inffast.s

测试过的

        cmpl    $32, bitslong_r
        ja      .L_get_length_code_mmx  /* if (32 < bits) */

        movd    bitslong_r, tmp_mm
        movd    (in_r), %mm7
        addl    $4, in_r
        psllq   tmp_mm, %mm7
        addl    $32, bitslong_r
        por     %mm7, hold_mm           /* hold_mm |= *((uint *)in)++ << bits */

.L_get_length_code_mmx:
        pand    hold_mm, lmask_mm
        movd    lmask_mm, %eax
        movq    lmask2_mm, lmask_mm
        movl    (%ebx,%eax,4), %eax     /* eax = lcode[hold & lmask] */

.L_dolen_mmx:
        movzbl  %ah, %ecx               /* ecx = this.bits */
        movd    %ecx, used_mm
        subl    %ecx, bitslong_r        /* bits -= this.bits */

        testb   %al, %al
        jnz     .L_test_for_length_base_mmx /* if (op != 0) 45.7% */

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

.L_test_for_length_base_mmx:
#define len_r  %edx
        movl    %eax, len_r             /* len = this */
        shrl    $16, len_r              /* len = this.val */

        testb   $16, %al
        jz      .L_test_for_second_level_length_mmx /* if ((op & 16) == 0) 8% */
        andl    $15, %eax               /* op &= 15 */
        jz      .L_decode_distance_mmx  /* if (!op) */

        psrlq   used_mm, hold_mm        /* hold_mm >>= last bit length */
        movd    %eax, used_mm
        movd    hold_mm, %ecx
        subl    %eax, bitslong_r
        andl    .L_mask(,%eax,4), %ecx
        addl    %ecx, len_r             /* len += hold & mask[op] */

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

        cmpl    $32, bitslong_r
        ja      .L_get_dist_code_mmx    /* if (32 < bits) */

        movd    bitslong_r, tmp_mm
        movd    (in_r), %mm7
        addl    $4, in_r
        psllq   tmp_mm, %mm7
        addl    $32, bitslong_r
        por     %mm7, hold_mm           /* hold_mm |= *((uint *)in)++ << bits */

.L_get_dist_code_mmx:
        movl    dcode(%esp), %ebx       /* ebx = dcode */
        pand    hold_mm, dmask_mm
        movd    dmask_mm, %eax
        movq    dmask2_mm, dmask_mm
        movl    (%ebx,%eax,4), %eax     /* eax = dcode[hold & lmask] */

.L_dodist_mmx:
#define dist_r %ebx
        movzbl  %ah, %ecx               /* ecx = this.bits */
        movl    %eax, dist_r
        shrl    $16, dist_r             /* dist  = this.val */
        subl    %ecx, bitslong_r        /* bits -= this.bits */
        movd    %ecx, used_mm

        testb   $16, %al                /* if ((op & 16) == 0) */
        jz      .L_test_for_second_level_dist_mmx
        andl    $15, %eax               /* op &= 15 */
        jz      .L_check_dist_one_mmx

.L_add_bits_to_dist_mmx:
        psrlq   used_mm, hold_mm        /* hold_mm >>= last bit length */
        movd    %eax, used_mm           /* save bit length of current op */
        movd    hold_mm, %ecx           /* get the next bits on input stream */
        subl    %eax, bitslong_r        /* bits -= op bits */
        andl    .L_mask(,%eax,4), %ecx  /* ecx   = hold & mask[op] */
        addl    %ecx, dist_r            /* dist += hold & mask[op] */

.L_check_window_mmx:
        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_mmx      /* if (dist > nbytes) 4.2% */

        movl    len_r, %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 */
        movl    lcode(%esp), %ebx       /* move lcode back to %ebx, toss dist */
        jmp     .L_while_test_mmx

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

        decl    out_r
        movl    len_r, %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

        movl    lcode(%esp), %ebx       /* move lcode back to %ebx, toss dist */
        jmp     .L_while_test_mmx

.align 16,0x90
.L_test_for_second_level_length_mmx:
        testb   $64, %al
        jnz     .L_test_for_end_of_block  /* if ((op & 64) != 0) */

        andl    $15, %eax
        psrlq   used_mm, hold_mm        /* hold_mm >>= last bit length */
        movd    hold_mm, %ecx
        andl    .L_mask(,%eax,4), %ecx
        addl    len_r, %ecx
        movl    (%ebx,%ecx,4), %eax     /* eax = lcode[hold & lmask] */
        jmp     .L_dolen_mmx

.align 16,0x90
.L_test_for_second_level_dist_mmx:
        testb   $64, %al
        jnz     .L_invalid_distance_code  /* if ((op & 64) != 0) */

        andl    $15, %eax
        psrlq   used_mm, hold_mm        /* hold_mm >>= last bit length */
        movd    hold_mm, %ecx
        andl    .L_mask(,%eax,4), %ecx
        movl    dcode(%esp), %eax       /* ecx = dcode */
        addl    dist_r, %ecx
        movl    (%eax,%ecx,4), %eax     /* eax = lcode[hold & lmask] */
        jmp     .L_dodist_mmx

.align 16,0x90
.L_clip_window_mmx:
#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_mmx /* if (write != 0) */

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

        cmpl    nbytes_r, len_r
        jbe     .L_do_copy1_mmx         /* 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_mmx

        cmpl    nbytes_r, len_r
        jbe     .L_do_copy1_mmx         /* 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_mmx

.L_wrap_around_window_mmx:
#define write_r %eax
        movl    write(%esp), write_r
        cmpl    write_r, nbytes_r
        jbe     .L_contiguous_in_window_mmx /* 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

        cmpl    nbytes_r, len_r
        jbe     .L_do_copy1_mmx         /* 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_mmx         /* 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_mmx

.L_contiguous_in_window_mmx:
#define write_r %eax
        addl    write_r, from_r
        subl    nbytes_r, from_r        /* from += write - nbytes */
#undef write_r

        cmpl    nbytes_r, len_r
        jbe     .L_do_copy1_mmx         /* 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_mmx:
#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 */
        movl    lcode(%esp), %ebx       /* move lcode back to %ebx, toss dist */
        jmp     .L_while_test_mmx

#undef hold_r
#undef bitslong_r

#endif /* USE_MMX || RUN_TIME_MMX */


/*** USE_MMX, NO_MMX, and RUNTIME_MMX from here on ***/

.L_invalid_distance_code:
        /* else {
         *   strm->msg = "invalid distance code";
         *   state->mode = BAD;
         * }
         */
        movl    $.L_invalid_distance_code_msg, %ecx
        movl    $INFLATE_MODE_BAD, %edx
        jmp     .L_update_stream_state

.L_test_for_end_of_block:
        /* else if (op & 32) {
         *   state->mode = TYPE;
         *   break;
         * }
         */
        testb   $32, %al
        jz      .L_invalid_literal_length_code  /* if ((op & 32) == 0) */

        movl    $0, %ecx
        movl    $INFLATE_MODE_TYPE, %edx
        jmp     .L_update_stream_state

.L_invalid_literal_length_code:
        /* else {
         *   strm->msg = "invalid literal/length code";
         *   state->mode = BAD;
         * }
         */
        movl    $.L_invalid_literal_length_code_msg, %ecx
        movl    $INFLATE_MODE_BAD, %edx
        jmp     .L_update_stream_state

.L_invalid_distance_too_far:
        /* strm->msg = "invalid distance too far back";
         * state->mode = BAD;
         */
        movl    in(%esp), in_r          /* from_r has in's reg, put in back */
        movl    $.L_invalid_distance_too_far_msg, %ecx
        movl    $INFLATE_MODE_BAD, %edx
        jmp     .L_update_stream_state

.L_update_stream_state:
        /* set strm->msg = %ecx, strm->state->mode = %edx */
        movl    strm_sp(%esp), %eax
        testl   %ecx, %ecx              /* if (msg != NULL) */
        jz      .L_skip_msg
        movl    %ecx, msg_strm(%eax)    /* strm->msg = msg */
.L_skip_msg:
        movl    state_strm(%eax), %eax  /* state = strm->state */
        movl    %edx, mode_state(%eax)  /* state->mode = edx (BAD | TYPE) */
        jmp     .L_break_loop

.align 32,0x90
.L_break_loop:

/*
 * Regs:
 *
 * bits = %ebp when mmx, and in %ebx when non-mmx
 * hold = %hold_mm when mmx, and in %ebp when non-mmx
 * in   = %esi
 * out  = %edi
 */

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

#if defined( RUN_TIME_MMX )

        cmpl    $DO_USE_MMX, inflate_fast_use_mmx
        jne     .L_update_next_in

#endif /* RUN_TIME_MMX */

        movl    %ebp, %ebx

.L_update_next_in:

#endif

#define strm_r  %eax
#define state_r %edx

        /* len = bits >> 3;
         * in -= len;
         * bits -= len << 3;
         * hold &= (1U << bits) - 1;
         * state->hold = hold;
         * state->bits = bits;
         * strm->next_in = in;
         * strm->next_out = out;
         */
        movl    strm_sp(%esp), strm_r
        movl    %ebx, %ecx
        movl    state_strm(strm_r), state_r
        shrl    $3, %ecx
        subl    %ecx, in_r
        shll    $3, %ecx
        subl    %ecx, %ebx
        movl    out_r, next_out_strm(strm_r)
        movl    %ebx, bits_state(state_r)
        movl    %ebx, %ecx

        leal    buf(%esp), %ebx
        cmpl    %ebx, last(%esp)
        jne     .L_buf_not_used         /* if buf != last */

        subl    %ebx, in_r              /* in -= buf */
        movl    next_in_strm(strm_r), %ebx
        movl    %ebx, last(%esp)        /* last = strm->next_in */
        addl    %ebx, in_r              /* in += strm->next_in */
        movl    avail_in_strm(strm_r), %ebx
        subl    $11, %ebx
        addl    %ebx, last(%esp)    /* last = &strm->next_in[ avail_in - 11 ] */

.L_buf_not_used:
        movl    in_r, next_in_strm(strm_r)

        movl    $1, %ebx
        shll    %cl, %ebx
        decl    %ebx

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

#if defined( RUN_TIME_MMX )

        cmpl    $DO_USE_MMX, inflate_fast_use_mmx
        jne     .L_update_hold

#endif /* RUN_TIME_MMX */

        psrlq   used_mm, hold_mm        /* hold_mm >>= last bit length */
        movd    hold_mm, %ebp

        emms

.L_update_hold:

#endif /* USE_MMX || RUN_TIME_MMX */

        andl    %ebx, %ebp
        movl    %ebp, hold_state(state_r)

#define last_r %ebx

        /* strm->avail_in = in < last ? 11 + (last - in) : 11 - (in - last) */
        movl    last(%esp), last_r
        cmpl    in_r, last_r
        jbe     .L_last_is_smaller     /* if (in >= last) */

        subl    in_r, last_r           /* last -= in */
        addl    $11, last_r            /* last += 11 */
        movl    last_r, avail_in_strm(strm_r)
        jmp     .L_fixup_out
.L_last_is_smaller:
        subl    last_r, in_r           /* in -= last */
        negl    in_r                   /* in = -in */
        addl    $11, in_r              /* in += 11 */
        movl    in_r, avail_in_strm(strm_r)

#undef last_r
#define end_r %ebx

.L_fixup_out:
        /* strm->avail_out = out < end ? 257 + (end - out) : 257 - (out - end)*/
        movl    end(%esp), end_r
        cmpl    out_r, end_r
        jbe     .L_end_is_smaller      /* if (out >= end) */

        subl    out_r, end_r           /* end -= out */
        addl    $257, end_r            /* end += 257 */
        movl    end_r, avail_out_strm(strm_r)
        jmp     .L_done
.L_end_is_smaller:
        subl    end_r, out_r           /* out -= end */
        negl    out_r                  /* out = -out */
        addl    $257, out_r            /* out += 257 */
        movl    out_r, avail_out_strm(strm_r)

#undef end_r
#undef strm_r
#undef state_r

.L_done:
        addl    $local_var_size, %esp
        popf
        popl    %ebx
        popl    %ebp
        popl    %esi
        popl    %edi
        ret

#if defined( GAS_ELF )
/* elf info */
.type inflate_fast,@function
.size inflate_fast,.-inflate_fast
#endif