--- a/media/libopus/README_MOZILLA
+++ b/media/libopus/README_MOZILLA
@@ -3,9 +3,9 @@ IETF Opus audio codec reference implemen
 The source in this directory was copied from an opus
 repository checkout by running the ./update.sh script.
 Any changes made to this version of the source should
 be reflected in that script, e.g. by applying patch
 files after the copy step.
 
 The upstream repository is https://git.xiph.org/opus.git
 
-The git tag/revision used was v1.1-beta-23-gf2446c2.
+The git tag/revision used was v1.1-rc2-1-g35a44c6.

--- a/media/libopus/celt/_kiss_fft_guts.h
+++ b/media/libopus/celt/_kiss_fft_guts.h
@@ -89,21 +89,21 @@
 #define C_ADDTO( res , a)\
     do {(res).r = ADD32((res).r, (a).r);  (res).i = ADD32((res).i,(a).i);\
     }while(0)
 
 #define C_SUBFROM( res , a)\
     do {(res).r = ADD32((res).r,(a).r);  (res).i = SUB32((res).i,(a).i); \
     }while(0)
 
-#if defined(ARMv4_ASM)
+#if defined(OPUS_ARM_INLINE_ASM)
 #include "arm/kiss_fft_armv4.h"
 #endif
 
-#if defined(ARMv5E_ASM)
+#if defined(OPUS_ARM_INLINE_EDSP)
 #include "arm/kiss_fft_armv5e.h"
 #endif
 
 #else  /* not FIXED_POINT*/
 
 #   define S_MUL(a,b) ( (a)*(b) )
 #define C_MUL(m,a,b) \
     do{ (m).r = (a).r*(b).r - (a).i*(b).i;\

--- a/media/libopus/celt/arch.h
+++ b/media/libopus/celt/arch.h
@@ -30,16 +30,17 @@
    NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */
 
 #ifndef ARCH_H
 #define ARCH_H
 
 #include "opus_types.h"
+#include "opus_defines.h"
 
 # if !defined(__GNUC_PREREQ)
 #  if defined(__GNUC__)&&defined(__GNUC_MINOR__)
 #   define __GNUC_PREREQ(_maj,_min) \
  ((__GNUC__<<16)+__GNUC_MINOR__>=((_maj)<<16)+(_min))
 #  else
 #   define __GNUC_PREREQ(_maj,_min) 0
 #  endif
@@ -49,17 +50,17 @@
 
 #define celt_fatal(str) _celt_fatal(str, __FILE__, __LINE__);
 #ifdef ENABLE_ASSERTIONS
 #include <stdio.h>
 #include <stdlib.h>
 #ifdef __GNUC__
 __attribute__((noreturn))
 #endif
-static inline void _celt_fatal(const char *str, const char *file, int line)
+static OPUS_INLINE 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);}}
 #else
 #define celt_assert(cond)
@@ -108,19 +109,19 @@ typedef opus_val32 celt_ener;
 #define SCALEOUT(a)     (a)
 
 #ifdef FIXED_DEBUG
 #include "fixed_debug.h"
 #else
 
 #include "fixed_generic.h"
 
-#ifdef ARMv5E_ASM
+#ifdef OPUS_ARM_INLINE_EDSP
 #include "arm/fixed_armv5e.h"
-#elif defined (ARMv4_ASM)
+#elif defined (OPUS_ARM_INLINE_ASM)
 #include "arm/fixed_armv4.h"
 #elif defined (BFIN_ASM)
 #include "fixed_bfin.h"
 #elif defined (TI_C5X_ASM)
 #include "fixed_c5x.h"
 #elif defined (TI_C6X_ASM)
 #include "fixed_c6x.h"
 #endif
@@ -180,16 +181,17 @@ typedef float celt_ener;
 #define MULT16_32_Q15(a,b)     ((a)*(b))
 #define MULT16_32_Q16(a,b)     ((a)*(b))
 
 #define MULT32_32_Q31(a,b)     ((a)*(b))
 
 #define MAC16_32_Q15(c,a,b)     ((c)+(a)*(b))
 
 #define MULT16_16_Q11_32(a,b)     ((a)*(b))
+#define MULT16_16_Q11(a,b)     ((a)*(b))
 #define MULT16_16_Q13(a,b)     ((a)*(b))
 #define MULT16_16_Q14(a,b)     ((a)*(b))
 #define MULT16_16_Q15(a,b)     ((a)*(b))
 #define MULT16_16_P15(a,b)     ((a)*(b))
 #define MULT16_16_P13(a,b)     ((a)*(b))
 #define MULT16_16_P14(a,b)     ((a)*(b))
 #define MULT16_32_P16(a,b)     ((a)*(b))
 

new file mode 100644
--- /dev/null
+++ b/media/libopus/celt/arm/arm_celt_map.c
@@ -0,0 +1,49 @@
+/* Copyright (c) 2010 Xiph.Org Foundation
+ * Copyright (c) 2013 Parrot */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "pitch.h"
+
+#if defined(OPUS_HAVE_RTCD)
+
+# if defined(FIXED_POINT)
+opus_val32 (*const CELT_PITCH_XCORR_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *,
+    const opus_val16 *, opus_val32 *, int , int) = {
+  celt_pitch_xcorr_c,               /* ARMv4 */
+  MAY_HAVE_EDSP(celt_pitch_xcorr),  /* EDSP */
+  MAY_HAVE_MEDIA(celt_pitch_xcorr), /* Media */
+  MAY_HAVE_NEON(celt_pitch_xcorr)   /* NEON */
+};
+# else
+#  error "Floating-point implementation is not supported by ARM asm yet." \
+ "Reconfigure with --disable-rtcd or send patches."
+# endif
+
+#endif

--- a/media/libopus/celt/arm/armcpu.c
+++ b/media/libopus/celt/arm/armcpu.c
@@ -44,41 +44,41 @@
 #define OPUS_CPU_ARM_NEON  (1<<3)
 
 #if defined(_MSC_VER)
 /*For GetExceptionCode() and EXCEPTION_ILLEGAL_INSTRUCTION.*/
 # define WIN32_LEAN_AND_MEAN
 # define WIN32_EXTRA_LEAN
 # include <windows.h>
 
-static inline opus_uint32 opus_cpu_capabilities(void){
+static OPUS_INLINE opus_uint32 opus_cpu_capabilities(void){
   opus_uint32 flags;
   flags=0;
-  /* MSVC has no inline __asm support for ARM, but it does let you __emit
+  /* MSVC has no OPUS_INLINE __asm support for ARM, but it does let you __emit
    * instructions via their assembled hex code.
    * All of these instructions should be essentially nops. */
-# if defined(ARMv5E_ASM)
+# if defined(OPUS_ARM_MAY_HAVE_EDSP)
   __try{
     /*PLD [r13]*/
     __emit(0xF5DDF000);
     flags|=OPUS_CPU_ARM_EDSP;
   }
   __except(GetExceptionCode()==EXCEPTION_ILLEGAL_INSTRUCTION){
     /*Ignore exception.*/
   }
-#  if defined(ARMv6E_ASM)
+#  if defined(OPUS_ARM_MAY_HAVE_MEDIA)
   __try{
     /*SHADD8 r3,r3,r3*/
     __emit(0xE6333F93);
     flags|=OPUS_CPU_ARM_MEDIA;
   }
   __except(GetExceptionCode()==EXCEPTION_ILLEGAL_INSTRUCTION){
     /*Ignore exception.*/
   }
-#   if defined(ARM_HAVE_NEON)
+#   if defined(OPUS_ARM_MAY_HAVE_NEON)
   __try{
     /*VORR q0,q0,q0*/
     __emit(0xF2200150);
     flags|=OPUS_CPU_ARM_NEON;
   }
   __except(GetExceptionCode()==EXCEPTION_ILLEGAL_INSTRUCTION){
     /*Ignore exception.*/
   }
@@ -102,38 +102,46 @@ opus_uint32 opus_cpu_capabilities(void)
   if(cpuinfo != NULL)
   {
     /* 512 should be enough for anybody (it's even enough for all the flags that
      * x86 has accumulated... so far). */
     char buf[512];
 
     while(fgets(buf, 512, cpuinfo) != NULL)
     {
+# if defined(OPUS_ARM_MAY_HAVE_EDSP) || defined(OPUS_ARM_MAY_HAVE_NEON)
       /* Search for edsp and neon flag */
       if(memcmp(buf, "Features", 8) == 0)
       {
         char *p;
+#  if defined(OPUS_ARM_MAY_HAVE_EDSP)
         p = strstr(buf, " edsp");
         if(p != NULL && (p[5] == ' ' || p[5] == '\n'))
           flags |= OPUS_CPU_ARM_EDSP;
+#  endif
 
+#  if defined(OPUS_ARM_MAY_HAVE_NEON)
         p = strstr(buf, " neon");
         if(p != NULL && (p[5] == ' ' || p[5] == '\n'))
           flags |= OPUS_CPU_ARM_NEON;
+#  endif
       }
+# endif
 
+# if defined(OPUS_ARM_MAY_HAVE_MEDIA)
       /* Search for media capabilities (>= ARMv6) */
       if(memcmp(buf, "CPU architecture:", 17) == 0)
       {
         int version;
         version = atoi(buf+17);
 
         if(version >= 6)
           flags |= OPUS_CPU_ARM_MEDIA;
       }
+# endif
     }
 
     fclose(cpuinfo);
   }
   return flags;
 }
 #else
 /* The feature registers which can tell us what the processor supports are

--- a/media/libopus/celt/arm/armcpu.h
+++ b/media/libopus/celt/arm/armcpu.h
@@ -20,16 +20,52 @@
    EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
    PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
    PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
    LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
    NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */
 
-/* Original code from libtheora modified to suit to Opus */
+#if !defined(ARMCPU_H)
+# define ARMCPU_H
+
+# if defined(OPUS_ARM_MAY_HAVE_EDSP)
+#  define MAY_HAVE_EDSP(name) name ## _edsp
+# else
+#  define MAY_HAVE_EDSP(name) name ## _c
+# endif
+
+# if defined(OPUS_ARM_MAY_HAVE_MEDIA)
+#  define MAY_HAVE_MEDIA(name) name ## _media
+# else
+#  define MAY_HAVE_MEDIA(name) MAY_HAVE_EDSP(name)
+# endif
+
+# if defined(OPUS_ARM_MAY_HAVE_NEON)
+#  define MAY_HAVE_NEON(name) name ## _neon
+# else
+#  define MAY_HAVE_NEON(name) MAY_HAVE_MEDIA(name)
+# endif
 
-#ifndef ARMCPU_H
-#define ARMCPU_H
+# if defined(OPUS_ARM_PRESUME_EDSP)
+#  define PRESUME_EDSP(name) name ## _edsp
+# else
+#  define PRESUME_EDSP(name) name ## _c
+# endif
 
+# if defined(OPUS_ARM_PRESUME_MEDIA)
+#  define PRESUME_MEDIA(name) name ## _media
+# else
+#  define PRESUME_MEDIA(name) PRESUME_EDSP(name)
+# endif
+
+# if defined(OPUS_ARM_PRESUME_NEON)
+#  define PRESUME_NEON(name) name ## _neon
+# else
+#  define PRESUME_NEON(name) PRESUME_MEDIA(name)
+# endif
+
+# if defined(OPUS_HAVE_RTCD)
 int opus_select_arch(void);
+# endif
 
 #endif

new file mode 100644
--- /dev/null
+++ b/media/libopus/celt/arm/armopts.s.in
@@ -0,0 +1,37 @@
+/* Copyright (C) 2013 Mozilla Corporation */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+; Set the following to 1 if we have EDSP instructions
+;  (LDRD/STRD, etc., ARMv5E and later).
+OPUS_ARM_MAY_HAVE_EDSP  * @OPUS_ARM_MAY_HAVE_EDSP@
+
+; Set the following to 1 if we have ARMv6 media instructions.
+OPUS_ARM_MAY_HAVE_MEDIA * @OPUS_ARM_MAY_HAVE_MEDIA@
+
+; Set the following to 1 if we have NEON (some ARMv7)
+OPUS_ARM_MAY_HAVE_NEON  * @OPUS_ARM_MAY_HAVE_NEON@
+
+END

new file mode 100644
--- /dev/null
+++ b/media/libopus/celt/arm/celt_pitch_xcorr_arm.s
@@ -0,0 +1,545 @@
+; Copyright (c) 2007-2008 CSIRO
+; Copyright (c) 2007-2009 Xiph.Org Foundation
+; Copyright (c) 2013      Parrot
+; Written by Aurélien Zanelli
+;
+; Redistribution and use in source and binary forms, with or without
+; modification, are permitted provided that the following conditions
+; are met:
+;
+; - Redistributions of source code must retain the above copyright
+; notice, this list of conditions and the following disclaimer.
+;
+; - Redistributions in binary form must reproduce the above copyright
+; notice, this list of conditions and the following disclaimer in the
+; documentation and/or other materials provided with the distribution.
+;
+; THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+; ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+; LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+; A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+; OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+; EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+; PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+; PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+; LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+; NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+; SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+  AREA  |.text|, CODE, READONLY
+
+  GET    celt/arm/armopts.s
+
+IF OPUS_ARM_MAY_HAVE_EDSP
+  EXPORT celt_pitch_xcorr_edsp
+ENDIF
+
+IF OPUS_ARM_MAY_HAVE_NEON
+  EXPORT celt_pitch_xcorr_neon
+ENDIF
+
+IF OPUS_ARM_MAY_HAVE_NEON
+
+; Compute sum[k]=sum(x[j]*y[j+k],j=0...len-1), k=0...3
+xcorr_kernel_neon PROC
+  ; input:
+  ;   r3     = int         len
+  ;   r4     = opus_val16 *x
+  ;   r5     = opus_val16 *y
+  ;   q0     = opus_val32  sum[4]
+  ; output:
+  ;   q0     = opus_val32  sum[4]
+  ; preserved: r0-r3, r6-r11, d2, q4-q7, q9-q15
+  ; internal usage:
+  ;   r12 = int j
+  ;   d3  = y_3|y_2|y_1|y_0
+  ;   q2  = y_B|y_A|y_9|y_8|y_7|y_6|y_5|y_4
+  ;   q3  = x_7|x_6|x_5|x_4|x_3|x_2|x_1|x_0
+  ;   q8  = scratch
+  ;
+  ; Load y[0...3]
+  ; This requires len>0 to always be valid (which we assert in the C code).
+  VLD1.16      {d5}, [r5]!
+  SUBS         r12, r3, #8
+  BLE xcorr_kernel_neon_process4
+; Process 8 samples at a time.
+; This loop loads one y value more than we actually need. Therefore we have to
+; stop as soon as there are 8 or fewer samples left (instead of 7), to avoid
+; reading past the end of the array.
+xcorr_kernel_neon_process8
+  ; This loop has 19 total instructions (10 cycles to issue, minimum), with
+  ; - 2 cycles of ARM insrtuctions,
+  ; - 10 cycles of load/store/byte permute instructions, and
+  ; - 9 cycles of data processing instructions.
+  ; On a Cortex A8, we dual-issue the maximum amount (9 cycles) between the
+  ; latter two categories, meaning the whole loop should run in 10 cycles per
+  ; iteration, barring cache misses.
+  ;
+  ; Load x[0...7]
+  VLD1.16      {d6, d7}, [r4]!
+  ; Unlike VMOV, VAND is a data processsing instruction (and doesn't get
+  ; assembled to VMOV, like VORR would), so it dual-issues with the prior VLD1.
+  VAND         d3, d5, d5
+  SUBS         r12, r12, #8
+  ; Load y[4...11]
+  VLD1.16      {d4, d5}, [r5]!
+  VMLAL.S16    q0, d3, d6[0]
+  VEXT.16      d16, d3, d4, #1
+  VMLAL.S16    q0, d4, d7[0]
+  VEXT.16      d17, d4, d5, #1
+  VMLAL.S16    q0, d16, d6[1]
+  VEXT.16      d16, d3, d4, #2
+  VMLAL.S16    q0, d17, d7[1]
+  VEXT.16      d17, d4, d5, #2
+  VMLAL.S16    q0, d16, d6[2]
+  VEXT.16      d16, d3, d4, #3
+  VMLAL.S16    q0, d17, d7[2]
+  VEXT.16      d17, d4, d5, #3
+  VMLAL.S16    q0, d16, d6[3]
+  VMLAL.S16    q0, d17, d7[3]
+  BGT xcorr_kernel_neon_process8
+; Process 4 samples here if we have > 4 left (still reading one extra y value).
+xcorr_kernel_neon_process4
+  ADDS         r12, r12, #4
+  BLE xcorr_kernel_neon_process2
+  ; Load x[0...3]
+  VLD1.16      d6, [r4]!
+  ; Use VAND since it's a data processing instruction again.
+  VAND         d4, d5, d5
+  SUB          r12, r12, #4
+  ; Load y[4...7]
+  VLD1.16      d5, [r5]!
+  VMLAL.S16    q0, d4, d6[0]
+  VEXT.16      d16, d4, d5, #1
+  VMLAL.S16    q0, d16, d6[1]
+  VEXT.16      d16, d4, d5, #2
+  VMLAL.S16    q0, d16, d6[2]
+  VEXT.16      d16, d4, d5, #3
+  VMLAL.S16    q0, d16, d6[3]
+; Process 2 samples here if we have > 2 left (still reading one extra y value).
+xcorr_kernel_neon_process2
+  ADDS         r12, r12, #2
+  BLE xcorr_kernel_neon_process1
+  ; Load x[0...1]
+  VLD2.16      {d6[],d7[]}, [r4]!
+  ; Use VAND since it's a data processing instruction again.
+  VAND         d4, d5, d5
+  SUB          r12, r12, #2
+  ; Load y[4...5]
+  VLD1.32      {d5[]}, [r5]!
+  VMLAL.S16    q0, d4, d6
+  VEXT.16      d16, d4, d5, #1
+  ; Replace bottom copy of {y5,y4} in d5 with {y3,y2} from d4, using VSRI
+  ; instead of VEXT, since it's a data-processing instruction.
+  VSRI.64      d5, d4, #32
+  VMLAL.S16    q0, d16, d7
+; Process 1 sample using the extra y value we loaded above.
+xcorr_kernel_neon_process1
+  ; Load next *x
+  VLD1.16      {d6[]}, [r4]!
+  ADDS         r12, r12, #1
+  ; y[0...3] are left in d5 from prior iteration(s) (if any)
+  VMLAL.S16    q0, d5, d6
+  MOVLE        pc, lr
+; Now process 1 last sample, not reading ahead.
+  ; Load last *y
+  VLD1.16      {d4[]}, [r5]!
+  VSRI.64      d4, d5, #16
+  ; Load last *x
+  VLD1.16      {d6[]}, [r4]!
+  VMLAL.S16    q0, d4, d6
+  MOV          pc, lr
+  ENDP
+
+; opus_val32 celt_pitch_xcorr_neon(opus_val16 *_x, opus_val16 *_y,
+;  opus_val32 *xcorr, int len, int max_pitch)
+celt_pitch_xcorr_neon PROC
+  ; input:
+  ;   r0  = opus_val16 *_x
+  ;   r1  = opus_val16 *_y
+  ;   r2  = opus_val32 *xcorr
+  ;   r3  = int         len
+  ; output:
+  ;   r0  = int         maxcorr
+  ; internal usage:
+  ;   r4  = opus_val16 *x (for xcorr_kernel_neon())
+  ;   r5  = opus_val16 *y (for xcorr_kernel_neon())
+  ;   r6  = int         max_pitch
+  ;   r12 = int         j
+  ;   q15 = int         maxcorr[4] (q15 is not used by xcorr_kernel_neon())
+  STMFD        sp!, {r4-r6, lr}
+  LDR          r6, [sp, #16]
+  VMOV.S32     q15, #1
+  ; if (max_pitch < 4) goto celt_pitch_xcorr_neon_process4_done
+  SUBS         r6, r6, #4
+  BLT celt_pitch_xcorr_neon_process4_done
+celt_pitch_xcorr_neon_process4
+  ; xcorr_kernel_neon parameters:
+  ; r3 = len, r4 = _x, r5 = _y, q0 = {0, 0, 0, 0}
+  MOV          r4, r0
+  MOV          r5, r1
+  VEOR         q0, q0, q0
+  ; xcorr_kernel_neon only modifies r4, r5, r12, and q0...q3.
+  ; So we don't save/restore any other registers.
+  BL xcorr_kernel_neon
+  SUBS         r6, r6, #4
+  VST1.32      {q0}, [r2]!
+  ; _y += 4
+  ADD          r1, r1, #8
+  VMAX.S32     q15, q15, q0
+  ; if (max_pitch < 4) goto celt_pitch_xcorr_neon_process4_done
+  BGE celt_pitch_xcorr_neon_process4
+; We have less than 4 sums left to compute.
+celt_pitch_xcorr_neon_process4_done
+  ADDS         r6, r6, #4
+  ; Reduce maxcorr to a single value
+  VMAX.S32     d30, d30, d31
+  VPMAX.S32    d30, d30, d30
+  ; if (max_pitch <= 0) goto celt_pitch_xcorr_neon_done
+  BLE celt_pitch_xcorr_neon_done
+; Now compute each remaining sum one at a time.
+celt_pitch_xcorr_neon_process_remaining
+  MOV          r4, r0
+  MOV          r5, r1
+  VMOV.I32     q0, #0
+  SUBS         r12, r3, #8
+  BLT celt_pitch_xcorr_neon_process_remaining4
+; Sum terms 8 at a time.
+celt_pitch_xcorr_neon_process_remaining_loop8
+  ; Load x[0...7]
+  VLD1.16      {q1}, [r4]!
+  ; Load y[0...7]
+  VLD1.16      {q2}, [r5]!
+  SUBS         r12, r12, #8
+  VMLAL.S16    q0, d4, d2
+  VMLAL.S16    q0, d5, d3
+  BGE celt_pitch_xcorr_neon_process_remaining_loop8
+; Sum terms 4 at a time.
+celt_pitch_xcorr_neon_process_remaining4
+  ADDS         r12, r12, #4
+  BLT celt_pitch_xcorr_neon_process_remaining4_done
+  ; Load x[0...3]
+  VLD1.16      {d2}, [r4]!
+  ; Load y[0...3]
+  VLD1.16      {d3}, [r5]!
+  SUB          r12, r12, #4
+  VMLAL.S16    q0, d3, d2
+celt_pitch_xcorr_neon_process_remaining4_done
+  ; Reduce the sum to a single value.
+  VADD.S32     d0, d0, d1
+  VPADDL.S32   d0, d0
+  ADDS         r12, r12, #4
+  BLE celt_pitch_xcorr_neon_process_remaining_loop_done
+; Sum terms 1 at a time.
+celt_pitch_xcorr_neon_process_remaining_loop1
+  VLD1.16      {d2[]}, [r4]!
+  VLD1.16      {d3[]}, [r5]!
+  SUBS         r12, r12, #1
+  VMLAL.S16    q0, d2, d3
+  BGT celt_pitch_xcorr_neon_process_remaining_loop1
+celt_pitch_xcorr_neon_process_remaining_loop_done
+  VST1.32      {d0[0]}, [r2]!
+  VMAX.S32     d30, d30, d0
+  SUBS         r6, r6, #1
+  ; _y++
+  ADD          r1, r1, #2
+  ; if (--max_pitch > 0) goto celt_pitch_xcorr_neon_process_remaining
+  BGT celt_pitch_xcorr_neon_process_remaining
+celt_pitch_xcorr_neon_done
+  VMOV.32      r0, d30[0]
+  LDMFD        sp!, {r4-r6, pc}
+  ENDP
+
+ENDIF
+
+IF OPUS_ARM_MAY_HAVE_EDSP
+
+; This will get used on ARMv7 devices without NEON, so it has been optimized
+; to take advantage of dual-issuing where possible.
+xcorr_kernel_edsp PROC
+  ; input:
+  ;   r3      = int         len
+  ;   r4      = opus_val16 *_x (must be 32-bit aligned)
+  ;   r5      = opus_val16 *_y (must be 32-bit aligned)
+  ;   r6...r9 = opus_val32  sum[4]
+  ; output:
+  ;   r6...r9 = opus_val32  sum[4]
+  ; preserved: r0-r5
+  ; internal usage
+  ;   r2      = int         j
+  ;   r12,r14 = opus_val16  x[4]
+  ;   r10,r11 = opus_val16  y[4]
+  STMFD        sp!, {r2,r4,r5,lr}
+  LDR          r10, [r5], #4      ; Load y[0...1]
+  SUBS         r2, r3, #4         ; j = len-4
+  LDR          r11, [r5], #4      ; Load y[2...3]
+  BLE xcorr_kernel_edsp_process4_done
+  LDR          r12, [r4], #4      ; Load x[0...1]
+  ; Stall
+xcorr_kernel_edsp_process4
+  ; The multiplies must issue from pipeline 0, and can't dual-issue with each
+  ; other. Every other instruction here dual-issues with a multiply, and is
+  ; thus "free". There should be no stalls in the body of the loop.
+  SMLABB       r6, r12, r10, r6   ; sum[0] = MAC16_16(sum[0],x_0,y_0)
+  LDR          r14, [r4], #4      ; Load x[2...3]
+  SMLABT       r7, r12, r10, r7   ; sum[1] = MAC16_16(sum[1],x_0,y_1)
+  SUBS         r2, r2, #4         ; j-=4
+  SMLABB       r8, r12, r11, r8   ; sum[2] = MAC16_16(sum[2],x_0,y_2)
+  SMLABT       r9, r12, r11, r9   ; sum[3] = MAC16_16(sum[3],x_0,y_3)
+  SMLATT       r6, r12, r10, r6   ; sum[0] = MAC16_16(sum[0],x_1,y_1)
+  LDR          r10, [r5], #4      ; Load y[4...5]
+  SMLATB       r7, r12, r11, r7   ; sum[1] = MAC16_16(sum[1],x_1,y_2)
+  SMLATT       r8, r12, r11, r8   ; sum[2] = MAC16_16(sum[2],x_1,y_3)
+  SMLATB       r9, r12, r10, r9   ; sum[3] = MAC16_16(sum[3],x_1,y_4)
+  LDRGT        r12, [r4], #4      ; Load x[0...1]
+  SMLABB       r6, r14, r11, r6   ; sum[0] = MAC16_16(sum[0],x_2,y_2)
+  SMLABT       r7, r14, r11, r7   ; sum[1] = MAC16_16(sum[1],x_2,y_3)
+  SMLABB       r8, r14, r10, r8   ; sum[2] = MAC16_16(sum[2],x_2,y_4)
+  SMLABT       r9, r14, r10, r9   ; sum[3] = MAC16_16(sum[3],x_2,y_5)
+  SMLATT       r6, r14, r11, r6   ; sum[0] = MAC16_16(sum[0],x_3,y_3)
+  LDR          r11, [r5], #4      ; Load y[6...7]
+  SMLATB       r7, r14, r10, r7   ; sum[1] = MAC16_16(sum[1],x_3,y_4)
+  SMLATT       r8, r14, r10, r8   ; sum[2] = MAC16_16(sum[2],x_3,y_5)
+  SMLATB       r9, r14, r11, r9   ; sum[3] = MAC16_16(sum[3],x_3,y_6)
+  BGT xcorr_kernel_edsp_process4
+xcorr_kernel_edsp_process4_done
+  ADDS         r2, r2, #4
+  BLE xcorr_kernel_edsp_done
+  LDRH         r12, [r4], #2      ; r12 = *x++
+  SUBS         r2, r2, #1         ; j--
+  ; Stall
+  SMLABB       r6, r12, r10, r6   ; sum[0] = MAC16_16(sum[0],x,y_0)
+  LDRGTH       r14, [r4], #2      ; r14 = *x++
+  SMLABT       r7, r12, r10, r7   ; sum[1] = MAC16_16(sum[1],x,y_1)
+  SMLABB       r8, r12, r11, r8   ; sum[2] = MAC16_16(sum[2],x,y_2)
+  SMLABT       r9, r12, r11, r9   ; sum[3] = MAC16_16(sum[3],x,y_3)
+  BLE xcorr_kernel_edsp_done
+  SMLABT       r6, r14, r10, r6   ; sum[0] = MAC16_16(sum[0],x,y_1)
+  SUBS         r2, r2, #1         ; j--
+  SMLABB       r7, r14, r11, r7   ; sum[1] = MAC16_16(sum[1],x,y_2)
+  LDRH         r10, [r5], #2      ; r10 = y_4 = *y++
+  SMLABT       r8, r14, r11, r8   ; sum[2] = MAC16_16(sum[2],x,y_3)
+  LDRGTH       r12, [r4], #2      ; r12 = *x++
+  SMLABB       r9, r14, r10, r9   ; sum[3] = MAC16_16(sum[3],x,y_4)
+  BLE xcorr_kernel_edsp_done
+  SMLABB       r6, r12, r11, r6   ; sum[0] = MAC16_16(sum[0],tmp,y_2)
+  CMP          r2, #1             ; j--
+  SMLABT       r7, r12, r11, r7   ; sum[1] = MAC16_16(sum[1],tmp,y_3)
+  LDRH         r2, [r5], #2       ; r2 = y_5 = *y++
+  SMLABB       r8, r12, r10, r8   ; sum[2] = MAC16_16(sum[2],tmp,y_4)
+  LDRGTH       r14, [r4]          ; r14 = *x
+  SMLABB       r9, r12, r2, r9    ; sum[3] = MAC16_16(sum[3],tmp,y_5)
+  BLE xcorr_kernel_edsp_done
+  SMLABT       r6, r14, r11, r6   ; sum[0] = MAC16_16(sum[0],tmp,y_3)
+  LDRH         r11, [r5]          ; r11 = y_6 = *y
+  SMLABB       r7, r14, r10, r7   ; sum[1] = MAC16_16(sum[1],tmp,y_4)
+  SMLABB       r8, r14, r2, r8    ; sum[2] = MAC16_16(sum[2],tmp,y_5)
+  SMLABB       r9, r14, r11, r9   ; sum[3] = MAC16_16(sum[3],tmp,y_6)
+xcorr_kernel_edsp_done
+  LDMFD        sp!, {r2,r4,r5,pc}
+  ENDP
+
+celt_pitch_xcorr_edsp PROC
+  ; input:
+  ;   r0  = opus_val16 *_x (must be 32-bit aligned)
+  ;   r1  = opus_val16 *_y (only needs to be 16-bit aligned)
+  ;   r2  = opus_val32 *xcorr
+  ;   r3  = int         len
+  ; output:
+  ;   r0  = maxcorr
+  ; internal usage
+  ;   r4  = opus_val16 *x
+  ;   r5  = opus_val16 *y
+  ;   r6  = opus_val32  sum0
+  ;   r7  = opus_val32  sum1
+  ;   r8  = opus_val32  sum2
+  ;   r9  = opus_val32  sum3
+  ;   r1  = int         max_pitch
+  ;   r12 = int         j
+  STMFD        sp!, {r4-r11, lr}
+  MOV          r5, r1
+  LDR          r1, [sp, #36]
+  MOV          r4, r0
+  TST          r5, #3
+  ; maxcorr = 1
+  MOV          r0, #1
+  BEQ          celt_pitch_xcorr_edsp_process1u_done
+; Compute one sum at the start to make y 32-bit aligned.
+  SUBS         r12, r3, #4
+  ; r14 = sum = 0
+  MOV          r14, #0
+  LDRH         r8, [r5], #2
+  BLE celt_pitch_xcorr_edsp_process1u_loop4_done
+  LDR          r6, [r4], #4
+  MOV          r8, r8, LSL #16
+celt_pitch_xcorr_edsp_process1u_loop4
+  LDR          r9, [r5], #4
+  SMLABT       r14, r6, r8, r14     ; sum = MAC16_16(sum, x_0, y_0)
+  LDR          r7, [r4], #4
+  SMLATB       r14, r6, r9, r14     ; sum = MAC16_16(sum, x_1, y_1)
+  LDR          r8, [r5], #4
+  SMLABT       r14, r7, r9, r14     ; sum = MAC16_16(sum, x_2, y_2)
+  SUBS         r12, r12, #4         ; j-=4
+  SMLATB       r14, r7, r8, r14     ; sum = MAC16_16(sum, x_3, y_3)
+  LDRGT        r6, [r4], #4
+  BGT celt_pitch_xcorr_edsp_process1u_loop4
+  MOV          r8, r8, LSR #16
+celt_pitch_xcorr_edsp_process1u_loop4_done
+  ADDS         r12, r12, #4
+celt_pitch_xcorr_edsp_process1u_loop1
+  LDRGEH       r6, [r4], #2
+  ; Stall
+  SMLABBGE     r14, r6, r8, r14    ; sum = MAC16_16(sum, *x, *y)
+  SUBGES       r12, r12, #1
+  LDRGTH       r8, [r5], #2
+  BGT celt_pitch_xcorr_edsp_process1u_loop1
+  ; Restore _x
+  SUB          r4, r4, r3, LSL #1
+  ; Restore and advance _y
+  SUB          r5, r5, r3, LSL #1
+  ; maxcorr = max(maxcorr, sum)
+  CMP          r0, r14
+  ADD          r5, r5, #2
+  MOVLT        r0, r14
+  SUBS         r1, r1, #1
+  ; xcorr[i] = sum
+  STR          r14, [r2], #4
+  BLE celt_pitch_xcorr_edsp_done
+celt_pitch_xcorr_edsp_process1u_done
+  ; if (max_pitch < 4) goto celt_pitch_xcorr_edsp_process2
+  SUBS         r1, r1, #4
+  BLT celt_pitch_xcorr_edsp_process2
+celt_pitch_xcorr_edsp_process4
+  ; xcorr_kernel_edsp parameters:
+  ; r3 = len, r4 = _x, r5 = _y, r6...r9 = sum[4] = {0, 0, 0, 0}
+  MOV          r6, #0
+  MOV          r7, #0
+  MOV          r8, #0
+  MOV          r9, #0
+  BL xcorr_kernel_edsp  ; xcorr_kernel_edsp(_x, _y+i, xcorr+i, len)
+  ; maxcorr = max(maxcorr, sum0, sum1, sum2, sum3)
+  CMP          r0, r6
+  ; _y+=4
+  ADD          r5, r5, #8
+  MOVLT        r0, r6
+  CMP          r0, r7
+  MOVLT        r0, r7
+  CMP          r0, r8
+  MOVLT        r0, r8
+  CMP          r0, r9
+  MOVLT        r0, r9
+  STMIA        r2!, {r6-r9}
+  SUBS         r1, r1, #4
+  BGE celt_pitch_xcorr_edsp_process4
+celt_pitch_xcorr_edsp_process2
+  ADDS         r1, r1, #2
+  BLT celt_pitch_xcorr_edsp_process1a
+  SUBS         r12, r3, #4
+  ; {r10, r11} = {sum0, sum1} = {0, 0}
+  MOV          r10, #0
+  MOV          r11, #0
+  LDR          r8, [r5], #4
+  BLE celt_pitch_xcorr_edsp_process2_loop_done
+  LDR          r6, [r4], #4
+  LDR          r9, [r5], #4
+celt_pitch_xcorr_edsp_process2_loop4
+  SMLABB       r10, r6, r8, r10     ; sum0 = MAC16_16(sum0, x_0, y_0)
+  LDR          r7, [r4], #4
+  SMLABT       r11, r6, r8, r11     ; sum1 = MAC16_16(sum1, x_0, y_1)
+  SUBS         r12, r12, #4         ; j-=4
+  SMLATT       r10, r6, r8, r10     ; sum0 = MAC16_16(sum0, x_1, y_1)
+  LDR          r8, [r5], #4
+  SMLATB       r11, r6, r9, r11     ; sum1 = MAC16_16(sum1, x_1, y_2)
+  LDRGT        r6, [r4], #4
+  SMLABB       r10, r7, r9, r10     ; sum0 = MAC16_16(sum0, x_2, y_2)
+  SMLABT       r11, r7, r9, r11     ; sum1 = MAC16_16(sum1, x_2, y_3)
+  SMLATT       r10, r7, r9, r10     ; sum0 = MAC16_16(sum0, x_3, y_3)
+  LDRGT        r9, [r5], #4
+  SMLATB       r11, r7, r8, r11     ; sum1 = MAC16_16(sum1, x_3, y_4)
+  BGT celt_pitch_xcorr_edsp_process2_loop4
+celt_pitch_xcorr_edsp_process2_loop_done
+  ADDS         r12, r12, #2
+  BLE  celt_pitch_xcorr_edsp_process2_1
+  LDR          r6, [r4], #4
+  ; Stall
+  SMLABB       r10, r6, r8, r10     ; sum0 = MAC16_16(sum0, x_0, y_0)
+  LDR          r9, [r5], #4
+  SMLABT       r11, r6, r8, r11     ; sum1 = MAC16_16(sum1, x_0, y_1)
+  SUB          r12, r12, #2
+  SMLATT       r10, r6, r8, r10     ; sum0 = MAC16_16(sum0, x_1, y_1)
+  MOV          r8, r9
+  SMLATB       r11, r6, r9, r11     ; sum1 = MAC16_16(sum1, x_1, y_2)
+celt_pitch_xcorr_edsp_process2_1
+  LDRH         r6, [r4], #2
+  ADDS         r12, r12, #1
+  ; Stall
+  SMLABB       r10, r6, r8, r10     ; sum0 = MAC16_16(sum0, x_0, y_0)
+  LDRGTH       r7, [r4], #2
+  SMLABT       r11, r6, r8, r11     ; sum1 = MAC16_16(sum1, x_0, y_1)
+  BLE celt_pitch_xcorr_edsp_process2_done
+  LDRH         r9, [r5], #2
+  SMLABT       r10, r7, r8, r10     ; sum0 = MAC16_16(sum0, x_0, y_1)
+  SMLABB       r11, r7, r9, r11     ; sum1 = MAC16_16(sum1, x_0, y_2)
+celt_pitch_xcorr_edsp_process2_done
+  ; Restore _x
+  SUB          r4, r4, r3, LSL #1
+  ; Restore and advance _y
+  SUB          r5, r5, r3, LSL #1
+  ; maxcorr = max(maxcorr, sum0)
+  CMP          r0, r10
+  ADD          r5, r5, #2
+  MOVLT        r0, r10
+  SUB          r1, r1, #2
+  ; maxcorr = max(maxcorr, sum1)
+  CMP          r0, r11
+  ; xcorr[i] = sum
+  STR          r10, [r2], #4
+  MOVLT        r0, r11
+  STR          r11, [r2], #4
+celt_pitch_xcorr_edsp_process1a
+  ADDS         r1, r1, #1
+  BLT celt_pitch_xcorr_edsp_done
+  SUBS         r12, r3, #4
+  ; r14 = sum = 0
+  MOV          r14, #0
+  BLT celt_pitch_xcorr_edsp_process1a_loop_done
+  LDR          r6, [r4], #4
+  LDR          r8, [r5], #4
+  LDR          r7, [r4], #4
+  LDR          r9, [r5], #4
+celt_pitch_xcorr_edsp_process1a_loop4
+  SMLABB       r14, r6, r8, r14     ; sum = MAC16_16(sum, x_0, y_0)
+  SUBS         r12, r12, #4         ; j-=4
+  SMLATT       r14, r6, r8, r14     ; sum = MAC16_16(sum, x_1, y_1)
+  LDRGE        r6, [r4], #4
+  SMLABB       r14, r7, r9, r14     ; sum = MAC16_16(sum, x_2, y_2)
+  LDRGE        r8, [r5], #4
+  SMLATT       r14, r7, r9, r14     ; sum = MAC16_16(sum, x_3, y_3)
+  LDRGE        r7, [r4], #4
+  LDRGE        r9, [r5], #4
+  BGE celt_pitch_xcorr_edsp_process1a_loop4
+celt_pitch_xcorr_edsp_process1a_loop_done
+  ADDS         r12, r12, #2
+  LDRGE        r6, [r4], #4
+  LDRGE        r8, [r5], #4
+  ; Stall
+  SMLABBGE     r14, r6, r8, r14     ; sum = MAC16_16(sum, x_0, y_0)
+  SUBGE        r12, r12, #2
+  SMLATTGE     r14, r6, r8, r14     ; sum = MAC16_16(sum, x_1, y_1)
+  ADDS         r12, r12, #1
+  LDRGEH       r6, [r4], #2
+  LDRGEH       r8, [r5], #2
+  ; Stall
+  SMLABBGE     r14, r6, r8, r14     ; sum = MAC16_16(sum, *x, *y)
+  ; maxcorr = max(maxcorr, sum)
+  CMP          r0, r14
+  ; xcorr[i] = sum
+  STR          r14, [r2], #4
+  MOVLT        r0, r14
+celt_pitch_xcorr_edsp_done
+  LDMFD        sp!, {r4-r11, pc}
+  ENDP
+
+ENDIF
+
+END

--- a/media/libopus/celt/arm/fixed_armv4.h
+++ b/media/libopus/celt/arm/fixed_armv4.h
@@ -24,34 +24,34 @@
    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */
 
 #ifndef FIXED_ARMv4_H
 #define FIXED_ARMv4_H
 
 /** 16x32 multiplication, followed by a 16-bit shift right. Results fits in 32 bits */
 #undef MULT16_32_Q16
-static inline opus_val32 MULT16_32_Q16_armv4(opus_val16 a, opus_val32 b)
+static OPUS_INLINE opus_val32 MULT16_32_Q16_armv4(opus_val16 a, opus_val32 b)
 {
   unsigned rd_lo;
   int rd_hi;
   __asm__(
       "#MULT16_32_Q16\n\t"
       "smull %0, %1, %2, %3\n\t"
       : "=&r"(rd_lo), "=&r"(rd_hi)
       : "%r"(b),"r"(a<<16)
   );
   return rd_hi;
 }
 #define MULT16_32_Q16(a, b) (MULT16_32_Q16_armv4(a, b))
 
 
 /** 16x32 multiplication, followed by a 15-bit shift right. Results fits in 32 bits */
 #undef MULT16_32_Q15
-static inline opus_val32 MULT16_32_Q15_armv4(opus_val16 a, opus_val32 b)
+static OPUS_INLINE opus_val32 MULT16_32_Q15_armv4(opus_val16 a, opus_val32 b)
 {
   unsigned rd_lo;
   int rd_hi;
   __asm__(
       "#MULT16_32_Q15\n\t"
       "smull %0, %1, %2, %3\n\t"
       : "=&r"(rd_lo), "=&r"(rd_hi)
       : "%r"(b), "r"(a<<16)

--- a/media/libopus/celt/arm/fixed_armv5e.h
+++ b/media/libopus/celt/arm/fixed_armv5e.h
@@ -29,33 +29,33 @@
 
 #ifndef FIXED_ARMv5E_H
 #define FIXED_ARMv5E_H
 
 #include "fixed_armv4.h"
 
 /** 16x32 multiplication, followed by a 16-bit shift right. Results fits in 32 bits */
 #undef MULT16_32_Q16
-static inline opus_val32 MULT16_32_Q16_armv5e(opus_val16 a, opus_val32 b)
+static OPUS_INLINE opus_val32 MULT16_32_Q16_armv5e(opus_val16 a, opus_val32 b)
 {
   int res;
   __asm__(
       "#MULT16_32_Q16\n\t"
       "smulwb %0, %1, %2\n\t"
       : "=r"(res)
       : "r"(b),"r"(a)
   );
   return res;
 }
 #define MULT16_32_Q16(a, b) (MULT16_32_Q16_armv5e(a, b))
 
 
 /** 16x32 multiplication, followed by a 15-bit shift right. Results fits in 32 bits */
 #undef MULT16_32_Q15
-static inline opus_val32 MULT16_32_Q15_armv5e(opus_val16 a, opus_val32 b)
+static OPUS_INLINE opus_val32 MULT16_32_Q15_armv5e(opus_val16 a, opus_val32 b)
 {
   int res;
   __asm__(
       "#MULT16_32_Q15\n\t"
       "smulwb %0, %1, %2\n\t"
       : "=r"(res)
       : "r"(b), "r"(a)
   );
@@ -63,49 +63,49 @@ static inline opus_val32 MULT16_32_Q15_a
 }
 #define MULT16_32_Q15(a, b) (MULT16_32_Q15_armv5e(a, b))
 
 
 /** 16x32 multiply, followed by a 15-bit shift right and 32-bit add.
     b must fit in 31 bits.
     Result fits in 32 bits. */
 #undef MAC16_32_Q15
-static inline opus_val32 MAC16_32_Q15_armv5e(opus_val32 c, opus_val16 a,
+static OPUS_INLINE opus_val32 MAC16_32_Q15_armv5e(opus_val32 c, opus_val16 a,
  opus_val32 b)
 {
   int res;
   __asm__(
       "#MAC16_32_Q15\n\t"
       "smlawb %0, %1, %2, %3;\n"
       : "=r"(res)
       : "r"(b<<1), "r"(a), "r"(c)
   );
   return res;
 }
 #define MAC16_32_Q15(c, a, b) (MAC16_32_Q15_armv5e(c, a, b))
 
 /** 16x16 multiply-add where the result fits in 32 bits */
 #undef MAC16_16
-static inline opus_val32 MAC16_16_armv5e(opus_val32 c, opus_val16 a,
+static OPUS_INLINE opus_val32 MAC16_16_armv5e(opus_val32 c, opus_val16 a,
  opus_val16 b)
 {
   int res;
   __asm__(
       "#MAC16_16\n\t"
       "smlabb %0, %1, %2, %3;\n"
       : "=r"(res)
       : "r"(a), "r"(b), "r"(c)
   );
   return res;
 }
 #define MAC16_16(c, a, b) (MAC16_16_armv5e(c, a, b))
 
 /** 16x16 multiplication where the result fits in 32 bits */
 #undef MULT16_16
-static inline opus_val32 MULT16_16_armv5e(opus_val16 a, opus_val16 b)
+static OPUS_INLINE opus_val32 MULT16_16_armv5e(opus_val16 a, opus_val16 b)
 {
   int res;
   __asm__(
       "#MULT16_16\n\t"
       "smulbb %0, %1, %2;\n"
       : "=r"(res)
       : "r"(a), "r"(b)
   );

new file mode 100644
--- /dev/null
+++ b/media/libopus/celt/arm/pitch_arm.h
@@ -0,0 +1,57 @@
+/* Copyright (c) 2010 Xiph.Org Foundation
+ * Copyright (c) 2013 Parrot */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#if !defined(PITCH_ARM_H)
+# define PITCH_ARM_H
+
+# include "armcpu.h"
+
+# if defined(FIXED_POINT)
+
+#  if defined(OPUS_ARM_MAY_HAVE_NEON)
+opus_val32 celt_pitch_xcorr_neon(const opus_val16 *_x, const opus_val16 *_y,
+    opus_val32 *xcorr, int len, int max_pitch);
+#  endif
+
+#  if defined(OPUS_ARM_MAY_HAVE_MEDIA)
+#   define celt_pitch_xcorr_media MAY_HAVE_EDSP(celt_pitch_xcorr)
+#  endif
+
+#  if defined(OPUS_ARM_MAY_HAVE_EDSP)
+opus_val32 celt_pitch_xcorr_edsp(const opus_val16 *_x, const opus_val16 *_y,
+    opus_val32 *xcorr, int len, int max_pitch);
+#  endif
+
+#  if !defined(OPUS_HAVE_RTCD)
+#   define OVERRIDE_PITCH_XCORR (1)
+#   define celt_pitch_xcorr(_x, _y, xcorr, len, max_pitch, arch) \
+  ((void)(arch),PRESUME_NEON(celt_pitch_xcorr)(_x, _y, xcorr, len, max_pitch))
+#  endif
+
+# endif
+
+#endif

--- a/media/libopus/celt/bands.c
+++ b/media/libopus/celt/bands.c
@@ -209,30 +209,46 @@ void denormalise_bands(const CELTMode *m
          opus_val16 g;
          opus_val16 lg;
 #ifdef FIXED_POINT
          int shift;
 #endif
          j=M*eBands[i];
          band_end = M*eBands[i+1];
          lg = ADD16(bandLogE[i+c*m->nbEBands], SHL16((opus_val16)eMeans[i],6));
-#ifdef FIXED_POINT
+#ifndef FIXED_POINT
+         g = celt_exp2(lg);
+#else
          /* Handle the integer part of the log energy */
          shift = 16-(lg>>DB_SHIFT);
          if (shift>31)
          {
             shift=0;
             g=0;
          } else {
             /* Handle the fractional part. */
             g = celt_exp2_frac(lg&((1<<DB_SHIFT)-1));
          }
-#else
-         g = celt_exp2(lg);
+         /* Handle extreme gains with negative shift. */
+         if (shift<0)
+         {
+            /* For shift < -2 we'd be likely to overflow, so we're capping
+               the gain here. This shouldn't happen unless the bitstream is
+               already corrupted. */
+            if (shift < -2)
+            {
+               g = 32767;
+               shift = -2;
+            }
+            do {
+               *f++ = SHL32(MULT16_16(*x++, g), -shift);
+            } while (++j<band_end);
+         } else
 #endif
+         /* Be careful of the fixed-point "else" just above when changing this code */
          do {
             *f++ = SHR32(MULT16_16(*x++, g), shift);
          } while (++j<band_end);
       }
       celt_assert(start <= end);
       for (i=M*eBands[end];i<N;i++)
          *f++ = 0;
    } while (++c<C);
@@ -487,17 +503,17 @@ int spreading_decision(const CELTMode *m
       if (hf_sum > 22)
          *tapset_decision=2;
       else if (hf_sum > 18)
          *tapset_decision=1;
       else
          *tapset_decision=0;
    }
    /*printf("%d %d %d\n", hf_sum, *hf_average, *tapset_decision);*/
-   celt_assert(nbBands>0); /*M*(eBands[end]-eBands[end-1]) <= 8 assures this*/
+   celt_assert(nbBands>0); /* end has to be non-zero */
    sum /= nbBands;
    /* Recursive averaging */
    sum = (sum+*average)>>1;
    *average = sum;
    /* Hysteresis */
    sum = (3*sum + (((3-last_decision)<<7) + 64) + 2)>>2;
    if (sum < 80)
    {
@@ -864,17 +880,16 @@ static unsigned quant_partition(struct b
       int N, int b, int B, celt_norm *lowband,
       int LM,
       opus_val16 gain, int fill)
 {
    const unsigned char *cache;
    int q;
    int curr_bits;
    int imid=0, iside=0;
-   int N_B=N;
    int B0=B;
    opus_val16 mid=0, side=0;
    unsigned cm=0;
 #ifdef RESYNTH
    int resynth = 1;
 #else
    int resynth = !ctx->encode;
 #endif
@@ -886,18 +901,16 @@ static unsigned quant_partition(struct b
    ec_ctx *ec;
 
    encode = ctx->encode;
    m = ctx->m;
    i = ctx->i;
    spread = ctx->spread;
    ec = ctx->ec;
 
-   N_B /= B;
-
    /* If we need 1.5 more bit than we can produce, split the band in two. */
    cache = m->cache.bits + m->cache.index[(LM+1)*m->nbEBands+i];
    if (LM != -1 && b > cache[cache[0]]+12 && N>2)
    {
       int mbits, sbits, delta;
       int itheta;
       int qalloc;
       struct split_ctx sctx;
@@ -1067,17 +1080,16 @@ static unsigned quant_band(struct band_c
    int tf_change;
 
    encode = ctx->encode;
    tf_change = ctx->tf_change;
 
    longBlocks = B0==1;
 
    N_B /= B;
-   N_B0 = N_B;
 
    /* Special case for one sample */
    if (N==1)
    {
       return quant_band_n1(ctx, X, NULL, b, lowband_out);
    }
 
    if (tf_change>0)

--- a/media/libopus/celt/celt.h
+++ b/media/libopus/celt/celt.h
@@ -47,30 +47,34 @@ extern "C" {
 #endif
 
 #define CELTEncoder OpusCustomEncoder
 #define CELTDecoder OpusCustomDecoder
 #define CELTMode OpusCustomMode
 
 typedef struct {
    int valid;
-   opus_val16 tonality;
-   opus_val16 tonality_slope;
-   opus_val16 noisiness;
-   opus_val16 activity;
-   opus_val16 music_prob;
+   float tonality;
+   float tonality_slope;
+   float noisiness;
+   float activity;
+   float music_prob;
    int        bandwidth;
 }AnalysisInfo;
 
 #define __celt_check_mode_ptr_ptr(ptr) ((ptr) + ((ptr) - (const CELTMode**)(ptr)))
 
 #define __celt_check_analysis_ptr(ptr) ((ptr) + ((ptr) - (const AnalysisInfo*)(ptr)))
 
 /* Encoder/decoder Requests */
 
+/* Expose this option again when variable framesize actually works */
+#define OPUS_FRAMESIZE_VARIABLE              5010 /**< Optimize the frame size dynamically */
+
+
 #define CELT_SET_PREDICTION_REQUEST    10002
 /** Controls the use of interframe prediction.
     0=Independent frames
     1=Short term interframe prediction allowed
     2=Long term prediction allowed
  */
 #define CELT_SET_PREDICTION(x) CELT_SET_PREDICTION_REQUEST, __opus_check_int(x)
 
@@ -104,29 +108,27 @@ typedef struct {
 #define CELT_SET_TONALITY_SLOPE(x) CELT_SET_TONALITY_SLOPE_REQUEST, __opus_check_int(x)
 
 #define CELT_SET_ANALYSIS_REQUEST    10022
 #define CELT_SET_ANALYSIS(x) CELT_SET_ANALYSIS_REQUEST, __celt_check_analysis_ptr(x)
 
 #define OPUS_SET_LFE_REQUEST    10024
 #define OPUS_SET_LFE(x) OPUS_SET_LFE_REQUEST, __opus_check_int(x)
 
-#define OPUS_SET_ENERGY_SAVE_REQUEST    10026
-#define OPUS_SET_ENERGY_SAVE(x) OPUS_SET_ENERGY_SAVE_REQUEST, __opus_check_val16_ptr(x)
-
-#define OPUS_SET_ENERGY_MASK_REQUEST    10028
+#define OPUS_SET_ENERGY_MASK_REQUEST    10026
 #define OPUS_SET_ENERGY_MASK(x) OPUS_SET_ENERGY_MASK_REQUEST, __opus_check_val16_ptr(x)
 
 /* Encoder stuff */
 
 int celt_encoder_get_size(int channels);
 
 int celt_encode_with_ec(OpusCustomEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, int frame_size, unsigned char *compressed, int nbCompressedBytes, ec_enc *enc);
 
-int celt_encoder_init(CELTEncoder *st, opus_int32 sampling_rate, int channels);
+int celt_encoder_init(CELTEncoder *st, opus_int32 sampling_rate, int channels,
+                      int arch);
 
 
 
 /* Decoder stuff */
 
 int celt_decoder_get_size(int channels);
 
 
@@ -136,17 +138,17 @@ int celt_decode_with_ec(OpusCustomDecode
 
 #define celt_encoder_ctl opus_custom_encoder_ctl
 #define celt_decoder_ctl opus_custom_decoder_ctl
 
 
 #ifdef CUSTOM_MODES
 #define OPUS_CUSTOM_NOSTATIC
 #else
-#define OPUS_CUSTOM_NOSTATIC static inline
+#define OPUS_CUSTOM_NOSTATIC static OPUS_INLINE
 #endif
 
 static const unsigned char trim_icdf[11] = {126, 124, 119, 109, 87, 41, 19, 9, 4, 2, 0};
 /* Probs: NONE: 21.875%, LIGHT: 6.25%, NORMAL: 65.625%, AGGRESSIVE: 6.25% */
 static const unsigned char spread_icdf[4] = {25, 23, 2, 0};
 
 static const unsigned char tapset_icdf[3]={2,1,0};
 
@@ -161,43 +163,46 @@ static const unsigned char toOpusTable[2
 
 static const unsigned char fromOpusTable[16] = {
       0x80, 0x88, 0x90, 0x98,
       0x40, 0x48, 0x50, 0x58,
       0x20, 0x28, 0x30, 0x38,
       0x00, 0x08, 0x10, 0x18
 };
 
-static inline int toOpus(unsigned char c)
+static OPUS_INLINE int toOpus(unsigned char c)
 {
    int ret=0;
    if (c<0xA0)
       ret = toOpusTable[c>>3];
    if (ret == 0)
       return -1;
    else
       return ret|(c&0x7);
 }
 
-static inline int fromOpus(unsigned char c)
+static OPUS_INLINE int fromOpus(unsigned char c)
 {
    if (c<0x80)
       return -1;
    else
       return fromOpusTable[(c>>3)-16] | (c&0x7);
 }
 #endif /* CUSTOM_MODES */
 
 #define COMBFILTER_MAXPERIOD 1024
 #define COMBFILTER_MINPERIOD 15
 
 extern const signed char tf_select_table[4][8];
 
 int resampling_factor(opus_int32 rate);
 
+void celt_preemphasis(const opus_val16 * OPUS_RESTRICT pcmp, celt_sig * OPUS_RESTRICT inp,
+                        int N, int CC, int upsample, const opus_val16 *coef, celt_sig *mem, int clip);
+
 void comb_filter(opus_val32 *y, opus_val32 *x, int T0, int T1, int N,
       opus_val16 g0, opus_val16 g1, int tapset0, int tapset1,
       const opus_val16 *window, int overlap);
 
 void init_caps(const CELTMode *m,int *cap,int LM,int C);
 
 #ifdef RESYNTH
 void deemphasis(celt_sig *in[], opus_val16 *pcm, int N, int C, int downsample, const opus_val16 *coef, celt_sig *mem, celt_sig * OPUS_RESTRICT scratch);

--- a/media/libopus/celt/celt_decoder.c
+++ b/media/libopus/celt/celt_decoder.c
@@ -170,17 +170,17 @@ OPUS_CUSTOM_NOSTATIC int opus_custom_dec
 
 #ifdef CUSTOM_MODES
 void opus_custom_decoder_destroy(CELTDecoder *st)
 {
    opus_free(st);
 }
 #endif /* CUSTOM_MODES */
 
-static inline opus_val16 SIG2WORD16(celt_sig x)
+static OPUS_INLINE opus_val16 SIG2WORD16(celt_sig x)
 {
 #ifdef FIXED_POINT
    x = PSHR32(x, SIG_SHIFT);
    x = MAX32(x, -32768);
    x = MIN32(x, 32767);
    return EXTRACT16(x);
 #else
    return (opus_val16)x;
@@ -442,20 +442,21 @@ static void celt_decode_lost(CELTDecoder
       int pitch_index;
       VARDECL(opus_val32, etmp);
       VARDECL(opus_val16, exc);
 
       if (loss_count == 0)
       {
          VARDECL( opus_val16, lp_pitch_buf );
          ALLOC( lp_pitch_buf, DECODE_BUFFER_SIZE>>1, opus_val16 );
-         pitch_downsample(decode_mem, lp_pitch_buf, DECODE_BUFFER_SIZE, C);
+         pitch_downsample(decode_mem, lp_pitch_buf,
+               DECODE_BUFFER_SIZE, C, st->arch);
          pitch_search(lp_pitch_buf+(PLC_PITCH_LAG_MAX>>1), lp_pitch_buf,
                DECODE_BUFFER_SIZE-PLC_PITCH_LAG_MAX,
-               PLC_PITCH_LAG_MAX-PLC_PITCH_LAG_MIN, &pitch_index);
+               PLC_PITCH_LAG_MAX-PLC_PITCH_LAG_MIN, &pitch_index, st->arch);
          pitch_index = PLC_PITCH_LAG_MAX-pitch_index;
          st->last_pitch_index = pitch_index;
       } else {
          pitch_index = st->last_pitch_index;
          fade = QCONST16(.8f,15);
       }
 
       ALLOC(etmp, overlap, opus_val32);
@@ -476,17 +477,18 @@ static void celt_decode_lost(CELTDecoder
             exc[i] = ROUND16(buf[DECODE_BUFFER_SIZE-MAX_PERIOD+i], SIG_SHIFT);
          }
 
          if (loss_count == 0)
          {
             opus_val32 ac[LPC_ORDER+1];
             /* Compute LPC coefficients for the last MAX_PERIOD samples before
                the first loss so we can work in the excitation-filter domain. */
-            _celt_autocorr(exc, ac, window, overlap, LPC_ORDER, MAX_PERIOD);
+            _celt_autocorr(exc, ac, window, overlap,
+                   LPC_ORDER, MAX_PERIOD, st->arch);
             /* Add a noise floor of -40 dB. */
 #ifdef FIXED_POINT
             ac[0] += SHR32(ac[0],13);
 #else
             ac[0] *= 1.0001f;
 #endif
             /* Use lag windowing to stabilize the Levinson-Durbin recursion. */
             for (i=1;i<=LPC_ORDER;i++)
@@ -660,17 +662,16 @@ int celt_decode_with_ec(CELTDecoder * OP
    VARDECL(celt_norm, X);
    VARDECL(int, fine_quant);
    VARDECL(int, pulses);
    VARDECL(int, cap);
    VARDECL(int, offsets);
    VARDECL(int, fine_priority);
    VARDECL(int, tf_res);
    VARDECL(unsigned char, collapse_masks);
-   celt_sig *out_mem[2];
    celt_sig *decode_mem[2];
    celt_sig *out_syn[2];
    opus_val16 *lpc;
    opus_val16 *oldBandE, *oldLogE, *oldLogE2, *backgroundLogE;
 
    int shortBlocks;
    int isTransient;
    int intra_ener;
@@ -701,17 +702,16 @@ int celt_decode_with_ec(CELTDecoder * OP
    mode = st->mode;
    nbEBands = mode->nbEBands;
    overlap = mode->overlap;
    eBands = mode->eBands;
    frame_size *= st->downsample;
 
    c=0; do {
       decode_mem[c] = st->_decode_mem + c*(DECODE_BUFFER_SIZE+overlap);
-      out_mem[c] = decode_mem[c]+DECODE_BUFFER_SIZE-MAX_PERIOD;
    } while (++c<CC);
    lpc = (opus_val16*)(st->_decode_mem+(DECODE_BUFFER_SIZE+overlap)*CC);
    oldBandE = lpc+CC*LPC_ORDER;
    oldLogE = oldBandE + 2*nbEBands;
    oldLogE2 = oldLogE + 2*nbEBands;
    backgroundLogE = oldLogE2  + 2*nbEBands;
 
 #ifdef CUSTOM_MODES
@@ -931,17 +931,17 @@ int celt_decode_with_ec(CELTDecoder * OP
       int bound = M*eBands[effEnd];
       if (st->downsample!=1)
          bound = IMIN(bound, N/st->downsample);
       for (i=bound;i<N;i++)
          freq[c*N+i] = 0;
    } while (++c<C);
 
    c=0; do {
-      out_syn[c] = out_mem[c]+MAX_PERIOD-N;
+      out_syn[c] = decode_mem[c]+DECODE_BUFFER_SIZE-N;
    } while (++c<CC);
 
    if (CC==2&&C==1)
    {
       for (i=0;i<N;i++)
          freq[N+i] = freq[i];
    }
    if (CC==1&&C==2)

--- a/media/libopus/celt/celt_encoder.c
+++ b/media/libopus/celt/celt_encoder.c
@@ -106,17 +106,16 @@ struct OpusCustomEncoder {
    /* VBR-related parameters */
    opus_int32 vbr_reservoir;
    opus_int32 vbr_drift;
    opus_int32 vbr_offset;
    opus_int32 vbr_count;
    opus_val32 overlap_max;
    opus_val16 stereo_saving;
    int intensity;
-   opus_val16 *energy_save;
    opus_val16 *energy_mask;
    opus_val16 spec_avg;
 
 #ifdef RESYNTH
    /* +MAX_PERIOD/2 to make space for overlap */
    celt_sig syn_mem[2][2*MAX_PERIOD+MAX_PERIOD/2];
 #endif
 
@@ -157,27 +156,18 @@ CELTEncoder *opus_custom_encoder_create(
       st = NULL;
    }
    if (error)
       *error = ret;
    return st;
 }
 #endif /* CUSTOM_MODES */
 
-int celt_encoder_init(CELTEncoder *st, opus_int32 sampling_rate, int channels)
-{
-   int ret;
-   ret = opus_custom_encoder_init(st, opus_custom_mode_create(48000, 960, NULL), channels);
-   if (ret != OPUS_OK)
-      return ret;
-   st->upsample = resampling_factor(sampling_rate);
-   return OPUS_OK;
-}
-
-OPUS_CUSTOM_NOSTATIC int opus_custom_encoder_init(CELTEncoder *st, const CELTMode *mode, int channels)
+static int opus_custom_encoder_init_arch(CELTEncoder *st, const CELTMode *mode,
+                                         int channels, int arch)
 {
    if (channels < 0 || channels > 2)
       return OPUS_BAD_ARG;
 
    if (st==NULL || mode==NULL)
       return OPUS_ALLOC_FAIL;
 
    OPUS_CLEAR((char*)st, opus_custom_encoder_get_size(mode, channels));
@@ -186,33 +176,52 @@ OPUS_CUSTOM_NOSTATIC int opus_custom_enc
    st->overlap = mode->overlap;
    st->stream_channels = st->channels = channels;
 
    st->upsample = 1;
    st->start = 0;
    st->end = st->mode->effEBands;
    st->signalling = 1;
 
-   st->arch = opus_select_arch();
+   st->arch = arch;
 
    st->constrained_vbr = 1;
    st->clip = 1;
 
    st->bitrate = OPUS_BITRATE_MAX;
    st->vbr = 0;
    st->force_intra  = 0;
    st->complexity = 5;
    st->lsb_depth=24;
 
    opus_custom_encoder_ctl(st, OPUS_RESET_STATE);
 
    return OPUS_OK;
 }
 
 #ifdef CUSTOM_MODES
+int opus_custom_encoder_init(CELTEncoder *st, const CELTMode *mode, int channels)
+{
+   return opus_custom_encoder_init_arch(st, mode, channels, opus_select_arch());
+}
+#endif
+
+int celt_encoder_init(CELTEncoder *st, opus_int32 sampling_rate, int channels,
+                      int arch)
+{
+   int ret;
+   ret = opus_custom_encoder_init_arch(st,
+           opus_custom_mode_create(48000, 960, NULL), channels, arch);
+   if (ret != OPUS_OK)
+      return ret;
+   st->upsample = resampling_factor(sampling_rate);
+   return OPUS_OK;
+}
+
+#ifdef CUSTOM_MODES
 void opus_custom_encoder_destroy(CELTEncoder *st)
 {
    opus_free(st);
 }
 #endif /* CUSTOM_MODES */
 
 
 static int transient_analysis(const opus_val32 * OPUS_RESTRICT in, int len, int C,
@@ -236,17 +245,16 @@ static int transient_analysis(const opus
            5,  5,  5,  5,  5,  4,  4,  4,  4,  4,  4,  4,  4,  4,  4,  4,
            4,  4,  4,  4,  4,  4,  4,  4,  4,  4,  4,  4,  4,  4,  3,  3,
            3,  3,  3,  3,  3,  3,  3,  3,  3,  3,  3,  3,  3,  3,  3,  2,
    };
    SAVE_STACK;
    ALLOC(tmp, len, opus_val16);
 
    len2=len/2;
-   tf_max = 0;
    for (c=0;c<C;c++)
    {
       opus_val32 mean;
       opus_int32 unmask=0;
       opus_val32 norm;
       opus_val16 maxE;
       mem0=0;
       mem1=0;
@@ -365,43 +373,43 @@ static int transient_analysis(const opus
    is_transient = rand()&0x1;
 #endif
    /*printf("%d %f %d\n", is_transient, (float)*tf_estimate, tf_max);*/
    return is_transient;
 }
 
 /* Looks for sudden increases of energy to decide whether we need to patch
    the transient decision */
-int patch_transient_decision(opus_val16 *new, opus_val16 *old, int nbEBands,
+int patch_transient_decision(opus_val16 *newE, opus_val16 *oldE, int nbEBands,
       int end, int C)
 {
    int i, c;
    opus_val32 mean_diff=0;
    opus_val16 spread_old[26];
    /* Apply an aggressive (-6 dB/Bark) spreading function to the old frame to
       avoid false detection caused by irrelevant bands */
    if (C==1)
    {
-      spread_old[0] = old[0];
+      spread_old[0] = oldE[0];
       for (i=1;i<end;i++)
-         spread_old[i] = MAX16(spread_old[i-1]-QCONST16(1.0f, DB_SHIFT), old[i]);
+         spread_old[i] = MAX16(spread_old[i-1]-QCONST16(1.0f, DB_SHIFT), oldE[i]);
    } else {
-      spread_old[0] = MAX16(old[0],old[nbEBands]);
+      spread_old[0] = MAX16(oldE[0],oldE[nbEBands]);
       for (i=1;i<end;i++)
          spread_old[i] = MAX16(spread_old[i-1]-QCONST16(1.0f, DB_SHIFT),
-                               MAX16(old[i],old[i+nbEBands]));
+                               MAX16(oldE[i],oldE[i+nbEBands]));
    }
    for (i=end-2;i>=0;i--)
       spread_old[i] = MAX16(spread_old[i], spread_old[i+1]-QCONST16(1.0f, DB_SHIFT));
    /* Compute mean increase */
    c=0; do {
       for (i=2;i<end-1;i++)
       {
          opus_val16 x1, x2;
-         x1 = MAX16(0, new[i]);
+         x1 = MAX16(0, newE[i]);
          x2 = MAX16(0, spread_old[i]);
          mean_diff = ADD32(mean_diff, EXTEND32(MAX16(0, SUB16(x1, x2))));
       }
    } while (++c<C);
    mean_diff = DIV32(mean_diff, C*(end-3));
    /*printf("%f %f %d\n", mean_diff, max_diff, count);*/
    return mean_diff > QCONST16(1.f, DB_SHIFT);
 }
@@ -447,17 +455,17 @@ static void compute_mdcts(const CELTMode
             out[c*B*N+i] *= upsample;
          for (;i<B*N;i++)
             out[c*B*N+i] = 0;
       } while (++c<C);
    }
 }
 
 
-static void preemphasis(const opus_val16 * OPUS_RESTRICT pcmp, celt_sig * OPUS_RESTRICT inp,
+void celt_preemphasis(const opus_val16 * OPUS_RESTRICT pcmp, celt_sig * OPUS_RESTRICT inp,
                         int N, int CC, int upsample, const opus_val16 *coef, celt_sig *mem, int clip)
 {
    int i;
    opus_val16 coef0;
    celt_sig m;
    int Nu;
 
    coef0 = coef[0];
@@ -484,16 +492,18 @@ static void preemphasis(const opus_val16
 
 #ifndef FIXED_POINT
    if (clip)
    {
       /* Clip input to avoid encoding non-portable files */
       for (i=0;i<Nu;i++)
          inp[i*upsample] = MAX32(-65536.f, MIN32(65536.f,inp[i*upsample]));
    }
+#else
+   (void)clip; /* Avoids a warning about clip being unused. */
 #endif
    m = *mem;
 #ifdef CUSTOM_MODES
    if (coef[1] != 0)
    {
       opus_val16 coef1 = coef[1];
       opus_val16 coef2 = coef[2];
       for (i=0;i<N;i++)
@@ -739,17 +749,17 @@ static void tf_encode(int start, int end
       tf_res[i] = tf_select_table[LM][4*isTransient+2*tf_select+tf_res[i]];
    /*for(i=0;i<end;i++)printf("%d ", isTransient ? tf_res[i] : LM+tf_res[i]);printf("\n");*/
 }
 
 
 static int alloc_trim_analysis(const CELTMode *m, const celt_norm *X,
       const opus_val16 *bandLogE, int end, int LM, int C, int N0,
       AnalysisInfo *analysis, opus_val16 *stereo_saving, opus_val16 tf_estimate,
-      int intensity)
+      int intensity, opus_val16 surround_trim)
 {
    int i;
    opus_val32 diff=0;
    int c;
    int trim_index = 5;
    opus_val16 trim = QCONST16(5.f, 8);
    opus_val16 logXC, logXC2;
    if (C==2)
@@ -813,21 +823,23 @@ static int alloc_trim_analysis(const CEL
       trim_index--;
    if (diff > QCONST16(8.f, DB_SHIFT))
       trim_index--;
    if (diff < -QCONST16(4.f, DB_SHIFT))
       trim_index++;
    if (diff < -QCONST16(10.f, DB_SHIFT))
       trim_index++;
    trim -= MAX16(-QCONST16(2.f, 8), MIN16(QCONST16(2.f, 8), SHR16(diff+QCONST16(1.f, DB_SHIFT),DB_SHIFT-8)/6 ));
+   trim -= SHR16(surround_trim, DB_SHIFT-8);
    trim -= 2*SHR16(tf_estimate, 14-8);
-#ifndef FIXED_POINT
+#ifndef DISABLE_FLOAT_API
    if (analysis->valid)
    {
-      trim -= MAX16(-QCONST16(2.f, 8), MIN16(QCONST16(2.f, 8), 2*(analysis->tonality_slope+.05f)));
+      trim -= MAX16(-QCONST16(2.f, 8), MIN16(QCONST16(2.f, 8),
+            (opus_val16)(QCONST16(2.f, 8)*(analysis->tonality_slope+.05f))));
    }
 #endif
 
 #ifdef FIXED_POINT
    trim_index = PSHR32(trim, 8);
 #else
    trim_index = (int)floor(.5f+trim);
 #endif
@@ -872,17 +884,17 @@ static int stereo_analysis(const CELTMod
       thetas -= 8;
    return MULT16_32_Q15((m->eBands[13]<<(LM+1))+thetas, sumMS)
          > MULT16_32_Q15(m->eBands[13]<<(LM+1), sumLR);
 }
 
 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)
+      int effectiveBytes, opus_int32 *tot_boost_, int lfe, opus_val16 *surround_dynalloc)
 {
    int i, c;
    opus_int32 tot_boost=0;
    opus_val16 maxDepth;
    VARDECL(opus_val16, follower);
    VARDECL(opus_val16, noise_floor);
    SAVE_STACK;
    ALLOC(follower, C*nbEBands, opus_val16);
@@ -935,16 +947,18 @@ static opus_val16 dynalloc_analysis(cons
             follower[i] = HALF16(MAX16(0, bandLogE[i]-follower[i]) + MAX16(0, bandLogE[nbEBands+i]-follower[nbEBands+i]));
          }
       } else {
          for (i=start;i<end;i++)
          {
             follower[i] = MAX16(0, bandLogE[i]-follower[i]);
          }
       }
+      for (i=start;i<end;i++)
+         follower[i] = MAX16(follower[i], surround_dynalloc[i]);
       /* For non-transient CBR/CVBR frames, halve the dynalloc contribution */
       if ((!vbr || constrained_vbr)&&!isTransient)
       {
          for (i=start;i<end;i++)
             follower[i] = HALF16(follower[i]);
       }
       for (i=start;i<end;i++)
       {
@@ -1016,21 +1030,22 @@ static int run_prefilter(CELTEncoder *st
       OPUS_COPY(pre[c]+COMBFILTER_MAXPERIOD, in+c*(N+st->overlap)+st->overlap, N);
    } while (++c<CC);
 
    if (enabled)
    {
       VARDECL(opus_val16, pitch_buf);
       ALLOC(pitch_buf, (COMBFILTER_MAXPERIOD+N)>>1, opus_val16);
 
-      pitch_downsample(pre, pitch_buf, COMBFILTER_MAXPERIOD+N, CC);
+      pitch_downsample(pre, pitch_buf, COMBFILTER_MAXPERIOD+N, CC, st->arch);
       /* Don't search for the fir last 1.5 octave of the range because
          there's too many false-positives due to short-term correlation */
       pitch_search(pitch_buf+(COMBFILTER_MAXPERIOD>>1), pitch_buf, N,
-            COMBFILTER_MAXPERIOD-3*COMBFILTER_MINPERIOD, &pitch_index);
+            COMBFILTER_MAXPERIOD-3*COMBFILTER_MINPERIOD, &pitch_index,
+            st->arch);
       pitch_index = COMBFILTER_MAXPERIOD-pitch_index;
 
       gain1 = remove_doubling(pitch_buf, COMBFILTER_MAXPERIOD, COMBFILTER_MINPERIOD,
             N, &pitch_index, st->prefilter_period, st->prefilter_gain);
       if (pitch_index > COMBFILTER_MAXPERIOD-2)
          pitch_index = COMBFILTER_MAXPERIOD-2;
       gain1 = MULT16_16_Q15(QCONST16(.7f,15),gain1);
       /*printf("%d %d %f %f\n", pitch_change, pitch_index, gain1, st->analysis.tonality);*/
@@ -1135,54 +1150,55 @@ static int compute_vbr(const CELTMode *m
    coded_bands = lastCodedBands ? lastCodedBands : nbEBands;
    coded_bins = eBands[coded_bands]<<LM;
    if (C==2)
       coded_bins += eBands[IMIN(intensity, coded_bands)]<<LM;
 
    target = base_target;
 
    /*printf("%f %f %f %f %d %d ", st->analysis.activity, st->analysis.tonality, tf_estimate, st->stereo_saving, tot_boost, coded_bands);*/
-#ifndef FIXED_POINT
+#ifndef DISABLE_FLOAT_API
    if (analysis->valid && analysis->activity<.4)
       target -= (opus_int32)((coded_bins<<BITRES)*(.4f-analysis->activity));
 #endif
    /* Stereo savings */
    if (C==2)
    {
       int coded_stereo_bands;
       int coded_stereo_dof;
       opus_val16 max_frac;
       coded_stereo_bands = IMIN(intensity, coded_bands);
       coded_stereo_dof = (eBands[coded_stereo_bands]<<LM)-coded_stereo_bands;
       /* Maximum fraction of the bits we can save if the signal is mono. */
       max_frac = DIV32_16(MULT16_16(QCONST16(0.8f, 15), coded_stereo_dof), coded_bins);
+      stereo_saving = MIN16(stereo_saving, QCONST16(1.f, 8));
       /*printf("%d %d %d ", coded_stereo_dof, coded_bins, tot_boost);*/
       target -= (opus_int32)MIN32(MULT16_32_Q15(max_frac,target),
                       SHR32(MULT16_16(stereo_saving-QCONST16(0.1f,8),(coded_stereo_dof<<BITRES)),8));
    }
    /* Boost the rate according to dynalloc (minus the dynalloc average for calibration). */
    target += tot_boost-(16<<LM);
    /* Apply transient boost, compensating for average boost. */
    tf_calibration = variable_duration==OPUS_FRAMESIZE_VARIABLE ?
                     QCONST16(0.02f,14) : QCONST16(0.04f,14);
    target += (opus_int32)SHL32(MULT16_32_Q15(tf_estimate-tf_calibration, target),1);
 
-#ifndef FIXED_POINT
+#ifndef DISABLE_FLOAT_API
    /* Apply tonality boost */
    if (analysis->valid && !lfe)
    {
       opus_int32 tonal_target;
       float tonal;
 
       /* Tonality boost (compensating for the average). */
       tonal = MAX16(0.f,analysis->tonality-.15f)-0.09f;
       tonal_target = target + (opus_int32)((coded_bins<<BITRES)*1.2f*tonal);
       if (pitch_change)
          tonal_target +=  (opus_int32)((coded_bins<<BITRES)*.8f);
-      /*printf("%f %f ", st->analysis.tonality, tonal);*/
+      /*printf("%f %f ", analysis->tonality, tonal);*/
       target = tonal_target;
    }
 #endif
 
    if (has_surround_mask&&!lfe)
    {
       opus_int32 surround_target = target + (opus_int32)SHR32(MULT16_16(surround_masking,coded_bins<<BITRES), DB_SHIFT);
       /*printf("%f %d %d %d %d %d %d ", surround_masking, coded_bins, st->end, st->intensity, surround_target, target, st->bitrate);*/
@@ -1286,32 +1302,41 @@ int celt_encode_with_ec(CELTEncoder * OP
    int nbEBands;
    int overlap;
    const opus_int16 *eBands;
    int secondMdct;
    int signalBandwidth;
    int transient_got_disabled=0;
    opus_val16 surround_masking=0;
    opus_val16 temporal_vbr=0;
+   opus_val16 surround_trim = 0;
+   opus_int32 equiv_rate = 510000;
+   VARDECL(opus_val16, surround_dynalloc);
    ALLOC_STACK;
 
    mode = st->mode;
    nbEBands = mode->nbEBands;
    overlap = mode->overlap;
    eBands = mode->eBands;
    tf_estimate = 0;
    if (nbCompressedBytes<2 || pcm==NULL)
-     return OPUS_BAD_ARG;
+   {
+      RESTORE_STACK;
+      return OPUS_BAD_ARG;
+   }
 
    frame_size *= st->upsample;
    for (LM=0;LM<=mode->maxLM;LM++)
       if (mode->shortMdctSize<<LM==frame_size)
          break;
    if (LM>mode->maxLM)
+   {
+      RESTORE_STACK;
       return OPUS_BAD_ARG;
+   }
    M=1<<LM;
    N = M*mode->shortMdctSize;
 
    prefilter_mem = st->in_mem+CC*(st->overlap);
    oldBandE = (opus_val16*)(st->in_mem+CC*(st->overlap+COMBFILTER_MAXPERIOD));
    oldLogE = oldBandE + CC*nbEBands;
    oldLogE2 = oldLogE + CC*nbEBands;
 
@@ -1332,17 +1357,20 @@ int celt_encode_with_ec(CELTEncoder * OP
       compressed[0] = tmp<<5;
       compressed[0] |= LM<<3;
       compressed[0] |= (C==2)<<2;
       /* Convert "standard mode" to Opus header */
       if (mode->Fs==48000 && mode->shortMdctSize==120)
       {
          int c0 = toOpus(compressed[0]);
          if (c0<0)
+         {
+            RESTORE_STACK;
             return OPUS_BAD_ARG;
+         }
          compressed[0] = c0;
       }
       compressed++;
       nbCompressedBytes--;
    }
 #else
    celt_assert(st->signalling==0);
 #endif
@@ -1366,16 +1394,18 @@ int celt_encode_with_ec(CELTEncoder * OP
       tmp = st->bitrate*frame_size;
       if (tell>1)
          tmp += tell;
       if (st->bitrate!=OPUS_BITRATE_MAX)
          nbCompressedBytes = IMAX(2, IMIN(nbCompressedBytes,
                (tmp+4*mode->Fs)/(8*mode->Fs)-!!st->signalling));
       effectiveBytes = nbCompressedBytes;
    }
+   if (st->bitrate != OPUS_BITRATE_MAX)
+      equiv_rate = st->bitrate - (40*C+20)*((400>>LM) - 50);
 
    if (enc==NULL)
    {
       ec_enc_init(&_enc, compressed, nbCompressedBytes);
       enc = &_enc;
    }
 
    if (vbr_rate>0)
@@ -1439,27 +1469,27 @@ int celt_encode_with_ec(CELTEncoder * OP
          ec_enc_shrink(enc, nbCompressedBytes);
       }
       /* Pretend we've filled all the remaining bits with zeros
             (that's what the initialiser did anyway) */
       tell = nbCompressedBytes*8;
       enc->nbits_total+=tell-ec_tell(enc);
    }
    c=0; do {
-      preemphasis(pcm+c, in+c*(N+st->overlap)+st->overlap, N, CC, st->upsample,
+      celt_preemphasis(pcm+c, in+c*(N+st->overlap)+st->overlap, N, CC, st->upsample,
                   mode->preemph, st->preemph_memE+c, st->clip);
    } while (++c<CC);
 
 
 
    /* Find pitch period and gain */
    {
       int enabled;
       int qg;
-      enabled = (st->lfe || nbAvailableBytes>12*C) && st->start==0 && !silence && !st->disable_pf
+      enabled = ((st->lfe&&nbAvailableBytes>3) || nbAvailableBytes>12*C) && st->start==0 && !silence && !st->disable_pf
             && st->complexity >= 5 && !(st->consec_transient && LM!=3 && st->variable_duration==OPUS_FRAMESIZE_VARIABLE);
 
       prefilter_tapset = st->tapset_decision;
       pf_on = run_prefilter(st, in, prefilter_mem, CC, N, prefilter_tapset, &pitch_index, &gain1, &qg, enabled, nbAvailableBytes);
       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;
       if (pf_on==0)
@@ -1521,53 +1551,99 @@ int celt_encode_with_ec(CELTEncoder * OP
    {
       for (i=2;i<st->end;i++)
       {
          bandE[i] = IMIN(bandE[i], MULT16_32_Q15(QCONST16(1e-4f,15),bandE[0]));
          bandE[i] = MAX32(bandE[i], EPSILON);
       }
    }
    amp2Log2(mode, effEnd, st->end, bandE, bandLogE, C);
-   if (st->energy_save)
+
+   ALLOC(surround_dynalloc, C*nbEBands, opus_val16);
+   for(i=0;i<st->end;i++)
+      surround_dynalloc[i] = 0;
+   /* This computes how much masking takes place between surround channels */
+   if (st->start==0&&st->energy_mask&&!st->lfe)
    {
-      opus_val16 offset = shortBlocks?HALF16(SHL16(LM, DB_SHIFT)):0;
-#ifdef FIXED_POINT
-      /* Compensate for the 1/8 gain we apply in the fixed-point downshift to avoid overflows. */
-      offset -= QCONST16(3.0f, DB_SHIFT);
-#endif
-      for(i=0;i<C*nbEBands;i++)
-         st->energy_save[i]=bandLogE[i]-offset;
-      st->energy_save=NULL;
-   }
-   /* This computes how much masking takes place between surround channels */
-   if (st->energy_mask&&!st->lfe)
-   {
+      int mask_end;
+      int midband;
+      int count_dynalloc;
       opus_val32 mask_avg=0;
-      opus_val16 offset = shortBlocks?HALF16(SHL16(LM, DB_SHIFT)):0;
+      opus_val32 diff=0;
+      int count=0;
+      mask_end = IMAX(2,st->lastCodedBands);
       for (c=0;c<C;c++)
       {
-         opus_val16 followE, followMask;
-         followE = followMask = -QCONST16(14.f, DB_SHIFT);
-         for(i=0;i<st->end;i++)
+         for(i=0;i<mask_end;i++)
          {
-            /* We use a simple follower to approximate the masking spreading function. */
-            followE = MAX16(followE-QCONST16(1.f, DB_SHIFT), bandLogE[nbEBands*c+i]-offset);
-            followMask = MAX16(followMask-QCONST16(1.f, DB_SHIFT), st->energy_mask[nbEBands*c+i]);
-            mask_avg += followE-followMask;
+            opus_val16 mask;
+            mask = MAX16(MIN16(st->energy_mask[nbEBands*c+i],
+                   QCONST16(.25f, DB_SHIFT)), -QCONST16(2.0f, DB_SHIFT));
+            if (mask > 0)
+               mask = HALF16(mask);
+            mask_avg += MULT16_16(mask, eBands[i+1]-eBands[i]);
+            count += eBands[i+1]-eBands[i];
+            diff += MULT16_16(mask, 1+2*i-mask_end);
          }
       }
-      surround_masking = DIV32_16(mask_avg,C*st->end) + QCONST16(.7f, DB_SHIFT);
-      surround_masking = MIN16(MAX16(surround_masking, -QCONST16(2.f, DB_SHIFT)), QCONST16(.2f, DB_SHIFT));
-      surround_masking -= HALF16(HALF16(surround_masking));
+      mask_avg = DIV32_16(mask_avg,count);
+      mask_avg += QCONST16(.2f, DB_SHIFT);
+      diff = diff*6/(C*(mask_end-1)*(mask_end+1)*mask_end);
+      /* Again, being conservative */
+      diff = HALF32(diff);
+      diff = MAX32(MIN32(diff, QCONST32(.031f, DB_SHIFT)), -QCONST32(.031f, DB_SHIFT));
+      /* Find the band that's in the middle of the coded spectrum */
+      for (midband=0;eBands[midband+1] < eBands[mask_end]/2;midband++);
+      count_dynalloc=0;
+      for(i=0;i<mask_end;i++)
+      {
+         opus_val32 lin;
+         opus_val16 unmask;
+         lin = mask_avg + diff*(i-midband);
+         if (C==2)
+            unmask = MAX16(st->energy_mask[i], st->energy_mask[nbEBands+i]);
+         else
+            unmask = st->energy_mask[i];
+         unmask = MIN16(unmask, QCONST16(.0f, DB_SHIFT));
+         unmask -= lin;
+         if (unmask > QCONST16(.25f, DB_SHIFT))
+         {
+            surround_dynalloc[i] = unmask - QCONST16(.25f, DB_SHIFT);
+            count_dynalloc++;
+         }
+      }
+      if (count_dynalloc>=3)
+      {
+         /* If we need dynalloc in many bands, it's probably because our
+            initial masking rate was too low. */
+         mask_avg += QCONST16(.25f, DB_SHIFT);
+         if (mask_avg>0)
+         {
+            /* Something went really wrong in the original calculations,
+               disabling masking. */
+            mask_avg = 0;
+            diff = 0;
+            for(i=0;i<mask_end;i++)
+               surround_dynalloc[i] = 0;
+         } else {
+            for(i=0;i<mask_end;i++)
+               surround_dynalloc[i] = MAX16(0, surround_dynalloc[i]-QCONST16(.25f, DB_SHIFT));
+         }
+      }
+      mask_avg += QCONST16(.2f, DB_SHIFT);
+      /* Convert to 1/64th units used for the trim */
+      surround_trim = 64*diff;
+      /*printf("%d %d ", mask_avg, surround_trim);*/
+      surround_masking = mask_avg;
    }
    /* Temporal VBR (but not for LFE) */
    if (!st->lfe)
    {
       opus_val16 follow=-QCONST16(10.0f,DB_SHIFT);
-      float frame_avg=0;
+      opus_val32 frame_avg=0;
       opus_val16 offset = shortBlocks?HALF16(SHL16(LM, DB_SHIFT)):0;
       for(i=st->start;i<st->end;i++)
       {
          follow = MAX16(follow-QCONST16(1.f, DB_SHIFT), bandLogE[i]-offset);
          if (C==2)
             follow = MAX16(follow, bandLogE[i+nbEBands]-offset);
          frame_avg += follow;
       }
@@ -1655,40 +1731,43 @@ int celt_encode_with_ec(CELTEncoder * OP
          if (st->complexity == 0)
             st->spread_decision = SPREAD_NONE;
          else
             st->spread_decision = SPREAD_NORMAL;
       } else {
          /* Disable new spreading+tapset estimator until we can show it works
             better than the old one. So far it seems like spreading_decision()
             works best. */
-         if (0&&st->analysis.valid)
+#if 0
+         if (st->analysis.valid)
          {
             static const opus_val16 spread_thresholds[3] = {-QCONST16(.6f, 15), -QCONST16(.2f, 15), -QCONST16(.07f, 15)};
             static const opus_val16 spread_histeresis[3] = {QCONST16(.15f, 15), QCONST16(.07f, 15), QCONST16(.02f, 15)};
             static const opus_val16 tapset_thresholds[2] = {QCONST16(.0f, 15), QCONST16(.15f, 15)};
             static const opus_val16 tapset_histeresis[2] = {QCONST16(.1f, 15), QCONST16(.05f, 15)};
             st->spread_decision = hysteresis_decision(-st->analysis.tonality, spread_thresholds, spread_histeresis, 3, st->spread_decision);
             st->tapset_decision = hysteresis_decision(st->analysis.tonality_slope, tapset_thresholds, tapset_histeresis, 2, st->tapset_decision);
-         } else {
+         } else
+#endif
+         {
             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);
          }
          /*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);*/
       }
       ec_enc_icdf(enc, st->spread_decision, spread_icdf, 5);
    }
 
    ALLOC(offsets, nbEBands, int);
 
    maxDepth = dynalloc_analysis(bandLogE, bandLogE2, nbEBands, st->start, st->end, C, offsets,
          st->lsb_depth, mode->logN, isTransient, st->vbr, st->constrained_vbr,
-         eBands, LM, effectiveBytes, &tot_boost, st->lfe);
+         eBands, LM, effectiveBytes, &tot_boost, st->lfe, surround_dynalloc);
    /* For LFE, everything interesting is in the first band */
    if (st->lfe)
       offsets[0] = IMIN(8, effectiveBytes/3);
    ALLOC(cap, nbEBands, int);
    init_caps(mode,cap,LM,C);
 
    dynalloc_logp = 6;
    total_bits<<=BITRES;
@@ -1722,46 +1801,39 @@ int celt_encode_with_ec(CELTEncoder * OP
       /* Making dynalloc more likely */
       if (j)
          dynalloc_logp = IMAX(2, dynalloc_logp-1);
       offsets[i] = boost;
    }
 
    if (C==2)
    {
-      int effectiveRate;
-
       static const opus_val16 intensity_thresholds[21]=
       /* 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19  20  off*/
-        { 16,21,23,25,27,29,31,33,35,38,42,46,50,54,58,63,68,75,84,102,130};
+        {  1, 2, 3, 4, 5, 6, 7, 8,16,24,36,44,50,56,62,67,72,79,88,106,134};
       static const opus_val16 intensity_histeresis[21]=
-        {  2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 3, 4, 5, 6,  8, 12};
+        {  1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 3, 3, 4, 5, 6,  8, 8};
 
       /* Always use MS for 2.5 ms frames until we can do a better analysis */
       if (LM!=0)
          dual_stereo = stereo_analysis(mode, X, LM, N);
 
-      /* Account for coarse energy */
-      effectiveRate = (8*effectiveBytes - 80)>>LM;
-
-      /* effectiveRate in kb/s */
-      effectiveRate = 2*effectiveRate/5;
-
-      st->intensity = hysteresis_decision((opus_val16)effectiveRate, intensity_thresholds, intensity_histeresis, 21, st->intensity);
+      st->intensity = hysteresis_decision((opus_val16)equiv_rate/1000,
+            intensity_thresholds, intensity_histeresis, 21, st->intensity);
       st->intensity = IMIN(st->end,IMAX(st->start, st->intensity));
    }
 
    alloc_trim = 5;
    if (tell+(6<<BITRES) <= total_bits - total_boost)
    {
       if (st->lfe)
          alloc_trim = 5;
       else
          alloc_trim = alloc_trim_analysis(mode, X, bandLogE,
-            st->end, LM, C, N, &st->analysis, &st->stereo_saving, tf_estimate, st->intensity);
+            st->end, LM, C, N, &st->analysis, &st->stereo_saving, tf_estimate, st->intensity, surround_trim);
       ec_enc_icdf(enc, alloc_trim, trim_icdf, 7);
       tell = ec_tell_frac(enc);
    }
 
    /* Variable bitrate */
    if (vbr_rate>0)
    {
      opus_val16 alpha;
@@ -1774,17 +1846,17 @@ int celt_encode_with_ec(CELTEncoder * OP
      /* Don't attempt to use more than 510 kb/s, even for frames smaller than 20 ms.
         The CELT allocator will just not be able to use more than that anyway. */
      nbCompressedBytes = IMIN(nbCompressedBytes,1275>>(3-LM));
      base_target = vbr_rate - ((40*C+20)<<BITRES);
 
      if (st->constrained_vbr)
         base_target += (st->vbr_offset>>lm_diff);
 
-     target = compute_vbr(mode, &st->analysis, base_target, LM, st->bitrate,
+     target = compute_vbr(mode, &st->analysis, base_target, LM, equiv_rate,
            st->lastCodedBands, C, st->intensity, st->constrained_vbr,
            st->stereo_saving, tot_boost, tf_estimate, pitch_change, maxDepth,
            st->variable_duration, st->lfe, st->energy_mask!=NULL, surround_masking,
            temporal_vbr);
 
      /* The current offset is removed from the target and the space used
         so far is added*/
      target=target+tell;
@@ -1854,27 +1926,27 @@ int celt_encode_with_ec(CELTEncoder * OP
    ALLOC(pulses, nbEBands, int);
    ALLOC(fine_priority, nbEBands, int);
 
    /* bits =           packet size                    - where we are - safety*/
    bits = (((opus_int32)nbCompressedBytes*8)<<BITRES) - ec_tell_frac(enc) - 1;
    anti_collapse_rsv = isTransient&&LM>=2&&bits>=((LM+2)<<BITRES) ? (1<<BITRES) : 0;
    bits -= anti_collapse_rsv;
    signalBandwidth = st->end-1;
-#ifndef FIXED_POINT
+#ifndef DISABLE_FLOAT_API
    if (st->analysis.valid)
    {
       int min_bandwidth;
-      if (st->bitrate < (opus_int32)32000*C)
+      if (equiv_rate < (opus_int32)32000*C)
          min_bandwidth = 13;
-      else if (st->bitrate < (opus_int32)48000*C)
+      else if (equiv_rate < (opus_int32)48000*C)
          min_bandwidth = 16;
-      else if (st->bitrate < (opus_int32)60000*C)
+      else if (equiv_rate < (opus_int32)60000*C)
          min_bandwidth = 18;
-      else  if (st->bitrate < (opus_int32)80000*C)
+      else  if (equiv_rate < (opus_int32)80000*C)
          min_bandwidth = 19;
       else
          min_bandwidth = 20;
       signalBandwidth = IMAX(st->analysis.bandwidth, min_bandwidth);
    }
 #endif
    if (st->lfe)
       signalBandwidth = 1;
@@ -2256,22 +2328,16 @@ int opus_custom_encoder_ctl(CELTEncoder 
       }
       break;
       case OPUS_SET_LFE_REQUEST:
       {
           opus_int32 value = va_arg(ap, opus_int32);
           st->lfe = value;
       }
       break;
-      case OPUS_SET_ENERGY_SAVE_REQUEST:
-      {
-          opus_val16 *value = va_arg(ap, opus_val16*);
-          st->energy_save=value;
-      }
-      break;
       case OPUS_SET_ENERGY_MASK_REQUEST:
       {
           opus_val16 *value = va_arg(ap, opus_val16*);
           st->energy_mask = value;
       }
       break;
       default:
          goto bad_request;

--- a/media/libopus/celt/celt_lpc.c
+++ b/media/libopus/celt/celt_lpc.c
@@ -221,17 +221,18 @@ void celt_iir(const opus_val32 *_x,
 }
 
 int _celt_autocorr(
                    const opus_val16 *x,   /*  in: [0...n-1] samples x   */
                    opus_val32       *ac,  /* out: [0...lag-1] ac values */
                    const opus_val16       *window,
                    int          overlap,
                    int          lag,
-                   int          n
+                   int          n,
+                   int          arch
                   )
 {
    opus_val32 d;
    int i, k;
    int fastN=n-lag;
    int shift;
    const opus_val16 *xptr;
    VARDECL(opus_val16, xx);
@@ -270,17 +271,17 @@ int _celt_autocorr(
       {
          for(i=0;i<n;i++)
             xx[i] = PSHR32(xptr[i], shift);
          xptr = xx;
       } else
          shift = 0;
    }
 #endif
-   celt_pitch_xcorr(xptr, xptr, ac, fastN, lag+1);
+   celt_pitch_xcorr(xptr, xptr, ac, fastN, lag+1, arch);
    for (k=0;k<=lag;k++)
    {
       for (i = k+fastN, d = 0; i < n; i++)
          d = MAC16_16(d, xptr[i], xptr[i-k]);
       ac[k] += d;
    }
 #ifdef FIXED_POINT
    shift = 2*shift;

--- a/media/libopus/celt/celt_lpc.h
+++ b/media/libopus/celt/celt_lpc.h
@@ -43,11 +43,12 @@ void celt_fir(const opus_val16 *x,
 
 void celt_iir(const opus_val32 *x,
          const opus_val16 *den,
          opus_val32 *y,
          int N,
          int ord,
          opus_val16 *mem);
 
-int _celt_autocorr(const opus_val16 *x, opus_val32 *ac, const opus_val16 *window, int overlap, int lag, int n);
+int _celt_autocorr(const opus_val16 *x, opus_val32 *ac,
+         const opus_val16 *window, int overlap, int lag, int n, int arch);
 
 #endif /* PLC_H */

--- a/media/libopus/celt/cpu_support.h
+++ b/media/libopus/celt/cpu_support.h
@@ -23,29 +23,32 @@
    LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
    NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */
 
 #ifndef CPU_SUPPORT_H
 #define CPU_SUPPORT_H
 
-#if defined(OPUS_HAVE_RTCD) && defined(ARMv4_ASM)
+#include "opus_types.h"
+#include "opus_defines.h"
+
+#if defined(OPUS_HAVE_RTCD) && defined(OPUS_ARM_ASM)
 #include "arm/armcpu.h"
 
 /* We currently support 4 ARM variants:
  * arch[0] -> ARMv4
  * arch[1] -> ARMv5E
  * arch[2] -> ARMv6
  * arch[3] -> NEON
  */
 #define OPUS_ARCHMASK 3
 
 #else
 #define OPUS_ARCHMASK 0
 
-static inline int opus_select_arch(void)
+static OPUS_INLINE int opus_select_arch(void)
 {
   return 0;
 }
 #endif
 
 #endif

--- a/media/libopus/celt/cwrs.c
+++ b/media/libopus/celt/cwrs.c
@@ -405,25 +405,25 @@ static const opus_uint32 CELT_PVQ_U_DATA
   3248227095U,
   /*N=13, K=13...16:*/
   251595969, 579168825, 1267854873, 2653649025U,
   /*N=14, K=14:*/
   1409933619
 };
 
 #if defined(CUSTOM_MODES)
-const opus_uint32 *const CELT_PVQ_U_ROW[15]={
+static const opus_uint32 *const CELT_PVQ_U_ROW[15]={
   CELT_PVQ_U_DATA+   0,CELT_PVQ_U_DATA+ 208,CELT_PVQ_U_DATA+ 415,
   CELT_PVQ_U_DATA+ 621,CELT_PVQ_U_DATA+ 826,CELT_PVQ_U_DATA+1030,
   CELT_PVQ_U_DATA+1233,CELT_PVQ_U_DATA+1336,CELT_PVQ_U_DATA+1389,
   CELT_PVQ_U_DATA+1421,CELT_PVQ_U_DATA+1441,CELT_PVQ_U_DATA+1455,
   CELT_PVQ_U_DATA+1464,CELT_PVQ_U_DATA+1470,CELT_PVQ_U_DATA+1473
 };
 #else
-const opus_uint32 *const CELT_PVQ_U_ROW[15]={
+static const opus_uint32 *const CELT_PVQ_U_ROW[15]={
   CELT_PVQ_U_DATA+   0,CELT_PVQ_U_DATA+ 176,CELT_PVQ_U_DATA+ 351,
   CELT_PVQ_U_DATA+ 525,CELT_PVQ_U_DATA+ 698,CELT_PVQ_U_DATA+ 870,
   CELT_PVQ_U_DATA+1041,CELT_PVQ_U_DATA+1131,CELT_PVQ_U_DATA+1178,
   CELT_PVQ_U_DATA+1207,CELT_PVQ_U_DATA+1226,CELT_PVQ_U_DATA+1240,
   CELT_PVQ_U_DATA+1248,CELT_PVQ_U_DATA+1254,CELT_PVQ_U_DATA+1257
 };
 #endif
 
@@ -529,33 +529,33 @@ void decode_pulses(int *_y,int _n,int _k
   cwrsi(_n,_k,ec_dec_uint(_dec,CELT_PVQ_V(_n,_k)),_y);
 }
 
 #else /* SMALL_FOOTPRINT */
 
 /*Computes the next row/column of any recurrence that obeys the relation
    u[i][j]=u[i-1][j]+u[i][j-1]+u[i-1][j-1].
   _ui0 is the base case for the new row/column.*/
-static inline void unext(opus_uint32 *_ui,unsigned _len,opus_uint32 _ui0){
+static OPUS_INLINE void unext(opus_uint32 *_ui,unsigned _len,opus_uint32 _ui0){
   opus_uint32 ui1;
   unsigned      j;
   /*This do-while will overrun the array if we don't have storage for at least
      2 values.*/
   j=1; do {
     ui1=UADD32(UADD32(_ui[j],_ui[j-1]),_ui0);
     _ui[j-1]=_ui0;
     _ui0=ui1;
   } while (++j<_len);
   _ui[j-1]=_ui0;
 }
 
 /*Computes the previous row/column of any recurrence that obeys the relation
    u[i-1][j]=u[i][j]-u[i][j-1]-u[i-1][j-1].
   _ui0 is the base case for the new row/column.*/
-static inline void uprev(opus_uint32 *_ui,unsigned _n,opus_uint32 _ui0){
+static OPUS_INLINE void uprev(opus_uint32 *_ui,unsigned _n,opus_uint32 _ui0){
   opus_uint32 ui1;
   unsigned      j;
   /*This do-while will overrun the array if we don't have storage for at least
      2 values.*/
   j=1; do {
     ui1=USUB32(USUB32(_ui[j],_ui[j-1]),_ui0);
     _ui[j-1]=_ui0;
     _ui0=ui1;
@@ -612,26 +612,26 @@ static void cwrsi(int _n,int _k,opus_uin
   }
   while(++j<_n);
 }
 
 /*Returns the index of the given combination of K elements chosen from a set
    of size 1 with associated sign bits.
   _y: The vector of pulses, whose sum of absolute values is K.
   _k: Returns K.*/
-static inline opus_uint32 icwrs1(const int *_y,int *_k){
+static OPUS_INLINE opus_uint32 icwrs1(const int *_y,int *_k){
   *_k=abs(_y[0]);
   return _y[0]<0;
 }
 
 /*Returns the index of the given combination of K elements chosen from a set
    of size _n with associated sign bits.
   _y:  The vector of pulses, whose sum of absolute values must be _k.
   _nc: Returns V(_n,_k).*/
-static inline opus_uint32 icwrs(int _n,int _k,opus_uint32 *_nc,const int *_y,
+static OPUS_INLINE opus_uint32 icwrs(int _n,int _k,opus_uint32 *_nc,const int *_y,
  opus_uint32 *_u){
   opus_uint32 i;
   int         j;
   int         k;
   /*We can't unroll the first two iterations of the loop unless _n>=2.*/
   celt_assert(_n>=2);
   _u[0]=0;
   for(k=1;k<=_k+1;k++)_u[k]=(k<<1)-1;

--- a/media/libopus/celt/ecintrin.h
+++ b/media/libopus/celt/ecintrin.h
@@ -28,17 +28,17 @@
 /*Some common macros for potential platform-specific optimization.*/
 #include "opus_types.h"
 #include <math.h>
 #include <limits.h>
 #include "arch.h"
 #if !defined(_ecintrin_H)
 # define _ecintrin_H (1)
 
-/*Some specific platforms may have optimized intrinsic or inline assembly
+/*Some specific platforms may have optimized intrinsic or OPUS_INLINE assembly
    versions of these functions which can substantially improve performance.
   We define macros for them to allow easy incorporation of these non-ANSI
    features.*/
 
 /*Modern gcc (4.x) can compile the naive versions of min and max with cmov if
    given an appropriate architecture, but the branchless bit-twiddling versions
    are just as fast, and do not require any special target architecture.
   Earlier gcc versions (3.x) compiled both code to the same assembly

--- a/media/libopus/celt/entcode.h
+++ b/media/libopus/celt/entcode.h
@@ -21,16 +21,17 @@
    PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
    PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
    LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
    NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */
 
 #include "opus_types.h"
+#include "opus_defines.h"
 
 #if !defined(_entcode_H)
 # define _entcode_H (1)
 # include <limits.h>
 # include <stddef.h>
 # include "ecintrin.h"
 
 /*OPT: ec_window must be at least 32 bits, but if you have fast arithmetic on a
@@ -78,35 +79,35 @@ struct ec_ctx{
      In the encoder: the number of oustanding carry propagating symbols.*/
    opus_uint32    ext;
    /*A buffered input/output symbol, awaiting carry propagation.*/
    int            rem;
    /*Nonzero if an error occurred.*/
    int            error;
 };
 
-static inline opus_uint32 ec_range_bytes(ec_ctx *_this){
+static OPUS_INLINE opus_uint32 ec_range_bytes(ec_ctx *_this){
   return _this->offs;
 }
 
-static inline unsigned char *ec_get_buffer(ec_ctx *_this){
+static OPUS_INLINE unsigned char *ec_get_buffer(ec_ctx *_this){
   return _this->buf;
 }
 
-static inline int ec_get_error(ec_ctx *_this){
+static OPUS_INLINE int ec_get_error(ec_ctx *_this){
   return _this->error;
 }
 
 /*Returns the number of bits "used" by the encoded or decoded symbols so far.
   This same number can be computed in either the encoder or the decoder, and is
    suitable for making coding decisions.
   Return: The number of bits.
           This will always be slightly larger than the exact value (e.g., all
            rounding error is in the positive direction).*/
-static inline int ec_tell(ec_ctx *_this){
+static OPUS_INLINE int ec_tell(ec_ctx *_this){
   return _this->nbits_total-EC_ILOG(_this->rng);
 }
 
 /*Returns the number of bits "used" by the encoded or decoded symbols so far.
   This same number can be computed in either the encoder or the decoder, and is
    suitable for making coding decisions.
   Return: The number of bits scaled by 2**BITRES.
           This will always be slightly larger than the exact value (e.g., all

--- a/media/libopus/celt/fixed_debug.h
+++ b/media/libopus/celt/fixed_debug.h
@@ -28,19 +28,19 @@
    NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */
 
 #ifndef FIXED_DEBUG_H
 #define FIXED_DEBUG_H
 
 #include <stdio.h>
+#include "opus_defines.h"
 
 #ifdef CELT_C
-#include "opus_defines.h"
 OPUS_EXPORT opus_int64 celt_mips=0;
 #else
 extern opus_int64 celt_mips;
 #endif
 
 #define MULT16_16SU(a,b) ((opus_val32)(opus_val16)(a)*(opus_val32)(opus_uint16)(b))
 #define MULT32_32_Q31(a,b) ADD32(ADD32(SHL32(MULT16_16(SHR32((a),16),SHR((b),16)),1), SHR32(MULT16_16SU(SHR32((a),16),((b)&0x0000ffff)),15)), SHR32(MULT16_16SU(SHR32((b),16),((a)&0x0000ffff)),15))
 
@@ -54,17 +54,17 @@ extern opus_int64 celt_mips;
 
 #define VERIFY_SHORT(x) ((x)<=32767&&(x)>=-32768)
 #define VERIFY_INT(x) ((x)<=2147483647LL&&(x)>=-2147483648LL)
 #define VERIFY_UINT(x) ((x)<=(2147483647LLU<<1))
 
 #define SHR(a,b) SHR32(a,b)
 #define PSHR(a,b) PSHR32(a,b)
 
-static inline short NEG16(int x)
+static OPUS_INLINE short NEG16(int x)
 {
    int res;
    if (!VERIFY_SHORT(x))
    {
       fprintf (stderr, "NEG16: input is not short: %d\n", (int)x);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
@@ -75,17 +75,17 @@ static inline short NEG16(int x)
       fprintf (stderr, "NEG16: output is not short: %d\n", (int)res);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
    }
    celt_mips++;
    return res;
 }
-static inline int NEG32(opus_int64 x)
+static OPUS_INLINE int NEG32(opus_int64 x)
 {
    opus_int64 res;
    if (!VERIFY_INT(x))
    {
       fprintf (stderr, "NEG16: input is not int: %d\n", (int)x);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
@@ -98,49 +98,49 @@ static inline int NEG32(opus_int64 x)
       celt_assert(0);
 #endif
    }
    celt_mips+=2;
    return res;
 }
 
 #define EXTRACT16(x) EXTRACT16_(x, __FILE__, __LINE__)
-static inline short EXTRACT16_(int x, char *file, int line)
+static OPUS_INLINE short EXTRACT16_(int x, char *file, int line)
 {
    int res;
    if (!VERIFY_SHORT(x))
    {
       fprintf (stderr, "EXTRACT16: input is not short: %d in %s: line %d\n", x, file, line);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
    }
    res = x;
    celt_mips++;
    return res;
 }
 
 #define EXTEND32(x) EXTEND32_(x, __FILE__, __LINE__)
-static inline int EXTEND32_(int x, char *file, int line)
+static OPUS_INLINE int EXTEND32_(int x, char *file, int line)
 {
    int res;
    if (!VERIFY_SHORT(x))
    {
       fprintf (stderr, "EXTEND32: input is not short: %d in %s: line %d\n", x, file, line);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
    }
    res = x;
    celt_mips++;
    return res;
 }
 
 #define SHR16(a, shift) SHR16_(a, shift, __FILE__, __LINE__)
-static inline short SHR16_(int a, int shift, char *file, int line)
+static OPUS_INLINE short SHR16_(int a, int shift, char *file, int line)
 {
    int res;
    if (!VERIFY_SHORT(a) || !VERIFY_SHORT(shift))
    {
       fprintf (stderr, "SHR16: inputs are not short: %d >> %d in %s: line %d\n", a, shift, file, line);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
@@ -152,17 +152,17 @@ static inline short SHR16_(int a, int sh
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
    }
    celt_mips++;
    return res;
 }
 #define SHL16(a, shift) SHL16_(a, shift, __FILE__, __LINE__)
-static inline short SHL16_(int a, int shift, char *file, int line)
+static OPUS_INLINE short SHL16_(int a, int shift, char *file, int line)
 {
    int res;
    if (!VERIFY_SHORT(a) || !VERIFY_SHORT(shift))
    {
       fprintf (stderr, "SHL16: inputs are not short: %d %d in %s: line %d\n", a, shift, file, line);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
@@ -174,17 +174,17 @@ static inline short SHL16_(int a, int sh
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
    }
    celt_mips++;
    return res;
 }
 
-static inline int SHR32(opus_int64 a, int shift)
+static OPUS_INLINE int SHR32(opus_int64 a, int shift)
 {
    opus_int64  res;
    if (!VERIFY_INT(a) || !VERIFY_SHORT(shift))
    {
       fprintf (stderr, "SHR32: inputs are not int: %d %d\n", (int)a, shift);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
@@ -196,17 +196,17 @@ static inline int SHR32(opus_int64 a, in
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
    }
    celt_mips+=2;
    return res;
 }
 #define SHL32(a, shift) SHL32_(a, shift, __FILE__, __LINE__)
-static inline int SHL32_(opus_int64 a, int shift, char *file, int line)
+static OPUS_INLINE int SHL32_(opus_int64 a, int shift, char *file, int line)
 {
    opus_int64  res;
    if (!VERIFY_INT(a) || !VERIFY_SHORT(shift))
    {
       fprintf (stderr, "SHL32: inputs are not int: %lld %d in %s: line %d\n", a, shift, file, line);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
@@ -229,17 +229,17 @@ static inline int SHL32_(opus_int64 a, i
 #define ROUND16(x,a) (celt_mips--,EXTRACT16(PSHR32((x),(a))))
 #define HALF16(x)  (SHR16(x,1))
 #define HALF32(x)  (SHR32(x,1))
 
 //#define SHR(a,shift) ((a) >> (shift))
 //#define SHL(a,shift) ((a) << (shift))
 
 #define ADD16(a, b) ADD16_(a, b, __FILE__, __LINE__)
-static inline short ADD16_(int a, int b, char *file, int line)
+static OPUS_INLINE short ADD16_(int a, int b, char *file, int line)
 {
    int res;
    if (!VERIFY_SHORT(a) || !VERIFY_SHORT(b))
    {
       fprintf (stderr, "ADD16: inputs are not short: %d %d in %s: line %d\n", a, b, file, line);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
@@ -252,17 +252,17 @@ static inline short ADD16_(int a, int b,
       celt_assert(0);
 #endif
    }
    celt_mips++;
    return res;
 }
 
 #define SUB16(a, b) SUB16_(a, b, __FILE__, __LINE__)
-static inline short SUB16_(int a, int b, char *file, int line)
+static OPUS_INLINE short SUB16_(int a, int b, char *file, int line)
 {
    int res;
    if (!VERIFY_SHORT(a) || !VERIFY_SHORT(b))
    {
       fprintf (stderr, "SUB16: inputs are not short: %d %d in %s: line %d\n", a, b, file, line);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
@@ -275,17 +275,17 @@ static inline short SUB16_(int a, int b,
       celt_assert(0);
 #endif
    }
    celt_mips++;
    return res;
 }
 
 #define ADD32(a, b) ADD32_(a, b, __FILE__, __LINE__)
-static inline int ADD32_(opus_int64 a, opus_int64 b, char *file, int line)
+static OPUS_INLINE int ADD32_(opus_int64 a, opus_int64 b, char *file, int line)
 {
    opus_int64 res;
    if (!VERIFY_INT(a) || !VERIFY_INT(b))
    {
       fprintf (stderr, "ADD32: inputs are not int: %d %d in %s: line %d\n", (int)a, (int)b, file, line);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
@@ -298,17 +298,17 @@ static inline int ADD32_(opus_int64 a, o
       celt_assert(0);
 #endif
    }
    celt_mips+=2;
    return res;
 }
 
 #define SUB32(a, b) SUB32_(a, b, __FILE__, __LINE__)
-static inline int SUB32_(opus_int64 a, opus_int64 b, char *file, int line)
+static OPUS_INLINE int SUB32_(opus_int64 a, opus_int64 b, char *file, int line)
 {
    opus_int64 res;
    if (!VERIFY_INT(a) || !VERIFY_INT(b))
    {
       fprintf (stderr, "SUB32: inputs are not int: %d %d in %s: line %d\n", (int)a, (int)b, file, line);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
@@ -322,17 +322,17 @@ static inline int SUB32_(opus_int64 a, o
 #endif
    }
    celt_mips+=2;
    return res;
 }
 
 #undef UADD32
 #define UADD32(a, b) UADD32_(a, b, __FILE__, __LINE__)
-static inline unsigned int UADD32_(opus_uint64 a, opus_uint64 b, char *file, int line)
+static OPUS_INLINE unsigned int UADD32_(opus_uint64 a, opus_uint64 b, char *file, int line)
 {
    opus_uint64 res;
    if (!VERIFY_UINT(a) || !VERIFY_UINT(b))
    {
       fprintf (stderr, "UADD32: inputs are not uint32: %llu %llu in %s: line %d\n", a, b, file, line);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
@@ -346,17 +346,17 @@ static inline unsigned int UADD32_(opus_
 #endif
    }
    celt_mips+=2;
    return res;
 }
 
 #undef USUB32
 #define USUB32(a, b) USUB32_(a, b, __FILE__, __LINE__)
-static inline unsigned int USUB32_(opus_uint64 a, opus_uint64 b, char *file, int line)
+static OPUS_INLINE unsigned int USUB32_(opus_uint64 a, opus_uint64 b, char *file, int line)
 {
    opus_uint64 res;
    if (!VERIFY_UINT(a) || !VERIFY_UINT(b))
    {
       fprintf (stderr, "USUB32: inputs are not uint32: %llu %llu in %s: line %d\n", a, b, file, line);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
@@ -376,17 +376,17 @@ static inline unsigned int USUB32_(opus_
       celt_assert(0);
 #endif
    }
    celt_mips+=2;
    return res;
 }
 
 /* result fits in 16 bits */
-static inline short MULT16_16_16(int a, int b)
+static OPUS_INLINE short MULT16_16_16(int a, int b)
 {
    int res;
    if (!VERIFY_SHORT(a) || !VERIFY_SHORT(b))
    {
       fprintf (stderr, "MULT16_16_16: inputs are not short: %d %d\n", a, b);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
@@ -399,17 +399,17 @@ static inline short MULT16_16_16(int a, 
       celt_assert(0);
 #endif
    }
    celt_mips++;
    return res;
 }
 
 #define MULT16_16(a, b) MULT16_16_(a, b, __FILE__, __LINE__)
-static inline int MULT16_16_(int a, int b, char *file, int line)
+static OPUS_INLINE int MULT16_16_(int a, int b, char *file, int line)
 {
    opus_int64 res;
    if (!VERIFY_SHORT(a) || !VERIFY_SHORT(b))
    {
       fprintf (stderr, "MULT16_16: inputs are not short: %d %d in %s: line %d\n", a, b, file, line);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
@@ -424,17 +424,17 @@ static inline int MULT16_16_(int a, int 
    }
    celt_mips++;
    return res;
 }
 
 #define MAC16_16(c,a,b)     (celt_mips-=2,ADD32((c),MULT16_16((a),(b))))
 
 #define MULT16_32_QX(a, b, Q) MULT16_32_QX_(a, b, Q, __FILE__, __LINE__)
-static inline int MULT16_32_QX_(int a, opus_int64 b, int Q, char *file, int line)
+static OPUS_INLINE int MULT16_32_QX_(int a, opus_int64 b, int Q, char *file, int line)
 {
    opus_int64 res;
    if (!VERIFY_SHORT(a) || !VERIFY_INT(b))
    {
       fprintf (stderr, "MULT16_32_Q%d: inputs are not short+int: %d %d in %s: line %d\n", Q, (int)a, (int)b, file, line);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
@@ -457,17 +457,17 @@ static inline int MULT16_32_QX_(int a, o
    if (Q==15)
       celt_mips+=3;
    else
       celt_mips+=4;
    return res;
 }
 
 #define MULT16_32_PX(a, b, Q) MULT16_32_PX_(a, b, Q, __FILE__, __LINE__)
-static inline int MULT16_32_PX_(int a, opus_int64 b, int Q, char *file, int line)
+static OPUS_INLINE int MULT16_32_PX_(int a, opus_int64 b, int Q, char *file, int line)
 {
    opus_int64 res;
    if (!VERIFY_SHORT(a) || !VERIFY_INT(b))
    {
       fprintf (stderr, "MULT16_32_P%d: inputs are not short+int: %d %d in %s: line %d\n\n", Q, (int)a, (int)b, file, line);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
@@ -492,27 +492,37 @@ static inline int MULT16_32_PX_(int a, o
    else
       celt_mips+=5;
    return res;
 }
 
 #define MULT16_32_Q15(a,b) MULT16_32_QX(a,b,15)
 #define MAC16_32_Q15(c,a,b) (celt_mips-=2,ADD32((c),MULT16_32_Q15((a),(b))))
 
-static inline int SATURATE(int a, int b)
+static OPUS_INLINE int SATURATE(int a, int b)
 {
    if (a>b)
       a=b;
    if (a<-b)
       a = -b;
    celt_mips+=3;
    return a;
 }
 
-static inline int MULT16_16_Q11_32(int a, int b)
+static OPUS_INLINE opus_int16 SATURATE16(opus_int32 a)
+{
+   celt_mips+=3;
+   if (a>32767)
+      return 32767;
+   else if (a<-32768)
+      return -32768;
+   else return a;
+}
+
+static OPUS_INLINE int MULT16_16_Q11_32(int a, int b)
 {
    opus_int64 res;
    if (!VERIFY_SHORT(a) || !VERIFY_SHORT(b))
    {
       fprintf (stderr, "MULT16_16_Q11: inputs are not short: %d %d\n", a, b);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
@@ -524,17 +534,17 @@ static inline int MULT16_16_Q11_32(int a
       fprintf (stderr, "MULT16_16_Q11: output is not short: %d*%d=%d\n", (int)a, (int)b, (int)res);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
    }
    celt_mips+=3;
    return res;
 }
-static inline short MULT16_16_Q13(int a, int b)
+static OPUS_INLINE short MULT16_16_Q13(int a, int b)
 {
    opus_int64 res;
    if (!VERIFY_SHORT(a) || !VERIFY_SHORT(b))
    {
       fprintf (stderr, "MULT16_16_Q13: inputs are not short: %d %d\n", a, b);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
@@ -546,17 +556,17 @@ static inline short MULT16_16_Q13(int a,
       fprintf (stderr, "MULT16_16_Q13: output is not short: %d*%d=%d\n", a, b, (int)res);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
    }
    celt_mips+=3;
    return res;
 }
-static inline short MULT16_16_Q14(int a, int b)
+static OPUS_INLINE short MULT16_16_Q14(int a, int b)
 {
    opus_int64 res;
    if (!VERIFY_SHORT(a) || !VERIFY_SHORT(b))
    {
       fprintf (stderr, "MULT16_16_Q14: inputs are not short: %d %d\n", a, b);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
@@ -570,17 +580,17 @@ static inline short MULT16_16_Q14(int a,
       celt_assert(0);
 #endif
    }
    celt_mips+=3;
    return res;
 }
 
 #define MULT16_16_Q15(a, b) MULT16_16_Q15_(a, b, __FILE__, __LINE__)
-static inline short MULT16_16_Q15_(int a, int b, char *file, int line)
+static OPUS_INLINE short MULT16_16_Q15_(int a, int b, char *file, int line)
 {
    opus_int64 res;
    if (!VERIFY_SHORT(a) || !VERIFY_SHORT(b))
    {
       fprintf (stderr, "MULT16_16_Q15: inputs are not short: %d %d in %s: line %d\n", a, b, file, line);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
@@ -593,17 +603,17 @@ static inline short MULT16_16_Q15_(int a
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
    }
    celt_mips+=1;
    return res;
 }
 
-static inline short MULT16_16_P13(int a, int b)
+static OPUS_INLINE short MULT16_16_P13(int a, int b)
 {
    opus_int64 res;
    if (!VERIFY_SHORT(a) || !VERIFY_SHORT(b))
    {
       fprintf (stderr, "MULT16_16_P13: inputs are not short: %d %d\n", a, b);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
@@ -623,17 +633,17 @@ static inline short MULT16_16_P13(int a,
       fprintf (stderr, "MULT16_16_P13: output is not short: %d*%d=%d\n", a, b, (int)res);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
    }
    celt_mips+=4;
    return res;
 }
-static inline short MULT16_16_P14(int a, int b)
+static OPUS_INLINE short MULT16_16_P14(int a, int b)
 {
    opus_int64 res;
    if (!VERIFY_SHORT(a) || !VERIFY_SHORT(b))
    {
       fprintf (stderr, "MULT16_16_P14: inputs are not short: %d %d\n", a, b);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
@@ -653,17 +663,17 @@ static inline short MULT16_16_P14(int a,
       fprintf (stderr, "MULT16_16_P14: output is not short: %d*%d=%d\n", a, b, (int)res);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
    }
    celt_mips+=4;
    return res;
 }
-static inline short MULT16_16_P15(int a, int b)
+static OPUS_INLINE short MULT16_16_P15(int a, int b)
 {
    opus_int64 res;
    if (!VERIFY_SHORT(a) || !VERIFY_SHORT(b))
    {
       fprintf (stderr, "MULT16_16_P15: inputs are not short: %d %d\n", a, b);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
@@ -686,17 +696,17 @@ static inline short MULT16_16_P15(int a,
 #endif
    }
    celt_mips+=2;
    return res;
 }
 
 #define DIV32_16(a, b) DIV32_16_(a, b, __FILE__, __LINE__)
 
-static inline int DIV32_16_(opus_int64 a, opus_int64 b, char *file, int line)
+static OPUS_INLINE int DIV32_16_(opus_int64 a, opus_int64 b, char *file, int line)
 {
    opus_int64 res;
    if (b==0)
    {
       fprintf(stderr, "DIV32_16: divide by zero: %d/%d in %s: line %d\n", (int)a, (int)b, file, line);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif
@@ -721,17 +731,17 @@ static inline int DIV32_16_(opus_int64 a
       celt_assert(0);
 #endif
    }
    celt_mips+=35;
    return res;
 }
 
 #define DIV32(a, b) DIV32_(a, b, __FILE__, __LINE__)
-static inline int DIV32_(opus_int64 a, opus_int64 b, char *file, int line)
+static OPUS_INLINE int DIV32_(opus_int64 a, opus_int64 b, char *file, int line)
 {
    opus_int64 res;
    if (b==0)
    {
       fprintf(stderr, "DIV32: divide by zero: %d/%d in %s: line %d\n", (int)a, (int)b, file, line);
 #ifdef FIXED_DEBUG_ASSERT
       celt_assert(0);
 #endif

--- a/media/libopus/celt/fixed_generic.h
+++ b/media/libopus/celt/fixed_generic.h
@@ -35,17 +35,17 @@
 
 /** Multiply a 16-bit signed value by a 16-bit unsigned value. The result is a 32-bit signed value */
 #define MULT16_16SU(a,b) ((opus_val32)(opus_val16)(a)*(opus_val32)(opus_uint16)(b))
 
 /** 16x32 multiplication, followed by a 16-bit shift right. Results fits in 32 bits */
 #define MULT16_32_Q16(a,b) ADD32(MULT16_16((a),SHR((b),16)), SHR(MULT16_16SU((a),((b)&0x0000ffff)),16))
 
 /** 16x32 multiplication, followed by a 16-bit shift right (round-to-nearest). Results fits in 32 bits */
-#define MULT16_32_P16(a,b) ADD32(MULT16_16((a),SHR((b),16)), PSHR(MULT16_16((a),((b)&0x0000ffff)),16))
+#define MULT16_32_P16(a,b) ADD32(MULT16_16((a),SHR((b),16)), PSHR(MULT16_16SU((a),((b)&0x0000ffff)),16))
 
 /** 16x32 multiplication, followed by a 15-bit shift right. Results fits in 32 bits */
 #define MULT16_32_Q15(a,b) ADD32(SHL(MULT16_16((a),SHR((b),16)),1), SHR(MULT16_16SU((a),((b)&0x0000ffff)),15))
 
 /** 32x32 multiplication, followed by a 31-bit shift right. Results fits in 32 bits */
 #define MULT32_32_Q31(a,b) ADD32(ADD32(SHL(MULT16_16(SHR((a),16),SHR((b),16)),1), SHR(MULT16_16SU(SHR((a),16),((b)&0x0000ffff)),15)), SHR(MULT16_16SU(SHR((b),16),((a)&0x0000ffff)),15))
 
 /** Compile-time conversion of float constant to 16-bit value */
@@ -111,16 +111,17 @@
 /** 16x16 multiply-add where the result fits in 32 bits */
 #define MAC16_16(c,a,b) (ADD32((c),MULT16_16((a),(b))))
 /** 16x32 multiply, followed by a 15-bit shift right and 32-bit add.
     b must fit in 31 bits.
     Result fits in 32 bits. */
 #define MAC16_32_Q15(c,a,b) ADD32(c,ADD32(MULT16_16((a),SHR((b),15)), SHR(MULT16_16((a),((b)&0x00007fff)),15)))
 
 #define MULT16_16_Q11_32(a,b) (SHR(MULT16_16((a),(b)),11))
+#define MULT16_16_Q11(a,b) (SHR(MULT16_16((a),(b)),11))
 #define MULT16_16_Q13(a,b) (SHR(MULT16_16((a),(b)),13))
 #define MULT16_16_Q14(a,b) (SHR(MULT16_16((a),(b)),14))
 #define MULT16_16_Q15(a,b) (SHR(MULT16_16((a),(b)),15))
 
 #define MULT16_16_P13(a,b) (SHR(ADD32(4096,MULT16_16((a),(b))),13))
 #define MULT16_16_P14(a,b) (SHR(ADD32(8192,MULT16_16((a),(b))),14))
 #define MULT16_16_P15(a,b) (SHR(ADD32(16384,MULT16_16((a),(b))),15))
 

--- a/media/libopus/celt/float_cast.h
+++ b/media/libopus/celt/float_cast.h
@@ -96,17 +96,17 @@
         __inline long int float2int(float value)
         {
                 return _mm_cvtss_si32(_mm_load_ss(&value));
         }
 #elif (defined(_MSC_VER) && _MSC_VER >= 1400) && (defined (WIN32) || defined (_WIN32))
         #include <math.h>
 
         /*      Win32 doesn't seem to have these functions.
-        **      Therefore implement inline versions of these functions here.
+        **      Therefore implement OPUS_INLINE versions of these functions here.
         */
 
         __inline long int
         float2int (float flt)
         {       int intgr;
 
                 _asm
                 {       fld flt
@@ -123,17 +123,17 @@
         #warning "Don't have the functions lrint() and lrintf ()."
         #warning "Replacing these functions with a standard C cast."
 #endif /* __STDC_VERSION__ >= 199901L */
         #include <math.h>
         #define float2int(flt) ((int)(floor(.5+flt)))
 #endif
 
 #ifndef DISABLE_FLOAT_API
-static inline opus_int16 FLOAT2INT16(float x)
+static OPUS_INLINE opus_int16 FLOAT2INT16(float x)
 {
    x = x*CELT_SIG_SCALE;
    x = MAX32(x, -32768);
    x = MIN32(x, 32767);
    return (opus_int16)float2int(x);
 }
 #endif /* DISABLE_FLOAT_API */
 

--- a/media/libopus/celt/mathops.c
+++ b/media/libopus/celt/mathops.c
@@ -134,17 +134,17 @@ opus_val32 celt_sqrt(opus_val32 x)
    return rt;
 }
 
 #define L1 32767
 #define L2 -7651
 #define L3 8277
 #define L4 -626
 
-static inline opus_val16 _celt_cos_pi_2(opus_val16 x)
+static OPUS_INLINE opus_val16 _celt_cos_pi_2(opus_val16 x)
 {
    opus_val16 x2;
 
    x2 = MULT16_16_P15(x,x);
    return ADD16(1,MIN16(32766,ADD32(SUB16(L1,x2), MULT16_16_P15(x2, ADD32(L2, MULT16_16_P15(x2, ADD32(L3, MULT16_16_P15(L4, x2
                                                                                 ))))))));
 }
 

--- a/media/libopus/celt/mathops.h
+++ b/media/libopus/celt/mathops.h
@@ -39,33 +39,33 @@
 #include "os_support.h"
 
 /* Multiplies two 16-bit fractional values. Bit-exactness of this macro is important */
 #define FRAC_MUL16(a,b) ((16384+((opus_int32)(opus_int16)(a)*(opus_int16)(b)))>>15)
 
 unsigned isqrt32(opus_uint32 _val);
 
 #ifndef OVERRIDE_CELT_MAXABS16
-static inline opus_val32 celt_maxabs16(const opus_val16 *x, int len)
+static OPUS_INLINE opus_val32 celt_maxabs16(const opus_val16 *x, int len)
 {
    int i;
    opus_val16 maxval = 0;
    opus_val16 minval = 0;
    for (i=0;i<len;i++)
    {
       maxval = MAX16(maxval, x[i]);
       minval = MIN16(minval, x[i]);
    }
    return MAX32(EXTEND32(maxval),-EXTEND32(minval));
 }
 #endif
 
 #ifndef OVERRIDE_CELT_MAXABS32
 #ifdef FIXED_POINT
-static inline opus_val32 celt_maxabs32(const opus_val32 *x, int len)
+static OPUS_INLINE opus_val32 celt_maxabs32(const opus_val32 *x, int len)
 {
    int i;
    opus_val32 maxval = 0;
    opus_val32 minval = 0;
    for (i=0;i<len;i++)
    {
       maxval = MAX32(maxval, x[i]);
       minval = MIN32(minval, x[i]);
@@ -90,17 +90,17 @@ static inline opus_val32 celt_maxabs32(c
 #define frac_div32(a,b) ((float)(a)/(b))
 
 #ifdef FLOAT_APPROX
 
 /* Note: This assumes radix-2 floating point with the exponent at bits 23..30 and an offset of 127
          denorm, +/- inf and NaN are *not* handled */
 
 /** Base-2 log approximation (log2(x)). */
-static inline float celt_log2(float x)
+static OPUS_INLINE float celt_log2(float x)
 {
    int integer;
    float frac;
    union {
       float f;
       opus_uint32 i;
    } in;
    in.f = x;
@@ -108,17 +108,17 @@ static inline float celt_log2(float x)
    in.i -= integer<<23;
    frac = in.f - 1.5f;
    frac = -0.41445418f + frac*(0.95909232f
           + frac*(-0.33951290f + frac*0.16541097f));
    return 1+integer+frac;
 }
 
 /** Base-2 exponential approximation (2^x). */
-static inline float celt_exp2(float x)
+static OPUS_INLINE float celt_exp2(float x)
 {
    int integer;
    float frac;
    union {
       float f;
       opus_uint32 i;
    } res;
    integer = floor(x);
@@ -140,37 +140,38 @@ static inline float celt_exp2(float x)
 #endif
 
 #ifdef FIXED_POINT
 
 #include "os_support.h"
 
 #ifndef OVERRIDE_CELT_ILOG2
 /** Integer log in base2. Undefined for zero and negative numbers */
-static inline opus_int16 celt_ilog2(opus_int32 x)
+static OPUS_INLINE opus_int16 celt_ilog2(opus_int32 x)
 {
    celt_assert2(x>0, "celt_ilog2() only defined for strictly positive numbers");
    return EC_ILOG(x)-1;
 }
 #endif
 
 
 /** Integer log in base2. Defined for zero, but not for negative numbers */
-static inline opus_int16 celt_zlog2(opus_val32 x)
+static OPUS_INLINE opus_int16 celt_zlog2(opus_val32 x)
 {
    return x <= 0 ? 0 : celt_ilog2(x);
 }
 
 opus_val16 celt_rsqrt_norm(opus_val32 x);
 
 opus_val32 celt_sqrt(opus_val32 x);
 
 opus_val16 celt_cos_norm(opus_val32 x);
 
-static inline opus_val16 celt_log2(opus_val32 x)
+/** Base-2 logarithm approximation (log2(x)). (Q14 input, Q10 output) */
+static OPUS_INLINE opus_val16 celt_log2(opus_val32 x)
 {
    int i;
    opus_val16 n, frac;
    /* -0.41509302963303146, 0.9609890551383969, -0.31836011537636605,
        0.15530808010959576, -0.08556153059057618 */
    static const opus_val16 C[5] = {-6801+(1<<(13-DB_SHIFT)), 15746, -5217, 2545, -1401};
    if (x==0)
       return -32767;
@@ -186,24 +187,24 @@ static inline opus_val16 celt_log2(opus_
  K2 = 3-4*log(2)
  K3 = 3*log(2) - 2
 */
 #define D0 16383
 #define D1 22804
 #define D2 14819
 #define D3 10204
 
-static inline opus_val32 celt_exp2_frac(opus_val16 x)
+static OPUS_INLINE opus_val32 celt_exp2_frac(opus_val16 x)
 {
    opus_val16 frac;
    frac = SHL16(x, 4);
    return ADD16(D0, MULT16_16_Q15(frac, ADD16(D1, MULT16_16_Q15(frac, ADD16(D2 , MULT16_16_Q15(D3,frac))))));
 }
 /** Base-2 exponential approximation (2^x). (Q10 input, Q16 output) */
-static inline opus_val32 celt_exp2(opus_val16 x)
+static OPUS_INLINE opus_val32 celt_exp2(opus_val16 x)
 {
    int integer;
    opus_val16 frac;
    integer = SHR16(x,10);
    if (integer>14)
       return 0x7f000000;
    else if (integer < -15)
       return 0;
@@ -219,28 +220,28 @@ opus_val32 frac_div32(opus_val32 a, opus
 
 #define M1 32767
 #define M2 -21
 #define M3 -11943
 #define M4 4936
 
 /* Atan approximation using a 4th order polynomial. Input is in Q15 format
    and normalized by pi/4. Output is in Q15 format */
-static inline opus_val16 celt_atan01(opus_val16 x)
+static OPUS_INLINE opus_val16 celt_atan01(opus_val16 x)
 {
    return MULT16_16_P15(x, ADD32(M1, MULT16_16_P15(x, ADD32(M2, MULT16_16_P15(x, ADD32(M3, MULT16_16_P15(M4, x)))))));
 }
 
 #undef M1
 #undef M2
 #undef M3
 #undef M4
 
 /* atan2() approximation valid for positive input values */
-static inline opus_val16 celt_atan2p(opus_val16 y, opus_val16 x)
+static OPUS_INLINE opus_val16 celt_atan2p(opus_val16 y, opus_val16 x)
 {
    if (y < x)
    {
       opus_val32 arg;
       arg = celt_div(SHL32(EXTEND32(y),15),x);
       if (arg >= 32767)
          arg = 32767;
       return SHR16(celt_atan01(EXTRACT16(arg)),1);