Upgrade Opus to v1.3android-wear-8.0.0_r2android-o-mr1-iot-release-smart-display-r9android-o-mr1-iot-release-smart-display-r8android-o-mr1-iot-release-smart-display-r5android-o-mr1-iot-release-smart-display-r40.1Jandroid-o-mr1-iot-release-smart-display-r4android-o-mr1-iot-release-smart-display-r39android-o-mr1-iot-release-smart-display-r30android-o-mr1-iot-release-smart-display-r22android-o-mr1-iot-release-smart-display-r14android-o-mr1-iot-release-smart-clock-r6android-o-mr1-iot-release-smart-clock-r2android-o-mr1-iot-release-smart-clock-fsiandroid-o-mr1-iot-release-smart-clock-fcsandroid-o-mr1-iot-release-cube_r2android-o-mr1-iot-release-cube-fsiandroid-o-mr1-iot-release-cube-fcsandroid-n-iot-release-ihome-igv1nougat-iot-release

Bug: 63932386
Test: - verified builds for arm*/x86*
      - checked functionality using an emulator and stagefright

Change-Id: I10c4b267be1c846d8992e3c5f6d2576c2cb258a9
Signed-off-by: Felicia Lim <flim@google.com>

28 files changed, 241 insertions, 111 deletions

diff --git a/celt/arch.h b/celt/arch.h
index d1e6457c..c627a744 100644
--- a/celt/arch.h
+++ b/celt/arch.h
@@ -56,23 +56,40 @@
 
 #define CELT_SIG_SCALE 32768.f
 
-#define celt_fatal(str) _celt_fatal(str, __FILE__, __LINE__);
-#ifdef ENABLE_ASSERTIONS
+#define CELT_FATAL(str) celt_fatal(str, __FILE__, __LINE__);
+
+#if defined(ENABLE_ASSERTIONS) || defined(ENABLE_HARDENING)
+#ifdef __GNUC__
+__attribute__((noreturn))
+#endif
+void celt_fatal(const char *str, const char *file, int line);
+
+#if defined(CELT_C) && !defined(OVERRIDE_celt_fatal)
 #include <stdio.h>
 #include <stdlib.h>
 #ifdef __GNUC__
 __attribute__((noreturn))
 #endif
-static OPUS_INLINE void _celt_fatal(const char *str, const char *file, int line)
+void celt_fatal(const char *str, const char *file, int line)
 {
    fprintf (stderr, "Fatal (internal) error in %s, line %d: %s\n", file, line, str);
    abort();
 }
-#define celt_assert(cond) {if (!(cond)) {celt_fatal("assertion failed: " #cond);}}
-#define celt_assert2(cond, message) {if (!(cond)) {celt_fatal("assertion failed: " #cond "\n" message);}}
+#endif
+
+#define celt_assert(cond) {if (!(cond)) {CELT_FATAL("assertion failed: " #cond);}}
+#define celt_assert2(cond, message) {if (!(cond)) {CELT_FATAL("assertion failed: " #cond "\n" message);}}
+#define MUST_SUCCEED(call) celt_assert((call) == OPUS_OK)
 #else
 #define celt_assert(cond)
 #define celt_assert2(cond, message)
+#define MUST_SUCCEED(call) do {if((call) != OPUS_OK) {RESTORE_STACK; return OPUS_INTERNAL_ERROR;} } while (0)
+#endif
+
+#if defined(ENABLE_ASSERTIONS)
+#define celt_sig_assert(cond) {if (!(cond)) {CELT_FATAL("signal assertion failed: " #cond);}}
+#else
+#define celt_sig_assert(cond)
 #endif
 
 #define IMUL32(a,b) ((a)*(b))
@@ -107,6 +124,8 @@ typedef opus_val32 celt_sig;
 typedef opus_val16 celt_norm;
 typedef opus_val32 celt_ener;
 
+#define celt_isnan(x) 0
+
 #define Q15ONE 32767
 
 #define SIG_SHIFT 12
diff --git a/celt/arm/celt_ne10_fft.c b/celt/arm/celt_fft_ne10.c
index b8c60abe..ea5fd780 100644
--- a/celt/arm/celt_ne10_fft.c
+++ b/celt/arm/celt_fft_ne10.c
@@ -1,7 +1,7 @@
 /* Copyright (c) 2015 Xiph.Org Foundation
    Written by Viswanath Puttagunta */
 /**
-   @file celt_ne10_fft.c
+   @file celt_fft_ne10.c
    @brief ARM Neon optimizations for fft using NE10 library
  */
 
diff --git a/celt/arm/celt_ne10_mdct.c b/celt/arm/celt_mdct_ne10.c
index 293c3efd..3531d02d 100644
--- a/celt/arm/celt_ne10_mdct.c
+++ b/celt/arm/celt_mdct_ne10.c
@@ -1,7 +1,7 @@
 /* Copyright (c) 2015 Xiph.Org Foundation
    Written by Viswanath Puttagunta */
 /**
-   @file celt_ne10_mdct.c
+   @file celt_mdct_ne10.c
    @brief ARM Neon optimizations for mdct using NE10 library
  */
 
diff --git a/celt/arm/celt_neon_intr.c b/celt/arm/celt_neon_intr.c
index cf443988..effda769 100644
--- a/celt/arm/celt_neon_intr.c
+++ b/celt/arm/celt_neon_intr.c
@@ -196,7 +196,7 @@ void celt_pitch_xcorr_float_neon(const opus_val16 *_x, const opus_val16 *_y,
    int i;
    (void)arch;
    celt_assert(max_pitch > 0);
-   celt_assert((((unsigned char *)_x-(unsigned char *)NULL)&3)==0);
+   celt_sig_assert((((unsigned char *)_x-(unsigned char *)NULL)&3)==0);
 
    for (i = 0; i < (max_pitch-3); i += 4) {
       xcorr_kernel_neon_float((const float32_t *)_x, (const float32_t *)_y+i,
diff --git a/celt/arm/pitch_arm.h b/celt/arm/pitch_arm.h
index 4ee13bd0..bed8b04e 100644
--- a/celt/arm/pitch_arm.h
+++ b/celt/arm/pitch_arm.h
@@ -90,7 +90,9 @@ extern opus_val32
 (*const CELT_PITCH_XCORR_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *,
       const opus_val16 *, opus_val32 *, int, int, int);
 #   define OVERRIDE_PITCH_XCORR (1)
-#   define celt_pitch_xcorr (*CELT_PITCH_XCORR_IMPL[(arch)&OPUS_ARCHMASK])
+#   define celt_pitch_xcorr(_x, _y, xcorr, len, max_pitch, arch) \
+  ((*CELT_PITCH_XCORR_IMPL[(arch)&OPUS_ARCHMASK])(_x, _y, \
+        xcorr, len, max_pitch, arch))
 
 #  elif defined(OPUS_ARM_PRESUME_EDSP) || \
     defined(OPUS_ARM_PRESUME_MEDIA) || \
@@ -142,7 +144,9 @@ extern void
       const opus_val16 *, opus_val32 *, int, int, int);
 
 #  define OVERRIDE_PITCH_XCORR (1)
-#  define celt_pitch_xcorr (*CELT_PITCH_XCORR_IMPL[(arch)&OPUS_ARCHMASK])
+#  define celt_pitch_xcorr(_x, _y, xcorr, len, max_pitch, arch) \
+  ((*CELT_PITCH_XCORR_IMPL[(arch)&OPUS_ARCHMASK])(_x, _y, \
+        xcorr, len, max_pitch, arch))
 
 #  elif defined(OPUS_ARM_PRESUME_NEON_INTR)
 
diff --git a/celt/bands.c b/celt/bands.c
index 3b1f5cfb..f7bb66a9 100644
--- a/celt/bands.c
+++ b/celt/bands.c
@@ -70,10 +70,10 @@ opus_int16 bitexact_cos(opus_int16 x)
    opus_int32 tmp;
    opus_int16 x2;
    tmp = (4096+((opus_int32)(x)*(x)))>>13;
-   celt_assert(tmp<=32767);
+   celt_sig_assert(tmp<=32767);
    x2 = tmp;
    x2 = (32767-x2) + FRAC_MUL16(x2, (-7651 + FRAC_MUL16(x2, (8277 + FRAC_MUL16(-626, x2)))));
-   celt_assert(x2<=32766);
+   celt_sig_assert(x2<=32766);
    return 1+x2;
 }
 
@@ -282,7 +282,7 @@ void anti_collapse(const CELTMode *m, celt_norm *X_, unsigned char *collapse_mas
 
       N0 = m->eBands[i+1]-m->eBands[i];
       /* depth in 1/8 bits */
-      celt_assert(pulses[i]>=0);
+      celt_sig_assert(pulses[i]>=0);
       depth = celt_udiv(1+pulses[i], (m->eBands[i+1]-m->eBands[i]))>>LM;
 
 #ifdef FIXED_POINT
@@ -478,7 +478,7 @@ static void stereo_merge(celt_norm * OPUS_RESTRICT X, celt_norm * OPUS_RESTRICT
 /* Decide whether we should spread the pulses in the current frame */
 int spreading_decision(const CELTMode *m, const celt_norm *X, int *average,
       int last_decision, int *hf_average, int *tapset_decision, int update_hf,
-      int end, int C, int M)
+      int end, int C, int M, const int *spread_weight)
 {
    int i, c, N0;
    int sum = 0, nbBands=0;
@@ -519,8 +519,8 @@ int spreading_decision(const CELTMode *m, const celt_norm *X, int *average,
          if (i>m->nbEBands-4)
             hf_sum += celt_udiv(32*(tcount[1]+tcount[0]), N);
          tmp = (2*tcount[2] >= N) + (2*tcount[1] >= N) + (2*tcount[0] >= N);
-         sum += tmp*256;
-         nbBands++;
+         sum += tmp*spread_weight[i];
+         nbBands+=spread_weight[i];
       }
    } while (++c<C);
 
@@ -544,7 +544,7 @@ int spreading_decision(const CELTMode *m, const celt_norm *X, int *average,
    /*printf("%d %d %d\n", hf_sum, *hf_average, *tapset_decision);*/
    celt_assert(nbBands>0); /* end has to be non-zero */
    celt_assert(sum>=0);
-   sum = celt_udiv(sum, nbBands);
+   sum = celt_udiv((opus_int32)sum<<8, nbBands);
    /* Recursive averaging */
    sum = (sum+*average)>>1;
    *average = sum;
@@ -1492,6 +1492,7 @@ void quant_all_bands(int encode, const CELTMode *m, int start, int end,
       else
          Y = NULL;
       N = M*eBands[i+1]-M*eBands[i];
+      celt_assert(N > 0);
       tell = ec_tell_frac(ec);
 
       /* Compute how many bits we want to allocate to this band */
@@ -1507,7 +1508,7 @@ void quant_all_bands(int encode, const CELTMode *m, int start, int end,
          b = 0;
       }
 
-#ifdef ENABLE_UPDATE_DRAFT
+#ifndef DISABLE_UPDATE_DRAFT
       if (resynth && (M*eBands[i]-N >= M*eBands[start] || i==start+1) && (update_lowband || lowband_offset==0))
             lowband_offset = i;
       if (i == start+1)
@@ -1541,7 +1542,7 @@ void quant_all_bands(int encode, const CELTMode *m, int start, int end,
          fold_start = lowband_offset;
          while(M*eBands[--fold_start] > effective_lowband+norm_offset);
          fold_end = lowband_offset-1;
-#ifdef ENABLE_UPDATE_DRAFT
+#ifndef DISABLE_UPDATE_DRAFT
          while(++fold_end < i && M*eBands[fold_end] < effective_lowband+norm_offset+N);
 #else
          while(M*eBands[++fold_end] < effective_lowband+norm_offset+N);
@@ -1621,8 +1622,10 @@ void quant_all_bands(int encode, const CELTMode *m, int start, int end,
                ctx = ctx_save;
                OPUS_COPY(X, X_save, N);
                OPUS_COPY(Y, Y_save, N);
+#ifndef DISABLE_UPDATE_DRAFT
                if (i == start+1)
                   special_hybrid_folding(m, norm, norm2, start, M, dual_stereo);
+#endif
                /* Encode and round up. */
                ctx.theta_round = 1;
                x_cm = quant_band_stereo(&ctx, X, Y, N, b, B,
diff --git a/celt/bands.h b/celt/bands.h
index 2488c185..422b32cf 100644
--- a/celt/bands.h
+++ b/celt/bands.h
@@ -72,7 +72,7 @@ void denormalise_bands(const CELTMode *m, const celt_norm * OPUS_RESTRICT X,
 
 int spreading_decision(const CELTMode *m, const celt_norm *X, int *average,
       int last_decision, int *hf_average, int *tapset_decision, int update_hf,
-      int end, int C, int M);
+      int end, int C, int M, const int *spread_weight);
 
 #ifdef MEASURE_NORM_MSE
 void measure_norm_mse(const CELTMode *m, float *X, float *X0, float *bandE, float *bandE0, int M, int N, int C);
diff --git a/celt/celt.h b/celt/celt.h
index 70175301..24b6b2b5 100644
--- a/celt/celt.h
+++ b/celt/celt.h
@@ -59,9 +59,11 @@ typedef struct {
    float noisiness;
    float activity;
    float music_prob;
-   float vad_prob;
+   float music_prob_min;
+   float music_prob_max;
    int   bandwidth;
    float activity_probability;
+   float max_pitch_ratio;
    /* Store as Q6 char to save space. */
    unsigned char leak_boost[LEAK_BANDS];
 } AnalysisInfo;
@@ -207,6 +209,13 @@ static OPUS_INLINE int fromOpus(unsigned char c)
 
 extern const signed char tf_select_table[4][8];
 
+#if defined(ENABLE_HARDENING) || defined(ENABLE_ASSERTIONS)
+void validate_celt_decoder(CELTDecoder *st);
+#define VALIDATE_CELT_DECODER(st) validate_celt_decoder(st)
+#else
+#define VALIDATE_CELT_DECODER(st)
+#endif
+
 int resampling_factor(opus_int32 rate);
 
 void celt_preemphasis(const opus_val16 * OPUS_RESTRICT pcmp, celt_sig * OPUS_RESTRICT inp,
diff --git a/celt/celt_decoder.c b/celt/celt_decoder.c
index 567d7456..e6efce93 100644
--- a/celt/celt_decoder.c
+++ b/celt/celt_decoder.c
@@ -51,6 +51,14 @@
 #include "celt_lpc.h"
 #include "vq.h"
 
+/* The maximum pitch lag to allow in the pitch-based PLC. It's possible to save
+   CPU time in the PLC pitch search by making this smaller than MAX_PERIOD. The
+   current value corresponds to a pitch of 66.67 Hz. */
+#define PLC_PITCH_LAG_MAX (720)
+/* The minimum pitch lag to allow in the pitch-based PLC. This corresponds to a
+   pitch of 480 Hz. */
+#define PLC_PITCH_LAG_MIN (100)
+
 #if defined(SMALL_FOOTPRINT) && defined(FIXED_POINT)
 #define NORM_ALIASING_HACK
 #endif
@@ -101,6 +109,38 @@ struct OpusCustomDecoder {
    /* opus_val16 backgroundLogE[], Size = 2*mode->nbEBands */
 };
 
+#if defined(ENABLE_HARDENING) || defined(ENABLE_ASSERTIONS)
+/* Make basic checks on the CELT state to ensure we don't end
+   up writing all over memory. */
+void validate_celt_decoder(CELTDecoder *st)
+{
+#ifndef CUSTOM_MODES
+   celt_assert(st->mode == opus_custom_mode_create(48000, 960, NULL));
+   celt_assert(st->overlap == 120);
+#endif
+   celt_assert(st->channels == 1 || st->channels == 2);
+   celt_assert(st->stream_channels == 1 || st->stream_channels == 2);
+   celt_assert(st->downsample > 0);
+   celt_assert(st->start == 0 || st->start == 17);
+   celt_assert(st->start < st->end);
+   celt_assert(st->end <= 21);
+#ifdef OPUS_ARCHMASK
+   celt_assert(st->arch >= 0);
+   celt_assert(st->arch <= OPUS_ARCHMASK);
+#endif
+   celt_assert(st->last_pitch_index <= PLC_PITCH_LAG_MAX);
+   celt_assert(st->last_pitch_index >= PLC_PITCH_LAG_MIN || st->last_pitch_index == 0);
+   celt_assert(st->postfilter_period < MAX_PERIOD);
+   celt_assert(st->postfilter_period >= COMBFILTER_MINPERIOD || st->postfilter_period == 0);
+   celt_assert(st->postfilter_period_old < MAX_PERIOD);
+   celt_assert(st->postfilter_period_old >= COMBFILTER_MINPERIOD || st->postfilter_period_old == 0);
+   celt_assert(st->postfilter_tapset <= 2);
+   celt_assert(st->postfilter_tapset >= 0);
+   celt_assert(st->postfilter_tapset_old <= 2);
+   celt_assert(st->postfilter_tapset_old >= 0);
+}
+#endif
+
 int celt_decoder_get_size(int channels)
 {
    const CELTMode *mode = opus_custom_mode_create(48000, 960, NULL);
@@ -164,7 +204,7 @@ OPUS_CUSTOM_NOSTATIC int opus_custom_decoder_init(CELTDecoder *st, const CELTMod
    st->start = 0;
    st->end = st->mode->effEBands;
    st->signalling = 1;
-#ifdef ENABLE_UPDATE_DRAFT
+#ifndef DISABLE_UPDATE_DRAFT
    st->disable_inv = channels == 1;
 #else
    st->disable_inv = 0;
@@ -437,14 +477,6 @@ static void tf_decode(int start, int end, int isTransient, int *tf_res, int LM,
    }
 }
 
-/* The maximum pitch lag to allow in the pitch-based PLC. It's possible to save
-   CPU time in the PLC pitch search by making this smaller than MAX_PERIOD. The
-   current value corresponds to a pitch of 66.67 Hz. */
-#define PLC_PITCH_LAG_MAX (720)
-/* The minimum pitch lag to allow in the pitch-based PLC. This corresponds to a
-   pitch of 480 Hz. */
-#define PLC_PITCH_LAG_MIN (100)
-
 static int celt_plc_pitch_search(celt_sig *decode_mem[2], int C, int arch)
 {
    int pitch_index;
@@ -554,6 +586,7 @@ static void celt_decode_lost(CELTDecoder * OPUS_RESTRICT st, int N, int LM)
 
       celt_synthesis(mode, X, out_syn, oldBandE, start, effEnd, C, C, 0, LM, st->downsample, 0, st->arch);
    } else {
+      int exc_length;
       /* Pitch-based PLC */
       const opus_val16 *window;
       opus_val16 *exc;
@@ -561,6 +594,7 @@ static void celt_decode_lost(CELTDecoder * OPUS_RESTRICT st, int N, int LM)
       int pitch_index;
       VARDECL(opus_val32, etmp);
       VARDECL(opus_val16, _exc);
+      VARDECL(opus_val16, fir_tmp);
 
       if (loss_count == 0)
       {
@@ -570,8 +604,13 @@ static void celt_decode_lost(CELTDecoder * OPUS_RESTRICT st, int N, int LM)
          fade = QCONST16(.8f,15);
       }
 
+      /* We want the excitation for 2 pitch periods in order to look for a
+         decaying signal, but we can't get more than MAX_PERIOD. */
+      exc_length = IMIN(2*pitch_index, MAX_PERIOD);
+
       ALLOC(etmp, overlap, opus_val32);
       ALLOC(_exc, MAX_PERIOD+LPC_ORDER, opus_val16);
+      ALLOC(fir_tmp, exc_length, opus_val16);
       exc = _exc+LPC_ORDER;
       window = mode->window;
       c=0; do {
@@ -581,13 +620,11 @@ static void celt_decode_lost(CELTDecoder * OPUS_RESTRICT st, int N, int LM)
          celt_sig *buf;
          int extrapolation_offset;
          int extrapolation_len;
-         int exc_length;
          int j;
 
          buf = decode_mem[c];
-         for (i=0;i<MAX_PERIOD;i++) {
-            exc[i] = ROUND16(buf[DECODE_BUFFER_SIZE-MAX_PERIOD+i], SIG_SHIFT);
-         }
+         for (i=0;i<MAX_PERIOD+LPC_ORDER;i++)
+            exc[i-LPC_ORDER] = ROUND16(buf[DECODE_BUFFER_SIZE-MAX_PERIOD-LPC_ORDER+i], SIG_SHIFT);
 
          if (loss_count == 0)
          {
@@ -631,20 +668,14 @@ static void celt_decode_lost(CELTDecoder * OPUS_RESTRICT st, int N, int LM)
          }
 #endif
          }
-         /* We want the excitation for 2 pitch periods in order to look for a
-            decaying signal, but we can't get more than MAX_PERIOD. */
-         exc_length = IMIN(2*pitch_index, MAX_PERIOD);
          /* Initialize the LPC history with the samples just before the start
             of the region for which we're computing the excitation. */
          {
-            for (i=0;i<LPC_ORDER;i++)
-            {
-               exc[MAX_PERIOD-exc_length-LPC_ORDER+i] =
-                     ROUND16(buf[DECODE_BUFFER_SIZE-exc_length-LPC_ORDER+i], SIG_SHIFT);
-            }
-            /* Compute the excitation for exc_length samples before the loss. */
+            /* Compute the excitation for exc_length samples before the loss. We need the copy
+               because celt_fir() cannot filter in-place. */
             celt_fir(exc+MAX_PERIOD-exc_length, lpc+c*LPC_ORDER,
-                  exc+MAX_PERIOD-exc_length, exc_length, LPC_ORDER, st->arch);
+                  fir_tmp, exc_length, LPC_ORDER, st->arch);
+            OPUS_COPY(exc+MAX_PERIOD-exc_length, fir_tmp, exc_length);
          }
 
          /* Check if the waveform is decaying, and if so how fast.
@@ -833,6 +864,7 @@ int celt_decode_with_ec(CELTDecoder * OPUS_RESTRICT st, const unsigned char *dat
    const opus_int16 *eBands;
    ALLOC_STACK;
 
+   VALIDATE_CELT_DECODER(st);
    mode = st->mode;
    nbEBands = mode->nbEBands;
    overlap = mode->overlap;
@@ -1027,7 +1059,7 @@ int celt_decode_with_ec(CELTDecoder * OPUS_RESTRICT st, const unsigned char *dat
    ALLOC(pulses, nbEBands, int);
    ALLOC(fine_priority, nbEBands, int);
 
-   codedBands = compute_allocation(mode, start, end, offsets, cap,
+   codedBands = clt_compute_allocation(mode, start, end, offsets, cap,
          alloc_trim, &intensity, &dual_stereo, bits, &balance, pulses,
          fine_quant, fine_priority, C, LM, dec, 0, 0, 0);
 
diff --git a/celt/celt_encoder.c b/celt/celt_encoder.c
index 053e5a3b..44cb0850 100644
--- a/celt/celt_encoder.c
+++ b/celt/celt_encoder.c
@@ -362,6 +362,12 @@ static int transient_analysis(const opus_val32 * OPUS_RESTRICT in, int len, int
       /* Compute harmonic mean discarding the unreliable boundaries
          The data is smooth, so we only take 1/4th of the samples */
       unmask=0;
+      /* We should never see NaNs here. If we find any, then something really bad happened and we better abort
+         before it does any damage later on. If these asserts are disabled (no hardening), then the table
+         lookup a few lines below (id = ...) is likely to crash dur to an out-of-bounds read. DO NOT FIX
+         that crash on NaN since it could result in a worse issue later on. */
+      celt_assert(!celt_isnan(tmp[0]));
+      celt_assert(!celt_isnan(norm));
       for (i=12;i<len2-5;i+=4)
       {
          int id;
@@ -577,7 +583,7 @@ static opus_val32 l1_metric(const celt_norm *tmp, int N, int LM, opus_val16 bias
 
 static int tf_analysis(const CELTMode *m, int len, int isTransient,
       int *tf_res, int lambda, celt_norm *X, int N0, int LM,
-      opus_val16 tf_estimate, int tf_chan)
+      opus_val16 tf_estimate, int tf_chan, int *importance)
 {
    int i;
    VARDECL(int, metric);
@@ -660,22 +666,22 @@ static int tf_analysis(const CELTMode *m, int len, int isTransient,
          biasing the decision */
       if (narrow && (metric[i]==0 || metric[i]==-2*LM))
          metric[i]-=1;
-      /*printf("%d ", metric[i]);*/
+      /*printf("%d ", metric[i]/2 + (!isTransient)*LM);*/
    }
    /*printf("\n");*/
    /* Search for the optimal tf resolution, including tf_select */
    tf_select = 0;
    for (sel=0;sel<2;sel++)
    {
-      cost0 = 0;
-      cost1 = isTransient ? 0 : lambda;
+      cost0 = importance[0]*abs(metric[0]-2*tf_select_table[LM][4*isTransient+2*sel+0]);
+      cost1 = importance[0]*abs(metric[0]-2*tf_select_table[LM][4*isTransient+2*sel+1]) + (isTransient ? 0 : lambda);
       for (i=1;i<len;i++)
       {
          int curr0, curr1;
          curr0 = IMIN(cost0, cost1 + lambda);
          curr1 = IMIN(cost0 + lambda, cost1);
-         cost0 = curr0 + abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*sel+0]);
-         cost1 = curr1 + abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*sel+1]);
+         cost0 = curr0 + importance[i]*abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*sel+0]);
+         cost1 = curr1 + importance[i]*abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*sel+1]);
       }
       cost0 = IMIN(cost0, cost1);
       selcost[sel]=cost0;
@@ -684,8 +690,8 @@ static int tf_analysis(const CELTMode *m, int len, int isTransient,
     * If tests confirm it's useful for non-transients, we could allow it. */
    if (selcost[1]<selcost[0] && isTransient)
       tf_select=1;
-   cost0 = 0;
-   cost1 = isTransient ? 0 : lambda;
+   cost0 = importance[0]*abs(metric[0]-2*tf_select_table[LM][4*isTransient+2*tf_select+0]);
+   cost1 = importance[0]*abs(metric[0]-2*tf_select_table[LM][4*isTransient+2*tf_select+1]) + (isTransient ? 0 : lambda);
    /* Viterbi forward pass */
    for (i=1;i<len;i++)
    {
@@ -713,8 +719,8 @@ static int tf_analysis(const CELTMode *m, int len, int isTransient,
          curr1 = from1;
          path1[i]= 1;
       }
-      cost0 = curr0 + abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*tf_select+0]);
-      cost1 = curr1 + abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*tf_select+1]);
+      cost0 = curr0 + importance[i]*abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*tf_select+0]);
+      cost1 = curr1 + importance[i]*abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*tf_select+1]);
    }
    tf_res[len-1] = cost0 < cost1 ? 0 : 1;
    /* Viterbi backward pass to check the decisions */
@@ -964,7 +970,8 @@ static opus_val16 median_of_3(const opus_val16 *x)
 static opus_val16 dynalloc_analysis(const opus_val16 *bandLogE, const opus_val16 *bandLogE2,
       int nbEBands, int start, int end, int C, int *offsets, int lsb_depth, const opus_int16 *logN,
       int isTransient, int vbr, int constrained_vbr, const opus_int16 *eBands, int LM,
-      int effectiveBytes, opus_int32 *tot_boost_, int lfe, opus_val16 *surround_dynalloc, AnalysisInfo *analysis)
+      int effectiveBytes, opus_int32 *tot_boost_, int lfe, opus_val16 *surround_dynalloc,
+      AnalysisInfo *analysis, int *importance, int *spread_weight)
 {
    int i, c;
    opus_int32 tot_boost=0;
@@ -990,6 +997,42 @@ static opus_val16 dynalloc_analysis(const opus_val16 *bandLogE, const opus_val16
       for (i=0;i<end;i++)
          maxDepth = MAX16(maxDepth, bandLogE[c*nbEBands+i]-noise_floor[i]);
    } while (++c<C);
+   {
+      /* Compute a really simple masking model to avoid taking into account completely masked
+         bands when computing the spreading decision. */
+      VARDECL(opus_val16, mask);
+      VARDECL(opus_val16, sig);
+      ALLOC(mask, nbEBands, opus_val16);
+      ALLOC(sig, nbEBands, opus_val16);
+      for (i=0;i<end;i++)
+         mask[i] = bandLogE[i]-noise_floor[i];
+      if (C==2)
+      {
+         for (i=0;i<end;i++)
+            mask[i] = MAX16(mask[i], bandLogE[nbEBands+i]-noise_floor[i]);
+      }
+      OPUS_COPY(sig, mask, end);
+      for (i=1;i<end;i++)
+         mask[i] = MAX16(mask[i], mask[i-1] - QCONST16(2.f, DB_SHIFT));
+      for (i=end-2;i>=0;i--)
+         mask[i] = MAX16(mask[i], mask[i+1] - QCONST16(3.f, DB_SHIFT));
+      for (i=0;i<end;i++)
+      {
+         /* Compute SMR: Mask is never more than 72 dB below the peak and never below the noise floor.*/
+         opus_val16 smr = sig[i]-MAX16(MAX16(0, maxDepth-QCONST16(12.f, DB_SHIFT)), mask[i]);
+         /* Clamp SMR to make sure we're not shifting by something negative or too large. */
+#ifdef FIXED_POINT
+         /* FIXME: Use PSHR16() instead */
+         int shift = -PSHR32(MAX16(-QCONST16(5.f, DB_SHIFT), MIN16(0, smr)), DB_SHIFT);
+#else
+         int shift = IMIN(5, IMAX(0, -(int)floor(.5f + smr)));
+#endif
+         spread_weight[i] = 32 >> shift;
+      }
+      /*for (i=0;i<end;i++)
+         printf("%d ", spread_weight[i]);
+      printf("\n");*/
+   }
    /* Make sure that dynamic allocation can't make us bust the budget */
    if (effectiveBytes > 50 && LM>=1 && !lfe)
    {
@@ -1046,6 +1089,14 @@ static opus_val16 dynalloc_analysis(const opus_val16 *bandLogE, const opus_val16
       }
       for (i=start;i<end;i++)
          follower[i] = MAX16(follower[i], surround_dynalloc[i]);
+      for (i=start;i<end;i++)
+      {
+#ifdef FIXED_POINT
+         importance[i] = PSHR32(13*celt_exp2(MIN16(follower[i], QCONST16(4.f, DB_SHIFT))), 16);
+#else
+         importance[i] = (int)floor(.5f+13*celt_exp2(MIN16(follower[i], QCONST16(4.f, DB_SHIFT))));
+#endif
+      }
       /* For non-transient CBR/CVBR frames, halve the dynalloc contribution */
       if ((!vbr || constrained_vbr)&&!isTransient)
       {
@@ -1101,6 +1152,9 @@ static opus_val16 dynalloc_analysis(const opus_val16 *bandLogE, const opus_val16
             tot_boost += boost_bits;
          }
       }
+   } else {
+      for (i=start;i<end;i++)
+         importance[i] = 13;
    }
    *tot_boost_ = tot_boost;
    RESTORE_STACK;
@@ -1109,7 +1163,7 @@ static opus_val16 dynalloc_analysis(const opus_val16 *bandLogE, const opus_val16
 
 
 static int run_prefilter(CELTEncoder *st, celt_sig *in, celt_sig *prefilter_mem, int CC, int N,
-      int prefilter_tapset, int *pitch, opus_val16 *gain, int *qgain, int enabled, int nbAvailableBytes)
+      int prefilter_tapset, int *pitch, opus_val16 *gain, int *qgain, int enabled, int nbAvailableBytes, AnalysisInfo *analysis)
 {
    int c;
    VARDECL(celt_sig, _pre);
@@ -1165,7 +1219,12 @@ static int run_prefilter(CELTEncoder *st, celt_sig *in, celt_sig *prefilter_mem,
       gain1 = 0;
       pitch_index = COMBFILTER_MINPERIOD;
    }
-
+#ifndef DISABLE_FLOAT_API
+   if (analysis->valid)
+      gain1 = (opus_val16)(gain1 * analysis->max_pitch_ratio);
+#else
+   (void)analysis;
+#endif
    /* Gain threshold for enabling the prefilter/postfilter */
    pf_threshold = QCONST16(.2f,15);
 
@@ -1362,6 +1421,8 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm,
    VARDECL(int, pulses);
    VARDECL(int, cap);
    VARDECL(int, offsets);
+   VARDECL(int, importance);
+   VARDECL(int, spread_weight);
    VARDECL(int, fine_priority);
    VARDECL(int, tf_res);
    VARDECL(unsigned char, collapse_masks);
@@ -1414,6 +1475,7 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm,
    opus_int32 equiv_rate;
    int hybrid;
    int weak_transient = 0;
+   int enable_tf_analysis;
    VARDECL(opus_val16, surround_dynalloc);
    ALLOC_STACK;
 
@@ -1454,7 +1516,7 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm,
       tell0_frac=tell=1;
       nbFilledBytes=0;
    } else {
-      tell0_frac=tell=ec_tell_frac(enc);
+      tell0_frac=ec_tell_frac(enc);
       tell=ec_tell(enc);
       nbFilledBytes=(tell+4)>>3;
    }
@@ -1603,7 +1665,7 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm,
             && st->complexity >= 5;
 
       prefilter_tapset = st->tapset_decision;
-      pf_on = run_prefilter(st, in, prefilter_mem, CC, N, prefilter_tapset, &pitch_index, &gain1, &qg, enabled, nbAvailableBytes);
+      pf_on = run_prefilter(st, in, prefilter_mem, CC, N, prefilter_tapset, &pitch_index, &gain1, &qg, enabled, nbAvailableBytes, &st->analysis);
       if ((gain1 > QCONST16(.4f,15) || st->prefilter_gain > QCONST16(.4f,15)) && (!st->analysis.valid || st->analysis.tonality > .3)
             && (pitch_index > 1.26*st->prefilter_period || pitch_index < .79*st->prefilter_period))
          pitch_change = 1;
@@ -1633,7 +1695,7 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm,
       /* Reduces the likelihood of energy instability on fricatives at low bitrate
          in hybrid mode. It seems like we still want to have real transients on vowels
          though (small SILK quantization offset value). */
-      int allow_weak_transients = hybrid && effectiveBytes<15 && st->silk_info.offset >= 100;
+      int allow_weak_transients = hybrid && effectiveBytes<15 && st->silk_info.signalType != 2;
       isTransient = transient_analysis(in, N+overlap, CC,
             &tf_estimate, &tf_chan, allow_weak_transients, &weak_transient);
    }
@@ -1662,6 +1724,9 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm,
    }
 
    compute_mdcts(mode, shortBlocks, in, freq, C, CC, LM, st->upsample, st->arch);
+   /* This should catch any NaN in the CELT input. Since we're not supposed to see any (they're filtered
+      at the Opus layer), just abort. */
+   celt_assert(!celt_isnan(freq[0]) && (C==1 || !celt_isnan(freq[N])));
    if (CC==2&&C==1)
       tf_chan = 0;
    compute_band_energies(mode, freq, bandE, effEnd, C, LM, st->arch);
@@ -1805,13 +1870,23 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm,
    /* Band normalisation */
    normalise_bands(mode, freq, X, bandE, effEnd, C, M);
 
+   enable_tf_analysis = effectiveBytes>=15*C && !hybrid && st->complexity>=2 && !st->lfe;
+
+   ALLOC(offsets, nbEBands, int);
+   ALLOC(importance, nbEBands, int);
+   ALLOC(spread_weight, nbEBands, int);
+
+   maxDepth = dynalloc_analysis(bandLogE, bandLogE2, nbEBands, start, end, C, offsets,
+         st->lsb_depth, mode->logN, isTransient, st->vbr, st->constrained_vbr,
+         eBands, LM, effectiveBytes, &tot_boost, st->lfe, surround_dynalloc, &st->analysis, importance, spread_weight);
+
    ALLOC(tf_res, nbEBands, int);
    /* Disable variable tf resolution for hybrid and at very low bitrate */
-   if (effectiveBytes>=15*C && !hybrid && st->complexity>=2 && !st->lfe)
+   if (enable_tf_analysis)
    {
       int lambda;
-      lambda = IMAX(5, 1280/effectiveBytes + 2);
-      tf_select = tf_analysis(mode, effEnd, isTransient, tf_res, lambda, X, N, LM, tf_estimate, tf_chan);
+      lambda = IMAX(80, 20480/effectiveBytes + 2);
+      tf_select = tf_analysis(mode, effEnd, isTransient, tf_res, lambda, X, N, LM, tf_estimate, tf_chan, importance);
       for (i=effEnd;i<end;i++)
          tf_res[i] = tf_res[effEnd-1];
    } else if (hybrid && weak_transient)
@@ -1822,7 +1897,7 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm,
       for (i=0;i<end;i++)
          tf_res[i] = 1;
       tf_select=0;
-   } else if (hybrid && effectiveBytes<15)
+   } else if (hybrid && effectiveBytes<15 && st->silk_info.signalType != 2)
    {
       /* For low bitrate hybrid, we force temporal resolution to 5 ms rather than 2.5 ms. */
       for (i=0;i<end;i++)
@@ -1893,7 +1968,7 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm,
          {
             st->spread_decision = spreading_decision(mode, X,
                   &st->tonal_average, st->spread_decision, &st->hf_average,
-                  &st->tapset_decision, pf_on&&!shortBlocks, effEnd, C, M);
+                  &st->tapset_decision, pf_on&&!shortBlocks, effEnd, C, M, spread_weight);
          }
          /*printf("%d %d\n", st->tapset_decision, st->spread_decision);*/
          /*printf("%f %d %f %d\n\n", st->analysis.tonality, st->spread_decision, st->analysis.tonality_slope, st->tapset_decision);*/
@@ -1901,11 +1976,6 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm,
       ec_enc_icdf(enc, st->spread_decision, spread_icdf, 5);
    }
 
-   ALLOC(offsets, nbEBands, int);
-
-   maxDepth = dynalloc_analysis(bandLogE, bandLogE2, nbEBands, start, end, C, offsets,
-         st->lsb_depth, mode->logN, isTransient, st->vbr, st->constrained_vbr,
-         eBands, LM, effectiveBytes, &tot_boost, st->lfe, surround_dynalloc, &st->analysis);
    /* For LFE, everything interesting is in the first band */
    if (st->lfe)
       offsets[0] = IMIN(8, effectiveBytes/3);
@@ -2121,7 +2191,7 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm,
 #endif
    if (st->lfe)
       signalBandwidth = 1;
-   codedBands = compute_allocation(mode, start, end, offsets, cap,
+   codedBands = clt_compute_allocation(mode, start, end, offsets, cap,
          alloc_trim, &st->intensity, &dual_stereo, bits, &balance, pulses,
          fine_quant, fine_priority, C, LM, enc, 1, st->lastCodedBands, signalBandwidth);
    if (st->lastCodedBands)
diff --git a/celt/celt_lpc.c b/celt/celt_lpc.c
index 57bc5f3d..8ecb693e 100644
--- a/celt/celt_lpc.c
+++ b/celt/celt_lpc.c
@@ -99,7 +99,7 @@ void celt_fir_c(
    int i,j;
    VARDECL(opus_val16, rnum);
    SAVE_STACK;
-
+   celt_assert(x != y);
    ALLOC(rnum, ord, opus_val16);
    for(i=0;i<ord;i++)
       rnum[i] = num[ord-i-1];
@@ -107,7 +107,7 @@ void celt_fir_c(
    {
       opus_val32 sum[4];
       sum[0] = SHL32(EXTEND32(x[i  ]), SIG_SHIFT);
-      sum[1] = SHL32(EXTEND32(x[i+1]), SIG_SHIFT),
+      sum[1] = SHL32(EXTEND32(x[i+1]), SIG_SHIFT);
       sum[2] = SHL32(EXTEND32(x[i+2]), SIG_SHIFT);
       sum[3] = SHL32(EXTEND32(x[i+3]), SIG_SHIFT);
       xcorr_kernel(rnum, x+i-ord, sum, ord, arch);
diff --git a/celt/cwrs.c b/celt/cwrs.c
index 9722f0ac..a552e4f0 100644
--- a/celt/cwrs.c
+++ b/celt/cwrs.c
@@ -482,7 +482,7 @@ static opus_val32 cwrsi(int _n,int _k,opus_uint32 _i,int *_y){
       k0=_k;
       q=row[_n];
       if(q>_i){
-        celt_assert(p>q);
+        celt_sig_assert(p>q);
         _k=_n;
         do p=CELT_PVQ_U_ROW[--_k][_n];
         while(p>_i);
diff --git a/celt/entcode.h b/celt/entcode.h
index 13d6c84e..3763e3f2 100644
--- a/celt/entcode.h
+++ b/celt/entcode.h
@@ -122,7 +122,7 @@ opus_uint32 ec_tell_frac(ec_ctx *_this);
 
 /* Tested exhaustively for all n and for 1<=d<=256 */
 static OPUS_INLINE opus_uint32 celt_udiv(opus_uint32 n, opus_uint32 d) {
-   celt_assert(d>0);
+   celt_sig_assert(d>0);
 #ifdef USE_SMALL_DIV_TABLE
    if (d>256)
       return n/d;
@@ -138,7 +138,7 @@ static OPUS_INLINE opus_uint32 celt_udiv(opus_uint32 n, opus_uint32 d) {
 }
 
 static OPUS_INLINE opus_int32 celt_sudiv(opus_int32 n, opus_int32 d) {
-   celt_assert(d>0);
+   celt_sig_assert(d>0);
 #ifdef USE_SMALL_DIV_TABLE
    if (n<0)
       return -(opus_int32)celt_udiv(-n, d);
diff --git a/celt/entdec.h b/celt/entdec.h
index d8ab3187..025fc187 100644
--- a/celt/entdec.h
+++ b/celt/entdec.h
@@ -85,7 +85,7 @@ int ec_dec_icdf(ec_dec *_this,const unsigned char *_icdf,unsigned _ftb);
   The bits must have been encoded with ec_enc_uint().
   No call to ec_dec_update() is necessary after this call.
   _ft: The number of integers that can be decoded (one more than the max).
-       This must be at least one, and no more than 2**32-1.
+       This must be at least 2, and no more than 2**32-1.
   Return: The decoded bits.*/
 opus_uint32 ec_dec_uint(ec_dec *_this,opus_uint32 _ft);
 
diff --git a/celt/entenc.h b/celt/entenc.h
index 796bc4d5..f502eaf6 100644
--- a/celt/entenc.h
+++ b/celt/entenc.h
@@ -67,7 +67,7 @@ void ec_enc_icdf(ec_enc *_this,int _s,const unsigned char *_icdf,unsigned _ftb);
 /*Encodes a raw unsigned integer in the stream.
   _fl: The integer to encode.
   _ft: The number of integers that can be encoded (one more than the max).
-       This must be at least one, and no more than 2**32-1.*/
+       This must be at least 2, and no more than 2**32-1.*/
 void ec_enc_uint(ec_enc *_this,opus_uint32 _fl,opus_uint32 _ft);
 
 /*Encodes a sequence of raw bits in the stream.
diff --git a/celt/float_cast.h b/celt/float_cast.h
index 98b40abc..f218e864 100644
--- a/celt/float_cast.h
+++ b/celt/float_cast.h
@@ -96,7 +96,7 @@ static OPUS_INLINE opus_int32 float2int(float x) {return _mm_cvt_ss2si(_mm_set_s
 #include <math.h>
 #define float2int(x) lrint(x)
 
-#elif (defined(_MSC_VER) && _MSC_VER >= 1400) && defined (_M_X64)
+#elif (defined(_MSC_VER) && _MSC_VER >= 1400) && (defined(_M_X64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1))
         #include <xmmintrin.h>
 
         __inline long int float2int(float value)
diff --git a/celt/mathops.c b/celt/mathops.c
index 21a01f52..6ee9b9e1 100644
--- a/celt/mathops.c
+++ b/celt/mathops.c
@@ -38,7 +38,8 @@
 #include "mathops.h"
 
 /*Compute floor(sqrt(_val)) with exact arithmetic.
-  This has been tested on all possible 32-bit inputs.*/
+  _val must be greater than 0.
+  This has been tested on all possible 32-bit inputs greater than 0.*/
 unsigned isqrt32(opus_uint32 _val){
   unsigned b;
   unsigned g;
@@ -182,7 +183,7 @@ opus_val32 celt_rcp(opus_val32 x)
    int i;
    opus_val16 n;
    opus_val16 r;
-   celt_assert2(x>0, "celt_rcp() only defined for positive values");
+   celt_sig_assert(x>0);
    i = celt_ilog2(x);
    /* n is Q15 with range [0,1). */
    n = VSHR32(x,i-15)-32768;
diff --git a/celt/mathops.h b/celt/mathops.h
index 1f8a20cb..5e86ff0d 100644
--- a/celt/mathops.h
+++ b/celt/mathops.h
@@ -72,7 +72,7 @@ static OPUS_INLINE float fast_atan2f(float y, float x) {
 #undef cA
 #undef cB
 #undef cC
-#undef cD
+#undef cE
 #endif
 
 
@@ -179,7 +179,7 @@ static OPUS_INLINE float celt_exp2(float x)
 /** Integer log in base2. Undefined for zero and negative numbers */
 static OPUS_INLINE opus_int16 celt_ilog2(opus_int32 x)
 {
-   celt_assert2(x>0, "celt_ilog2() only defined for strictly positive numbers");
+   celt_sig_assert(x>0);
    return EC_ILOG(x)-1;
 }
 #endif
diff --git a/celt/pitch.c b/celt/pitch.c
index 5b061d86..872582a4 100644
--- a/celt/pitch.c
+++ b/celt/pitch.c
@@ -102,11 +102,9 @@ static void find_best_pitch(opus_val32 *xcorr, opus_val16 *y, int len,
    }
 }
 
-static void celt_fir5(const opus_val16 *x,
+static void celt_fir5(opus_val16 *x,
          const opus_val16 *num,
-         opus_val16 *y,
-         int N,
-         opus_val16 *mem)
+         int N)
 {
    int i;
    opus_val16 num0, num1, num2, num3, num4;
@@ -116,11 +114,11 @@ static void celt_fir5(const opus_val16 *x,
    num2=num[2];
    num3=num[3];
    num4=num[4];
-   mem0=mem[0];
-   mem1=mem[1];
-   mem2=mem[2];
-   mem3=mem[3];
-   mem4=mem[4];
+   mem0=0;
+   mem1=0;
+   mem2=0;
+   mem3=0;
+   mem4=0;
    for (i=0;i<N;i++)
    {
       opus_val32 sum = SHL32(EXTEND32(x[i]), SIG_SHIFT);
@@ -134,13 +132,8 @@ static void celt_fir5(const opus_val16 *x,
       mem2 = mem1;
       mem1 = mem0;
       mem0 = x[i];
-      y[i] = ROUND16(sum, SIG_SHIFT);
+      x[i] = ROUND16(sum, SIG_SHIFT);
    }
-   mem[0]=mem0;
-   mem[1]=mem1;
-   mem[2]=mem2;
-   mem[3]=mem3;
-   mem[4]=mem4;
 }
 
 
@@ -150,7 +143,7 @@ void pitch_downsample(celt_sig * OPUS_RESTRICT x[], opus_val16 * OPUS_RESTRICT x
    int i;
    opus_val32 ac[5];
    opus_val16 tmp=Q15ONE;
-   opus_val16 lpc[4], mem[5]={0,0,0,0,0};
+   opus_val16 lpc[4];
    opus_val16 lpc2[5];
    opus_val16 c1 = QCONST16(.8f,15);
 #ifdef FIXED_POINT
@@ -211,7 +204,7 @@ void pitch_downsample(celt_sig * OPUS_RESTRICT x[], opus_val16 * OPUS_RESTRICT x
    lpc2[2] = lpc[2] + MULT16_16_Q15(c1,lpc[1]);
    lpc2[3] = lpc[3] + MULT16_16_Q15(c1,lpc[2]);
    lpc2[4] = MULT16_16_Q15(c1,lpc[3]);
-   celt_fir5(x_lp, lpc2, x_lp, len>>1, mem);
+   celt_fir5(x_lp, lpc2, len>>1);
 }
 
 /* Pure C implementation. */
@@ -256,7 +249,7 @@ celt_pitch_xcorr_c(const opus_val16 *_x, const opus_val16 *_y,
    opus_val32 maxcorr=1;
 #endif
    celt_assert(max_pitch>0);
-   celt_assert((((unsigned char *)_x-(unsigned char *)NULL)&3)==0);
+   celt_sig_assert((((unsigned char *)_x-(unsigned char *)NULL)&3)==0);
    for (i=0;i<max_pitch-3;i+=4)
    {
       opus_val32 sum[4]={0,0,0,0};
diff --git a/celt/quant_bands.c b/celt/quant_bands.c
index 56101b1c..39a221ed 100644
--- a/celt/quant_bands.c
+++ b/celt/quant_bands.c
@@ -457,7 +457,7 @@ void unquant_coarse_energy(const CELTMode *m, int start, int end, opus_val16 *ol
          /* It would be better to express this invariant as a
             test on C at function entry, but that isn't enough
             to make the static analyzer happy. */
-         celt_assert(c<2);
+         celt_sig_assert(c<2);
          tell = ec_tell(dec);
          if(budget-tell>=15)
          {
diff --git a/celt/rate.c b/celt/rate.c
index ca4cc870..465e1ba2 100644
--- a/celt/rate.c
+++ b/celt/rate.c
@@ -529,7 +529,7 @@ static OPUS_INLINE int interp_bits2pulses(const CELTMode *m, int start, int end,
    return codedBands;
 }
 
-int compute_allocation(const CELTMode *m, int start, int end, const int *offsets, const int *cap, int alloc_trim, int *intensity, int *dual_stereo,
+int clt_compute_allocation(const CELTMode *m, int start, int end, const int *offsets, const int *cap, int alloc_trim, int *intensity, int *dual_stereo,
       opus_int32 total, opus_int32 *balance, int *pulses, int *ebits, int *fine_priority, int C, int LM, ec_ctx *ec, int encode, int prev, int signalBandwidth)
 {
    int lo, hi, len, j;
diff --git a/celt/rate.h b/celt/rate.h
index 515f7687..852b9d6f 100644
--- a/celt/rate.h
+++ b/celt/rate.h
@@ -95,7 +95,7 @@ static OPUS_INLINE int pulses2bits(const CELTMode *m, int band, int LM, int puls
  @param pulses Number of pulses per band (returned)
  @return Total number of bits allocated
 */
-int compute_allocation(const CELTMode *m, int start, int end, const int *offsets, const int *cap, int alloc_trim, int *intensity, int *dual_stero,
+int clt_compute_allocation(const CELTMode *m, int start, int end, const int *offsets, const int *cap, int alloc_trim, int *intensity, int *dual_stero,
       opus_int32 total, opus_int32 *balance, int *pulses, int *ebits, int *fine_priority, int C, int LM, ec_ctx *ec, int encode, int prev, int signalBandwidth);
 
 #endif
diff --git a/celt/tests/test_unit_entropy.c b/celt/tests/test_unit_entropy.c
index ff926586..7f674529 100644
--- a/celt/tests/test_unit_entropy.c
+++ b/celt/tests/test_unit_entropy.c
@@ -34,6 +34,7 @@
 #include <stdio.h>
 #include <math.h>
 #include <time.h>
+#define CELT_C
 #include "entcode.h"
 #include "entenc.h"
 #include "entdec.h"
diff --git a/celt/tests/test_unit_laplace.c b/celt/tests/test_unit_laplace.c
index 22951e29..727bf012 100644
--- a/celt/tests/test_unit_laplace.c
+++ b/celt/tests/test_unit_laplace.c
@@ -31,8 +31,8 @@
 
 #include <stdio.h>
 #include <stdlib.h>
-#include "laplace.h"
 #define CELT_C
+#include "laplace.h"
 #include "stack_alloc.h"
 
 #include "entenc.c"
diff --git a/celt/tests/test_unit_rotation.c b/celt/tests/test_unit_rotation.c
index 267b9830..8a31b3f2 100644
--- a/celt/tests/test_unit_rotation.c
+++ b/celt/tests/test_unit_rotation.c
@@ -33,8 +33,6 @@
 #define CUSTOM_MODES
 #endif
 
-#define CELT_C
-
 #include <stdio.h>
 #include <stdlib.h>
 #include "vq.h"
diff --git a/celt/vq.c b/celt/vq.c
index 8ef80e50..a6b5552d 100644
--- a/celt/vq.c
+++ b/celt/vq.c
@@ -230,12 +230,12 @@ opus_val16 op_pvq_search_c(celt_norm *X, int *iy, int K, int N, int arch)
          pulsesLeft -= iy[j];
       }  while (++j<N);
    }
-   celt_assert2(pulsesLeft>=0, "Allocated too many pulses in the quick pass");
+   celt_sig_assert(pulsesLeft>=0);
 
    /* This should never happen, but just in case it does (e.g. on silence)
       we fill the first bin with pulses. */
 #ifdef FIXED_POINT_DEBUG
-   celt_assert2(pulsesLeft<=N+3, "Not enough pulses in the quick pass");
+   celt_sig_assert(pulsesLeft<=N+3);
 #endif
    if (pulsesLeft > N+3)
    {
diff --git a/celt/x86/celt_lpc_sse.c b/celt/x86/celt_lpc_sse4_1.c
index 54785688..54785688 100644
--- a/celt/x86/celt_lpc_sse.c
+++ b/celt/x86/celt_lpc_sse4_1.c
diff --git a/celt/x86/vq_sse2.c b/celt/x86/vq_sse2.c
index 6a317703..77504286 100644
--- a/celt/x86/vq_sse2.c
+++ b/celt/x86/vq_sse2.c
@@ -135,7 +135,7 @@ opus_val16 op_pvq_search_sse2(celt_norm *_X, int *iy, int K, int N, int arch)
    }
    X[N] = X[N+1] = X[N+2] = -100;
    y[N] = y[N+1] = y[N+2] = 100;
-   celt_assert2(pulsesLeft>=0, "Allocated too many pulses in the quick pass");
+   celt_sig_assert(pulsesLeft>=0);
 
    /* This should never happen, but just in case it does (e.g. on silence)
       we fill the first bin with pulses. */