inffast.s

测试过的

        movl    (%ecx,%edx,4), %eax     /* eax = lcode[hold & lmask] */

.L_dolen:
        /* regs: %esi = in, %ebp = hold, %bl = bits, %edi = out
         *
         * dolen:
         *    bits -= this.bits;
         *    hold >>= this.bits
         */
        movb    %ah, %cl                /* cl = this.bits */
        subb    %ah, bits_r             /* bits -= this.bits */
        shrl    %cl, hold_r             /* hold >>= this.bits */

        /* check if op is a literal
         * if (op == 0) {
         *    PUP(out) = this.val;
         *  }
         */
        testb   %al, %al
        jnz     .L_test_for_length_base /* if (op != 0) 45.7% */

        shrl    $16, %eax               /* output this.val char */
        stosb
        jmp     .L_while_test

.L_test_for_length_base:
        /* regs: %esi = in, %ebp = hold, %bl = bits, %edi = out, %edx = len
         *
         * else if (op & 16) {
         *   len = this.val
         *   op &= 15
         *   if (op) {
         *     if (op > bits) {
         *       hold |= *((unsigned short *)in)++ << bits;
         *       bits += 16
         *     }
         *     len += hold & mask[op];
         *     bits -= op;
         *     hold >>= op;
         *   }
         */
#define len_r %edx
        movl    %eax, len_r             /* len = this */
        shrl    $16, len_r              /* len = this.val */
        movb    %al, %cl

        testb   $16, %al
        jz      .L_test_for_second_level_length /* if ((op & 16) == 0) 8% */
        andb    $15, %cl                /* op &= 15 */
        jz      .L_save_len             /* if (!op) */
        cmpb    %cl, bits_r
        jae     .L_add_bits_to_len      /* if (op <= bits) */

        movb    %cl, %ch                /* stash op in ch, freeing cl */
        xorl    %eax, %eax
        lodsw                           /* al = *(ushort *)in++ */
        movb    bits_r, %cl             /* cl = bits, needs it for shifting */
        addb    $16, bits_r             /* bits += 16 */
        shll    %cl, %eax
        orl     %eax, hold_r            /* hold |= *((ushort *)in)++ << bits */
        movb    %ch, %cl                /* move op back to ecx */

.L_add_bits_to_len:
        movl    $1, %eax
        shll    %cl, %eax
        decl    %eax
        subb    %cl, bits_r
        andl    hold_r, %eax            /* eax &= hold */
        shrl    %cl, hold_r
        addl    %eax, len_r             /* len += hold & mask[op] */

.L_save_len:
        movl    len_r, len(%esp)        /* save len */
#undef  len_r

.L_decode_distance:
        /* regs: %esi = in, %ebp = hold, %bl = bits, %edi = out, %edx = dist
         *
         *   if (bits < 15) {
         *     hold |= *((unsigned short *)in)++ << bits;
         *     bits += 16
         *   }
         *   this = dcode[hold & dmask];
         * dodist:
         *   bits -= this.bits;
         *   hold >>= this.bits;
         *   op = this.op;
         */

        cmpb    $15, bits_r
        ja      .L_get_distance_code    /* if (15 < bits) */

        xorl    %eax, %eax
        lodsw                           /* al = *(ushort *)in++ */
        movb    bits_r, %cl             /* cl = bits, needs it for shifting */
        addb    $16, bits_r             /* bits += 16 */
        shll    %cl, %eax
        orl     %eax, hold_r            /* hold |= *((ushort *)in)++ << bits */

.L_get_distance_code:
        movl    dmask(%esp), %edx       /* edx = dmask */
        movl    dcode(%esp), %ecx       /* ecx = dcode */
        andl    hold_r, %edx            /* edx &= hold */
        movl    (%ecx,%edx,4), %eax     /* eax = dcode[hold & dmask] */

#define dist_r %edx
.L_dodist:
        movl    %eax, dist_r            /* dist = this */
        shrl    $16, dist_r             /* dist = this.val */
        movb    %ah, %cl
        subb    %ah, bits_r             /* bits -= this.bits */
        shrl    %cl, hold_r             /* hold >>= this.bits */

        /* if (op & 16) {
         *   dist = this.val
         *   op &= 15
         *   if (op > bits) {
         *     hold |= *((unsigned short *)in)++ << bits;
         *     bits += 16
         *   }
         *   dist += hold & mask[op];
         *   bits -= op;
         *   hold >>= op;
         */
        movb    %al, %cl                /* cl = this.op */

        testb   $16, %al                /* if ((op & 16) == 0) */
        jz      .L_test_for_second_level_dist
        andb    $15, %cl                /* op &= 15 */
        jz      .L_check_dist_one
        cmpb    %cl, bits_r
        jae     .L_add_bits_to_dist     /* if (op <= bits) 97.6% */

        movb    %cl, %ch                /* stash op in ch, freeing cl */
        xorl    %eax, %eax
        lodsw                           /* al = *(ushort *)in++ */
        movb    bits_r, %cl             /* cl = bits, needs it for shifting */
        addb    $16, bits_r             /* bits += 16 */
        shll    %cl, %eax
        orl     %eax, hold_r            /* hold |= *((ushort *)in)++ << bits */
        movb    %ch, %cl                /* move op back to ecx */

.L_add_bits_to_dist:
        movl    $1, %eax
        shll    %cl, %eax
        decl    %eax                    /* (1 << op) - 1 */
        subb    %cl, bits_r
        andl    hold_r, %eax            /* eax &= hold */
        shrl    %cl, hold_r
        addl    %eax, dist_r            /* dist += hold & ((1 << op) - 1) */
        jmp     .L_check_window

.L_check_window:
        /* regs: %esi = from, %ebp = hold, %bl = bits, %edi = out, %edx = dist
         *       %ecx = nbytes
         *
         * nbytes = out - beg;
         * if (dist <= nbytes) {
         *   from = out - dist;
         *   do {
         *     PUP(out) = PUP(from);
         *   } while (--len > 0) {
         * }
         */

        movl    in_r, in(%esp)          /* save in so from can use it's reg */
        movl    out_r, %eax
        subl    beg(%esp), %eax         /* nbytes = out - beg */

        cmpl    dist_r, %eax
        jb      .L_clip_window          /* if (dist > nbytes) 4.2% */

        movl    len(%esp), %ecx
        movl    out_r, from_r
        subl    dist_r, from_r          /* from = out - dist */

        subl    $3, %ecx
        movb    (from_r), %al
        movb    %al, (out_r)
        movb    1(from_r), %al
        movb    2(from_r), %dl
        addl    $3, from_r
        movb    %al, 1(out_r)
        movb    %dl, 2(out_r)
        addl    $3, out_r
        rep     movsb

        movl    in(%esp), in_r          /* move in back to %esi, toss from */
        jmp     .L_while_test

.align 16,0x90
.L_check_dist_one:
        cmpl    $1, dist_r
        jne     .L_check_window
        cmpl    out_r, beg(%esp)
        je      .L_check_window

        decl    out_r
        movl    len(%esp), %ecx
        movb    (out_r), %al
        subl    $3, %ecx

        movb    %al, 1(out_r)
        movb    %al, 2(out_r)
        movb    %al, 3(out_r)
        addl    $4, out_r
        rep     stosb

        jmp     .L_while_test

.align 16,0x90
.L_test_for_second_level_length:
        /* else if ((op & 64) == 0) {
         *   this = lcode[this.val + (hold & mask[op])];
         * }
         */
        testb   $64, %al
        jnz     .L_test_for_end_of_block  /* if ((op & 64) != 0) */

        movl    $1, %eax
        shll    %cl, %eax
        decl    %eax
        andl    hold_r, %eax            /* eax &= hold */
        addl    %edx, %eax              /* eax += this.val */
        movl    lcode(%esp), %edx       /* edx = lcode */
        movl    (%edx,%eax,4), %eax     /* eax = lcode[val + (hold&mask[op])] */
        jmp     .L_dolen

.align 16,0x90
.L_test_for_second_level_dist:
        /* else if ((op & 64) == 0) {
         *   this = dcode[this.val + (hold & mask[op])];
         * }
         */
        testb   $64, %al
        jnz     .L_invalid_distance_code  /* if ((op & 64) != 0) */

        movl    $1, %eax
        shll    %cl, %eax
        decl    %eax
        andl    hold_r, %eax            /* eax &= hold */
        addl    %edx, %eax              /* eax += this.val */
        movl    dcode(%esp), %edx       /* edx = dcode */
        movl    (%edx,%eax,4), %eax     /* eax = dcode[val + (hold&mask[op])] */
        jmp     .L_dodist

.align 16,0x90
.L_clip_window:
        /* regs: %esi = from, %ebp = hold, %bl = bits, %edi = out, %edx = dist
         *       %ecx = nbytes
         *
         * else {
         *   if (dist > wsize) {
         *     invalid distance
         *   }
         *   from = window;
         *   nbytes = dist - nbytes;
         *   if (write == 0) {
         *     from += wsize - nbytes;
         */
#define nbytes_r %ecx
        movl    %eax, nbytes_r
        movl    wsize(%esp), %eax       /* prepare for dist compare */
        negl    nbytes_r                /* nbytes = -nbytes */
        movl    window(%esp), from_r    /* from = window */

        cmpl    dist_r, %eax
        jb      .L_invalid_distance_too_far /* if (dist > wsize) */

        addl    dist_r, nbytes_r        /* nbytes = dist - nbytes */
        cmpl    $0, write(%esp)
        jne     .L_wrap_around_window   /* if (write != 0) */

        subl    nbytes_r, %eax
        addl    %eax, from_r            /* from += wsize - nbytes */

        /* regs: %esi = from, %ebp = hold, %bl = bits, %edi = out, %edx = dist
         *       %ecx = nbytes, %eax = len
         *
         *     if (nbytes < len) {
         *       len -= nbytes;
         *       do {
         *         PUP(out) = PUP(from);
         *       } while (--nbytes);
         *       from = out - dist;
         *     }
         *   }
         */
#define len_r %eax
        movl    len(%esp), len_r
        cmpl    nbytes_r, len_r
        jbe     .L_do_copy1             /* if (nbytes >= len) */

        subl    nbytes_r, len_r         /* len -= nbytes */
        rep     movsb
        movl    out_r, from_r
        subl    dist_r, from_r          /* from = out - dist */
        jmp     .L_do_copy1

        cmpl    nbytes_r, len_r
        jbe     .L_do_copy1             /* if (nbytes >= len) */

        subl    nbytes_r, len_r         /* len -= nbytes */
        rep     movsb
        movl    out_r, from_r
        subl    dist_r, from_r          /* from = out - dist */
        jmp     .L_do_copy1

.L_wrap_around_window:
        /* regs: %esi = from, %ebp = hold, %bl = bits, %edi = out, %edx = dist
         *       %ecx = nbytes, %eax = write, %eax = len
         *
         *   else if (write < nbytes) {
         *     from += wsize + write - nbytes;
         *     nbytes -= write;
         *     if (nbytes < len) {
         *       len -= nbytes;
         *       do {
         *         PUP(out) = PUP(from);
         *       } while (--nbytes);
         *       from = window;
         *       nbytes = write;
         *       if (nbytes < len) {
         *         len -= nbytes;
         *         do {
         *           PUP(out) = PUP(from);
         *         } while(--nbytes);
         *         from = out - dist;
         *       }
         *     }
         *   }
         */
#define write_r %eax
        movl    write(%esp), write_r
        cmpl    write_r, nbytes_r
        jbe     .L_contiguous_in_window /* if (write >= nbytes) */

        addl    wsize(%esp), from_r
        addl    write_r, from_r
        subl    nbytes_r, from_r        /* from += wsize + write - nbytes */
        subl    write_r, nbytes_r       /* nbytes -= write */
#undef write_r

        movl    len(%esp), len_r
        cmpl    nbytes_r, len_r
        jbe     .L_do_copy1             /* if (nbytes >= len) */

        subl    nbytes_r, len_r         /* len -= nbytes */
        rep     movsb
        movl    window(%esp), from_r    /* from = window */
        movl    write(%esp), nbytes_r   /* nbytes = write */
        cmpl    nbytes_r, len_r
        jbe     .L_do_copy1             /* if (nbytes >= len) */

        subl    nbytes_r, len_r         /* len -= nbytes */
        rep     movsb
        movl    out_r, from_r
        subl    dist_r, from_r          /* from = out - dist */
        jmp     .L_do_copy1

.L_contiguous_in_window:
        /* regs: %esi = from, %ebp = hold, %bl = bits, %edi = out, %edx = dist
         *       %ecx = nbytes, %eax = write, %eax = len
         *
         *   else {
         *     from += write - nbytes;
         *     if (nbytes < len) {
         *       len -= nbytes;
         *       do {
         *         PUP(out) = PUP(from);
         *       } while (--nbytes);
         *       from = out - dist;
         *     }
         *   }
         */
#define write_r %eax
        addl    write_r, from_r
        subl    nbytes_r, from_r        /* from += write - nbytes */
#undef write_r

        movl    len(%esp), len_r
        cmpl    nbytes_r, len_r
        jbe     .L_do_copy1             /* if (nbytes >= len) */

        subl    nbytes_r, len_r         /* len -= nbytes */
        rep     movsb
        movl    out_r, from_r
        subl    dist_r, from_r          /* from = out - dist */

.L_do_copy1:
        /* regs: %esi = from, %esi = in, %ebp = hold, %bl = bits, %edi = out
         *       %eax = len
         *
         *     while (len > 0) {
         *       PUP(out) = PUP(from);
         *       len--;
         *     }
         *   }
         * } while (in < last && out < end);
         */
#undef nbytes_r
#define in_r %esi
        movl    len_r, %ecx
        rep     movsb

        movl    in(%esp), in_r          /* move in back to %esi, toss from */
        jmp     .L_while_test

#undef len_r
#undef dist_r

#endif /* NO_MMX || RUN_TIME_MMX */


/*** MMX code ***/

#if defined( USE_MMX ) || defined( RUN_TIME_MMX )

.align 32,0x90
.L_init_mmx:
        emms

#undef  bits_r
#undef  bitslong_r
#define bitslong_r %ebp
#define hold_mm    %mm0
        movd    %ebp, hold_mm
        movl    %ebx, bitslong_r

#define used_mm   %mm1
#define dmask2_mm %mm2
#define lmask2_mm %mm3
#define lmask_mm  %mm4
#define dmask_mm  %mm5
#define tmp_mm    %mm6

        movd    lmask(%esp), lmask_mm
        movq    lmask_mm, lmask2_mm
        movd    dmask(%esp), dmask_mm
        movq    dmask_mm, dmask2_mm
        pxor    used_mm, used_mm
        movl    lcode(%esp), %ebx       /* ebx = lcode */
        jmp     .L_do_loop_mmx

.align 32,0x90
.L_while_test_mmx:
        /* while (in < last && out < end)
         */
        cmpl    out_r, end(%esp)
        jbe     .L_break_loop           /* if (out >= end) */

        cmpl    in_r, last(%esp)
        jbe     .L_break_loop

.L_do_loop_mmx:
        psrlq   used_mm, hold_mm        /* hold_mm >>= last bit length */