• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1;
2; jchuff-sse2.asm - Huffman entropy encoding (SSE2)
3;
4; Copyright (C) 2009-2011, 2014-2017, 2019, D. R. Commander.
5; Copyright (C) 2015, Matthieu Darbois.
6; Copyright (C) 2018, Matthias Räncker.
7;
8; Based on the x86 SIMD extension for IJG JPEG library
9; Copyright (C) 1999-2006, MIYASAKA Masaru.
10; For conditions of distribution and use, see copyright notice in jsimdext.inc
11;
12; This file should be assembled with NASM (Netwide Assembler),
13; can *not* be assembled with Microsoft's MASM or any compatible
14; assembler (including Borland's Turbo Assembler).
15; NASM is available from http://nasm.sourceforge.net/ or
16; http://sourceforge.net/project/showfiles.php?group_id=6208
17;
18; This file contains an SSE2 implementation for Huffman coding of one block.
19; The following code is based on jchuff.c; see jchuff.c for more details.
20
21%include "jsimdext.inc"
22
23struc working_state
24.next_output_byte:   resp 1     ; => next byte to write in buffer
25.free_in_buffer:     resp 1     ; # of byte spaces remaining in buffer
26.cur.put_buffer.simd resq 1     ; current bit accumulation buffer
27.cur.free_bits       resd 1     ; # of bits available in it
28.cur.last_dc_val     resd 4     ; last DC coef for each component
29.cinfo:              resp 1     ; dump_buffer needs access to this
30endstruc
31
32struc c_derived_tbl
33.ehufco:             resd 256   ; code for each symbol
34.ehufsi:             resb 256   ; length of code for each symbol
35; If no code has been allocated for a symbol S, ehufsi[S] contains 0
36endstruc
37
38; --------------------------------------------------------------------------
39    SECTION     SEG_CONST
40
41    GLOBAL_DATA(jconst_huff_encode_one_block)
42
43EXTN(jconst_huff_encode_one_block):
44
45    alignz      32
46
47jpeg_mask_bits dq 0x0000, 0x0001, 0x0003, 0x0007
48               dq 0x000f, 0x001f, 0x003f, 0x007f
49               dq 0x00ff, 0x01ff, 0x03ff, 0x07ff
50               dq 0x0fff, 0x1fff, 0x3fff, 0x7fff
51
52times 1 << 14 db 15
53times 1 << 13 db 14
54times 1 << 12 db 13
55times 1 << 11 db 12
56times 1 << 10 db 11
57times 1 <<  9 db 10
58times 1 <<  8 db  9
59times 1 <<  7 db  8
60times 1 <<  6 db  7
61times 1 <<  5 db  6
62times 1 <<  4 db  5
63times 1 <<  3 db  4
64times 1 <<  2 db  3
65times 1 <<  1 db  2
66times 1 <<  0 db  1
67times 1       db  0
68jpeg_nbits_table:
69times 1       db  0
70times 1 <<  0 db  1
71times 1 <<  1 db  2
72times 1 <<  2 db  3
73times 1 <<  3 db  4
74times 1 <<  4 db  5
75times 1 <<  5 db  6
76times 1 <<  6 db  7
77times 1 <<  7 db  8
78times 1 <<  8 db  9
79times 1 <<  9 db 10
80times 1 << 10 db 11
81times 1 << 11 db 12
82times 1 << 12 db 13
83times 1 << 13 db 14
84times 1 << 14 db 15
85
86    alignz      32
87
88%ifdef PIC
89%define NBITS(x)      nbits_base + x
90%else
91%define NBITS(x)      jpeg_nbits_table + x
92%endif
93%define MASK_BITS(x)  NBITS((x) * 8) + (jpeg_mask_bits - jpeg_nbits_table)
94
95; --------------------------------------------------------------------------
96    SECTION     SEG_TEXT
97    BITS        32
98
99%define mm_put_buffer     mm0
100%define mm_all_0xff       mm1
101%define mm_temp           mm2
102%define mm_nbits          mm3
103%define mm_code_bits      mm3
104%define mm_code           mm4
105%define mm_overflow_bits  mm5
106%define mm_save_nbits     mm6
107
108; Shorthand used to describe SIMD operations:
109; wN:  xmmN treated as eight signed 16-bit values
110; wN[i]:  perform the same operation on all eight signed 16-bit values, i=0..7
111; bN:  xmmN treated as 16 unsigned 8-bit values, or
112;      mmN treated as eight unsigned 8-bit values
113; bN[i]:  perform the same operation on all unsigned 8-bit values,
114;         i=0..15 (SSE register) or i=0..7 (MMX register)
115; Contents of SIMD registers are shown in memory order.
116
117; Fill the bit buffer to capacity with the leading bits from code, then output
118; the bit buffer and put the remaining bits from code into the bit buffer.
119;
120; Usage:
121; code - contains the bits to shift into the bit buffer (LSB-aligned)
122; %1 - temp register
123; %2 - low byte of temp register
124; %3 - second byte of temp register
125; %4-%8 (optional) - extra instructions to execute before the macro completes
126; %9 - the label to which to jump when the macro completes
127;
128; Upon completion, free_bits will be set to the number of remaining bits from
129; code, and put_buffer will contain those remaining bits.  temp and code will
130; be clobbered.
131;
132; This macro encodes any 0xFF bytes as 0xFF 0x00, as does the EMIT_BYTE()
133; macro in jchuff.c.
134
135%macro EMIT_QWORD 9
136%define %%temp   %1
137%define %%tempb  %2
138%define %%temph  %3
139    add         nbits, free_bits             ; nbits += free_bits;
140    neg         free_bits                    ; free_bits = -free_bits;
141    movq        mm_temp, mm_code             ; temp = code;
142    movd        mm_nbits, nbits              ; nbits --> MMX register
143    movd        mm_overflow_bits, free_bits  ; overflow_bits (temp register) = free_bits;
144    neg         free_bits                    ; free_bits = -free_bits;
145    psllq       mm_put_buffer, mm_nbits      ; put_buffer <<= nbits;
146    psrlq       mm_temp, mm_overflow_bits    ; temp >>= overflow_bits;
147    add         free_bits, 64                ; free_bits += 64;
148    por         mm_temp, mm_put_buffer       ; temp |= put_buffer;
149%ifidn %%temp, nbits_base
150    movd        mm_save_nbits, nbits_base    ; save nbits_base
151%endif
152    movq        mm_code_bits, mm_temp        ; code_bits (temp register) = temp;
153    movq        mm_put_buffer, mm_code       ; put_buffer = code;
154    pcmpeqb     mm_temp, mm_all_0xff         ; b_temp[i] = (b_temp[i] == 0xFF ? 0xFF : 0);
155    movq        mm_code, mm_code_bits        ; code = code_bits;
156    psrlq       mm_code_bits, 32             ; code_bits >>= 32;
157    pmovmskb    nbits, mm_temp               ; nbits = 0;  nbits |= ((b_temp[i] >> 7) << i);
158    movd        %%temp, mm_code_bits         ; temp = code_bits;
159    bswap       %%temp                       ; temp = htonl(temp);
160    test        nbits, nbits                 ; if (nbits != 0)  /* Some 0xFF bytes */
161    jnz         %%.SLOW                      ;   goto %%.SLOW
162    mov         dword [buffer], %%temp       ; *(uint32_t)buffer = temp;
163%ifidn %%temp, nbits_base
164    movd        nbits_base, mm_save_nbits    ; restore nbits_base
165%endif
166    %4
167    movd        nbits, mm_code               ; nbits = (uint32_t)(code);
168    %5
169    bswap       nbits                        ; nbits = htonl(nbits);
170    mov         dword [buffer + 4], nbits    ; *(uint32_t)(buffer + 4) = nbits;
171    lea         buffer, [buffer + 8]         ; buffer += 8;
172    %6
173    %7
174    %8
175    jmp %9                                   ; return
176%%.SLOW:
177    ; Execute the equivalent of the EMIT_BYTE() macro in jchuff.c for all 8
178    ; bytes in the qword.
179    mov         byte [buffer], %%tempb     ; buffer[0] = temp[0];
180    cmp         %%tempb, 0xFF              ; Set CF if temp[0] < 0xFF
181    mov         byte [buffer+1], 0         ; buffer[1] = 0;
182    sbb         buffer, -2                 ; buffer -= (-2 + (temp[0] < 0xFF ? 1 : 0));
183    mov         byte [buffer], %%temph     ; buffer[0] = temp[1];
184    cmp         %%temph, 0xFF              ; Set CF if temp[1] < 0xFF
185    mov         byte [buffer+1], 0         ; buffer[1] = 0;
186    sbb         buffer, -2                 ; buffer -= (-2 + (temp[1] < 0xFF ? 1 : 0));
187    shr         %%temp, 16                 ; temp >>= 16;
188    mov         byte [buffer], %%tempb     ; buffer[0] = temp[0];
189    cmp         %%tempb, 0xFF              ; Set CF if temp[0] < 0xFF
190    mov         byte [buffer+1], 0         ; buffer[1] = 0;
191    sbb         buffer, -2                 ; buffer -= (-2 + (temp[0] < 0xFF ? 1 : 0));
192    mov         byte [buffer], %%temph     ; buffer[0] = temp[1];
193    cmp         %%temph, 0xFF              ; Set CF if temp[1] < 0xFF
194    mov         byte [buffer+1], 0         ; buffer[1] = 0;
195    sbb         buffer, -2                 ; buffer -= (-2 + (temp[1] < 0xFF ? 1 : 0));
196    movd        nbits, mm_code             ; nbits (temp register) = (uint32_t)(code)
197%ifidn %%temp, nbits_base
198    movd        nbits_base, mm_save_nbits  ; restore nbits_base
199%endif
200    bswap       nbits                      ; nbits = htonl(nbits)
201    mov         byte [buffer], nbitsb      ; buffer[0] = nbits[0];
202    cmp         nbitsb, 0xFF               ; Set CF if nbits[0] < 0xFF
203    mov         byte [buffer+1], 0         ; buffer[1] = 0;
204    sbb         buffer, -2                 ; buffer -= (-2 + (nbits[0] < 0xFF ? 1 : 0));
205    mov         byte [buffer], nbitsh      ; buffer[0] = nbits[1];
206    cmp         nbitsh, 0xFF               ; Set CF if nbits[1] < 0xFF
207    mov         byte [buffer+1], 0         ; buffer[1] = 0;
208    sbb         buffer, -2                 ; buffer -= (-2 + (nbits[1] < 0xFF ? 1 : 0));
209    shr         nbits, 16                  ; nbits >>= 16;
210    mov         byte [buffer], nbitsb      ; buffer[0] = nbits[0];
211    cmp         nbitsb, 0xFF               ; Set CF if nbits[0] < 0xFF
212    mov         byte [buffer+1], 0         ; buffer[1] = 0;
213    sbb         buffer, -2                 ; buffer -= (-2 + (nbits[0] < 0xFF ? 1 : 0));
214    mov         byte [buffer], nbitsh      ; buffer[0] = nbits[1];
215    %4
216    cmp         nbitsh, 0xFF               ; Set CF if nbits[1] < 0xFF
217    mov         byte [buffer+1], 0         ; buffer[1] = 0;
218    sbb         buffer, -2                 ; buffer -= (-2 + (nbits[1] < 0xFF ? 1 : 0));
219    %5
220    %6
221    %7
222    %8
223    jmp %9                                 ; return;
224%endmacro
225
226%macro PUSH 1
227    push        %1
228%assign stack_offset  stack_offset + 4
229%endmacro
230
231%macro POP 1
232    pop         %1
233%assign stack_offset  stack_offset - 4
234%endmacro
235
236; If PIC is defined, load the address of a symbol defined in this file into a
237; register.  Equivalent to
238;   get_GOT     %1
239;   lea         %1, [GOTOFF(%1, %2)]
240; without using the GOT.
241;
242; Usage:
243; %1 - register into which to load the address of the symbol
244; %2 - symbol whose address should be loaded
245; %3 - optional multi-line macro to execute before the symbol address is loaded
246; %4 - optional multi-line macro to execute after the symbol address is loaded
247;
248; If PIC is not defined, then %3 and %4 are executed in order.
249
250%macro GET_SYM 2-4
251%ifdef PIC
252    call        %%.geteip
253%%.ref:
254    %4
255    add         %1, %2 - %%.ref
256    jmp         short %%.done
257    align       32
258%%.geteip:
259    %3          4               ; must adjust stack pointer because of call
260    mov         %1, POINTER [esp]
261    ret
262    align       32
263%%.done:
264%else
265    %3          0
266    %4
267%endif
268%endmacro
269
270;
271; Encode a single block's worth of coefficients.
272;
273; GLOBAL(JOCTET *)
274; jsimd_huff_encode_one_block_sse2(working_state *state, JOCTET *buffer,
275;                                  JCOEFPTR block, int last_dc_val,
276;                                  c_derived_tbl *dctbl, c_derived_tbl *actbl)
277;
278; Stack layout:
279; Function args
280; Return address
281; Saved ebx
282; Saved ebp
283; Saved esi
284; Saved edi <-- esp_save
285; ...
286; esp_save
287; t_ 64*2 bytes (aligned to 128 bytes)
288;
289; esp is used (as t) to point into t_ (data in lower indices is not used once
290; esp passes over them, so this is signal-safe.)  Aligning to 128 bytes allows
291; us to find the rest of the data again.
292;
293; NOTES:
294; When shuffling data, we try to avoid pinsrw as much as possible, since it is
295; slow on many CPUs.  Its reciprocal throughput (issue latency) is 1 even on
296; modern CPUs, so chains of pinsrw instructions (even with different outputs)
297; can limit performance.  pinsrw is a VectorPath instruction on AMD K8 and
298; requires 2 µops (with memory operand) on Intel.  In either case, only one
299; pinsrw instruction can be decoded per cycle (and nothing else if they are
300; back-to-back), so out-of-order execution cannot be used to work around long
301; pinsrw chains (though for Sandy Bridge and later, this may be less of a
302; problem if the code runs from the µop cache.)
303;
304; We use tzcnt instead of bsf without checking for support.  The instruction is
305; executed as bsf on CPUs that don't support tzcnt (encoding is equivalent to
306; rep bsf.)  The destination (first) operand of bsf (and tzcnt on some CPUs) is
307; an input dependency (although the behavior is not formally defined, Intel
308; CPUs usually leave the destination unmodified if the source is zero.)  This
309; can prevent out-of-order execution, so we clear the destination before
310; invoking tzcnt.
311;
312; Initial register allocation
313; eax - frame --> buffer
314; ebx - nbits_base (PIC) / emit_temp
315; ecx - dctbl --> size --> state
316; edx - block --> nbits
317; esi - code_temp --> state --> actbl
318; edi - index_temp --> free_bits
319; esp - t
320; ebp - index
321
322%define frame       eax
323%ifdef PIC
324%define nbits_base  ebx
325%endif
326%define emit_temp   ebx
327%define emit_tempb  bl
328%define emit_temph  bh
329%define dctbl       ecx
330%define block       edx
331%define code_temp   esi
332%define index_temp  edi
333%define t           esp
334%define index       ebp
335
336%assign save_frame  DCTSIZE2 * SIZEOF_WORD
337
338; Step 1: Re-arrange input data according to jpeg_natural_order
339; xx 01 02 03 04 05 06 07      xx 01 08 16 09 02 03 10
340; 08 09 10 11 12 13 14 15      17 24 32 25 18 11 04 05
341; 16 17 18 19 20 21 22 23      12 19 26 33 40 48 41 34
342; 24 25 26 27 28 29 30 31 ==>  27 20 13 06 07 14 21 28
343; 32 33 34 35 36 37 38 39      35 42 49 56 57 50 43 36
344; 40 41 42 43 44 45 46 47      29 22 15 23 30 37 44 51
345; 48 49 50 51 52 53 54 55      58 59 52 45 38 31 39 46
346; 56 57 58 59 60 61 62 63      53 60 61 54 47 55 62 63
347
348    align       32
349    GLOBAL_FUNCTION(jsimd_huff_encode_one_block_sse2)
350
351EXTN(jsimd_huff_encode_one_block_sse2):
352
353%assign stack_offset      0
354%define arg_state         4 + stack_offset
355%define arg_buffer        8 + stack_offset
356%define arg_block        12 + stack_offset
357%define arg_last_dc_val  16 + stack_offset
358%define arg_dctbl        20 + stack_offset
359%define arg_actbl        24 + stack_offset
360
361                                                          ;X: X = code stream
362    mov         block, [esp + arg_block]
363    PUSH        ebx
364    PUSH        ebp
365    movups      xmm3, XMMWORD [block + 0 * SIZEOF_WORD]   ;D: w3 = xx 01 02 03 04 05 06 07
366    PUSH        esi
367    PUSH        edi
368    movdqa      xmm0, xmm3                                ;A: w0 = xx 01 02 03 04 05 06 07
369    mov         frame, esp
370    lea         t, [frame - (save_frame + 4)]
371    movups      xmm1, XMMWORD [block + 8 * SIZEOF_WORD]   ;B: w1 = 08 09 10 11 12 13 14 15
372    and         t, -DCTSIZE2 * SIZEOF_WORD                                             ; t = &t_[0]
373    mov         [t + save_frame], frame
374    pxor        xmm4, xmm4                                ;A: w4[i] = 0;
375    punpckldq   xmm0, xmm1                                ;A: w0 = xx 01 08 09 02 03 10 11
376    pshuflw     xmm0, xmm0, 11001001b                     ;A: w0 = 01 08 xx 09 02 03 10 11
377    pinsrw      xmm0, word [block + 16 * SIZEOF_WORD], 2  ;A: w0 = 01 08 16 09 02 03 10 11
378    punpckhdq   xmm3, xmm1                                ;D: w3 = 04 05 12 13 06 07 14 15
379    punpcklqdq  xmm1, xmm3                                ;B: w1 = 08 09 10 11 04 05 12 13
380    pinsrw      xmm0, word [block + 17 * SIZEOF_WORD], 7  ;A: w0 = 01 08 16 09 02 03 10 17
381                                                          ;A:      (Row 0, offset 1)
382    pcmpgtw     xmm4, xmm0                                ;A: w4[i] = (w0[i] < 0 ? -1 : 0);
383    paddw       xmm0, xmm4                                ;A: w0[i] += w4[i];
384    movaps      XMMWORD [t + 0 * SIZEOF_WORD], xmm0       ;A: t[i] = w0[i];
385
386    movq        xmm2, qword [block + 24 * SIZEOF_WORD]    ;B: w2 = 24 25 26 27 -- -- -- --
387    pshuflw     xmm2, xmm2, 11011000b                     ;B: w2 = 24 26 25 27 -- -- -- --
388    pslldq      xmm1, 1 * SIZEOF_WORD                     ;B: w1 = -- 08 09 10 11 04 05 12
389    movups      xmm5, XMMWORD [block + 48 * SIZEOF_WORD]  ;H: w5 = 48 49 50 51 52 53 54 55
390    movsd       xmm1, xmm2                                ;B: w1 = 24 26 25 27 11 04 05 12
391    punpcklqdq  xmm2, xmm5                                ;C: w2 = 24 26 25 27 48 49 50 51
392    pinsrw      xmm1, word [block + 32 * SIZEOF_WORD], 1  ;B: w1 = 24 32 25 27 11 04 05 12
393    pxor        xmm4, xmm4                                ;A: w4[i] = 0;
394    psrldq      xmm3, 2 * SIZEOF_WORD                     ;D: w3 = 12 13 06 07 14 15 -- --
395    pcmpeqw     xmm0, xmm4                                ;A: w0[i] = (w0[i] == 0 ? -1 : 0);
396    pinsrw      xmm1, word [block + 18 * SIZEOF_WORD], 3  ;B: w1 = 24 32 25 18 11 04 05 12
397                                                          ;        (Row 1, offset 1)
398    pcmpgtw     xmm4, xmm1                                ;B: w4[i] = (w1[i] < 0 ? -1 : 0);
399    paddw       xmm1, xmm4                                ;B: w1[i] += w4[i];
400    movaps      XMMWORD [t + 8 * SIZEOF_WORD], xmm1       ;B: t[i+8] = w1[i];
401    pxor        xmm4, xmm4                                ;B: w4[i] = 0;
402    pcmpeqw     xmm1, xmm4                                ;B: w1[i] = (w1[i] == 0 ? -1 : 0);
403
404    packsswb    xmm0, xmm1                                ;AB: b0[i] = w0[i], b0[i+8] = w1[i]
405                                                          ;    w/ signed saturation
406
407    pinsrw      xmm3, word [block + 20 * SIZEOF_WORD], 0  ;D: w3 = 20 13 06 07 14 15 -- --
408    pinsrw      xmm3, word [block + 21 * SIZEOF_WORD], 5  ;D: w3 = 20 13 06 07 14 21 -- --
409    pinsrw      xmm3, word [block + 28 * SIZEOF_WORD], 6  ;D: w3 = 20 13 06 07 14 21 28 --
410    pinsrw      xmm3, word [block + 35 * SIZEOF_WORD], 7  ;D: w3 = 20 13 06 07 14 21 28 35
411                                                          ;        (Row 3, offset 1)
412    pcmpgtw     xmm4, xmm3                                ;D: w4[i] = (w3[i] < 0 ? -1 : 0);
413    paddw       xmm3, xmm4                                ;D: w3[i] += w4[i];
414    movaps      XMMWORD [t + 24 * SIZEOF_WORD], xmm3      ;D: t[i+24] = w3[i];
415    pxor        xmm4, xmm4                                ;D: w4[i] = 0;
416    pcmpeqw     xmm3, xmm4                                ;D: w3[i] = (w3[i] == 0 ? -1 : 0);
417
418    pinsrw      xmm2, word [block + 19 * SIZEOF_WORD], 0  ;C: w2 = 19 26 25 27 48 49 50 51
419    pinsrw      xmm2, word [block + 33 * SIZEOF_WORD], 2  ;C: w2 = 19 26 33 27 48 49 50 51
420    pinsrw      xmm2, word [block + 40 * SIZEOF_WORD], 3  ;C: w2 = 19 26 33 40 48 49 50 51
421    pinsrw      xmm2, word [block + 41 * SIZEOF_WORD], 5  ;C: w2 = 19 26 33 40 48 41 50 51
422    pinsrw      xmm2, word [block + 34 * SIZEOF_WORD], 6  ;C: w2 = 19 26 33 40 48 41 34 51
423    pinsrw      xmm2, word [block + 27 * SIZEOF_WORD], 7  ;C: w2 = 19 26 33 40 48 41 34 27
424                                                          ;        (Row 2, offset 1)
425    pcmpgtw     xmm4, xmm2                                ;C: w4[i] = (w2[i] < 0 ? -1 : 0);
426    paddw       xmm2, xmm4                                ;C: w2[i] += w4[i];
427    movsx       code_temp, word [block]                   ;Z:     code_temp = block[0];
428
429; %1 - stack pointer adjustment
430%macro GET_SYM_BEFORE 1
431    movaps      XMMWORD [t + 16 * SIZEOF_WORD + %1], xmm2
432                                                          ;C: t[i+16] = w2[i];
433    pxor        xmm4, xmm4                                ;C: w4[i] = 0;
434    pcmpeqw     xmm2, xmm4                                ;C: w2[i] = (w2[i] == 0 ? -1 : 0);
435    sub         code_temp, [frame + arg_last_dc_val]      ;Z:     code_temp -= last_dc_val;
436
437    packsswb    xmm2, xmm3                                ;CD: b2[i] = w2[i], b2[i+8] = w3[i]
438                                                          ;    w/ signed saturation
439
440    movdqa      xmm3, xmm5                                ;H: w3 = 48 49 50 51 52 53 54 55
441    pmovmskb    index_temp, xmm2                          ;Z:     index_temp = 0;  index_temp |= ((b2[i] >> 7) << i);
442    pmovmskb    index, xmm0                               ;Z:     index = 0;  index |= ((b0[i] >> 7) << i);
443    movups      xmm0, XMMWORD [block + 56 * SIZEOF_WORD]  ;H: w0 = 56 57 58 59 60 61 62 63
444    punpckhdq   xmm3, xmm0                                ;H: w3 = 52 53 60 61 54 55 62 63
445    shl         index_temp, 16                            ;Z:     index_temp <<= 16;
446    psrldq      xmm3, 1 * SIZEOF_WORD                     ;H: w3 = 53 60 61 54 55 62 63 --
447    pxor        xmm2, xmm2                                ;H: w2[i] = 0;
448    pshuflw     xmm3, xmm3, 00111001b                     ;H: w3 = 60 61 54 53 55 62 63 --
449    or          index, index_temp                         ;Z:     index |= index_temp;
450%undef index_temp
451%define free_bits  edi
452%endmacro
453
454%macro GET_SYM_AFTER 0
455    movq        xmm1, qword [block + 44 * SIZEOF_WORD]    ;G: w1 = 44 45 46 47 -- -- -- --
456    unpcklps    xmm5, xmm0                                ;E: w5 = 48 49 56 57 50 51 58 59
457    pxor        xmm0, xmm0                                ;H: w0[i] = 0;
458    not         index                                     ;Z:     index = ~index;
459    pinsrw      xmm3, word [block + 47 * SIZEOF_WORD], 3  ;H: w3 = 60 61 54 47 55 62 63 --
460                                                          ;        (Row 7, offset 1)
461    pcmpgtw     xmm2, xmm3                                ;H: w2[i] = (w3[i] < 0 ? -1 : 0);
462    mov         dctbl, [frame + arg_dctbl]
463    paddw       xmm3, xmm2                                ;H: w3[i] += w2[i];
464    movaps      XMMWORD [t + 56 * SIZEOF_WORD], xmm3      ;H: t[i+56] = w3[i];
465    movq        xmm4, qword [block + 36 * SIZEOF_WORD]    ;G: w4 = 36 37 38 39 -- -- -- --
466    pcmpeqw     xmm3, xmm0                                ;H: w3[i] = (w3[i] == 0 ? -1 : 0);
467    punpckldq   xmm4, xmm1                                ;G: w4 = 36 37 44 45 38 39 46 47
468    movdqa      xmm1, xmm4                                ;F: w1 = 36 37 44 45 38 39 46 47
469    pcmpeqw     mm_all_0xff, mm_all_0xff                  ;Z:     all_0xff[i] = 0xFF;
470%endmacro
471
472    GET_SYM     nbits_base, jpeg_nbits_table, GET_SYM_BEFORE, GET_SYM_AFTER
473
474    psrldq      xmm4, 1 * SIZEOF_WORD                     ;G: w4 = 37 44 45 38 39 46 47 --
475    shufpd      xmm1, xmm5, 10b                           ;F: w1 = 36 37 44 45 50 51 58 59
476    pshufhw     xmm4, xmm4, 11010011b                     ;G: w4 = 37 44 45 38 -- 39 46 --
477    pslldq      xmm1, 1 * SIZEOF_WORD                     ;F: w1 = -- 36 37 44 45 50 51 58
478    pinsrw      xmm4, word [block + 59 * SIZEOF_WORD], 0  ;G: w4 = 59 44 45 38 -- 39 46 --
479    pshufd      xmm1, xmm1, 11011000b                     ;F: w1 = -- 36 45 50 37 44 51 58
480    cmp         code_temp, 1 << 31                        ;Z:     Set CF if code_temp < 0x80000000,
481                                                          ;Z:     i.e. if code_temp is positive
482    pinsrw      xmm4, word [block + 52 * SIZEOF_WORD], 1  ;G: w4 = 59 52 45 38 -- 39 46 --
483    movlps      xmm1, qword [block + 20 * SIZEOF_WORD]    ;F: w1 = 20 21 22 23 37 44 51 58
484    pinsrw      xmm4, word [block + 31 * SIZEOF_WORD], 4  ;G: w4 = 59 52 45 38 31 39 46 --
485    pshuflw     xmm1, xmm1, 01110010b                     ;F: w1 = 22 20 23 21 37 44 51 58
486    pinsrw      xmm4, word [block + 53 * SIZEOF_WORD], 7  ;G: w4 = 59 52 45 38 31 39 46 53
487                                                          ;        (Row 6, offset 1)
488    adc         code_temp, -1                             ;Z:     code_temp += -1 + (code_temp >= 0 ? 1 : 0);
489    pxor        xmm2, xmm2                                ;G: w2[i] = 0;
490    pcmpgtw     xmm0, xmm4                                ;G: w0[i] = (w4[i] < 0 ? -1 : 0);
491    pinsrw      xmm1, word [block + 15 * SIZEOF_WORD], 1  ;F: w1 = 22 15 23 21 37 44 51 58
492    paddw       xmm4, xmm0                                ;G: w4[i] += w0[i];
493    movaps      XMMWORD [t + 48 * SIZEOF_WORD], xmm4      ;G: t[48+i] = w4[i];
494    movd        mm_temp, code_temp                        ;Z:     temp = code_temp
495    pinsrw      xmm1, word [block + 30 * SIZEOF_WORD], 3  ;F: w1 = 22 15 23 30 37 44 51 58
496                                                          ;        (Row 5, offset 1)
497    pcmpeqw     xmm4, xmm2                                ;G: w4[i] = (w4[i] == 0 ? -1 : 0);
498
499    packsswb    xmm4, xmm3                                ;GH: b4[i] = w4[i], b4[i+8] = w3[i]
500                                                          ;    w/ signed saturation
501
502    lea         t, [t - SIZEOF_WORD]                      ;Z:     t = &t[-1]
503    pxor        xmm0, xmm0                                ;F: w0[i] = 0;
504    pcmpgtw     xmm2, xmm1                                ;F: w2[i] = (w1[i] < 0 ? -1 : 0);
505    paddw       xmm1, xmm2                                ;F: w1[i] += w2[i];
506    movaps      XMMWORD [t + (40+1) * SIZEOF_WORD], xmm1  ;F: t[40+i] = w1[i];
507    pcmpeqw     xmm1, xmm0                                ;F: w1[i] = (w1[i] == 0 ? -1 : 0);
508    pinsrw      xmm5, word [block + 42 * SIZEOF_WORD], 0  ;E: w5 = 42 49 56 57 50 51 58 59
509    pinsrw      xmm5, word [block + 43 * SIZEOF_WORD], 5  ;E: w5 = 42 49 56 57 50 43 58 59
510    pinsrw      xmm5, word [block + 36 * SIZEOF_WORD], 6  ;E: w5 = 42 49 56 57 50 43 36 59
511    pinsrw      xmm5, word [block + 29 * SIZEOF_WORD], 7  ;E: w5 = 42 49 56 57 50 43 36 29
512                                                          ;        (Row 4, offset 1)
513%undef block
514%define nbits  edx
515%define nbitsb  dl
516%define nbitsh  dh
517    movzx       nbits, byte [NBITS(code_temp)]            ;Z:     nbits = JPEG_NBITS(code_temp);
518%undef code_temp
519%define state  esi
520    pxor        xmm2, xmm2                                ;E: w2[i] = 0;
521    mov         state, [frame + arg_state]
522    movd        mm_nbits, nbits                           ;Z:     nbits --> MMX register
523    pcmpgtw     xmm0, xmm5                                ;E: w0[i] = (w5[i] < 0 ? -1 : 0);
524    movd        mm_code, dword [dctbl + c_derived_tbl.ehufco + nbits * 4]
525                                                          ;Z:     code = dctbl->ehufco[nbits];
526%define size  ecx
527%define sizeb  cl
528%define sizeh  ch
529    paddw       xmm5, xmm0                                ;E: w5[i] += w0[i];
530    movaps      XMMWORD [t + (32+1) * SIZEOF_WORD], xmm5  ;E: t[32+i] = w5[i];
531    movzx       size, byte [dctbl + c_derived_tbl.ehufsi + nbits]
532                                                          ;Z:     size = dctbl->ehufsi[nbits];
533%undef dctbl
534    pcmpeqw     xmm5, xmm2                                ;E: w5[i] = (w5[i] == 0 ? -1 : 0);
535
536    packsswb    xmm5, xmm1                                ;EF: b5[i] = w5[i], b5[i+8] = w1[i]
537                                                          ;    w/ signed saturation
538
539    movq        mm_put_buffer, [state + working_state.cur.put_buffer.simd]
540                                                          ;Z:     put_buffer = state->cur.put_buffer.simd;
541    mov         free_bits, [state + working_state.cur.free_bits]
542                                                          ;Z:     free_bits = state->cur.free_bits;
543%undef state
544%define actbl  esi
545    mov         actbl, [frame + arg_actbl]
546%define buffer  eax
547    mov         buffer, [frame + arg_buffer]
548%undef frame
549    jmp        .BEGIN
550
551; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
552
553    align       16
554; size <= 32, so this is not really a loop
555.BRLOOP1:                                                 ; .BRLOOP1:
556    movzx       nbits, byte [actbl + c_derived_tbl.ehufsi + 0xf0]
557                                                          ; nbits = actbl->ehufsi[0xf0];
558    movd        mm_code, dword [actbl + c_derived_tbl.ehufco + 0xf0 * 4]
559                                                          ; code = actbl->ehufco[0xf0];
560    and         index, 0x7ffffff                          ; clear index if size == 32
561    sub         size, 16                                  ; size -= 16;
562    sub         free_bits, nbits                          ; if ((free_bits -= nbits) <= 0)
563    jle         .EMIT_BRLOOP1                             ;   goto .EMIT_BRLOOP1;
564    movd        mm_nbits, nbits                           ; nbits --> MMX register
565    psllq       mm_put_buffer, mm_nbits                   ; put_buffer <<= nbits;
566    por         mm_put_buffer, mm_code                    ; put_buffer |= code;
567    jmp         .ERLOOP1                                  ; goto .ERLOOP1;
568
569; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
570
571    align       16
572%ifdef PIC
573    times 6     nop
574%else
575    times 2     nop
576%endif
577.BLOOP1:                                                  ; do {  /* size = # of zero bits/elements to skip */
578; if size == 32, index remains unchanged.  Correct in .BRLOOP.
579    shr         index, sizeb                              ;   index >>= size;
580    lea         t, [t + size * SIZEOF_WORD]               ;   t += size;
581    cmp         size, 16                                  ;   if (size > 16)
582    jg          .BRLOOP1                                  ;     goto .BRLOOP1;
583.ERLOOP1:                                                 ; .ERLOOP1:
584    movsx       nbits, word [t]                           ;   nbits = *t;
585%ifdef PIC
586    add         size, size                                ;   size += size;
587%else
588    lea         size, [size * 2]                          ;   size += size;
589%endif
590    movd        mm_temp, nbits                            ;   temp = nbits;
591    movzx       nbits, byte [NBITS(nbits)]                ;   nbits = JPEG_NBITS(nbits);
592    lea         size, [size * 8 + nbits]                  ;   size = size * 8 + nbits;
593    movd        mm_nbits, nbits                           ;   nbits --> MMX register
594    movd        mm_code, dword [actbl + c_derived_tbl.ehufco + (size - 16) * 4]
595                                                          ;   code = actbl->ehufco[size-16];
596    movzx       size, byte [actbl + c_derived_tbl.ehufsi + (size - 16)]
597                                                          ;   size = actbl->ehufsi[size-16];
598.BEGIN:                                                   ; .BEGIN:
599    pand        mm_temp, [MASK_BITS(nbits)]               ;   temp &= (1 << nbits) - 1;
600    psllq       mm_code, mm_nbits                         ;   code <<= nbits;
601    add         nbits, size                               ;   nbits += size;
602    por         mm_code, mm_temp                          ;   code |= temp;
603    sub         free_bits, nbits                          ;   if ((free_bits -= nbits) <= 0)
604    jle         .EMIT_ERLOOP1                             ;     insert code, flush buffer, init size, goto .BLOOP1
605    xor         size, size                                ;   size = 0;  /* kill tzcnt input dependency */
606    tzcnt       size, index                               ;   size = # of trailing 0 bits in index
607    movd        mm_nbits, nbits                           ;   nbits --> MMX register
608    psllq       mm_put_buffer, mm_nbits                   ;   put_buffer <<= nbits;
609    inc         size                                      ;   ++size;
610    por         mm_put_buffer, mm_code                    ;   put_buffer |= code;
611    test        index, index
612    jnz         .BLOOP1                                   ; } while (index != 0);
613; Round 2
614; t points to the last used word, possibly below t_ if the previous index had 32 zero bits.
615.ELOOP1:                                                  ; .ELOOP1:
616    pmovmskb    size, xmm4                                ; size = 0;  size |= ((b4[i] >> 7) << i);
617    pmovmskb    index, xmm5                               ; index = 0;  index |= ((b5[i] >> 7) << i);
618    shl         size, 16                                  ; size <<= 16;
619    or          index, size                               ; index |= size;
620    not         index                                     ; index = ~index;
621    lea         nbits, [t + (1 + DCTSIZE2) * SIZEOF_WORD]
622                                                          ; nbits = t + 1 + 64;
623    and         nbits, -DCTSIZE2 * SIZEOF_WORD            ; nbits &= -128;  /* now points to &t_[64] */
624    sub         nbits, t                                  ; nbits -= t;
625    shr         nbits, 1                                  ; nbits >>= 1;  /* # of leading 0 bits in old index + 33 */
626    tzcnt       size, index                               ; size = # of trailing 0 bits in index
627    inc         size                                      ; ++size;
628    test        index, index                              ; if (index == 0)
629    jz          .ELOOP2                                   ;   goto .ELOOP2;
630; NOTE: size == 32 cannot happen, since the last element is always 0.
631    shr         index, sizeb                              ; index >>= size;
632    lea         size, [size + nbits - 33]                 ; size = size + nbits - 33;
633    lea         t, [t + size * SIZEOF_WORD]               ; t += size;
634    cmp         size, 16                                  ; if (size <= 16)
635    jle         .ERLOOP2                                  ;   goto .ERLOOP2;
636.BRLOOP2:                                                 ; do {
637    movzx       nbits, byte [actbl + c_derived_tbl.ehufsi + 0xf0]
638                                                          ;   nbits = actbl->ehufsi[0xf0];
639    sub         size, 16                                  ;   size -= 16;
640    movd        mm_code, dword [actbl + c_derived_tbl.ehufco + 0xf0 * 4]
641                                                          ;   code = actbl->ehufco[0xf0];
642    sub         free_bits, nbits                          ;   if ((free_bits -= nbits) <= 0)
643    jle         .EMIT_BRLOOP2                             ;     insert code and flush put_buffer
644    movd        mm_nbits, nbits                           ;   else { nbits --> MMX register
645    psllq       mm_put_buffer, mm_nbits                   ;     put_buffer <<= nbits;
646    por         mm_put_buffer, mm_code                    ;     put_buffer |= code;
647    cmp         size, 16                                  ;     if (size <= 16)
648    jle        .ERLOOP2                                   ;       goto .ERLOOP2;
649    jmp        .BRLOOP2                                   ; } while (1);
650
651; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
652
653    align      16
654.BLOOP2:                                                  ; do {  /* size = # of zero bits/elements to skip */
655    shr         index, sizeb                              ;   index >>= size;
656    lea         t, [t + size * SIZEOF_WORD]               ;   t += size;
657    cmp         size, 16                                  ;   if (size > 16)
658    jg          .BRLOOP2                                  ;     goto .BRLOOP2;
659.ERLOOP2:                                                 ; .ERLOOP2:
660    movsx       nbits, word [t]                           ;   nbits = *t;
661    add         size, size                                ;   size += size;
662    movd        mm_temp, nbits                            ;   temp = nbits;
663    movzx       nbits, byte [NBITS(nbits)]                ;   nbits = JPEG_NBITS(nbits);
664    movd        mm_nbits, nbits                           ;   nbits --> MMX register
665    lea         size, [size * 8 + nbits]                  ;   size = size * 8 + nbits;
666    movd        mm_code, dword [actbl + c_derived_tbl.ehufco + (size - 16) * 4]
667                                                          ;   code = actbl->ehufco[size-16];
668    movzx       size, byte [actbl + c_derived_tbl.ehufsi + (size - 16)]
669                                                          ;   size = actbl->ehufsi[size-16];
670    psllq       mm_code, mm_nbits                         ;   code <<= nbits;
671    pand        mm_temp, [MASK_BITS(nbits)]               ;   temp &= (1 << nbits) - 1;
672    lea         nbits, [nbits + size]                     ;   nbits += size;
673    por         mm_code, mm_temp                          ;   code |= temp;
674    xor         size, size                                ;   size = 0;  /* kill tzcnt input dependency */
675    sub         free_bits, nbits                          ;   if ((free_bits -= nbits) <= 0)
676    jle         .EMIT_ERLOOP2                             ;     insert code, flush buffer, init size, goto .BLOOP2
677    tzcnt       size, index                               ;   size = # of trailing 0 bits in index
678    movd        mm_nbits, nbits                           ;   nbits --> MMX register
679    psllq       mm_put_buffer, mm_nbits                   ;   put_buffer <<= nbits;
680    inc         size                                      ;   ++size;
681    por         mm_put_buffer, mm_code                    ;   put_buffer |= code;
682    test        index, index
683    jnz         .BLOOP2                                   ; } while (index != 0);
684.ELOOP2:                                                  ; .ELOOP2:
685    mov         nbits, t                                  ; nbits = t;
686    lea         t, [t + SIZEOF_WORD]                      ; t = &t[1];
687    and         nbits, DCTSIZE2 * SIZEOF_WORD - 1         ; nbits &= 127;
688    and         t, -DCTSIZE2 * SIZEOF_WORD                ; t &= -128;  /* t = &t_[0]; */
689    cmp         nbits, (DCTSIZE2 - 2) * SIZEOF_WORD       ; if (nbits != 62 * 2)
690    je          .EFN                                      ; {
691    movd        mm_code, dword [actbl + c_derived_tbl.ehufco + 0]
692                                                          ;   code = actbl->ehufco[0];
693    movzx       nbits, byte [actbl + c_derived_tbl.ehufsi + 0]
694                                                          ;   nbits = actbl->ehufsi[0];
695    sub         free_bits, nbits                          ;   if ((free_bits -= nbits) <= 0)
696    jg          .EFN_SKIP_EMIT_CODE                       ;   {
697    EMIT_QWORD  size, sizeb, sizeh, , , , , , .EFN        ;     insert code, flush put_buffer
698    align       16
699.EFN_SKIP_EMIT_CODE:                                      ;   } else {
700    movd        mm_nbits, nbits                           ;     nbits --> MMX register
701    psllq       mm_put_buffer, mm_nbits                   ;     put_buffer <<= nbits;
702    por         mm_put_buffer, mm_code                    ;     put_buffer |= code;
703.EFN:                                                     ; } }
704%define frame  esp
705    mov         frame, [t + save_frame]
706%define state  ecx
707    mov         state, [frame + arg_state]
708    movq        [state + working_state.cur.put_buffer.simd], mm_put_buffer
709                                                          ; state->cur.put_buffer.simd = put_buffer;
710    emms
711    mov         [state + working_state.cur.free_bits], free_bits
712                                                          ; state->cur.free_bits = free_bits;
713    POP         edi
714    POP         esi
715    POP         ebp
716    POP         ebx
717    ret
718
719; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
720
721    align       16
722.EMIT_BRLOOP1:
723    EMIT_QWORD  emit_temp, emit_tempb, emit_temph, , , , , , \
724      .ERLOOP1
725
726; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
727
728    align       16
729.EMIT_ERLOOP1:
730    EMIT_QWORD  size, sizeb, sizeh, \
731      { xor     size, size }, \
732      { tzcnt   size, index }, \
733      { inc     size }, \
734      { test    index, index }, \
735      { jnz     .BLOOP1 }, \
736      .ELOOP1
737
738; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
739
740    align       16
741.EMIT_BRLOOP2:
742    EMIT_QWORD  emit_temp, emit_tempb, emit_temph, , , , \
743      { cmp     size, 16 }, \
744      { jle     .ERLOOP2 }, \
745      .BRLOOP2
746
747; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
748
749    align       16
750.EMIT_ERLOOP2:
751    EMIT_QWORD  size, sizeb, sizeh, \
752      { xor     size, size }, \
753      { tzcnt   size, index }, \
754      { inc     size }, \
755      { test    index, index }, \
756      { jnz     .BLOOP2 }, \
757      .ELOOP2
758
759; For some reason, the OS X linker does not honor the request to align the
760; segment unless we do this.
761    align       32
762