unroll the kBands[] indirection to remove a dereference in GetCoeffs()

speed-up is small but visible.

Change-Id: Icff546adc3276f3c3d46b147c4a735b5eb8ff22e

src/dec/tree.c

@@ -494,6 +494,12 @@ static const uint8_t
 };
 
 // Paragraph 9.9
+
+static const int kBands[16 + 1] = {
+  0, 1, 2, 3, 6, 4, 5, 6, 6, 6, 6, 6, 6, 6, 6, 7,
+  0  // extra entry as sentinel
+};
+
 void VP8ParseProba(VP8BitReader* const br, VP8Decoder* const dec) {
   VP8Proba* const proba = &dec->proba_;
   int t, b, c, p;
@@ -507,6 +513,9 @@ void VP8ParseProba(VP8BitReader* const br, VP8Decoder* const dec) {
         }
       }
     }
+    for (b = 0; b < 16 + 1; ++b) {
+      proba->bands_ptr_[t][b] = &proba->bands_[t][kBands[b]];
+    }
   }
   dec->use_skip_proba_ = VP8Get(br);
   if (dec->use_skip_proba_) {

src/dec/vp8.c

@@ -371,11 +371,6 @@ int VP8GetHeaders(VP8Decoder* const dec, VP8Io* const io) {
 //------------------------------------------------------------------------------
 // Residual decoding (Paragraph 13.2 / 13.3)
 
-static const int kBands[16 + 1] = {
-  0, 1, 2, 3, 6, 4, 5, 6, 6, 6, 6, 6, 6, 6, 6, 7,
-  0  // extra entry as sentinel
-};
-
 static const uint8_t kCat3[] = { 173, 148, 140, 0 };
 static const uint8_t kCat4[] = { 176, 155, 140, 135, 0 };
 static const uint8_t kCat5[] = { 180, 157, 141, 134, 130, 0 };
@@ -419,20 +414,19 @@ static int GetLargeValue(VP8BitReader* const br, const uint8_t* const p) {
 }
 
 // Returns the position of the last non-zero coeff plus one
-static int GetCoeffs(VP8BitReader* const br, const VP8BandProbas* const prob,
+static int GetCoeffs(VP8BitReader* const br, const VP8BandProbas* const prob[],
                      int ctx, const quant_t dq, int n, int16_t* out) {
-  // n is either 0 or 1 here. kBands[n] is not necessary for extracting '*p'.
+  const uint8_t* p = prob[n]->probas_[ctx];
-  const uint8_t* p = prob[n].probas_[ctx];
   for (; n < 16; ++n) {
     if (!VP8GetBit(br, p[0])) {
       return n;  // previous coeff was last non-zero coeff
     }
     while (!VP8GetBit(br, p[1])) {       // sequence of zero coeffs
-      p = prob[kBands[++n]].probas_[0];
+      p = prob[++n]->probas_[0];
       if (n == 16) return 16;
     }
     {        // non zero coeff
-      const VP8ProbaArray* const p_ctx = &prob[kBands[n + 1]].probas_[0];
+      const VP8ProbaArray* const p_ctx = &prob[n + 1]->probas_[0];
       int v;
       if (!VP8GetBit(br, p[2])) {
         v = 1;
@@ -455,8 +449,8 @@ static WEBP_INLINE uint32_t NzCodeBits(uint32_t nz_coeffs, int nz, int dc_nz) {
 
 static int ParseResiduals(VP8Decoder* const dec,
                           VP8MB* const mb, VP8BitReader* const token_br) {
-  VP8BandProbas (* const bands)[NUM_BANDS] = dec->proba_.bands_;
+  const VP8BandProbas* (* const bands)[16 + 1] = dec->proba_.bands_ptr_;
-  const VP8BandProbas* ac_proba;
+  const VP8BandProbas* const * ac_proba;
   VP8MBData* const block = dec->mb_data_ + dec->mb_x_;
   const VP8QuantMatrix* const q = &dec->dqm_[block->segment_];
   int16_t* dst = block->coeffs_;

src/dec/vp8i.h

@@ -141,6 +141,7 @@ typedef struct {
   uint8_t segments_[MB_FEATURE_TREE_PROBS];
   // Type: 0:Intra16-AC  1:Intra16-DC   2:Chroma   3:Intra4
   VP8BandProbas bands_[NUM_TYPES][NUM_BANDS];
+  const VP8BandProbas* bands_ptr_[NUM_TYPES][16 + 1];
 } VP8Proba;
 
 // Filter parameters