Enable strict aliasing optimizations for codecs on gcc versions >= 4.0, fix alising violations that this uncovered, gives small speedups for most codecs, FS#10801

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@23784 a1c6a512-1295-4272-9138-f99709370657

apps/codecs/codecs.make

@@ -41,8 +41,8 @@ include $(APPSDIR)/codecs/librm/librm.make
 include $(APPSDIR)/codecs/libatrac/libatrac.make
 
 # compile flags for codecs
-CODECFLAGS = $(CFLAGS) -I$(APPSDIR)/codecs -I$(APPSDIR)/codecs/lib \
+CODECFLAGS = $(filter-out -fno-strict-aliasing,$(CFLAGS)) -fstrict-aliasing \
-	-DCODEC
+	-I$(APPSDIR)/codecs -I$(APPSDIR)/codecs/lib -DCODEC
 
 ifndef SIMVER
   CODEC_LDS := $(APPSDIR)/plugins/plugin.lds # codecs and plugins use same file

apps/codecs/libspc/spc_codec.h

@@ -137,16 +137,15 @@ static inline void set_le16( void* p, unsigned n )
 }
 
 #define GET_LE16( addr )        get_le16( addr )
+#define GET_LE16A( addr )       get_le16( addr )
 #define SET_LE16( addr, data )  set_le16( addr, data )
 #define INT16A( addr ) (*(uint16_t*) (addr))
 #define INT16SA( addr ) (*(int16_t*) (addr))
 
 #ifdef ROCKBOX_LITTLE_ENDIAN
-    #define GET_LE16A( addr )       (*(uint16_t*) (addr))
     #define GET_LE16SA( addr )      (*( int16_t*) (addr))
     #define SET_LE16A( addr, data ) (void) (*(uint16_t*) (addr) = (data))
 #else
-    #define GET_LE16A( addr )       get_le16 ( addr )
     #define GET_LE16SA( addr )      get_le16s( addr )
     #define SET_LE16A( addr, data ) set_le16 ( addr, data )
 #endif
@@ -166,13 +165,22 @@ struct cpu_regs_t
     uint8_t sp;
 };
 
+struct src_dir
+{
+    uint16_t start;
+    uint16_t loop;
+};
+
 struct cpu_ram_t
 {
     union {
         uint8_t padding1 [0x100];
         uint16_t align;
     } padding1 [1];
+    union {
         uint8_t ram       [0x10000];
+        struct src_dir sd [0x10000/sizeof(struct src_dir)];
+    };
     uint8_t padding2 [0x100];
 };
 
@@ -339,12 +347,6 @@ struct Spc_Dsp
 #endif
 };
 
-struct src_dir
-{
-    char start [2];
-    char loop  [2];
-};
-
 void DSP_run_( struct Spc_Dsp* this, long count, int32_t* out_buf ) ICODE_ATTR;
 void DSP_reset( struct Spc_Dsp* this );
 

apps/codecs/libspc/spc_dsp.c

@@ -77,8 +77,9 @@ static void decode_brr( struct Spc_Dsp* this, unsigned start_addr,
     /* setup same variables as where decode_brr() is called from */
     #undef RAM
     #define RAM ram.ram
+
     struct src_dir const* const sd =
-        (struct src_dir*) &RAM [this->r.g.wave_page * 0x100];
+        &ram.sd[this->r.g.wave_page * 0x100/sizeof(struct src_dir)];
     struct cache_entry_t* const wave_entry =
         &this->wave_entry [raw_voice->waveform];
 
@@ -106,7 +107,7 @@ static void decode_brr( struct Spc_Dsp* this, unsigned start_addr,
         wave_entry->start_addr = start_addr;
         
         uint8_t const* const loop_ptr =
-            RAM + GET_LE16A( sd [raw_voice->waveform].loop );
+            RAM + letoh16(sd[raw_voice->waveform].loop);
         short* loop_start = 0;
         
         short* out = BRRcache + start_addr * 2;
@@ -251,7 +252,7 @@ static void key_on(struct Spc_Dsp* const this, struct voice_t* const voice,
         voice->envx         = 0;
         voice->env_mode     = state_attack;
         voice->env_timer    = env_rate_init; /* TODO: inaccurate? */
-        unsigned start_addr = GET_LE16A(sd [raw_voice->waveform].start);
+        unsigned start_addr = letoh16(sd[raw_voice->waveform].start);
         #if !SPC_BRRCACHE
         {
             voice->addr = RAM + start_addr;
@@ -333,7 +334,7 @@ void DSP_run_( struct Spc_Dsp* this, long count, int32_t* out_buf )
     }
 
     struct src_dir const* const sd = 
-        (struct src_dir*) &RAM [this->r.g.wave_page * 0x100];
+        &ram.sd[this->r.g.wave_page * 0x100/sizeof(struct src_dir)];
 
     #ifdef ROCKBOX_BIG_ENDIAN
         /* Convert endiannesses before entering loops - these
@@ -358,7 +359,7 @@ void DSP_run_( struct Spc_Dsp* this, long count, int32_t* out_buf )
         const int echo_start = this->r.g.echo_page * 0x100;
     #endif /* CPU_COLDFIRE */
     #else
-        #define VOICE_RATE(x) (INT16A(raw_voice->rate) & 0x3FFF)
+        #define VOICE_RATE(x) (GET_LE16(raw_voice->rate) & 0x3FFF)
         #define IF_RBE(...)
     #endif /* ROCKBOX_BIG_ENDIAN */
    
@@ -590,7 +591,7 @@ void DSP_run_( struct Spc_Dsp* this, long count, int32_t* out_buf )
                 /* action based on previous block's header */
                 if ( voice->block_header & 1 )
                 {
-                    addr = RAM + GET_LE16A( sd [raw_voice->waveform].loop );
+                    addr = RAM + letoh16(sd[raw_voice->waveform].loop);
                     this->r.g.wave_ended |= vbit;
                     if ( !(voice->block_header & 2) ) /* 1% of the time */
                     {

apps/codecs/mod.c

@@ -1097,16 +1097,16 @@ static inline int clip(int i)
     else return(i);
 }
 
-STATICIRAM void synthrender(void *renderbuffer, int samplecount) ICODE_ATTR;
+STATICIRAM void synthrender(int32_t *renderbuffer, int samplecount) ICODE_ATTR;
-void synthrender(void *renderbuffer, int samplecount)
+void synthrender(int32_t *renderbuffer, int samplecount)
 {
     /* 125bpm equals to 50Hz (= 0.02s)
      * => one tick = mixingrate/50,
      * samples passing in one tick:
      * mixingrate/(bpm/2.5) = 2.5*mixingrate/bpm */
 
-    int *p_left = (int *) renderbuffer; /* int in rockbox */
+    int32_t *p_left = renderbuffer; /* int in rockbox */
-    int *p_right = p_left+1;
+    int32_t *p_right = p_left+1;
     signed short s;
     int qf_distance, qf_distance2;
 

apps/codecs/mp3_enc.c

@@ -2090,33 +2090,35 @@ STATICIRAM void to_mono_mm(void)
     /* |llllllllllllllll|rrrrrrrrrrrrrrrr| =>
      * |mmmmmmmmmmmmmmmm|mmmmmmmmmmmmmmmm|
      */
-    uint32_t *samp = (uint32_t *)&mfbuf[2*512];
+    uint16_t *samp = &mfbuf[2*512];
-    uint32_t *samp_end = samp + samp_per_frame;
+    uint16_t *samp_end = samp + 2*samp_per_frame;
 
-    inline void to_mono(uint32_t **samp)
+    inline void to_mono(uint16_t **samp)
     {
-        int32_t lr = **samp;
+        int16_t l = **samp;
+        int16_t r = **(samp+1);
         int32_t m;
 
         switch(cfg.rec_mono_mode)
         {
             case 1:
                 /* mono = L */
-                m  = lr >> 16;
+                m  = l;
                 break;
             case 2:
                 /* mono = R */
-                m  = (int16_t)lr;
+                m  = r;
                 break;
             case 0:
             default:
                 /* mono = (L+R)/2 */
-                m  = (int16_t)lr + (lr >> 16) + err;
+                m  = r + r + err;
                 err = m & 1;
                 m >>= 1;
                 break;
         }
-        *(*samp)++ = (m << 16) | (uint16_t)m;
+        *(*samp)++ = (uint16_t)m;
+        *(*samp)++ = (uint16_t)m;
     } /* to_mono */
 
     do