Branch optimisation in both C (giving hints to gcc - verified using -fprofile-arcs and gcov) and asm files. Biggest effect on coldfire (-c1000: +8%, -c2000: +5%), but ARM also profits a bit (less than 1% on ARM7TDMI, around 1% on ARM1136).

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@19199 a1c6a512-1295-4272-9138-f99709370657

6 files changed, 147 insertions, 125 deletions

diff --git a/apps/codecs/demac/libdemac/demac_config.h b/apps/codecs/demac/libdemac/demac_config.h
index dd3aaa3f9c..986e5376c4 100644
--- a/apps/codecs/demac/libdemac/demac_config.h
+++ b/apps/codecs/demac/libdemac/demac_config.h
@@ -70,6 +70,15 @@ Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110, USA
 #define ICODE_ATTR
 #define ICODE_ATTR_DEMAC
 
+/* Use to give gcc hints on which branch is most likely taken */
+#if defined(__GNUC__) && __GNUC__ >= 3
+#define LIKELY(x)   __builtin_expect(!!(x), 1)
+#define UNLIKELY(x) __builtin_expect(!!(x), 0)
+#else
+#define LIKELY(x)   (x)
+#define UNLIKELY(x) (x)
+#endif
+
 #endif /* !ROCKBOX */
 
 /* Defaults */
diff --git a/apps/codecs/demac/libdemac/entropy.c b/apps/codecs/demac/libdemac/entropy.c
index 54ff226bce..e8561122a7 100644
--- a/apps/codecs/demac/libdemac/entropy.c
+++ b/apps/codecs/demac/libdemac/entropy.c
@@ -283,13 +283,13 @@ static inline void update_rice(struct rice_t* rice, int x)
 {
     rice->ksum += ((x + 1) / 2) - ((rice->ksum + 16) >> 5);
 
-    if (rice->k == 0) {
+    if (UNLIKELY(rice->k == 0)) {
         rice->k = 1;
     } else {
         uint32_t lim = 1 << (rice->k + 4);
-        if (rice->ksum < lim) {
+        if (UNLIKELY(rice->ksum < lim)) {
             rice->k--;
-        } else if (rice->ksum >= 2 * lim) {
+        } else if (UNLIKELY(rice->ksum >= 2 * lim)) {
             rice->k++;
         }
     }
@@ -300,11 +300,12 @@ static inline int entropy_decode3980(struct rice_t* rice)
     int base, x, pivot, overflow;
 
     pivot = rice->ksum >> 5;
-    if (pivot == 0) pivot=1;
+    if (UNLIKELY(pivot == 0))
+        pivot=1;
 
     overflow = range_get_symbol_3980();
 
-    if (overflow == (MODEL_ELEMENTS-1)) {
+    if (UNLIKELY(overflow == (MODEL_ELEMENTS-1))) {
         overflow = range_decode_short() << 16;
         overflow |= range_decode_short();
     }
@@ -352,7 +353,7 @@ static inline int entropy_decode3970(struct rice_t* rice)
 
     int overflow = range_get_symbol_3970();
 
-    if (overflow == (MODEL_ELEMENTS - 1)) {
+    if (UNLIKELY(overflow == (MODEL_ELEMENTS - 1))) {
         tmpk = range_decode_bits(5);
         overflow = 0;
     } else {
@@ -435,13 +436,13 @@ int ICODE_ATTR_DEMAC entropy_decode(struct ape_ctx_t* ape_ctx,
         memset(decoded1, 0, blockstodecode * sizeof(int32_t));
     } else {
         if (ape_ctx->fileversion > 3970) {
-            while (blockstodecode--) {
+            while (LIKELY(blockstodecode--)) {
                 *(decoded0++) = entropy_decode3980(&riceY);
                 if (decoded1 != NULL)
                     *(decoded1++) = entropy_decode3980(&riceX);
             }
         } else {
-            while (blockstodecode--) {
+            while (LIKELY(blockstodecode--)) {
                 *(decoded0++) = entropy_decode3970(&riceY);
                 if (decoded1 != NULL)
                     *(decoded1++) = entropy_decode3970(&riceX);
diff --git a/apps/codecs/demac/libdemac/filter.c b/apps/codecs/demac/libdemac/filter.c
index 5601fffcd4..d66bdc69b0 100644
--- a/apps/codecs/demac/libdemac/filter.c
+++ b/apps/codecs/demac/libdemac/filter.c
@@ -100,7 +100,7 @@ struct filter_t {
 #if defined(CPU_ARM) && (ARM_ARCH >= 6)
 #define SATURATE(x) ({int __res; asm("ssat %0, #16, %1" : "=r"(__res) : "r"(x)); __res; })
 #else
-#define SATURATE(x) (((x) == (int16_t)(x)) ? (x) : ((x) >> 31) ^ 0x7FFF);
+#define SATURATE(x) (LIKELY((x) == (int16_t)(x)) ? (x) : ((x) >> 31) ^ 0x7FFF);
 #endif
 
 /* Apply the filter with state f to count entries in data[] */
@@ -109,20 +109,22 @@ static void ICODE_ATTR_DEMAC do_apply_filter_3980(struct filter_t* f,
                                                   int32_t* data, int count)
 {
     int res;
-    int absres;
+    int absres; 
 
 #ifdef PREPARE_SCALARPRODUCT
     PREPARE_SCALARPRODUCT
 #endif
 
-    while(count--)
+    while(LIKELY(count--))
     {
         res = FP_TO_INT(scalarproduct(f->coeffs, f->delay - ORDER));
 
-        if (*data < 0)
-            vector_add(f->coeffs, f->adaptcoeffs - ORDER);
-        else if (*data > 0)
-            vector_sub(f->coeffs, f->adaptcoeffs - ORDER);
+        if (LIKELY(*data != 0)) {
+            if (*data < 0)
+                vector_add(f->coeffs, f->adaptcoeffs - ORDER);
+            else
+                vector_sub(f->coeffs, f->adaptcoeffs - ORDER);
+        }
 
         res += *data;
 
@@ -136,11 +138,11 @@ static void ICODE_ATTR_DEMAC do_apply_filter_3980(struct filter_t* f,
         /* Update the adaption coefficients */
         absres = (res < 0 ? -res : res);
 
-        if (absres > (f->avg * 3))
+        if (UNLIKELY(absres > (f->avg * 3)))
             *f->adaptcoeffs = ((res >> 25) & 64) - 32;
         else if (absres > (f->avg * 4) / 3)
             *f->adaptcoeffs = ((res >> 26) & 32) - 16;
-        else if (absres > 0)
+        else if (LIKELY(absres > 0))
             *f->adaptcoeffs = ((res >> 27) & 16) - 8;
         else
             *f->adaptcoeffs = 0;
@@ -154,7 +156,7 @@ static void ICODE_ATTR_DEMAC do_apply_filter_3980(struct filter_t* f,
         f->adaptcoeffs++;
 
         /* Have we filled the history buffer? */
-        if (f->delay == f->history_end) {
+        if (UNLIKELY(f->delay == f->history_end)) {
             memmove(f->coeffs + ORDER, f->delay - (ORDER*2),
                     (ORDER*2) * sizeof(filter_int));
             f->adaptcoeffs = f->coeffs + ORDER*2;
@@ -172,14 +174,16 @@ static void ICODE_ATTR_DEMAC do_apply_filter_3970(struct filter_t* f,
     PREPARE_SCALARPRODUCT
 #endif
 
-    while(count--)
+    while(LIKELY(count--))
     {
         res = FP_TO_INT(scalarproduct(f->coeffs, f->delay - ORDER));
 
-        if (*data < 0)
-            vector_add(f->coeffs, f->adaptcoeffs - ORDER);
-        else if (*data > 0)
-            vector_sub(f->coeffs, f->adaptcoeffs - ORDER);
+        if (LIKELY(*data != 0)) {
+            if (*data < 0)
+                vector_add(f->coeffs, f->adaptcoeffs - ORDER);
+            else
+                vector_sub(f->coeffs, f->adaptcoeffs - ORDER);
+        }
 
         /* Convert res from (32-FRACBITS).FRACBITS fixed-point format to an
            integer (rounding to nearest) and add the input value to
@@ -199,7 +203,7 @@ static void ICODE_ATTR_DEMAC do_apply_filter_3970(struct filter_t* f,
         f->adaptcoeffs++;
 
         /* Have we filled the history buffer? */
-        if (f->delay == f->history_end) {
+        if (UNLIKELY(f->delay == f->history_end)) {
             memmove(f->coeffs + ORDER, f->delay - (ORDER*2),
                     (ORDER*2) * sizeof(filter_int));
             f->adaptcoeffs = f->coeffs + ORDER*2;
diff --git a/apps/codecs/demac/libdemac/predictor-arm.S b/apps/codecs/demac/libdemac/predictor-arm.S
index dfeba0dcc8..d62b6186f8 100644
--- a/apps/codecs/demac/libdemac/predictor-arm.S
+++ b/apps/codecs/demac/libdemac/predictor-arm.S
@@ -468,14 +468,24 @@ loop:
 
     add     r11, r12, #historybuffer    @ r11 := &p->historybuffer[0]
 
-    sub     r10, r14, #PREDICTOR_HISTORY_SIZE*4   
+    sub     r10, r14, #PREDICTOR_HISTORY_SIZE*4
                                        @ r10 := p->buf - PREDICTOR_HISTORY_SIZE
 
     cmp     r10, r11
-    bne     endofloop
+    beq     move_hist     @ The history buffer is full, we need to do a memmove
 
-    @ The history buffer is full, we need to do a memmove:
+    @ Check loop count
+    ldr     r0, [sp, #8]
+    subs    r0, r0, #1
+    strne   r0, [sp, #8]
+    bne     loop
+
+done:
+    str     r14, [r12]              @ Save value of p->buf
+    add     sp, sp, #12             @ Don't bother restoring r1-r3 
+    ldmia   sp!, {r4-r11, pc}
 
+move_hist:
     @ dest = r11 (p->historybuffer)
     @ src = r14 (p->buf)
     @ n = 200
@@ -493,15 +503,10 @@ loop:
 
     add     r14, r12, #historybuffer    @ p->buf = &p->historybuffer[0]
 
-
-endofloop:
-@ Check loop count
+    @ Check loop count
     ldr     r0, [sp, #8]
     subs    r0, r0, #1
     strne   r0, [sp, #8]
     bne     loop
-
-done:
-    str     r14, [r12]              @ Save value of p->buf
-    add     sp, sp, #12             @ Don't bother restoring r1-r3 
-    ldmia   sp!, {r4-r11, pc}
+    
+    b       done
diff --git a/apps/codecs/demac/libdemac/predictor-cf.S b/apps/codecs/demac/libdemac/predictor-cf.S
index b12d0932b0..0a1ffe9ead 100644
--- a/apps/codecs/demac/libdemac/predictor-cf.S
+++ b/apps/codecs/demac/libdemac/predictor-cf.S
@@ -486,10 +486,18 @@ predictor_decode_stereo:
                               | %a3 = &p->historybuffer[PREDICTOR_HISTORY_SIZE]
     
     cmp.l   %a3, %a5                    
-    bne.s   .endofloop
+    beq.s   .move_hist    | The history buffer is full, we need to do a memmove
                                               
-    | The history buffer is full, we need to do a memmove:
+    subq.l  #1, (8,%sp)                 | decrease loop count
+    bne.w   .loop
 
+.done:
+    move.l  %a5, (%a6)                  | Save value of p->buf
+    movem.l (3*4,%sp), %d2-%d7/%a2-%a6
+    lea.l   (14*4,%sp), %sp
+    rts
+    
+.move_hist:
     lea.l   (historybuffer,%a6), %a3    
 
     | dest = %a3 (p->historybuffer)
@@ -497,33 +505,19 @@ predictor_decode_stereo:
     | n = 200
     
     movem.l (%a5), %d0-%d7/%a0-%a1      | 40 bytes
-    lea.l   (40,%a5), %a5
     movem.l %d0-%d7/%a0-%a1, (%a3)
-    lea.l   (40,%a3), %a3
-    movem.l (%a5), %d0-%d7/%a0-%a1      | 40 bytes
-    lea.l   (40,%a5), %a5
-    movem.l %d0-%d7/%a0-%a1, (%a3)
-    lea.l   (40,%a3), %a3
-    movem.l (%a5), %d0-%d7/%a0-%a1      | 40 bytes
-    lea.l   (40,%a5), %a5
-    movem.l %d0-%d7/%a0-%a1, (%a3)
-    lea.l   (40,%a3), %a3
-    movem.l (%a5), %d0-%d7/%a0-%a1      | 40 bytes
-    lea.l   (40,%a5), %a5
-    movem.l %d0-%d7/%a0-%a1, (%a3)
-    lea.l   (40,%a3), %a3
-    movem.l (%a5), %d0-%d7/%a0-%a1      | 40 bytes
-    lea.l   (40,%a5), %a5
-    movem.l %d0-%d7/%a0-%a1, (%a3)
-    lea.l   (40,%a3), %a3
+    movem.l (40,%a5), %d0-%d7/%a0-%a1   | 40 bytes
+    movem.l %d0-%d7/%a0-%a1, (40,%a3)
+    movem.l (80,%a5), %d0-%d7/%a0-%a1   | 40 bytes
+    movem.l %d0-%d7/%a0-%a1, (80,%a3)
+    movem.l (120,%a5), %d0-%d7/%a0-%a1  | 40 bytes
+    movem.l %d0-%d7/%a0-%a1, (120,%a3)
+    movem.l (160,%a5), %d0-%d7/%a0-%a1  | 40 bytes
+    movem.l %d0-%d7/%a0-%a1, (160,%a3)
 
-    lea.l   (historybuffer,%a6), %a5    | p->buf = &p->historybuffer[0]
+    move.l  %a3, %a5                    | p->buf = &p->historybuffer[0]
 
-.endofloop:
     subq.l  #1, (8,%sp)                 | decrease loop count
     bne.w   .loop
 
-    move.l  %a5, (%a6)                  | Save value of p->buf
-    movem.l (3*4,%sp), %d2-%d7/%a2-%a6
-    lea.l   (14*4,%sp), %sp
-    rts
+    bra.s   .done
diff --git a/apps/codecs/demac/libdemac/predictor.c b/apps/codecs/demac/libdemac/predictor.c
index 1a9b48e0ca..67a17f4b5d 100644
--- a/apps/codecs/demac/libdemac/predictor.c
+++ b/apps/codecs/demac/libdemac/predictor.c
@@ -75,7 +75,7 @@ int ICODE_ATTR_DEMAC predictor_decode_stereo(struct predictor_t* p,
 {
     int32_t predictionA, predictionB; 
 
-    while (count--)
+    while (LIKELY(count--))
     {
         /* Predictor Y */
         p->buf[YDELAYA] = p->YlastA;
@@ -134,60 +134,66 @@ int ICODE_ATTR_DEMAC predictor_decode_stereo(struct predictor_t* p,
         p->XlastA = *decoded1 + ((predictionA + (predictionB >> 1)) >> 10); 
         p->XfilterA =  p->XlastA + ((p->XfilterA * 31) >> 5);
 
-        if (*decoded0 > 0) 
+        if (LIKELY(*decoded0 != 0))
         {
-            p->YcoeffsA[0] -= p->buf[YADAPTCOEFFSA];
-            p->YcoeffsA[1] -= p->buf[YADAPTCOEFFSA-1];
-            p->YcoeffsA[2] -= p->buf[YADAPTCOEFFSA-2];
-            p->YcoeffsA[3] -= p->buf[YADAPTCOEFFSA-3];
-
-            p->YcoeffsB[0] -= p->buf[YADAPTCOEFFSB];
-            p->YcoeffsB[1] -= p->buf[YADAPTCOEFFSB-1];
-            p->YcoeffsB[2] -= p->buf[YADAPTCOEFFSB-2];
-            p->YcoeffsB[3] -= p->buf[YADAPTCOEFFSB-3];
-            p->YcoeffsB[4] -= p->buf[YADAPTCOEFFSB-4];
-        }
-        else if (*decoded0 < 0) 
-        {
-            p->YcoeffsA[0] += p->buf[YADAPTCOEFFSA];
-            p->YcoeffsA[1] += p->buf[YADAPTCOEFFSA-1];
-            p->YcoeffsA[2] += p->buf[YADAPTCOEFFSA-2];
-            p->YcoeffsA[3] += p->buf[YADAPTCOEFFSA-3];
-
-            p->YcoeffsB[0] += p->buf[YADAPTCOEFFSB];
-            p->YcoeffsB[1] += p->buf[YADAPTCOEFFSB-1];
-            p->YcoeffsB[2] += p->buf[YADAPTCOEFFSB-2];
-            p->YcoeffsB[3] += p->buf[YADAPTCOEFFSB-3];
-            p->YcoeffsB[4] += p->buf[YADAPTCOEFFSB-4];
+            if (*decoded0 > 0)
+            {
+                p->YcoeffsA[0] -= p->buf[YADAPTCOEFFSA];
+                p->YcoeffsA[1] -= p->buf[YADAPTCOEFFSA-1];
+                p->YcoeffsA[2] -= p->buf[YADAPTCOEFFSA-2];
+                p->YcoeffsA[3] -= p->buf[YADAPTCOEFFSA-3];
+
+                p->YcoeffsB[0] -= p->buf[YADAPTCOEFFSB];
+                p->YcoeffsB[1] -= p->buf[YADAPTCOEFFSB-1];
+                p->YcoeffsB[2] -= p->buf[YADAPTCOEFFSB-2];
+                p->YcoeffsB[3] -= p->buf[YADAPTCOEFFSB-3];
+                p->YcoeffsB[4] -= p->buf[YADAPTCOEFFSB-4];
+            }
+            else
+            {
+                p->YcoeffsA[0] += p->buf[YADAPTCOEFFSA];
+                p->YcoeffsA[1] += p->buf[YADAPTCOEFFSA-1];
+                p->YcoeffsA[2] += p->buf[YADAPTCOEFFSA-2];
+                p->YcoeffsA[3] += p->buf[YADAPTCOEFFSA-3];
+
+                p->YcoeffsB[0] += p->buf[YADAPTCOEFFSB];
+                p->YcoeffsB[1] += p->buf[YADAPTCOEFFSB-1];
+                p->YcoeffsB[2] += p->buf[YADAPTCOEFFSB-2];
+                p->YcoeffsB[3] += p->buf[YADAPTCOEFFSB-3];
+                p->YcoeffsB[4] += p->buf[YADAPTCOEFFSB-4];
+            }
         }
 
         *(decoded0++) = p->YfilterA;
 
-        if (*decoded1 > 0) 
-        {
-            p->XcoeffsA[0] -= p->buf[XADAPTCOEFFSA];
-            p->XcoeffsA[1] -= p->buf[XADAPTCOEFFSA-1];
-            p->XcoeffsA[2] -= p->buf[XADAPTCOEFFSA-2];
-            p->XcoeffsA[3] -= p->buf[XADAPTCOEFFSA-3];
-
-            p->XcoeffsB[0] -= p->buf[XADAPTCOEFFSB];
-            p->XcoeffsB[1] -= p->buf[XADAPTCOEFFSB-1];
-            p->XcoeffsB[2] -= p->buf[XADAPTCOEFFSB-2];
-            p->XcoeffsB[3] -= p->buf[XADAPTCOEFFSB-3];
-            p->XcoeffsB[4] -= p->buf[XADAPTCOEFFSB-4];
-        }
-        else if (*decoded1 < 0) 
+        if (LIKELY(*decoded1 != 0))
         {
-            p->XcoeffsA[0] += p->buf[XADAPTCOEFFSA];
-            p->XcoeffsA[1] += p->buf[XADAPTCOEFFSA-1];
-            p->XcoeffsA[2] += p->buf[XADAPTCOEFFSA-2];
-            p->XcoeffsA[3] += p->buf[XADAPTCOEFFSA-3];
-
-            p->XcoeffsB[0] += p->buf[XADAPTCOEFFSB];
-            p->XcoeffsB[1] += p->buf[XADAPTCOEFFSB-1];
-            p->XcoeffsB[2] += p->buf[XADAPTCOEFFSB-2];
-            p->XcoeffsB[3] += p->buf[XADAPTCOEFFSB-3];
-            p->XcoeffsB[4] += p->buf[XADAPTCOEFFSB-4];
+            if (*decoded1 > 0)
+            {
+                p->XcoeffsA[0] -= p->buf[XADAPTCOEFFSA];
+                p->XcoeffsA[1] -= p->buf[XADAPTCOEFFSA-1];
+                p->XcoeffsA[2] -= p->buf[XADAPTCOEFFSA-2];
+                p->XcoeffsA[3] -= p->buf[XADAPTCOEFFSA-3];
+
+                p->XcoeffsB[0] -= p->buf[XADAPTCOEFFSB];
+                p->XcoeffsB[1] -= p->buf[XADAPTCOEFFSB-1];
+                p->XcoeffsB[2] -= p->buf[XADAPTCOEFFSB-2];
+                p->XcoeffsB[3] -= p->buf[XADAPTCOEFFSB-3];
+                p->XcoeffsB[4] -= p->buf[XADAPTCOEFFSB-4];
+            }
+            else
+            {
+                p->XcoeffsA[0] += p->buf[XADAPTCOEFFSA];
+                p->XcoeffsA[1] += p->buf[XADAPTCOEFFSA-1];
+                p->XcoeffsA[2] += p->buf[XADAPTCOEFFSA-2];
+                p->XcoeffsA[3] += p->buf[XADAPTCOEFFSA-3];
+
+                p->XcoeffsB[0] += p->buf[XADAPTCOEFFSB];
+                p->XcoeffsB[1] += p->buf[XADAPTCOEFFSB-1];
+                p->XcoeffsB[2] += p->buf[XADAPTCOEFFSB-2];
+                p->XcoeffsB[3] += p->buf[XADAPTCOEFFSB-3];
+                p->XcoeffsB[4] += p->buf[XADAPTCOEFFSB-4];
+            }
         }
 
         *(decoded1++) = p->XfilterA;
@@ -196,7 +202,7 @@ int ICODE_ATTR_DEMAC predictor_decode_stereo(struct predictor_t* p,
         p->buf++;
 
         /* Have we filled the history buffer? */
-        if (p->buf == p->historybuffer + PREDICTOR_HISTORY_SIZE) {
+        if (UNLIKELY(p->buf == p->historybuffer + PREDICTOR_HISTORY_SIZE)) {
             memmove(p->historybuffer, p->buf, 
                     PREDICTOR_SIZE * sizeof(int32_t));
             p->buf = p->historybuffer;
@@ -215,7 +221,7 @@ int ICODE_ATTR_DEMAC predictor_decode_mono(struct predictor_t* p,
 
     currentA = p->YlastA;
 
-    while (count--)
+    while (LIKELY(count--))
     {
         A = *decoded0;
 
@@ -232,25 +238,28 @@ int ICODE_ATTR_DEMAC predictor_decode_mono(struct predictor_t* p,
         p->buf[YADAPTCOEFFSA] = SIGN(p->buf[YDELAYA]);
         p->buf[YADAPTCOEFFSA-1] = SIGN(p->buf[YDELAYA-1]);
         
-        if (A > 0) 
-        {
-            p->YcoeffsA[0] -= p->buf[YADAPTCOEFFSA];
-            p->YcoeffsA[1] -= p->buf[YADAPTCOEFFSA-1];
-            p->YcoeffsA[2] -= p->buf[YADAPTCOEFFSA-2];
-            p->YcoeffsA[3] -= p->buf[YADAPTCOEFFSA-3];
-        }
-        else if (A < 0) 
+        if (LIKELY(A != 0))
         {
-            p->YcoeffsA[0] += p->buf[YADAPTCOEFFSA];
-            p->YcoeffsA[1] += p->buf[YADAPTCOEFFSA-1];
-            p->YcoeffsA[2] += p->buf[YADAPTCOEFFSA-2];
-            p->YcoeffsA[3] += p->buf[YADAPTCOEFFSA-3];
+            if (A > 0)
+            {
+                p->YcoeffsA[0] -= p->buf[YADAPTCOEFFSA];
+                p->YcoeffsA[1] -= p->buf[YADAPTCOEFFSA-1];
+                p->YcoeffsA[2] -= p->buf[YADAPTCOEFFSA-2];
+                p->YcoeffsA[3] -= p->buf[YADAPTCOEFFSA-3];
+            }
+            else
+            {
+                p->YcoeffsA[0] += p->buf[YADAPTCOEFFSA];
+                p->YcoeffsA[1] += p->buf[YADAPTCOEFFSA-1];
+                p->YcoeffsA[2] += p->buf[YADAPTCOEFFSA-2];
+                p->YcoeffsA[3] += p->buf[YADAPTCOEFFSA-3];
+            }
         }
 
         p->buf++;
 
         /* Have we filled the history buffer? */
-        if (p->buf == p->historybuffer + PREDICTOR_HISTORY_SIZE) {
+        if (UNLIKELY(p->buf == p->historybuffer + PREDICTOR_HISTORY_SIZE)) {
             memmove(p->historybuffer, p->buf, 
                     PREDICTOR_SIZE * sizeof(int32_t));
             p->buf = p->historybuffer;