Merge remote-tracking branch 'qatar/master'

* qatar/master:
  aacdec: Drop some unused function arguments
  x86: Fix assembly with NASM
  x86: fmtconvert: Refactor cvtps2pi emulation through cpuflags

Conflicts:
	libavcodec/aacdec.c

Merged-by: Michael Niedermayer <michaelni@gmx.at>

2 files changed, 47 insertions, 51 deletions

diff --git a/libavcodec/aacdec.c b/libavcodec/aacdec.c
index 44869c5e7d..b132cbac46 100644
--- a/libavcodec/aacdec.c
+++ b/libavcodec/aacdec.c
@@ -175,7 +175,7 @@ struct elem_to_channel {
 };
 
 static int assign_pair(struct elem_to_channel e2c_vec[MAX_ELEM_ID],
-    uint8_t (*layout_map)[3], int offset, int tags, uint64_t left,
+                       uint8_t (*layout_map)[3], int offset, uint64_t left,
     uint64_t right, int pos)
 {
     if (layout_map[offset][0] == TYPE_CPE) {
@@ -257,21 +257,21 @@ static uint64_t sniff_channel_order(uint8_t (*layout_map)[3], int tags)
         num_front_channels--;
     }
     if (num_front_channels >= 4) {
-        i += assign_pair(e2c_vec, layout_map, i, tags,
+        i += assign_pair(e2c_vec, layout_map, i,
                          AV_CH_FRONT_LEFT_OF_CENTER,
                          AV_CH_FRONT_RIGHT_OF_CENTER,
                          AAC_CHANNEL_FRONT);
         num_front_channels -= 2;
     }
     if (num_front_channels >= 2) {
-        i += assign_pair(e2c_vec, layout_map, i, tags,
+        i += assign_pair(e2c_vec, layout_map, i,
                          AV_CH_FRONT_LEFT,
                          AV_CH_FRONT_RIGHT,
                          AAC_CHANNEL_FRONT);
         num_front_channels -= 2;
     }
     while (num_front_channels >= 2) {
-        i += assign_pair(e2c_vec, layout_map, i, tags,
+        i += assign_pair(e2c_vec, layout_map, i,
                          UINT64_MAX,
                          UINT64_MAX,
                          AAC_CHANNEL_FRONT);
@@ -279,14 +279,14 @@ static uint64_t sniff_channel_order(uint8_t (*layout_map)[3], int tags)
     }
 
     if (num_side_channels >= 2) {
-        i += assign_pair(e2c_vec, layout_map, i, tags,
+        i += assign_pair(e2c_vec, layout_map, i,
                          AV_CH_SIDE_LEFT,
                          AV_CH_SIDE_RIGHT,
                          AAC_CHANNEL_FRONT);
         num_side_channels -= 2;
     }
     while (num_side_channels >= 2) {
-        i += assign_pair(e2c_vec, layout_map, i, tags,
+        i += assign_pair(e2c_vec, layout_map, i,
                          UINT64_MAX,
                          UINT64_MAX,
                          AAC_CHANNEL_SIDE);
@@ -294,14 +294,14 @@ static uint64_t sniff_channel_order(uint8_t (*layout_map)[3], int tags)
     }
 
     while (num_back_channels >= 4) {
-        i += assign_pair(e2c_vec, layout_map, i, tags,
+        i += assign_pair(e2c_vec, layout_map, i,
                          UINT64_MAX,
                          UINT64_MAX,
                          AAC_CHANNEL_BACK);
         num_back_channels -= 2;
     }
     if (num_back_channels >= 2) {
-        i += assign_pair(e2c_vec, layout_map, i, tags,
+        i += assign_pair(e2c_vec, layout_map, i,
                          AV_CH_BACK_LEFT,
                          AV_CH_BACK_RIGHT,
                          AAC_CHANNEL_BACK);
@@ -383,7 +383,7 @@ static void pop_output_configuration(AACContext *ac) {
  */
 static int output_configure(AACContext *ac,
                                     uint8_t layout_map[MAX_ELEM_ID*4][3], int tags,
-                                    int channel_config, enum OCStatus oc_type)
+                            enum OCStatus oc_type)
 {
     AVCodecContext *avctx = ac->avctx;
     int i, channels = 0, ret;
@@ -481,7 +481,7 @@ static ChannelElement *get_che(AACContext *ac, int type, int elem_id)
                                        2) < 0)
             return NULL;
         if (output_configure(ac, layout_map, layout_map_tags,
-                             2, OC_TRIAL_FRAME) < 0)
+                             OC_TRIAL_FRAME) < 0)
             return NULL;
 
         ac->oc[1].m4ac.chan_config = 2;
@@ -499,7 +499,7 @@ static ChannelElement *get_che(AACContext *ac, int type, int elem_id)
                                        1) < 0)
             return NULL;
         if (output_configure(ac, layout_map, layout_map_tags,
-                             1, OC_TRIAL_FRAME) < 0)
+                             OC_TRIAL_FRAME) < 0)
             return NULL;
 
         ac->oc[1].m4ac.chan_config = 1;
@@ -692,8 +692,7 @@ static int decode_ga_specific_config(AACContext *ac, AVCodecContext *avctx,
     } else if (m4ac->sbr == 1 && m4ac->ps == -1)
         m4ac->ps = 1;
 
-    if (ac && (ret = output_configure(ac, layout_map, tags,
-                                      channel_config, OC_GLOBAL_HDR)))
+    if (ac && (ret = output_configure(ac, layout_map, tags, OC_GLOBAL_HDR)))
         return ret;
 
     if (extension_flag) {
@@ -877,7 +876,7 @@ static av_cold int aac_decode_init(AVCodecContext *avctx)
                 &layout_map_tags, ac->oc[1].m4ac.chan_config);
             if (!ret)
                 output_configure(ac, layout_map, layout_map_tags,
-                                 ac->oc[1].m4ac.chan_config, OC_GLOBAL_HDR);
+                                 OC_GLOBAL_HDR);
             else if (avctx->err_recognition & AV_EF_EXPLODE)
                 return AVERROR_INVALIDDATA;
         }
@@ -967,7 +966,7 @@ static int decode_prediction(AACContext *ac, IndividualChannelStream *ics,
 /**
  * Decode Long Term Prediction data; reference: table 4.xx.
  */
-static void decode_ltp(AACContext *ac, LongTermPrediction *ltp,
+static void decode_ltp(LongTermPrediction *ltp,
                        GetBitContext *gb, uint8_t max_sfb)
 {
     int sfb;
@@ -1028,7 +1027,7 @@ static int decode_ics_info(AACContext *ac, IndividualChannelStream *ics,
                 goto fail;
             } else {
                 if ((ics->ltp.present = get_bits(gb, 1)))
-                    decode_ltp(ac, &ics->ltp, gb, ics->max_sfb);
+                    decode_ltp(&ics->ltp, gb, ics->max_sfb);
             }
         }
     }
@@ -1769,7 +1768,7 @@ static int decode_cpe(AACContext *ac, GetBitContext *gb, ChannelElement *cpe)
         cpe->ch[1].ics.use_kb_window[1] = i;
         if (cpe->ch[1].ics.predictor_present && (ac->oc[1].m4ac.object_type != AOT_AAC_MAIN))
             if ((cpe->ch[1].ics.ltp.present = get_bits(gb, 1)))
-                decode_ltp(ac, &cpe->ch[1].ics.ltp, gb, cpe->ch[1].ics.max_sfb);
+                decode_ltp(&cpe->ch[1].ics.ltp, gb, cpe->ch[1].ics.max_sfb);
         ms_present = get_bits(gb, 2);
         if (ms_present == 3) {
             av_log(ac->avctx, AV_LOG_ERROR, "ms_present = 3 is reserved.\n");
@@ -1896,12 +1895,10 @@ static int decode_drc_channel_exclusions(DynamicRangeControl *che_drc,
 /**
  * Decode dynamic range information; reference: table 4.52.
  *
- * @param   cnt length of TYPE_FIL syntactic element in bytes
- *
  * @return  Returns number of bytes consumed.
  */
 static int decode_dynamic_range(DynamicRangeControl *che_drc,
-                                GetBitContext *gb, int cnt)
+                                GetBitContext *gb)
 {
     int n             = 1;
     int drc_num_bands = 1;
@@ -2004,14 +2001,14 @@ static int decode_extension_payload(AACContext *ac, GetBitContext *gb, int cnt,
             ac->oc[1].m4ac.sbr = 1;
             ac->oc[1].m4ac.ps = 1;
             output_configure(ac, ac->oc[1].layout_map, ac->oc[1].layout_map_tags,
-                             ac->oc[1].m4ac.chan_config, ac->oc[1].status);
+                             ac->oc[1].status);
         } else {
             ac->oc[1].m4ac.sbr = 1;
         }
         res = ff_decode_sbr_extension(ac, &che->sbr, gb, crc_flag, cnt, elem_type);
         break;
     case EXT_DYNAMIC_RANGE:
-        res = decode_dynamic_range(&ac->che_drc, gb, cnt);
+        res = decode_dynamic_range(&ac->che_drc, gb);
         break;
     case EXT_FILL:
         decode_fill(ac, gb, 8 * cnt - 4);
@@ -2398,7 +2395,6 @@ static int parse_adts_frame_header(AACContext *ac, GetBitContext *gb)
                     &layout_map_tags, hdr_info.chan_config))
                 return -7;
             if (output_configure(ac, layout_map, layout_map_tags,
-                                 hdr_info.chan_config,
                                  FFMAX(ac->oc[1].status, OC_TRIAL_FRAME)))
                 return -7;
         } else {
@@ -2415,7 +2411,7 @@ static int parse_adts_frame_header(AACContext *ac, GetBitContext *gb)
                 layout_map[0][1] = 0;
                 layout_map[1][1] = 1;
                 if (output_configure(ac, layout_map, layout_map_tags,
-                                     0, OC_TRIAL_FRAME))
+                                     OC_TRIAL_FRAME))
                     return -7;
             }
         }
@@ -2513,7 +2509,7 @@ static int aac_decode_frame_int(AVCodecContext *avctx, void *data,
                        "Not evaluating a further program_config_element as this construct is dubious at best.\n");
                 pop_output_configuration(ac);
             } else {
-                err = output_configure(ac, layout_map, tags, 0, OC_TRIAL_PCE);
+                err = output_configure(ac, layout_map, tags, OC_TRIAL_PCE);
                 if (!err)
                     ac->oc[1].m4ac.chan_config = 0;
                 pce_found = 1;
diff --git a/libavcodec/x86/fmtconvert.asm b/libavcodec/x86/fmtconvert.asm
index f6b2c943e1..8313d261d9 100644
--- a/libavcodec/x86/fmtconvert.asm
+++ b/libavcodec/x86/fmtconvert.asm
@@ -23,6 +23,14 @@
 
 SECTION_TEXT
 
+%macro CVTPS2PI 2
+%if cpuflag(sse)
+    cvtps2pi %1, %2
+%elif cpuflag(3dnow)
+    pf2id %1, %2
+%endif
+%endmacro
+
 ;---------------------------------------------------------------------------------
 ; void int32_to_float_fmul_scalar(float *dst, const int *src, float mul, int len);
 ;---------------------------------------------------------------------------------
@@ -88,10 +96,10 @@ cglobal float_to_int16, 3, 3, %1, dst, src, len
     packssdw    m0, m1
     mova  [dstq+lenq], m0
 %else
-    cvtps2pi    m0, [srcq+2*lenq   ]
-    cvtps2pi    m1, [srcq+2*lenq+ 8]
-    cvtps2pi    m2, [srcq+2*lenq+16]
-    cvtps2pi    m3, [srcq+2*lenq+24]
+    CVTPS2PI    m0, [srcq+2*lenq   ]
+    CVTPS2PI    m1, [srcq+2*lenq+ 8]
+    CVTPS2PI    m2, [srcq+2*lenq+16]
+    CVTPS2PI    m3, [srcq+2*lenq+24]
     packssdw    m0, m1
     packssdw    m2, m3
     mova  [dstq+lenq  ], m0
@@ -109,10 +117,8 @@ INIT_XMM sse2
 FLOAT_TO_INT16 2
 INIT_MMX sse
 FLOAT_TO_INT16 0
-%define cvtps2pi pf2id
 INIT_MMX 3dnow
 FLOAT_TO_INT16 0
-%undef cvtps2pi
 
 ;------------------------------------------------------------------------------
 ; void ff_float_to_int16_step(int16_t *dst, const float *src, long len, long step);
@@ -150,10 +156,10 @@ cglobal float_to_int16_step, 4, 7, %1, dst, src, len, step, step3, v1, v2
     mov  [dstq+step3q*2], v2w
     lea       dstq, [dstq+stepq*8]
 %else
-    cvtps2pi    m0, [srcq+2*lenq   ]
-    cvtps2pi    m1, [srcq+2*lenq+ 8]
-    cvtps2pi    m2, [srcq+2*lenq+16]
-    cvtps2pi    m3, [srcq+2*lenq+24]
+    CVTPS2PI    m0, [srcq+2*lenq   ]
+    CVTPS2PI    m1, [srcq+2*lenq+ 8]
+    CVTPS2PI    m2, [srcq+2*lenq+16]
+    CVTPS2PI    m3, [srcq+2*lenq+24]
     packssdw    m0, m1
     packssdw    m2, m3
     movd       v1d, m0
@@ -189,10 +195,8 @@ INIT_XMM sse2
 FLOAT_TO_INT16_STEP 2
 INIT_MMX sse
 FLOAT_TO_INT16_STEP 0
-%define cvtps2pi pf2id
 INIT_MMX 3dnow
 FLOAT_TO_INT16_STEP 0
-%undef cvtps2pi
 
 ;-------------------------------------------------------------------------------
 ; void ff_float_to_int16_interleave2(int16_t *dst, const float **src, long len);
@@ -215,10 +219,10 @@ cglobal float_to_int16_interleave2, 3, 4, 2, dst, src0, src1, len
     punpcklwd  m0, m1
     mova  [dstq+lenq], m0
 %else
-    cvtps2pi   m0, [src0q+lenq  ]
-    cvtps2pi   m1, [src0q+lenq+8]
-    cvtps2pi   m2, [src1q+lenq  ]
-    cvtps2pi   m3, [src1q+lenq+8]
+    CVTPS2PI   m0, [src0q+lenq  ]
+    CVTPS2PI   m1, [src0q+lenq+8]
+    CVTPS2PI   m2, [src1q+lenq  ]
+    CVTPS2PI   m3, [src1q+lenq+8]
     packssdw   m0, m1
     packssdw   m2, m3
     mova       m1, m0
@@ -236,9 +240,7 @@ cglobal float_to_int16_interleave2, 3, 4, 2, dst, src0, src1, len
 %endmacro
 
 INIT_MMX 3dnow
-%define cvtps2pi pf2id
 FLOAT_TO_INT16_INTERLEAVE2
-%undef cvtps2pi
 INIT_MMX sse
 FLOAT_TO_INT16_INTERLEAVE2
 INIT_XMM sse2
@@ -274,12 +276,12 @@ cglobal float_to_int16_interleave6, 2, 8, 0, dst, src, src1, src2, src3, src4, s
     sub src4q, srcq
     sub src5q, srcq
 .loop:
-    cvtps2pi   mm0, [srcq]
-    cvtps2pi   mm1, [srcq+src1q]
-    cvtps2pi   mm2, [srcq+src2q]
-    cvtps2pi   mm3, [srcq+src3q]
-    cvtps2pi   mm4, [srcq+src4q]
-    cvtps2pi   mm5, [srcq+src5q]
+    CVTPS2PI   mm0, [srcq]
+    CVTPS2PI   mm1, [srcq+src1q]
+    CVTPS2PI   mm2, [srcq+src2q]
+    CVTPS2PI   mm3, [srcq+src3q]
+    CVTPS2PI   mm4, [srcq+src4q]
+    CVTPS2PI   mm5, [srcq+src5q]
     packssdw   mm0, mm3
     packssdw   mm1, mm4
     packssdw   mm2, mm5
@@ -306,13 +308,11 @@ INIT_MMX sse
 %define pswapd PSWAPD_SSE
 FLOAT_TO_INT16_INTERLEAVE6
 INIT_MMX 3dnow
-%define cvtps2pi pf2id
 %define pswapd PSWAPD_3DNOW
 FLOAT_TO_INT16_INTERLEAVE6
 %undef pswapd
 INIT_MMX 3dnowext
 FLOAT_TO_INT16_INTERLEAVE6
-%undef cvtps2pi
 
 ;-----------------------------------------------------------------------------
 ; void ff_float_interleave6(float *dst, const float **src, unsigned int len);