p2align \p2align
+	.global \f
+	.type \f, %function
+\f:
+	.endm
 
-        /* Initialize the shift queue.  */
-        ldr     r5, [r1], #4
+#if defined (__OPTIMIZE_SIZE__) || defined (PREFER_SIZE_OVER_SPEED) \
+	|| (__ARM_ARCH == 6 && __ARM_ARCH_PROFILE == 'M')
 
-        /* Compare one word from each string in every loop iteration.  */
-        .p2align        2
-7:
-        ldr     r3, [r0], #4
-        S2LOMEM r5, r5, #\offsetlo
-        magic_find_zero_bytes w1=r3
-        cmp     r7, ip, S2HIMEM #\offsetlo
-        and     r2, r3, r6, S2LOMEM #\offsetlo
-        it      eq
-        cmpeq   r2, r5
-        bne     .Lreturn_25
-        ldr     r5, [r1], #4
-        cmp     ip, #0
-        eor	r3, r2, r3
-        S2HIMEM r2, r5, #\offsethi
-        it      eq
-        cmpeq   r3, r2
-        bne     .Lreturn_32
-        b       7b
-        .endm /* miscmp_word */
+# if defined (__thumb__) && !defined (__thumb2__)
+/* Thumb1 only variant.  */
+#  include "strcmp-armv4t.S"
+# else
+#  include "strcmp-arm-tiny.S"
+# endif
 
-.Lstrcmp_unaligned:
-        /* r0 is word-aligned, r1 is at offset ip from a word.  */
-        /* Align r1 to the (previous) word-boundary.  */
-        bic     r1, r1, #3
+#elif __ARM_ARCH >= 7
 
-        /* Unaligned comparison word by word using LDRs. */
-        cmp     ip, #2
-        beq     .Lmiscmp_word_16                    /* If ip == 2.  */
-        bge     .Lmiscmp_word_24                    /* If ip == 3.  */
-        miscmp_word offsetlo=8 offsethi=24        /* If ip == 1.  */
-.Lmiscmp_word_16:  miscmp_word offsetlo=16 offsethi=16
-.Lmiscmp_word_24:  miscmp_word offsetlo=24 offsethi=8
+# ifdef __ARM_FEATURE_SIMD32
+#  include "strcmp-armv7.S"
+# else
+#  include "strcmp-armv7m.S"
+# endif
 
+#elif __ARM_ARCH >= 6
 
-.Lreturn_32:
-        setup_return w1=r3, w2=r2
-        b       .Ldo_return
-.Lreturn_34:
-        setup_return w1=r3, w2=r4
-        b       .Ldo_return
-.Lreturn_25:
-        setup_return w1=r2, w2=r5
-        b       .Ldo_return
-.Lreturn_35:
-        setup_return w1=r3, w2=r5
-        b       .Ldo_return
-.Lreturn_24:
-        setup_return w1=r2, w2=r4
+# include "strcmp-armv6.S"
 
-.Ldo_return:
-
-#ifdef __ARMEB__
-        mov     r0, ip
-#else /* not  __ARMEB__ */
-        rev     r0, ip
-#endif /* not  __ARMEB__ */
-
-        /* Restore temporaries early, before computing the return value.  */
-        ldrd    r6, r7, [sp]
-        ldrd    r4, r5, [sp, #8]
-        adds    sp, sp, #16
-
-        /* There is a zero or a different byte between r1 and r2.  */
-        /* r0 contains a mask of all-zero bytes in r1.  */
-        /* Using r0 and not ip here because cbz requires low register.  */
-        m_cbz   reg=r0, label=.Lcompute_return_value
-        clz     r0, r0
-        /* r0 contains the number of bits on the left of the first all-zero byte in r1.  */
-        rsb     r0, r0, #24
-        /* Here, r0 contains the number of bits on the right of the first all-zero byte in r1.  */
-        lsr     r1, r1, r0
-        lsr     r2, r2, r0
-
-.Lcompute_return_value:
-        movs    r0, #1
-        cmp     r1, r2
-        /* The return value is computed as follows.
-        If r1>r2 then (C==1 and Z==0) and LS doesn't hold and r0 is #1 at return.
-        If r1<r2 then (C==0 and Z==0) and we execute SBC with carry_in=0,
-        which means r0:=r0-r0-1 and r0 is #-1 at return.
-        If r1=r2 then (C==1 and Z==1) and we execute SBC with carry_in=1,
-        which means r0:=r0-r0 and r0 is #0 at return.
-        (C==0 and Z==1) cannot happen because the carry bit is "not borrow".  */
-        it      ls
-        sbcls   r0, r0, r0
-        bx      lr
-
-
-#else   /* !(defined (_ISA_THUMB_2) || defined (_ISA_ARM_6)
-             defined (__OPTIMIZE_SIZE__) || defined (PREFER_SIZE_OVER_SPEED) ||
-             (defined (__thumb__) && !defined (__thumb2__))) */
-
-        /* Use LDR whenever possible. */
-
-#ifdef __thumb2__
-#define magic1(REG) 0x01010101
-#define magic2(REG) 0x80808080
 #else
-#define magic1(REG) REG
-#define magic2(REG) REG, lsl #7
-#endif
-
-        optpld  r0
-        optpld  r1
-        eor     r2, r0, r1
-        tst     r2, #3
-        /* Strings not at same byte offset from a word boundary.  */
-        bne     .Lstrcmp_unaligned
-        ands    r2, r0, #3
-        bic     r0, r0, #3
-        bic     r1, r1, #3
-        ldr     ip, [r0], #4
-        it      eq
-        ldreq   r3, [r1], #4
-        beq     1f
-        /* Although s1 and s2 have identical initial alignment, they are
-        not currently word aligned.  Rather than comparing bytes,
-	make sure that any bytes fetched from before the addressed
-	bytes are forced to 0xff.  Then they will always compare
-	equal.  */
-        eor     r2, r2, #3
-        lsl     r2, r2, #3
-        mvn     r3, MSB
-        S2LOMEM        r2, r3, r2
-        ldr     r3, [r1], #4
-        orr     ip, ip, r2
-        orr     r3, r3, r2
-1:
-#ifndef __thumb2__
-              /* Load the 'magic' constant 0x01010101.  */
-        str     r4, [sp, #-4]!
-        mov     r4, #1
-        orr     r4, r4, r4, lsl #8
-        orr     r4, r4, r4, lsl #16
-#endif
-        .p2align        2
-4:
-        optpld  r0, #8
-        optpld  r1, #8
-        sub     r2, ip, magic1(r4)
-        cmp     ip, r3
-        itttt   eq
-        /* check for any zero bytes in first word */
-        biceq   r2, r2, ip
-        tsteq   r2, magic2(r4)
-        ldreq   ip, [r0], #4
-        ldreq   r3, [r1], #4
-        beq     4b
-2:
-        /* There's a zero or a different byte in the word */
-        S2HIMEM  r0, ip, #24
-        S2LOMEM  ip, ip, #8
-        cmp     r0, #1
-        it      cs
-        cmpcs   r0, r3, S2HIMEM #24
-        it      eq
-        S2LOMEMEQ r3, r3, #8
-        beq     2b
-        /* On a big-endian machine, r0 contains the desired byte in bits
-        0-7; on a little-endian machine they are in bits 24-31.  In
-        both cases the other bits in r0 are all zero.  For r3 the
-        interesting byte is at the other end of the word, but the
-        other bits are not necessarily zero.  We need a signed result
-        representing the differnece in the unsigned bytes, so for the
-        little-endian case we can't just shift the interesting bits
-        up.  */
-#ifdef __ARMEB__
-        sub     r0, r0, r3, lsr #24
-#else
-        and     r3, r3, #255
-#ifdef __thumb2__
-        /* No RSB instruction in Thumb2 */
-        lsr     r0, r0, #24
-        sub     r0, r0, r3
-#else
-        rsb     r0, r3, r0, lsr #24
-#endif
-#endif
-#ifndef __thumb2__
-        ldr     r4, [sp], #4
-#endif
-        RETURN
 
+# include "strcmp-armv4.S"
 
-.Lstrcmp_unaligned:
-
-#if 0
-        /* The assembly code below is based on the following alogrithm.  */
-#ifdef __ARMEB__
-#define RSHIFT <<
-#define LSHIFT >>
-#else
-#define RSHIFT >>
-#define LSHIFT <<
 #endif
-
-#define body(shift)							\
-  mask = 0xffffffffU RSHIFT shift;					\
-  w1 = *wp1++;								\
-  w2 = *wp2++;								\
-  do									\
-    {									\
-      t1 = w1 & mask;							\
-      if (__builtin_expect(t1 != w2 RSHIFT shift, 0))			\
-	{								\
-	  w2 RSHIFT= shift;						\
-	  break;							\
-	}								\
-      if (__builtin_expect(((w1 - b1) & ~w1) & (b1 << 7), 0))		\
-	{								\
-	  /* See comment in assembler below re syndrome on big-endian */\
-	  if ((((w1 - b1) & ~w1) & (b1 << 7)) & mask)			\
-	    w2 RSHIFT= shift;						\
-	  else								\
-	    {								\
-	      w2 = *wp2;						\
-	      t1 = w1 RSHIFT (32 - shift);				\
-	      w2 = (w2 LSHIFT (32 - shift)) RSHIFT (32 - shift);	\
-	    }								\
-	  break;							\
-	}								\
-      w2 = *wp2++;							\
-      t1 ^= w1;								\
-      if (__builtin_expect(t1 != w2 LSHIFT (32 - shift), 0))		\
-	{								\
-	  t1 = w1 >> (32 - shift);					\
-	  w2 = (w2 << (32 - shift)) RSHIFT (32 - shift);		\
-	  break;							\
-	}								\
-      w1 = *wp1++;							\
-    } while (1)
-
-  const unsigned* wp1;
-  const unsigned* wp2;
-  unsigned w1, w2;
-  unsigned mask;
-  unsigned shift;
-  unsigned b1 = 0x01010101;
-  char c1, c2;
-  unsigned t1;
-
-  while (((unsigned) s1) & 3)
-    {
-      c1 = *s1++;
-      c2 = *s2++;
-      if (c1 == 0 || c1 != c2)
-	return c1 - (int)c2;
-    }
-  wp1 = (unsigned*) (((unsigned)s1) & ~3);
-  wp2 = (unsigned*) (((unsigned)s2) & ~3);
-  t1 = ((unsigned) s2) & 3;
-  if (t1 == 1)
-    {
-      body(8);
-    }
-  else if (t1 == 2)
-    {
-      body(16);
-    }
-  else
-    {
-      body (24);
-    }
-
-  do
-    {
-#ifdef __ARMEB__
-      c1 = (char) t1 >> 24;
-      c2 = (char) w2 >> 24;
-#else /* not  __ARMEB__ */
-      c1 = (char) t1;
-      c2 = (char) w2;
-#endif /* not  __ARMEB__ */
-      t1 RSHIFT= 8;
-      w2 RSHIFT= 8;
-    } while (c1 != 0 && c1 == c2);
-  return c1 - c2;
-#endif /* 0 */
-
-
-        wp1 .req r0
-        wp2 .req r1
-        b1  .req r2
-        w1  .req r4
-        w2  .req r5
-        t1  .req ip
-        @ r3 is scratch
-
-        /* First of all, compare bytes until wp1(sp1) is word-aligned. */
-1:
-        tst     wp1, #3
-        beq     2f
-        ldrb    r2, [wp1], #1
-        ldrb    r3, [wp2], #1
-        cmp     r2, #1
-        it      cs
-        cmpcs   r2, r3
-        beq     1b
-        sub     r0, r2, r3
-        RETURN
-
-2:
-        str     r5, [sp, #-4]!
-        str     r4, [sp, #-4]!
-        //stmfd   sp!, {r4, r5}
-        mov     b1, #1
-        orr     b1, b1, b1, lsl #8
-        orr     b1, b1, b1, lsl #16
-
-        and     t1, wp2, #3
-        bic     wp2, wp2, #3
-        ldr     w1, [wp1], #4
-        ldr     w2, [wp2], #4
-        cmp     t1, #2
-        beq     2f
-        bhi     3f
-
-        /* Critical inner Loop: Block with 3 bytes initial overlap */
-        .p2align        2
-1:
-        bic     t1, w1, MSB
-        cmp     t1, w2, S2LOMEM #8
-        sub     r3, w1, b1
-        bic     r3, r3, w1
-        bne     4f
-        ands    r3, r3, b1, lsl #7
-        it      eq
-        ldreq   w2, [wp2], #4
-        bne     5f
-        eor     t1, t1, w1
-        cmp     t1, w2, S2HIMEM #24
-        bne     6f
-        ldr     w1, [wp1], #4
-        b       1b
-4:
-        S2LOMEM        w2, w2, #8
-        b       8f
-
-5:
-#ifdef __ARMEB__
-        /* The syndrome value may contain false ones if the string ends
-        with the bytes 0x01 0x00 */
-        tst     w1, #0xff000000
-        itt     ne
-        tstne   w1, #0x00ff0000
-        tstne   w1, #0x0000ff00
-        beq     7f
-#else
-        bics    r3, r3, #0xff000000
-        bne     7f
-#endif
-        ldrb    w2, [wp2]
-        S2LOMEM  t1, w1, #24
-#ifdef __ARMEB__
-        lsl     w2, w2, #24
-#endif
-        b       8f
-
-6:
-        S2LOMEM  t1, w1, #24
-        and     w2, w2, LSB
-        b       8f
-
-        /* Critical inner Loop: Block with 2 bytes initial overlap */
-        .p2align        2
-2:
-        S2HIMEM  t1, w1, #16
-        sub     r3, w1, b1
-        S2LOMEM  t1, t1, #16
-        bic     r3, r3, w1
-        cmp     t1, w2, S2LOMEM #16
-        bne     4f
-        ands    r3, r3, b1, lsl #7
-        it      eq
-        ldreq   w2, [wp2], #4
-        bne     5f
-        eor     t1, t1, w1
-        cmp     t1, w2, S2HIMEM #16
-        bne     6f
-        ldr     w1, [wp1], #4
-        b       2b
-
-5:
-#ifdef __ARMEB__
-        /* The syndrome value may contain false ones if the string ends
-        with the bytes 0x01 0x00 */
-        tst     w1, #0xff000000
-        it      ne
-        tstne   w1, #0x00ff0000
-        beq     7f
-#else
-        lsls    r3, r3, #16
-        bne     7f
-#endif
-        ldrh    w2, [wp2]
-        S2LOMEM  t1, w1, #16
-#ifdef __ARMEB__
-        lsl     w2, w2, #16
-#endif
-        b       8f
-
-6:
-        S2HIMEM  w2, w2, #16
-        S2LOMEM  t1, w1, #16
-4:
-        S2LOMEM  w2, w2, #16
-        b       8f
-
-        /* Critical inner Loop: Block with 1 byte initial overlap */
-        .p2align        2
-3:
-        and     t1, w1, LSB
-        cmp     t1, w2, S2LOMEM #24
-        sub     r3, w1, b1
-        bic     r3, r3, w1
-        bne     4f
-        ands    r3, r3, b1, lsl #7
-        it      eq
-        ldreq   w2, [wp2], #4
-        bne     5f
-        eor     t1, t1, w1
-        cmp     t1, w2, S2HIMEM #8
-        bne     6f
-        ldr     w1, [wp1], #4
-        b       3b
-4:
-        S2LOMEM  w2, w2, #24
-        b       8f
-5:
-        /* The syndrome value may contain false ones if the string ends
-        with the bytes 0x01 0x00 */
-        tst     w1, LSB
-        beq     7f
-        ldr     w2, [wp2], #4
-6:
-        S2LOMEM  t1, w1, #8
-        bic     w2, w2, MSB
-        b       8f
-7:
-        mov     r0, #0
-        //ldmfd   sp!, {r4, r5}
-        ldr     r4, [sp], #4
-        ldr     r5, [sp], #4
-        RETURN
-8:
-        and     r2, t1, LSB
-        and     r0, w2, LSB
-        cmp     r0, #1
-        it      cs
-        cmpcs   r0, r2
-        itt     eq
-        S2LOMEMEQ        t1, t1, #8
-        S2LOMEMEQ        w2, w2, #8
-        beq     8b
-        sub     r0, r2, r0
-        //ldmfd   sp!, {r4, r5}
-        ldr     r4, [sp], #4
-        ldr     r5, [sp], #4
-        RETURN
-
-#endif /* !(defined (_ISA_THUMB_2) || defined (_ISA_ARM_6)
-            defined (__OPTIMIZE_SIZE__) || defined (PREFER_SIZE_OVER_SPEED) ||
-            (defined (__thumb__) && !defined (__thumb2__))) */