SSSE3 optimized strcasecmp and strncasecmp for x86-32

[thirdparty/glibc.git] / sysdeps / i386 / i686 / multiarch / strcmp-ssse3.S

diff --git a/sysdeps/i386/i686/multiarch/strcmp-ssse3.S b/sysdeps/i386/i686/multiarch/strcmp-ssse3.S
index 0131e8d80bb48619a2d57d4623cb9fc5b04b070e..f8a2c7de835723a5510301cd233ef2bf376d1c48 100644 (file)
--- a/sysdeps/i386/i686/multiarch/strcmp-ssse3.S
+++ b/sysdeps/i386/i686/multiarch/strcmp-ssse3.S
@@ -34,43 +34,173 @@
 #define PUSH(REG)      pushl REG; CFI_PUSH (REG)
 #define POP(REG)       popl REG; CFI_POP (REG)
 
-#ifndef USE_AS_STRNCMP
+#ifdef USE_AS_STRNCMP
 # ifndef STRCMP
-#  define STRCMP       __strcmp_ssse3
+#  define STRCMP       __strncmp_ssse3
 # endif
-# define STR1          4
+# define STR1          8
 # define STR2          STR1+4
-# define RETURN                ret; .p2align 4
-# define UPDATE_STRNCMP_COUNTER
-#else
+# define CNT           STR2+4
+# define RETURN                POP (REM); ret; .p2align 4; CFI_PUSH (REM)
+# define UPDATE_STRNCMP_COUNTER                                \
+       /* calculate left number to compare */          \
+       mov     $16, %esi;                              \
+       sub     %ecx, %esi;                             \
+       cmp     %esi, REM;                              \
+       jbe     L(more8byteseq);                        \
+       sub     %esi, REM
+# define FLAGS         %ebx
+# define REM           %ebp
+#elif defined USE_AS_STRCASECMP_L
+# include "locale-defines.h"
 # ifndef STRCMP
-#  define STRCMP       __strncmp_ssse3
+#  define STRCMP       __strcasecmp_l_ssse3
 # endif
 # define STR1          8
 # define STR2          STR1+4
+# define LOCALE                12      /* Loaded before the adjustement.  */
+# ifdef PIC
+#  define RETURN       POP (%ebx); ret; .p2align 4; CFI_PUSH (%ebx)
+# else
+#  define RETURN       ret; .p2align 4
+# endif
+# define UPDATE_STRNCMP_COUNTER
+# define FLAGS         (%esp)
+# define NONASCII      __strcasecmp_nonascii
+#elif defined USE_AS_STRNCASECMP_L
+# include "locale-defines.h"
+# ifndef STRCMP
+#  define STRCMP       __strncasecmp_l_ssse3
+# endif
+# define STR1          12
+# define STR2          STR1+4
 # define CNT           STR2+4
-# define RETURN                POP (%ebp); ret; .p2align 4; CFI_PUSH (%ebp)
+# define LOCALE                16      /* Loaded before the adjustement.  */
+# ifdef PIC
+#  define RETURN       POP (REM); POP (%ebx); ret; \
+                       .p2align 4; CFI_PUSH (%ebx); CFI_PUSH (REM)
+# else
+#  define RETURN       POP (REM); ret; .p2align 4; CFI_PUSH (REM)
+# endif
 # define UPDATE_STRNCMP_COUNTER                                \
        /* calculate left number to compare */          \
        mov     $16, %esi;                              \
        sub     %ecx, %esi;                             \
-       cmp     %esi, %ebp;                             \
+       cmp     %esi, REM;                              \
        jbe     L(more8byteseq);                        \
-       sub     %esi, %ebp
+       sub     %esi, REM
+# define FLAGS         (%esp)
+# define REM           %ebp
+# define NONASCII      __strncasecmp_nonascii
+#else
+# ifndef STRCMP
+#  define STRCMP       __strcmp_ssse3
+# endif
+# define STR1          4
+# define STR2          STR1+4
+# define RETURN                ret; .p2align 4
+# define UPDATE_STRNCMP_COUNTER
+# define FLAGS         %ebx
 #endif
 
        .section .text.ssse3,"ax",@progbits
+
+#ifdef USE_AS_STRCASECMP_L
+ENTRY (__strcasecmp_ssse3)
+# ifdef PIC
+       PUSH    (%ebx)
+       call    __i686.get_pc_thunk.bx
+       addl    $_GLOBAL_OFFSET_TABLE_, %ebx
+       movl    __libc_tsd_LOCALE@GOTNTPOFF(%ebx), %eax
+# else
+       movl    __libc_tsd_LOCALE@NTPOFF, %eax
+# endif
+       movl    %gs:(%eax), %eax
+# if LOCALE_T___LOCALES != 0 || LC_CTYPE != 0
+       movl    LOCALE_T___LOCALES+LC_CTYPE*4(%eax), %eax
+# else
+       movl    (%eax), %eax
+# endif
+       testl   $1, LOCALE_DATA_VALUES+_NL_CTYPE_NONASCII_CASE*SIZEOF_VALUES(%eax)
+       jne     __strcasecmp_nonascii
+       jmp     .Lascii
+END (__strcasecmp_ssse3)
+#endif
+
+#ifdef USE_AS_STRNCASECMP_L
+ENTRY (__strncasecmp_ssse3)
+# ifdef PIC
+       PUSH    (%ebx)
+       call    __i686.get_pc_thunk.bx
+       addl    $_GLOBAL_OFFSET_TABLE_, %ebx
+       movl    __libc_tsd_LOCALE@GOTNTPOFF(%ebx), %eax
+# else
+       movl    __libc_tsd_LOCALE@NTPOFF, %eax
+# endif
+       movl    %gs:(%eax), %eax
+# if LOCALE_T___LOCALES != 0 || LC_CTYPE != 0
+       movl    LOCALE_T___LOCALES+LC_CTYPE*4(%eax), %eax
+# else
+       movl    (%eax), %eax
+# endif
+       testl   $1, LOCALE_DATA_VALUES+_NL_CTYPE_NONASCII_CASE*SIZEOF_VALUES(%eax)
+       jne     __strncasecmp_nonascii
+       jmp     .Lascii
+END (__strncasecmp_ssse3)
+#endif
+
 ENTRY (STRCMP)
-#ifdef USE_AS_STRNCMP
-       PUSH    (%ebp)
+#if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
+       movl    LOCALE(%esp), %eax
+# if LOCALE_T___LOCALES != 0 || LC_CTYPE != 0
+       movl    LOCALE_T___LOCALES+LC_CTYPE*4(%eax), %eax
+# else
+       movl    (%eax), %eax
+# endif
+       testl   $1, LOCALE_DATA_VALUES+_NL_CTYPE_NONASCII_CASE*SIZEOF_VALUES(%eax)
+       jne     NONASCII
+
+# ifdef PIC
+       PUSH    (%ebx)
+       call    __i686.get_pc_thunk.bx
+       addl    $_GLOBAL_OFFSET_TABLE_, %ebx
+# endif
+.Lascii:
+       .section .rodata.cst16,"aM",@progbits,16
+       .align 16
+.Lbelowupper:
+       .quad   0x4040404040404040
+       .quad   0x4040404040404040
+.Ltopupper:
+       .quad   0x5b5b5b5b5b5b5b5b
+       .quad   0x5b5b5b5b5b5b5b5b
+.Ltouppermask:
+       .quad   0x2020202020202020
+       .quad   0x2020202020202020
+       .previous
+
+# ifdef PIC
+#  define UCLOW_reg .Lbelowupper@GOTOFF(%ebx)
+#  define UCHIGH_reg .Ltopupper@GOTOFF(%ebx)
+#  define LCQWORD_reg .Ltouppermask@GOTOFF(%ebx)
+# else
+#  define UCLOW_reg .Lbelowupper
+#  define UCHIGH_reg .Ltopupper
+#  define LCQWORD_reg .Ltouppermask
+# endif
 #endif
+
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       PUSH    (REM)
+#endif
+
        movl    STR1(%esp), %edx
        movl    STR2(%esp), %eax
-#ifdef USE_AS_STRNCMP
-       movl    CNT(%esp), %ebp
-       cmp     $16, %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       movl    CNT(%esp), REM
+       cmp     $16, REM
        jb      L(less16bytes_sncmp)
-#else
+#elif !defined USE_AS_STRCASECMP_L
        movzbl  (%eax), %ecx
        cmpb    %cl, (%edx)
        jne     L(neq)
@@ -135,15 +265,35 @@ ENTRY (STRCMP)
        movlpd  (%edx), %xmm2
        movhpd  8(%eax), %xmm1
        movhpd  8(%edx), %xmm2
+#if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
+# define TOLOWER(reg1, reg2) \
+       movdqa  reg1, %xmm5;                                    \
+       movdqa  reg2, %xmm7;                                    \
+       movdqa  UCHIGH_reg, %xmm6;                              \
+       pcmpgtb UCLOW_reg, %xmm5;                               \
+       pcmpgtb UCLOW_reg, %xmm7;                               \
+       pcmpgtb reg1, %xmm6;                                    \
+       pand    %xmm6, %xmm5;                                   \
+       movdqa  UCHIGH_reg, %xmm6;                              \
+       pcmpgtb reg2, %xmm6;                                    \
+       pand    %xmm6, %xmm7;                                   \
+       pand    LCQWORD_reg, %xmm5;                             \
+       por     %xmm5, reg1;                                    \
+       pand    LCQWORD_reg, %xmm7;                             \
+       por     %xmm7, reg2
+       TOLOWER (%xmm1, %xmm2)
+#else
+# define TOLOWER(reg1, reg2)
+#endif
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
        psubb   %xmm0, %xmm1
        pmovmskb %xmm1, %ecx
        sub     $0xffff, %ecx
        jnz     L(less16bytes)
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(eq)
 #endif
        add     $16, %eax
@@ -151,10 +301,16 @@ ENTRY (STRCMP)
 
 L(crosspage):
 
-       PUSH    (%ebx)
+#if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
+       PUSH    (FLAGS)
+#endif
        PUSH    (%edi)
        PUSH    (%esi)
-#ifdef USE_AS_STRNCMP
+#if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
+       pushl   $0
+       cfi_adjust_cfa_offset (4)
+#endif
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
        cfi_remember_state
 #endif
 
@@ -164,11 +320,13 @@ L(crosspage):
        and     $0xf, %edi
        xor     %ecx, %eax
        xor     %edi, %edx
-       xor     %ebx, %ebx
+#if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
+       xor     FLAGS, FLAGS
+#endif
        cmp     %edi, %ecx
        je      L(ashr_0)
        ja      L(bigger)
-       or      $0x20, %ebx
+       orl     $0x20, FLAGS
        xchg    %edx, %eax
        xchg    %ecx, %edi
 L(bigger):
@@ -218,7 +376,13 @@ L(ashr_0):
        movdqa  (%eax), %xmm1
        pxor    %xmm0, %xmm0
        pcmpeqb %xmm1, %xmm0
+#if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
+       movdqa  (%edx), %xmm2
+       TOLOWER (%xmm1, %xmm2)
+       pcmpeqb %xmm2, %xmm1
+#else
        pcmpeqb (%edx), %xmm1
+#endif
        psubb   %xmm0, %xmm1
        pmovmskb %xmm1, %edi
        shr     %cl, %esi
@@ -227,23 +391,29 @@ L(ashr_0):
        mov     %ecx, %edi
        jne     L(less32bytes)
        UPDATE_STRNCMP_COUNTER
-       mov     $0x10, %ebx
+       movl    $0x10, FLAGS
        mov     $0x10, %ecx
        pxor    %xmm0, %xmm0
        .p2align 4
 L(loop_ashr_0):
        movdqa  (%eax, %ecx), %xmm1
+#if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
        movdqa  (%edx, %ecx), %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
+#else
+       pcmpeqb %xmm1, %xmm0
+       pcmpeqb (%edx, %ecx), %xmm1
+#endif
        psubb   %xmm0, %xmm1
        pmovmskb %xmm1, %esi
        sub     $0xffff, %esi
        jnz     L(exit)
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
        add     $16, %ecx
@@ -262,6 +432,7 @@ L(ashr_1):
        movdqa  (%eax), %xmm1
        pcmpeqb %xmm1, %xmm0
        pslldq  $15, %xmm2
+       TOLOWER (%xmm1, %xmm2)
        pcmpeqb %xmm1, %xmm2
        psubb   %xmm0, %xmm2
        pmovmskb %xmm2, %edi
@@ -276,7 +447,7 @@ L(ashr_1):
        movdqa  (%edx), %xmm3
        pxor    %xmm0, %xmm0
        mov     $16, %ecx
-       or      $1, %ebx
+       orl     $1, FLAGS
        lea     1(%edx), %edi
        and     $0xfff, %edi
        sub     $0x1000, %edi
@@ -292,6 +463,7 @@ L(gobble_ashr_1):
        movdqa  %xmm2, %xmm4
 
        palignr $1, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -299,9 +471,9 @@ L(gobble_ashr_1):
        pmovmskb %xmm1, %esi
        sub     $0xffff, %esi
        jnz     L(exit)
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
 
@@ -316,6 +488,7 @@ L(gobble_ashr_1):
        movdqa  %xmm2, %xmm4
 
        palignr $1, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -324,9 +497,9 @@ L(gobble_ashr_1):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
        add     $16, %ecx
@@ -340,8 +513,8 @@ L(nibble_ashr_1):
        test    $0xfffe, %esi
        jnz     L(ashr_1_exittail)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $15, %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $15, REM
        jbe     L(ashr_1_exittail)
 #endif
        pxor    %xmm0, %xmm0
@@ -368,6 +541,7 @@ L(ashr_2):
        movdqa  (%eax), %xmm1
        pcmpeqb %xmm1, %xmm0
        pslldq  $14, %xmm2
+       TOLOWER (%xmm1, %xmm2)
        pcmpeqb %xmm1, %xmm2
        psubb   %xmm0, %xmm2
        pmovmskb %xmm2, %edi
@@ -382,7 +556,7 @@ L(ashr_2):
        movdqa  (%edx), %xmm3
        pxor    %xmm0, %xmm0
        mov     $16, %ecx
-       or      $2, %ebx
+       orl     $2, FLAGS
        lea     2(%edx), %edi
        and     $0xfff, %edi
        sub     $0x1000, %edi
@@ -398,6 +572,7 @@ L(gobble_ashr_2):
        movdqa  %xmm2, %xmm4
 
        palignr $2, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -406,9 +581,9 @@ L(gobble_ashr_2):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
        add     $16, %ecx
@@ -422,6 +597,7 @@ L(gobble_ashr_2):
        movdqa  %xmm2, %xmm4
 
        palignr $2, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -430,9 +606,9 @@ L(gobble_ashr_2):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
        add     $16, %ecx
@@ -446,8 +622,8 @@ L(nibble_ashr_2):
        test    $0xfffc, %esi
        jnz     L(ashr_2_exittail)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $14, %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $14, REM
        jbe     L(ashr_2_exittail)
 #endif
 
@@ -475,6 +651,7 @@ L(ashr_3):
        movdqa  (%eax), %xmm1
        pcmpeqb %xmm1, %xmm0
        pslldq  $13, %xmm2
+       TOLOWER (%xmm1, %xmm2)
        pcmpeqb %xmm1, %xmm2
        psubb   %xmm0, %xmm2
        pmovmskb %xmm2, %edi
@@ -489,7 +666,7 @@ L(ashr_3):
        movdqa  (%edx), %xmm3
        pxor    %xmm0, %xmm0
        mov     $16, %ecx
-       or      $3, %ebx
+       orl     $3, FLAGS
        lea     3(%edx), %edi
        and     $0xfff, %edi
        sub     $0x1000, %edi
@@ -505,6 +682,7 @@ L(gobble_ashr_3):
        movdqa  %xmm2, %xmm4
 
        palignr $3, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -513,9 +691,9 @@ L(gobble_ashr_3):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
        add     $16, %ecx
@@ -529,6 +707,7 @@ L(gobble_ashr_3):
        movdqa  %xmm2, %xmm4
 
        palignr $3, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -537,9 +716,9 @@ L(gobble_ashr_3):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
        add     $16, %ecx
@@ -553,8 +732,8 @@ L(nibble_ashr_3):
        test    $0xfff8, %esi
        jnz     L(ashr_3_exittail)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $13, %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $13, REM
        jbe     L(ashr_3_exittail)
 #endif
        pxor    %xmm0, %xmm0
@@ -581,6 +760,7 @@ L(ashr_4):
        movdqa  (%eax), %xmm1
        pcmpeqb %xmm1, %xmm0
        pslldq  $12, %xmm2
+       TOLOWER (%xmm1, %xmm2)
        pcmpeqb %xmm1, %xmm2
        psubb   %xmm0, %xmm2
        pmovmskb %xmm2, %edi
@@ -595,7 +775,7 @@ L(ashr_4):
        movdqa  (%edx), %xmm3
        pxor    %xmm0, %xmm0
        mov     $16, %ecx
-       or      $4, %ebx
+       orl     $4, FLAGS
        lea     4(%edx), %edi
        and     $0xfff, %edi
        sub     $0x1000, %edi
@@ -611,6 +791,7 @@ L(gobble_ashr_4):
        movdqa  %xmm2, %xmm4
 
        palignr $4, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -619,9 +800,9 @@ L(gobble_ashr_4):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
 
@@ -636,6 +817,7 @@ L(gobble_ashr_4):
        movdqa  %xmm2, %xmm4
 
        palignr $4, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -644,9 +826,9 @@ L(gobble_ashr_4):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
 
@@ -661,8 +843,8 @@ L(nibble_ashr_4):
        test    $0xfff0, %esi
        jnz     L(ashr_4_exittail)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $12, %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $12, REM
        jbe     L(ashr_4_exittail)
 #endif
 
@@ -690,6 +872,7 @@ L(ashr_5):
        movdqa  (%eax), %xmm1
        pcmpeqb %xmm1, %xmm0
        pslldq  $11, %xmm2
+       TOLOWER (%xmm1, %xmm2)
        pcmpeqb %xmm1, %xmm2
        psubb   %xmm0, %xmm2
        pmovmskb %xmm2, %edi
@@ -704,7 +887,7 @@ L(ashr_5):
        movdqa  (%edx), %xmm3
        pxor    %xmm0, %xmm0
        mov     $16, %ecx
-       or      $5, %ebx
+       orl     $5, FLAGS
        lea     5(%edx), %edi
        and     $0xfff, %edi
        sub     $0x1000, %edi
@@ -720,6 +903,7 @@ L(gobble_ashr_5):
        movdqa  %xmm2, %xmm4
 
        palignr $5, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -728,9 +912,9 @@ L(gobble_ashr_5):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
        add     $16, %ecx
@@ -744,6 +928,7 @@ L(gobble_ashr_5):
        movdqa  %xmm2, %xmm4
 
        palignr $5, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -752,9 +937,9 @@ L(gobble_ashr_5):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
        add     $16, %ecx
@@ -768,8 +953,8 @@ L(nibble_ashr_5):
        test    $0xffe0, %esi
        jnz     L(ashr_5_exittail)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $11, %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $11, REM
        jbe     L(ashr_5_exittail)
 #endif
        pxor    %xmm0, %xmm0
@@ -797,6 +982,7 @@ L(ashr_6):
        movdqa  (%eax), %xmm1
        pcmpeqb %xmm1, %xmm0
        pslldq  $10, %xmm2
+       TOLOWER (%xmm1, %xmm2)
        pcmpeqb %xmm1, %xmm2
        psubb   %xmm0, %xmm2
        pmovmskb %xmm2, %edi
@@ -811,7 +997,7 @@ L(ashr_6):
        movdqa  (%edx), %xmm3
        pxor    %xmm0, %xmm0
        mov     $16, %ecx
-       or      $6, %ebx
+       orl     $6, FLAGS
        lea     6(%edx), %edi
        and     $0xfff, %edi
        sub     $0x1000, %edi
@@ -827,6 +1013,7 @@ L(gobble_ashr_6):
        movdqa  %xmm2, %xmm4
 
        palignr $6, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -835,9 +1022,9 @@ L(gobble_ashr_6):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
 
@@ -852,6 +1039,7 @@ L(gobble_ashr_6):
        movdqa  %xmm2, %xmm4
 
        palignr $6, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -859,9 +1047,9 @@ L(gobble_ashr_6):
        pmovmskb %xmm1, %esi
        sub     $0xffff, %esi
        jnz     L(exit)
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
 
@@ -876,8 +1064,8 @@ L(nibble_ashr_6):
        test    $0xffc0, %esi
        jnz     L(ashr_6_exittail)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $10, %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $10, REM
        jbe     L(ashr_6_exittail)
 #endif
        pxor    %xmm0, %xmm0
@@ -905,6 +1093,7 @@ L(ashr_7):
        movdqa  (%eax), %xmm1
        pcmpeqb %xmm1, %xmm0
        pslldq  $9, %xmm2
+       TOLOWER (%xmm1, %xmm2)
        pcmpeqb %xmm1, %xmm2
        psubb   %xmm0, %xmm2
        pmovmskb %xmm2, %edi
@@ -919,7 +1108,7 @@ L(ashr_7):
        movdqa  (%edx), %xmm3
        pxor    %xmm0, %xmm0
        mov     $16, %ecx
-       or      $7, %ebx
+       orl     $7, FLAGS
        lea     8(%edx), %edi
        and     $0xfff, %edi
        sub     $0x1000, %edi
@@ -935,6 +1124,7 @@ L(gobble_ashr_7):
        movdqa  %xmm2, %xmm4
 
        palignr $7, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -943,9 +1133,9 @@ L(gobble_ashr_7):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
 
@@ -960,6 +1150,7 @@ L(gobble_ashr_7):
        movdqa  %xmm2, %xmm4
 
        palignr $7, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -968,9 +1159,9 @@ L(gobble_ashr_7):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
 
@@ -985,8 +1176,8 @@ L(nibble_ashr_7):
        test    $0xff80, %esi
        jnz     L(ashr_7_exittail)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $9, %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $9, REM
        jbe     L(ashr_7_exittail)
 #endif
        pxor    %xmm0, %xmm0
@@ -1014,6 +1205,7 @@ L(ashr_8):
        movdqa  (%eax), %xmm1
        pcmpeqb %xmm1, %xmm0
        pslldq  $8, %xmm2
+       TOLOWER (%xmm1, %xmm2)
        pcmpeqb %xmm1, %xmm2
        psubb   %xmm0, %xmm2
        pmovmskb %xmm2, %edi
@@ -1028,7 +1220,7 @@ L(ashr_8):
        movdqa  (%edx), %xmm3
        pxor    %xmm0, %xmm0
        mov     $16, %ecx
-       or      $8, %ebx
+       orl     $8, FLAGS
        lea     8(%edx), %edi
        and     $0xfff, %edi
        sub     $0x1000, %edi
@@ -1044,6 +1236,7 @@ L(gobble_ashr_8):
        movdqa  %xmm2, %xmm4
 
        palignr $8, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -1052,9 +1245,9 @@ L(gobble_ashr_8):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
        add     $16, %ecx
@@ -1068,6 +1261,7 @@ L(gobble_ashr_8):
        movdqa  %xmm2, %xmm4
 
        palignr $8, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -1076,9 +1270,9 @@ L(gobble_ashr_8):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
        add     $16, %ecx
@@ -1092,8 +1286,8 @@ L(nibble_ashr_8):
        test    $0xff00, %esi
        jnz     L(ashr_8_exittail)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $8, %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $8, REM
        jbe     L(ashr_8_exittail)
 #endif
        pxor    %xmm0, %xmm0
@@ -1121,6 +1315,7 @@ L(ashr_9):
        movdqa  (%eax), %xmm1
        pcmpeqb %xmm1, %xmm0
        pslldq  $7, %xmm2
+       TOLOWER (%xmm1, %xmm2)
        pcmpeqb %xmm1, %xmm2
        psubb   %xmm0, %xmm2
        pmovmskb %xmm2, %edi
@@ -1135,7 +1330,7 @@ L(ashr_9):
        movdqa  (%edx), %xmm3
        pxor    %xmm0, %xmm0
        mov     $16, %ecx
-       or      $9, %ebx
+       orl     $9, FLAGS
        lea     9(%edx), %edi
        and     $0xfff, %edi
        sub     $0x1000, %edi
@@ -1151,6 +1346,7 @@ L(gobble_ashr_9):
        movdqa  %xmm2, %xmm4
 
        palignr $9, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -1159,9 +1355,9 @@ L(gobble_ashr_9):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
        add     $16, %ecx
@@ -1175,6 +1371,7 @@ L(gobble_ashr_9):
        movdqa  %xmm2, %xmm4
 
        palignr $9, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -1183,9 +1380,9 @@ L(gobble_ashr_9):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
        add     $16, %ecx
@@ -1199,8 +1396,8 @@ L(nibble_ashr_9):
        test    $0xfe00, %esi
        jnz     L(ashr_9_exittail)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $7, %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $7, REM
        jbe     L(ashr_9_exittail)
 #endif
        pxor    %xmm0, %xmm0
@@ -1227,6 +1424,7 @@ L(ashr_10):
        movdqa  (%eax), %xmm1
        pcmpeqb %xmm1, %xmm0
        pslldq  $6, %xmm2
+       TOLOWER (%xmm1, %xmm2)
        pcmpeqb %xmm1, %xmm2
        psubb   %xmm0, %xmm2
        pmovmskb %xmm2, %edi
@@ -1241,7 +1439,7 @@ L(ashr_10):
        movdqa  (%edx), %xmm3
        pxor    %xmm0, %xmm0
        mov     $16, %ecx
-       or      $10, %ebx
+       orl     $10, FLAGS
        lea     10(%edx), %edi
        and     $0xfff, %edi
        sub     $0x1000, %edi
@@ -1257,6 +1455,7 @@ L(gobble_ashr_10):
        movdqa  %xmm2, %xmm4
 
        palignr $10, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -1265,9 +1464,9 @@ L(gobble_ashr_10):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
        add     $16, %ecx
@@ -1281,6 +1480,7 @@ L(gobble_ashr_10):
        movdqa  %xmm2, %xmm4
 
        palignr $10, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -1289,9 +1489,9 @@ L(gobble_ashr_10):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
        add     $16, %ecx
@@ -1305,8 +1505,8 @@ L(nibble_ashr_10):
        test    $0xfc00, %esi
        jnz     L(ashr_10_exittail)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $6, %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $6, REM
        jbe     L(ashr_10_exittail)
 #endif
        pxor    %xmm0, %xmm0
@@ -1333,6 +1533,7 @@ L(ashr_11):
        movdqa  (%eax), %xmm1
        pcmpeqb %xmm1, %xmm0
        pslldq  $5, %xmm2
+       TOLOWER (%xmm1, %xmm2)
        pcmpeqb %xmm1, %xmm2
        psubb   %xmm0, %xmm2
        pmovmskb %xmm2, %edi
@@ -1347,7 +1548,7 @@ L(ashr_11):
        movdqa  (%edx), %xmm3
        pxor    %xmm0, %xmm0
        mov     $16, %ecx
-       or      $11, %ebx
+       orl     $11, FLAGS
        lea     11(%edx), %edi
        and     $0xfff, %edi
        sub     $0x1000, %edi
@@ -1363,6 +1564,7 @@ L(gobble_ashr_11):
        movdqa  %xmm2, %xmm4
 
        palignr $11, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -1371,9 +1573,9 @@ L(gobble_ashr_11):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
        add     $16, %ecx
@@ -1387,6 +1589,7 @@ L(gobble_ashr_11):
        movdqa  %xmm2, %xmm4
 
        palignr $11, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -1395,9 +1598,9 @@ L(gobble_ashr_11):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
        add     $16, %ecx
@@ -1411,8 +1614,8 @@ L(nibble_ashr_11):
        test    $0xf800, %esi
        jnz     L(ashr_11_exittail)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $5, %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $5, REM
        jbe     L(ashr_11_exittail)
 #endif
        pxor    %xmm0, %xmm0
@@ -1439,6 +1642,7 @@ L(ashr_12):
        movdqa  (%eax), %xmm1
        pcmpeqb %xmm1, %xmm0
        pslldq  $4, %xmm2
+       TOLOWER (%xmm1, %xmm2)
        pcmpeqb %xmm1, %xmm2
        psubb   %xmm0, %xmm2
        pmovmskb %xmm2, %edi
@@ -1453,7 +1657,7 @@ L(ashr_12):
        movdqa  (%edx), %xmm3
        pxor    %xmm0, %xmm0
        mov     $16, %ecx
-       or      $12, %ebx
+       orl     $12, FLAGS
        lea     12(%edx), %edi
        and     $0xfff, %edi
        sub     $0x1000, %edi
@@ -1469,6 +1673,7 @@ L(gobble_ashr_12):
        movdqa  %xmm2, %xmm4
 
        palignr $12, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -1477,9 +1682,9 @@ L(gobble_ashr_12):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
 
@@ -1494,6 +1699,7 @@ L(gobble_ashr_12):
        movdqa  %xmm2, %xmm4
 
        palignr $12, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -1502,9 +1708,9 @@ L(gobble_ashr_12):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
        add     $16, %ecx
@@ -1518,8 +1724,8 @@ L(nibble_ashr_12):
        test    $0xf000, %esi
        jnz     L(ashr_12_exittail)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $4, %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $4, REM
        jbe     L(ashr_12_exittail)
 #endif
        pxor    %xmm0, %xmm0
@@ -1546,6 +1752,7 @@ L(ashr_13):
        movdqa  (%eax), %xmm1
        pcmpeqb %xmm1, %xmm0
        pslldq  $3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
        pcmpeqb %xmm1, %xmm2
        psubb   %xmm0, %xmm2
        pmovmskb %xmm2, %edi
@@ -1560,7 +1767,7 @@ L(ashr_13):
        movdqa  (%edx), %xmm3
        pxor    %xmm0, %xmm0
        mov     $16, %ecx
-       or      $13, %ebx
+       orl     $13, FLAGS
        lea     13(%edx), %edi
        and     $0xfff, %edi
        sub     $0x1000, %edi
@@ -1576,6 +1783,7 @@ L(gobble_ashr_13):
        movdqa  %xmm2, %xmm4
 
        palignr $13, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -1584,9 +1792,9 @@ L(gobble_ashr_13):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
        add     $16, %ecx
@@ -1600,6 +1808,7 @@ L(gobble_ashr_13):
        movdqa  %xmm2, %xmm4
 
        palignr $13, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -1608,9 +1817,9 @@ L(gobble_ashr_13):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
        add     $16, %ecx
@@ -1624,8 +1833,8 @@ L(nibble_ashr_13):
        test    $0xe000, %esi
        jnz     L(ashr_13_exittail)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $3, %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $3, REM
        jbe     L(ashr_13_exittail)
 #endif
        pxor    %xmm0, %xmm0
@@ -1652,6 +1861,7 @@ L(ashr_14):
        movdqa  (%eax), %xmm1
        pcmpeqb %xmm1, %xmm0
        pslldq  $2, %xmm2
+       TOLOWER (%xmm1, %xmm2)
        pcmpeqb %xmm1, %xmm2
        psubb   %xmm0, %xmm2
        pmovmskb %xmm2, %edi
@@ -1666,7 +1876,7 @@ L(ashr_14):
        movdqa  (%edx), %xmm3
        pxor    %xmm0, %xmm0
        mov     $16, %ecx
-       or      $14, %ebx
+       orl     $14, FLAGS
        lea     14(%edx), %edi
        and     $0xfff, %edi
        sub     $0x1000, %edi
@@ -1682,6 +1892,7 @@ L(gobble_ashr_14):
        movdqa  %xmm2, %xmm4
 
        palignr $14, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -1690,9 +1901,9 @@ L(gobble_ashr_14):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
        add     $16, %ecx
@@ -1706,6 +1917,7 @@ L(gobble_ashr_14):
        movdqa  %xmm2, %xmm4
 
        palignr $14, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -1714,9 +1926,9 @@ L(gobble_ashr_14):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
        add     $16, %ecx
@@ -1730,8 +1942,8 @@ L(nibble_ashr_14):
        test    $0xc000, %esi
        jnz     L(ashr_14_exittail)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $2, %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $2, REM
        jbe     L(ashr_14_exittail)
 #endif
        pxor    %xmm0, %xmm0
@@ -1759,6 +1971,7 @@ L(ashr_15):
        movdqa  (%eax), %xmm1
        pcmpeqb %xmm1, %xmm0
        pslldq  $1, %xmm2
+       TOLOWER (%xmm1, %xmm2)
        pcmpeqb %xmm1, %xmm2
        psubb   %xmm0, %xmm2
        pmovmskb %xmm2, %edi
@@ -1773,7 +1986,7 @@ L(ashr_15):
        movdqa  (%edx), %xmm3
        pxor    %xmm0, %xmm0
        mov     $16, %ecx
-       or      $15, %ebx
+       orl     $15, FLAGS
        lea     15(%edx), %edi
        and     $0xfff, %edi
        sub     $0x1000, %edi
@@ -1789,6 +2002,7 @@ L(gobble_ashr_15):
        movdqa  %xmm2, %xmm4
 
        palignr $15, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -1797,9 +2011,9 @@ L(gobble_ashr_15):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
        add     $16, %ecx
@@ -1813,6 +2027,7 @@ L(gobble_ashr_15):
        movdqa  %xmm2, %xmm4
 
        palignr $15, %xmm3, %xmm2
+       TOLOWER (%xmm1, %xmm2)
 
        pcmpeqb %xmm1, %xmm0
        pcmpeqb %xmm2, %xmm1
@@ -1821,9 +2036,9 @@ L(gobble_ashr_15):
        sub     $0xffff, %esi
        jnz     L(exit)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $16, %ebp
-       lea     -16(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $16, REM
+       lea     -16(REM), REM
        jbe     L(more8byteseq)
 #endif
        add     $16, %ecx
@@ -1837,8 +2052,8 @@ L(nibble_ashr_15):
        test    $0x8000, %esi
        jnz     L(ashr_15_exittail)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $1, %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $1, REM
        jbe     L(ashr_15_exittail)
 #endif
        pxor    %xmm0, %xmm0
@@ -1854,27 +2069,34 @@ L(ashr_15_exittail):
 
        .p2align 4
 L(aftertail):
+       TOLOWER (%xmm1, %xmm3)
        pcmpeqb %xmm3, %xmm1
        psubb   %xmm0, %xmm1
        pmovmskb %xmm1, %esi
        not     %esi
 L(exit):
-       mov     %ebx, %edi
+       mov     FLAGS, %edi
        and     $0x1f, %edi
        lea     -16(%edi, %ecx), %edi
 L(less32bytes):
        add     %edi, %edx
        add     %ecx, %eax
-       test    $0x20, %ebx
+       testl   $0x20, FLAGS
        jz      L(ret2)
        xchg    %eax, %edx
 
        .p2align 4
 L(ret2):
        mov     %esi, %ecx
+#if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
+       addl    $4, %esp
+       cfi_adjust_cfa_offset (-4)
+#endif
        POP     (%esi)
        POP     (%edi)
-       POP     (%ebx)
+#if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
+       POP     (FLAGS)
+#endif
 L(less16bytes):
        test    %cl, %cl
        jz      L(2next_8_bytes)
@@ -1899,100 +2121,179 @@ L(less16bytes):
 
        test    $0x40, %cl
        jnz     L(Byte6)
-#ifdef USE_AS_STRNCMP
-       cmp     $7, %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $7, REM
        jbe     L(eq)
 #endif
 
        movzx   7(%eax), %ecx
        movzx   7(%edx), %eax
+#if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
+# ifdef PIC
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%eax,4), %eax
+# else
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%eax,4), %eax
+# endif
+#endif
 
        sub     %ecx, %eax
        RETURN
 
 L(Byte0):
-#ifdef USE_AS_STRNCMP
-       cmp     $0, %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $0, REM
        jbe     L(eq)
 #endif
        movzx   (%eax), %ecx
        movzx   (%edx), %eax
 
+#if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
+# ifdef PIC
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%eax,4), %eax
+# else
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%eax,4), %eax
+# endif
+#endif
+
        sub     %ecx, %eax
        RETURN
 
 L(Byte1):
-#ifdef USE_AS_STRNCMP
-       cmp     $1, %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $1, REM
        jbe     L(eq)
 #endif
        movzx   1(%eax), %ecx
        movzx   1(%edx), %eax
 
+#if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
+# ifdef PIC
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%eax,4), %eax
+# else
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%eax,4), %eax
+# endif
+#endif
+
        sub     %ecx, %eax
        RETURN
 
 L(Byte2):
-#ifdef USE_AS_STRNCMP
-       cmp     $2, %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $2, REM
        jbe     L(eq)
 #endif
        movzx   2(%eax), %ecx
        movzx   2(%edx), %eax
 
+#if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
+# ifdef PIC
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%eax,4), %eax
+# else
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%eax,4), %eax
+# endif
+#endif
+
        sub     %ecx, %eax
        RETURN
 
 L(Byte3):
-#ifdef USE_AS_STRNCMP
-       cmp     $3, %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $3, REM
        jbe     L(eq)
 #endif
        movzx   3(%eax), %ecx
        movzx   3(%edx), %eax
 
+#if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
+# ifdef PIC
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%eax,4), %eax
+# else
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%eax,4), %eax
+# endif
+#endif
+
        sub     %ecx, %eax
        RETURN
 
 L(Byte4):
-#ifdef USE_AS_STRNCMP
-       cmp     $4, %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $4, REM
        jbe     L(eq)
 #endif
        movzx   4(%eax), %ecx
        movzx   4(%edx), %eax
 
+#if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
+# ifdef PIC
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%eax,4), %eax
+# else
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%eax,4), %eax
+# endif
+#endif
+
        sub     %ecx, %eax
        RETURN
 
 L(Byte5):
-#ifdef USE_AS_STRNCMP
-       cmp     $5, %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $5, REM
        jbe     L(eq)
 #endif
        movzx   5(%eax), %ecx
        movzx   5(%edx), %eax
 
+#if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
+# ifdef PIC
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%eax,4), %eax
+# else
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%eax,4), %eax
+# endif
+#endif
+
        sub     %ecx, %eax
        RETURN
 
 L(Byte6):
-#ifdef USE_AS_STRNCMP
-       cmp     $6, %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $6, REM
        jbe     L(eq)
 #endif
        movzx   6(%eax), %ecx
        movzx   6(%edx), %eax
 
+#if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
+# ifdef PIC
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%eax,4), %eax
+# else
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%eax,4), %eax
+# endif
+#endif
+
        sub     %ecx, %eax
        RETURN
 
 L(2next_8_bytes):
        add     $8, %eax
        add     $8, %edx
-#ifdef USE_AS_STRNCMP
-       cmp     $8, %ebp
-       lea     -8(%ebp), %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $8, REM
+       lea     -8(REM), REM
        jbe     L(eq)
 #endif
 
@@ -2017,194 +2318,455 @@ L(2next_8_bytes):
        test    $0x40, %ch
        jnz     L(Byte6)
 
-#ifdef USE_AS_STRNCMP
-       cmp     $7, %ebp
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       cmp     $7, REM
        jbe     L(eq)
 #endif
        movzx   7(%eax), %ecx
        movzx   7(%edx), %eax
 
+#if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
+# ifdef PIC
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%eax,4), %eax
+# else
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%eax,4), %eax
+# endif
+#endif
+
        sub     %ecx, %eax
        RETURN
 
+#ifdef USE_AS_STRNCMP
+L(neq_sncmp):
+#endif
 L(neq):
        mov     $1, %eax
        ja      L(neq_bigger)
        neg     %eax
 L(neq_bigger):
-#ifdef USE_AS_STRNCMP
-       POP     (%ebp)
+#if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
+       addl    $4, %esp
+       cfi_adjust_cfa_offset (-4)
+#endif
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       POP     (REM)
+#endif
+#if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
+# ifdef PIC
+       POP     (%ebx)
+# endif
 #endif
        ret
 
-#ifdef USE_AS_STRNCMP
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
        .p2align 4
        cfi_restore_state
 L(more8byteseq):
+
+# ifdef USE_AS_STRNCASECMP_L
+       addl    $4, %esp
+       cfi_adjust_cfa_offset (-4)
+# endif
        POP     (%esi)
        POP     (%edi)
-       POP     (%ebx)
+# ifdef USE_AS_STRNCMP
+       POP     (FLAGS)
+# endif
 #endif
 
+#ifdef USE_AS_STRNCMP
+L(eq_sncmp):
+#endif
 L(eq):
 
-#ifdef USE_AS_STRNCMP
-       POP     (%ebp)
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
+       POP     (REM)
+#endif
+#if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
+# ifdef PIC
+       POP     (%ebx)
+# endif
 #endif
        xorl    %eax, %eax
        ret
 
-#ifdef USE_AS_STRNCMP
+#if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
        .p2align 4
-       CFI_PUSH (%ebp)
+# if defined USE_AS_STRNCASECMP_L && defined PIC
+       CFI_PUSH (%ebx)
+# endif
+       CFI_PUSH (REM)
 L(less16bytes_sncmp):
-       test    %ebp, %ebp
-       jz      L(eq)
+# ifdef USE_AS_STRNCASECMP_L
+       PUSH    (%esi)
+# endif
+       test    REM, REM
+       jz      L(eq_sncmp)
 
        movzbl  (%eax), %ecx
+# ifdef USE_AS_STRNCASECMP_L
+       movzbl  (%edx), %esi
+#  ifdef PIC
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%esi,4), %esi
+#  else
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%esi,4), %esi
+#  endif
+       cmpl    %ecx, %esi
+# else
        cmpb    %cl, (%edx)
-       jne     L(neq)
+# endif
+       jne     L(neq_sncmp)
        test    %cl, %cl
        je      L(eq)
 
-       cmp     $1, %ebp
-       je      L(eq)
+       cmp     $1, REM
+       je      L(eq_sncmp)
 
        movzbl  1(%eax), %ecx
+# ifdef USE_AS_STRNCASECMP_L
+       movzbl  1(%edx), %esi
+#  ifdef PIC
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%esi,4), %esi
+#  else
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%esi,4), %esi
+#  endif
+       cmpl    %ecx, %esi
+# else
        cmpb    %cl, 1(%edx)
-       jne     L(neq)
+# endif
+       jne     L(neq_sncmp)
        test    %cl, %cl
-       je      L(eq)
+       je      L(eq_sncmp)
 
-       cmp     $2, %ebp
-       je      L(eq)
+       cmp     $2, REM
+       je      L(eq_sncmp)
 
        movzbl  2(%eax), %ecx
+# ifdef USE_AS_STRNCASECMP_L
+       movzbl  2(%edx), %esi
+#  ifdef PIC
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%esi,4), %esi
+#  else
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%esi,4), %esi
+#  endif
+       cmpl    %ecx, %esi
+# else
        cmpb    %cl, 2(%edx)
-       jne     L(neq)
+# endif
+       jne     L(neq_sncmp)
        test    %cl, %cl
-       je      L(eq)
+       je      L(eq_sncmp)
 
-       cmp     $3, %ebp
-       je      L(eq)
+       cmp     $3, REM
+       je      L(eq_sncmp)
 
        movzbl  3(%eax), %ecx
+# ifdef USE_AS_STRNCASECMP_L
+       movzbl  3(%edx), %esi
+#  ifdef PIC
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%esi,4), %esi
+#  else
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%esi,4), %esi
+#  endif
+       cmpl    %ecx, %esi
+# else
        cmpb    %cl, 3(%edx)
-       jne     L(neq)
+# endif
+       jne     L(neq_sncmp)
        test    %cl, %cl
-       je      L(eq)
+       je      L(eq_sncmp)
 
-       cmp     $4, %ebp
-       je      L(eq)
+       cmp     $4, REM
+       je      L(eq_sncmp)
 
        movzbl  4(%eax), %ecx
+# ifdef USE_AS_STRNCASECMP_L
+       movzbl  4(%edx), %esi
+#  ifdef PIC
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%esi,4), %esi
+#  else
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%esi,4), %esi
+#  endif
+       cmpl    %ecx, %esi
+# else
        cmpb    %cl, 4(%edx)
-       jne     L(neq)
+# endif
+       jne     L(neq_sncmp)
        test    %cl, %cl
-       je      L(eq)
+       je      L(eq_sncmp)
 
-       cmp     $5, %ebp
-       je      L(eq)
+       cmp     $5, REM
+       je      L(eq_sncmp)
 
        movzbl  5(%eax), %ecx
+# ifdef USE_AS_STRNCASECMP_L
+       movzbl  5(%edx), %esi
+#  ifdef PIC
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%esi,4), %esi
+#  else
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%esi,4), %esi
+#  endif
+       cmpl    %ecx, %esi
+# else
        cmpb    %cl, 5(%edx)
-       jne     L(neq)
+# endif
+       jne     L(neq_sncmp)
        test    %cl, %cl
-       je      L(eq)
+       je      L(eq_sncmp)
 
-       cmp     $6, %ebp
-       je      L(eq)
+       cmp     $6, REM
+       je      L(eq_sncmp)
 
        movzbl  6(%eax), %ecx
+# ifdef USE_AS_STRNCASECMP_L
+       movzbl  6(%edx), %esi
+#  ifdef PIC
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%esi,4), %esi
+#  else
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%esi,4), %esi
+#  endif
+       cmpl    %ecx, %esi
+# else
        cmpb    %cl, 6(%edx)
-       jne     L(neq)
+# endif
+       jne     L(neq_sncmp)
        test    %cl, %cl
-       je      L(eq)
+       je      L(eq_sncmp)
 
-       cmp     $7, %ebp
-       je      L(eq)
+       cmp     $7, REM
+       je      L(eq_sncmp)
 
        movzbl  7(%eax), %ecx
+# ifdef USE_AS_STRNCASECMP_L
+       movzbl  7(%edx), %esi
+#  ifdef PIC
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%esi,4), %esi
+#  else
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%esi,4), %esi
+#  endif
+       cmpl    %ecx, %esi
+# else
        cmpb    %cl, 7(%edx)
-       jne     L(neq)
+# endif
+       jne     L(neq_sncmp)
        test    %cl, %cl
-       je      L(eq)
+       je      L(eq_sncmp)
 
 
-       cmp     $8, %ebp
-       je      L(eq)
+       cmp     $8, REM
+       je      L(eq_sncmp)
 
        movzbl  8(%eax), %ecx
+# ifdef USE_AS_STRNCASECMP_L
+       movzbl  8(%edx), %esi
+#  ifdef PIC
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%esi,4), %esi
+#  else
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%esi,4), %esi
+#  endif
+       cmpl    %ecx, %esi
+# else
        cmpb    %cl, 8(%edx)
-       jne     L(neq)
+# endif
+       jne     L(neq_sncmp)
        test    %cl, %cl
-       je      L(eq)
+       je      L(eq_sncmp)
 
-       cmp     $9, %ebp
-       je      L(eq)
+       cmp     $9, REM
+       je      L(eq_sncmp)
 
        movzbl  9(%eax), %ecx
+# ifdef USE_AS_STRNCASECMP_L
+       movzbl  9(%edx), %esi
+#  ifdef PIC
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%esi,4), %esi
+#  else
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%esi,4), %esi
+#  endif
+       cmpl    %ecx, %esi
+# else
        cmpb    %cl, 9(%edx)
-       jne     L(neq)
+# endif
+       jne     L(neq_sncmp)
        test    %cl, %cl
-       je      L(eq)
+       je      L(eq_sncmp)
 
-       cmp     $10, %ebp
-       je      L(eq)
+       cmp     $10, REM
+       je      L(eq_sncmp)
 
        movzbl  10(%eax), %ecx
+# ifdef USE_AS_STRNCASECMP_L
+       movzbl  10(%edx), %esi
+#  ifdef PIC
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%esi,4), %esi
+#  else
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%esi,4), %esi
+#  endif
+       cmpl    %ecx, %esi
+# else
        cmpb    %cl, 10(%edx)
-       jne     L(neq)
+# endif
+       jne     L(neq_sncmp)
        test    %cl, %cl
-       je      L(eq)
+       je      L(eq_sncmp)
 
-       cmp     $11, %ebp
-       je      L(eq)
+       cmp     $11, REM
+       je      L(eq_sncmp)
 
        movzbl  11(%eax), %ecx
+# ifdef USE_AS_STRNCASECMP_L
+       movzbl  11(%edx), %esi
+#  ifdef PIC
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%esi,4), %esi
+#  else
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%esi,4), %esi
+#  endif
+       cmpl    %ecx, %esi
+# else
        cmpb    %cl, 11(%edx)
-       jne     L(neq)
+# endif
+       jne     L(neq_sncmp)
        test    %cl, %cl
-       je      L(eq)
+       je      L(eq_sncmp)
 
 
-       cmp     $12, %ebp
-       je      L(eq)
+       cmp     $12, REM
+       je      L(eq_sncmp)
 
        movzbl  12(%eax), %ecx
+# ifdef USE_AS_STRNCASECMP_L
+       movzbl  12(%edx), %esi
+#  ifdef PIC
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%esi,4), %esi
+#  else
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%esi,4), %esi
+#  endif
+       cmpl    %ecx, %esi
+# else
        cmpb    %cl, 12(%edx)
-       jne     L(neq)
+# endif
+       jne     L(neq_sncmp)
        test    %cl, %cl
-       je      L(eq)
+       je      L(eq_sncmp)
 
-       cmp     $13, %ebp
-       je      L(eq)
+       cmp     $13, REM
+       je      L(eq_sncmp)
 
        movzbl  13(%eax), %ecx
+# ifdef USE_AS_STRNCASECMP_L
+       movzbl  13(%edx), %esi
+#  ifdef PIC
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%esi,4), %esi
+#  else
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%esi,4), %esi
+#  endif
+       cmpl    %ecx, %esi
+# else
        cmpb    %cl, 13(%edx)
-       jne     L(neq)
+# endif
+       jne     L(neq_sncmp)
        test    %cl, %cl
-       je      L(eq)
+       je      L(eq_sncmp)
 
-       cmp     $14, %ebp
-       je      L(eq)
+       cmp     $14, REM
+       je      L(eq_sncmp)
 
        movzbl  14(%eax), %ecx
+# ifdef USE_AS_STRNCASECMP_L
+       movzbl  14(%edx), %esi
+#  ifdef PIC
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%esi,4), %esi
+#  else
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%esi,4), %esi
+#  endif
+       cmpl    %ecx, %esi
+# else
        cmpb    %cl, 14(%edx)
-       jne     L(neq)
+# endif
+       jne     L(neq_sncmp)
        test    %cl, %cl
-       je      L(eq)
+       je      L(eq_sncmp)
 
-       cmp     $15, %ebp
-       je      L(eq)
+       cmp     $15, REM
+       je      L(eq_sncmp)
 
        movzbl  15(%eax), %ecx
+# ifdef USE_AS_STRNCASECMP_L
+       movzbl  15(%edx), %esi
+#  ifdef PIC
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower@GOTOFF+128*4(%ebx,%esi,4), %esi
+#  else
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%ecx,4), %ecx
+       movl    _nl_C_LC_CTYPE_tolower+128*4(,%esi,4), %esi
+#  endif
+       cmpl    %ecx, %esi
+# else
        cmpb    %cl, 15(%edx)
-       jne     L(neq)
+# endif
+       jne     L(neq_sncmp)
 
-       POP     (%ebp)
+# ifdef USE_AS_STRNCASECMP_L
+L(eq_sncmp):
+       POP     (%esi)
+# endif
+       POP     (REM)
+# if defined USE_AS_STRNCASECMP_L && defined PIC
+       POP     (%ebx)
+# endif
        xor     %eax, %eax
        ret
+
+# ifdef USE_AS_STRNCASECMP_L
+       .p2align 4
+#  ifdef PIC
+       CFI_PUSH (%ebx)
+#  endif
+       CFI_PUSH (REM)
+       CFI_PUSH (%esi)
+L(neq_sncmp):
+       mov     $1, %eax
+       mov     $-1, %edx
+       cmovna  %edx, %eax
+       POP     (%esi)
+       POP     (REM)
+#  ifdef PIC
+       POP     (%ebx)
+#  endif
+       ret
+# endif
 #endif
 
 END (STRCMP)