From 02eac8a741a3d6b9bfacbceb750c44fba7ed4662 Mon Sep 17 00:00:00 2001
From: James Zern <jzern@google.com>
Date: Sat, 3 Jul 2021 17:45:13 -0700
Subject: [PATCH 1/8] dsp/cost*: use WEBP_RESTRICT qualifier

on SetResidualCoeffs_*. This results in some minor code reordering when
targeting arvm7 with ndk r27 and other recent versions of clang. No
changes in the x86 compilations with clang-16 / gcc-13.

This only affects non-vector pointers; any vector pointers are left as a
follow up.

Change-Id: I7c3554ece848fafbc5ac9c4944f1dc85129f6fd8
---
 src/dsp/cost.c        | 4 ++--
 src/dsp/cost_mips32.c | 4 ++--
 src/dsp/cost_neon.c   | 4 ++--
 src/dsp/cost_sse2.c   | 4 ++--
 src/dsp/dsp.h         | 5 +++--
 5 files changed, 11 insertions(+), 10 deletions(-)

diff --git a/src/dsp/cost.c b/src/dsp/cost.c
index 73d21401..609f9264 100644
--- a/src/dsp/cost.c
+++ b/src/dsp/cost.c
@@ -354,8 +354,8 @@ static int GetResidualCost_C(int ctx0, const VP8Residual* const res) {
   return cost;
 }
 
-static void SetResidualCoeffs_C(const int16_t* const coeffs,
-                                VP8Residual* const res) {
+static void SetResidualCoeffs_C(const int16_t* WEBP_RESTRICT const coeffs,
+                                VP8Residual* WEBP_RESTRICT const res) {
   int n;
   res->last = -1;
   assert(res->first == 0 || coeffs[0] == 0);
diff --git a/src/dsp/cost_mips32.c b/src/dsp/cost_mips32.c
index 0500f88c..54586576 100644
--- a/src/dsp/cost_mips32.c
+++ b/src/dsp/cost_mips32.c
@@ -96,8 +96,8 @@ static int GetResidualCost_MIPS32(int ctx0, const VP8Residual* const res) {
   return cost;
 }
 
-static void SetResidualCoeffs_MIPS32(const int16_t* const coeffs,
-                                     VP8Residual* const res) {
+static void SetResidualCoeffs_MIPS32(const int16_t* WEBP_RESTRICT const coeffs,
+                                     VP8Residual* WEBP_RESTRICT const res) {
   const int16_t* p_coeffs = (int16_t*)coeffs;
   int temp0, temp1, temp2, n, n1;
   assert(res->first == 0 || coeffs[0] == 0);
diff --git a/src/dsp/cost_neon.c b/src/dsp/cost_neon.c
index 6582669c..e1bf3657 100644
--- a/src/dsp/cost_neon.c
+++ b/src/dsp/cost_neon.c
@@ -19,8 +19,8 @@
 static const uint8_t position[16] = { 1, 2,  3,  4,  5,  6,  7,  8,
                                       9, 10, 11, 12, 13, 14, 15, 16 };
 
-static void SetResidualCoeffs_NEON(const int16_t* const coeffs,
-                                   VP8Residual* const res) {
+static void SetResidualCoeffs_NEON(const int16_t* WEBP_RESTRICT const coeffs,
+                                   VP8Residual* WEBP_RESTRICT const res) {
   const int16x8_t minus_one = vdupq_n_s16(-1);
   const int16x8_t coeffs_0 = vld1q_s16(coeffs);
   const int16x8_t coeffs_1 = vld1q_s16(coeffs + 8);
diff --git a/src/dsp/cost_sse2.c b/src/dsp/cost_sse2.c
index 487a0799..a869b48d 100644
--- a/src/dsp/cost_sse2.c
+++ b/src/dsp/cost_sse2.c
@@ -22,8 +22,8 @@
 
 //------------------------------------------------------------------------------
 
-static void SetResidualCoeffs_SSE2(const int16_t* const coeffs,
-                                   VP8Residual* const res) {
+static void SetResidualCoeffs_SSE2(const int16_t* WEBP_RESTRICT const coeffs,
+                                   VP8Residual* WEBP_RESTRICT const res) {
   const __m128i c0 = _mm_loadu_si128((const __m128i*)(coeffs + 0));
   const __m128i c1 = _mm_loadu_si128((const __m128i*)(coeffs + 8));
   // Use SSE2 to compare 16 values with a single instruction.
diff --git a/src/dsp/dsp.h b/src/dsp/dsp.h
index 23bc2965..82029533 100644
--- a/src/dsp/dsp.h
+++ b/src/dsp/dsp.h
@@ -138,8 +138,9 @@ extern const uint16_t VP8LevelFixedCosts[2047 /*MAX_LEVEL*/ + 1];
 extern const uint8_t VP8EncBands[16 + 1];
 
 struct VP8Residual;
-typedef void (*VP8SetResidualCoeffsFunc)(const int16_t* const coeffs,
-                                         struct VP8Residual* const res);
+typedef void (*VP8SetResidualCoeffsFunc)(
+    const int16_t* WEBP_RESTRICT const coeffs,
+    struct VP8Residual* WEBP_RESTRICT const res);
 extern VP8SetResidualCoeffsFunc VP8SetResidualCoeffs;
 
 // Cost calculation function.

From 201894ef248f745288290e1a676c679b998d6673 Mon Sep 17 00:00:00 2001
From: James Zern <jzern@google.com>
Date: Sat, 3 Jul 2021 17:52:50 -0700
Subject: [PATCH 2/8] dsp/dec*: use WEBP_RESTRICT qualifier

A minor improvement for arm targets with ndk r27/gcc-13 in H/VFilter8 (a
couple fewer moves w/aarch64) and much better vectorization of
DitherCombine8x8_C in most targets.

This only affects non-vector pointers; any vector pointers are left as a
follow up.

Change-Id: I03e73e6d6404261bb8408a9ae76a4b6ef142f8f0
---
 src/dsp/dec.c             | 47 ++++++++++++++++++++++-----------------
 src/dsp/dec_mips32.c      | 22 +++++++++---------
 src/dsp/dec_mips_dsp_r2.c | 28 +++++++++++++----------
 src/dsp/dec_msa.c         | 29 +++++++++++++++---------
 src/dsp/dec_neon.c        | 32 ++++++++++++++++----------
 src/dsp/dec_sse2.c        | 20 ++++++++++-------
 src/dsp/dsp.h             | 23 ++++++++++++-------
 src/dsp/enc.c             |  2 +-
 8 files changed, 121 insertions(+), 82 deletions(-)

diff --git a/src/dsp/dec.c b/src/dsp/dec.c
index 451d649d..51067f45 100644
--- a/src/dsp/dec.c
+++ b/src/dsp/dec.c
@@ -38,7 +38,8 @@ static WEBP_INLINE uint8_t clip_8b(int v) {
 } while (0)
 
 #if !WEBP_NEON_OMIT_C_CODE
-static void TransformOne_C(const int16_t* in, uint8_t* dst) {
+static void TransformOne_C(const int16_t* WEBP_RESTRICT in,
+                           uint8_t* WEBP_RESTRICT dst) {
   int C[4 * 4], *tmp;
   int i;
   tmp = C;
@@ -82,7 +83,8 @@ static void TransformOne_C(const int16_t* in, uint8_t* dst) {
 }
 
 // Simplified transform when only in[0], in[1] and in[4] are non-zero
-static void TransformAC3_C(const int16_t* in, uint8_t* dst) {
+static void TransformAC3_C(const int16_t* WEBP_RESTRICT in,
+                           uint8_t* WEBP_RESTRICT dst) {
   const int a = in[0] + 4;
   const int c4 = WEBP_TRANSFORM_AC3_MUL2(in[4]);
   const int d4 = WEBP_TRANSFORM_AC3_MUL1(in[4]);
@@ -95,7 +97,8 @@ static void TransformAC3_C(const int16_t* in, uint8_t* dst) {
 }
 #undef STORE2
 
-static void TransformTwo_C(const int16_t* in, uint8_t* dst, int do_two) {
+static void TransformTwo_C(const int16_t* WEBP_RESTRICT in,
+                           uint8_t* WEBP_RESTRICT dst, int do_two) {
   TransformOne_C(in, dst);
   if (do_two) {
     TransformOne_C(in + 16, dst + 4);
@@ -103,13 +106,15 @@ static void TransformTwo_C(const int16_t* in, uint8_t* dst, int do_two) {
 }
 #endif  // !WEBP_NEON_OMIT_C_CODE
 
-static void TransformUV_C(const int16_t* in, uint8_t* dst) {
+static void TransformUV_C(const int16_t* WEBP_RESTRICT in,
+                          uint8_t* WEBP_RESTRICT dst) {
   VP8Transform(in + 0 * 16, dst, 1);
   VP8Transform(in + 2 * 16, dst + 4 * BPS, 1);
 }
 
 #if !WEBP_NEON_OMIT_C_CODE
-static void TransformDC_C(const int16_t* in, uint8_t* dst) {
+static void TransformDC_C(const int16_t* WEBP_RESTRICT in,
+                          uint8_t* WEBP_RESTRICT dst) {
   const int DC = in[0] + 4;
   int i, j;
   for (j = 0; j < 4; ++j) {
@@ -120,7 +125,8 @@ static void TransformDC_C(const int16_t* in, uint8_t* dst) {
 }
 #endif  // !WEBP_NEON_OMIT_C_CODE
 
-static void TransformDCUV_C(const int16_t* in, uint8_t* dst) {
+static void TransformDCUV_C(const int16_t* WEBP_RESTRICT in,
+                            uint8_t* WEBP_RESTRICT dst) {
   if (in[0 * 16]) VP8TransformDC(in + 0 * 16, dst);
   if (in[1 * 16]) VP8TransformDC(in + 1 * 16, dst + 4);
   if (in[2 * 16]) VP8TransformDC(in + 2 * 16, dst + 4 * BPS);
@@ -133,7 +139,8 @@ static void TransformDCUV_C(const int16_t* in, uint8_t* dst) {
 // Paragraph 14.3
 
 #if !WEBP_NEON_OMIT_C_CODE
-static void TransformWHT_C(const int16_t* in, int16_t* out) {
+static void TransformWHT_C(const int16_t* WEBP_RESTRICT in,
+                           int16_t* WEBP_RESTRICT out) {
   int tmp[16];
   int i;
   for (i = 0; i < 4; ++i) {
@@ -161,7 +168,7 @@ static void TransformWHT_C(const int16_t* in, int16_t* out) {
 }
 #endif  // !WEBP_NEON_OMIT_C_CODE
 
-void (*VP8TransformWHT)(const int16_t* in, int16_t* out);
+VP8IWHT VP8TransformWHT;
 
 //------------------------------------------------------------------------------
 // Intra predictions
@@ -661,32 +668,32 @@ static void HFilter16i_C(uint8_t* p, int stride,
 
 #if !WEBP_NEON_OMIT_C_CODE
 // 8-pixels wide variant, for chroma filtering
-static void VFilter8_C(uint8_t* u, uint8_t* v, int stride,
-                       int thresh, int ithresh, int hev_thresh) {
+static void VFilter8_C(uint8_t* WEBP_RESTRICT u, uint8_t* WEBP_RESTRICT v,
+                       int stride, int thresh, int ithresh, int hev_thresh) {
   FilterLoop26_C(u, stride, 1, 8, thresh, ithresh, hev_thresh);
   FilterLoop26_C(v, stride, 1, 8, thresh, ithresh, hev_thresh);
 }
 #endif  // !WEBP_NEON_OMIT_C_CODE
 
 #if !WEBP_NEON_OMIT_C_CODE || WEBP_NEON_WORK_AROUND_GCC
-static void HFilter8_C(uint8_t* u, uint8_t* v, int stride,
-                       int thresh, int ithresh, int hev_thresh) {
+static void HFilter8_C(uint8_t* WEBP_RESTRICT u, uint8_t* WEBP_RESTRICT v,
+                       int stride, int thresh, int ithresh, int hev_thresh) {
   FilterLoop26_C(u, 1, stride, 8, thresh, ithresh, hev_thresh);
   FilterLoop26_C(v, 1, stride, 8, thresh, ithresh, hev_thresh);
 }
 #endif  // !WEBP_NEON_OMIT_C_CODE || WEBP_NEON_WORK_AROUND_GCC
 
 #if !WEBP_NEON_OMIT_C_CODE
-static void VFilter8i_C(uint8_t* u, uint8_t* v, int stride,
-                        int thresh, int ithresh, int hev_thresh) {
+static void VFilter8i_C(uint8_t* WEBP_RESTRICT u, uint8_t* WEBP_RESTRICT v,
+                        int stride, int thresh, int ithresh, int hev_thresh) {
   FilterLoop24_C(u + 4 * stride, stride, 1, 8, thresh, ithresh, hev_thresh);
   FilterLoop24_C(v + 4 * stride, stride, 1, 8, thresh, ithresh, hev_thresh);
 }
 #endif  // !WEBP_NEON_OMIT_C_CODE
 
 #if !WEBP_NEON_OMIT_C_CODE || WEBP_NEON_WORK_AROUND_GCC
-static void HFilter8i_C(uint8_t* u, uint8_t* v, int stride,
-                        int thresh, int ithresh, int hev_thresh) {
+static void HFilter8i_C(uint8_t* WEBP_RESTRICT u, uint8_t* WEBP_RESTRICT v,
+                        int stride, int thresh, int ithresh, int hev_thresh) {
   FilterLoop24_C(u + 4, 1, stride, 8, thresh, ithresh, hev_thresh);
   FilterLoop24_C(v + 4, 1, stride, 8, thresh, ithresh, hev_thresh);
 }
@@ -694,8 +701,8 @@ static void HFilter8i_C(uint8_t* u, uint8_t* v, int stride,
 
 //------------------------------------------------------------------------------
 
-static void DitherCombine8x8_C(const uint8_t* dither, uint8_t* dst,
-                               int dst_stride) {
+static void DitherCombine8x8_C(const uint8_t* WEBP_RESTRICT dither,
+                               uint8_t* WEBP_RESTRICT dst, int dst_stride) {
   int i, j;
   for (j = 0; j < 8; ++j) {
     for (i = 0; i < 8; ++i) {
@@ -730,8 +737,8 @@ VP8SimpleFilterFunc VP8SimpleHFilter16;
 VP8SimpleFilterFunc VP8SimpleVFilter16i;
 VP8SimpleFilterFunc VP8SimpleHFilter16i;
 
-void (*VP8DitherCombine8x8)(const uint8_t* dither, uint8_t* dst,
-                            int dst_stride);
+void (*VP8DitherCombine8x8)(const uint8_t* WEBP_RESTRICT dither,
+                            uint8_t* WEBP_RESTRICT dst, int dst_stride);
 
 extern VP8CPUInfo VP8GetCPUInfo;
 extern void VP8DspInitSSE2(void);
diff --git a/src/dsp/dec_mips32.c b/src/dsp/dec_mips32.c
index f0e7de4a..89fe9009 100644
--- a/src/dsp/dec_mips32.c
+++ b/src/dsp/dec_mips32.c
@@ -133,26 +133,26 @@ static void HFilter16(uint8_t* p, int stride,
 }
 
 // 8-pixels wide variant, for chroma filtering
-static void VFilter8(uint8_t* u, uint8_t* v, int stride,
-                     int thresh, int ithresh, int hev_thresh) {
+static void VFilter8(uint8_t* WEBP_RESTRICT u, uint8_t* WEBP_RESTRICT v,
+                     int stride, int thresh, int ithresh, int hev_thresh) {
   FilterLoop26(u, stride, 1, 8, thresh, ithresh, hev_thresh);
   FilterLoop26(v, stride, 1, 8, thresh, ithresh, hev_thresh);
 }
 
-static void HFilter8(uint8_t* u, uint8_t* v, int stride,
-                     int thresh, int ithresh, int hev_thresh) {
+static void HFilter8(uint8_t* WEBP_RESTRICT u, uint8_t* WEBP_RESTRICT v,
+                     int stride, int thresh, int ithresh, int hev_thresh) {
   FilterLoop26(u, 1, stride, 8, thresh, ithresh, hev_thresh);
   FilterLoop26(v, 1, stride, 8, thresh, ithresh, hev_thresh);
 }
 
-static void VFilter8i(uint8_t* u, uint8_t* v, int stride,
-                      int thresh, int ithresh, int hev_thresh) {
+static void VFilter8i(uint8_t* WEBP_RESTRICT u, uint8_t* WEBP_RESTRICT v,
+                      int stride, int thresh, int ithresh, int hev_thresh) {
   FilterLoop24(u + 4 * stride, stride, 1, 8, thresh, ithresh, hev_thresh);
   FilterLoop24(v + 4 * stride, stride, 1, 8, thresh, ithresh, hev_thresh);
 }
 
-static void HFilter8i(uint8_t* u, uint8_t* v, int stride,
-                      int thresh, int ithresh, int hev_thresh) {
+static void HFilter8i(uint8_t* WEBP_RESTRICT u, uint8_t* WEBP_RESTRICT v,
+                      int stride, int thresh, int ithresh, int hev_thresh) {
   FilterLoop24(u + 4, 1, stride, 8, thresh, ithresh, hev_thresh);
   FilterLoop24(v + 4, 1, stride, 8, thresh, ithresh, hev_thresh);
 }
@@ -215,7 +215,8 @@ static void SimpleHFilter16i(uint8_t* p, int stride, int thresh) {
   }
 }
 
-static void TransformOne(const int16_t* in, uint8_t* dst) {
+static void TransformOne(const int16_t* WEBP_RESTRICT in,
+                         uint8_t* WEBP_RESTRICT dst) {
   int temp0, temp1, temp2, temp3, temp4;
   int temp5, temp6, temp7, temp8, temp9;
   int temp10, temp11, temp12, temp13, temp14;
@@ -532,7 +533,8 @@ static void TransformOne(const int16_t* in, uint8_t* dst) {
   );
 }
 
-static void TransformTwo(const int16_t* in, uint8_t* dst, int do_two) {
+static void TransformTwo(const int16_t* WEBP_RESTRICT in,
+                         uint8_t* WEBP_RESTRICT dst, int do_two) {
   TransformOne(in, dst);
   if (do_two) {
     TransformOne(in + 16, dst + 4);
diff --git a/src/dsp/dec_mips_dsp_r2.c b/src/dsp/dec_mips_dsp_r2.c
index 0ba706a2..03b5f122 100644
--- a/src/dsp/dec_mips_dsp_r2.c
+++ b/src/dsp/dec_mips_dsp_r2.c
@@ -21,7 +21,8 @@
 static const int kC1 = WEBP_TRANSFORM_AC3_C1;
 static const int kC2 = WEBP_TRANSFORM_AC3_C2;
 
-static void TransformDC(const int16_t* in, uint8_t* dst) {
+static void TransformDC(const int16_t* WEBP_RESTRICT in,
+                        uint8_t* WEBP_RESTRICT dst) {
   int temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8, temp9, temp10;
 
   __asm__ volatile (
@@ -45,7 +46,8 @@ static void TransformDC(const int16_t* in, uint8_t* dst) {
   );
 }
 
-static void TransformAC3(const int16_t* in, uint8_t* dst) {
+static void TransformAC3(const int16_t* WEBP_RESTRICT in,
+                         uint8_t* WEBP_RESTRICT dst) {
   const int a = in[0] + 4;
   int c4 = WEBP_TRANSFORM_AC3_MUL2(in[4]);
   const int d4 = WEBP_TRANSFORM_AC3_MUL1(in[4]);
@@ -81,7 +83,8 @@ static void TransformAC3(const int16_t* in, uint8_t* dst) {
   );
 }
 
-static void TransformOne(const int16_t* in, uint8_t* dst) {
+static void TransformOne(const int16_t* WEBP_RESTRICT in,
+                         uint8_t* WEBP_RESTRICT dst) {
   int temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8, temp9;
   int temp10, temp11, temp12, temp13, temp14, temp15, temp16, temp17, temp18;
 
@@ -148,7 +151,8 @@ static void TransformOne(const int16_t* in, uint8_t* dst) {
   );
 }
 
-static void TransformTwo(const int16_t* in, uint8_t* dst, int do_two) {
+static void TransformTwo(const int16_t* WEBP_RESTRICT in,
+                         uint8_t* WEBP_RESTRICT dst, int do_two) {
   TransformOne(in, dst);
   if (do_two) {
     TransformOne(in + 16, dst + 4);
@@ -434,14 +438,14 @@ static void HFilter16(uint8_t* p, int stride,
 }
 
 // 8-pixels wide variant, for chroma filtering
-static void VFilter8(uint8_t* u, uint8_t* v, int stride,
-                     int thresh, int ithresh, int hev_thresh) {
+static void VFilter8(uint8_t* WEBP_RESTRICT u, uint8_t* WEBP_RESTRICT v,
+                     int stride, int thresh, int ithresh, int hev_thresh) {
   FilterLoop26(u, stride, 1, 8, thresh, ithresh, hev_thresh);
   FilterLoop26(v, stride, 1, 8, thresh, ithresh, hev_thresh);
 }
 
-static void HFilter8(uint8_t* u, uint8_t* v, int stride,
-                     int thresh, int ithresh, int hev_thresh) {
+static void HFilter8(uint8_t* WEBP_RESTRICT u, uint8_t* WEBP_RESTRICT v,
+                     int stride, int thresh, int ithresh, int hev_thresh) {
   FilterLoop26(u, 1, stride, 8, thresh, ithresh, hev_thresh);
   FilterLoop26(v, 1, stride, 8, thresh, ithresh, hev_thresh);
 }
@@ -465,14 +469,14 @@ static void HFilter16i(uint8_t* p, int stride,
   }
 }
 
-static void VFilter8i(uint8_t* u, uint8_t* v, int stride,
-                      int thresh, int ithresh, int hev_thresh) {
+static void VFilter8i(uint8_t* WEBP_RESTRICT u, uint8_t* WEBP_RESTRICT v,
+                      int stride, int thresh, int ithresh, int hev_thresh) {
   FilterLoop24(u + 4 * stride, stride, 1, 8, thresh, ithresh, hev_thresh);
   FilterLoop24(v + 4 * stride, stride, 1, 8, thresh, ithresh, hev_thresh);
 }
 
-static void HFilter8i(uint8_t* u, uint8_t* v, int stride,
-                      int thresh, int ithresh, int hev_thresh) {
+static void HFilter8i(uint8_t* WEBP_RESTRICT u, uint8_t* WEBP_RESTRICT v,
+                      int stride, int thresh, int ithresh, int hev_thresh) {
   FilterLoop24(u + 4, 1, stride, 8, thresh, ithresh, hev_thresh);
   FilterLoop24(v + 4, 1, stride, 8, thresh, ithresh, hev_thresh);
 }
diff --git a/src/dsp/dec_msa.c b/src/dsp/dec_msa.c
index 58d17301..422b3632 100644
--- a/src/dsp/dec_msa.c
+++ b/src/dsp/dec_msa.c
@@ -38,7 +38,8 @@
   BUTTERFLY_4(a1_m, b1_m, c1_m, d1_m, out0, out1, out2, out3);   \
 }
 
-static void TransformOne(const int16_t* in, uint8_t* dst) {
+static void TransformOne(const int16_t* WEBP_RESTRICT in,
+                         uint8_t* WEBP_RESTRICT dst) {
   v8i16 input0, input1;
   v4i32 in0, in1, in2, in3, hz0, hz1, hz2, hz3, vt0, vt1, vt2, vt3;
   v4i32 res0, res1, res2, res3;
@@ -65,14 +66,16 @@ static void TransformOne(const int16_t* in, uint8_t* dst) {
   ST4x4_UB(res0, res0, 3, 2, 1, 0, dst, BPS);
 }
 
-static void TransformTwo(const int16_t* in, uint8_t* dst, int do_two) {
+static void TransformTwo(const int16_t* WEBP_RESTRICT in,
+                         uint8_t* WEBP_RESTRICT dst, int do_two) {
   TransformOne(in, dst);
   if (do_two) {
     TransformOne(in + 16, dst + 4);
   }
 }
 
-static void TransformWHT(const int16_t* in, int16_t* out) {
+static void TransformWHT(const int16_t* WEBP_RESTRICT in,
+                         int16_t* WEBP_RESTRICT out) {
   v8i16 input0, input1;
   const v8i16 mask0 = { 0, 1, 2, 3, 8, 9, 10, 11 };
   const v8i16 mask1 = { 4, 5, 6, 7, 12, 13, 14, 15 };
@@ -114,13 +117,15 @@ static void TransformWHT(const int16_t* in, int16_t* out) {
   out[240] = __msa_copy_s_h(out1, 7);
 }
 
-static void TransformDC(const int16_t* in, uint8_t* dst) {
+static void TransformDC(const int16_t* WEBP_RESTRICT in,
+                        uint8_t* WEBP_RESTRICT dst) {
   const int DC = (in[0] + 4) >> 3;
   const v8i16 tmp0 = __msa_fill_h(DC);
   ADDBLK_ST4x4_UB(tmp0, tmp0, tmp0, tmp0, dst, BPS);
 }
 
-static void TransformAC3(const int16_t* in, uint8_t* dst) {
+static void TransformAC3(const int16_t* WEBP_RESTRICT in,
+                         uint8_t* WEBP_RESTRICT dst) {
   const int a = in[0] + 4;
   const int c4 = WEBP_TRANSFORM_AC3_MUL2(in[4]);
   const int d4 = WEBP_TRANSFORM_AC3_MUL1(in[4]);
@@ -475,8 +480,8 @@ static void HFilter16i(uint8_t* src_y, int stride,
 }
 
 // 8-pixels wide variants, for chroma filtering
-static void VFilter8(uint8_t* src_u, uint8_t* src_v, int stride,
-                     int b_limit_in, int limit_in, int thresh_in) {
+static void VFilter8(uint8_t* WEBP_RESTRICT src_u, uint8_t* WEBP_RESTRICT src_v,
+                     int stride, int b_limit_in, int limit_in, int thresh_in) {
   uint8_t* ptmp_src_u = src_u - 4 * stride;
   uint8_t* ptmp_src_v = src_v - 4 * stride;
   uint64_t p2_d, p1_d, p0_d, q0_d, q1_d, q2_d;
@@ -520,8 +525,8 @@ static void VFilter8(uint8_t* src_u, uint8_t* src_v, int stride,
   SD(q2_d, ptmp_src_v);
 }
 
-static void HFilter8(uint8_t* src_u, uint8_t* src_v, int stride,
-                     int b_limit_in, int limit_in, int thresh_in) {
+static void HFilter8(uint8_t* WEBP_RESTRICT src_u, uint8_t* WEBP_RESTRICT src_v,
+                     int stride, int b_limit_in, int limit_in, int thresh_in) {
   uint8_t* ptmp_src_u = src_u - 4;
   uint8_t* ptmp_src_v = src_v - 4;
   v16u8 p3, p2, p1, p0, q3, q2, q1, q0, mask, hev;
@@ -556,7 +561,8 @@ static void HFilter8(uint8_t* src_u, uint8_t* src_v, int stride,
   ST6x4_UB(tmp7, 0, tmp5, 4, ptmp_src_v, stride);
 }
 
-static void VFilter8i(uint8_t* src_u, uint8_t* src_v, int stride,
+static void VFilter8i(uint8_t* WEBP_RESTRICT src_u,
+                      uint8_t* WEBP_RESTRICT src_v, int stride,
                       int b_limit_in, int limit_in, int thresh_in) {
   uint64_t p1_d, p0_d, q0_d, q1_d;
   v16u8 p3, p2, p1, p0, q3, q2, q1, q0, mask, hev;
@@ -587,7 +593,8 @@ static void VFilter8i(uint8_t* src_u, uint8_t* src_v, int stride,
   SD4(q1_d, q0_d, p0_d, p1_d, src_v, -stride);
 }
 
-static void HFilter8i(uint8_t* src_u, uint8_t* src_v, int stride,
+static void HFilter8i(uint8_t* WEBP_RESTRICT src_u,
+                      uint8_t* WEBP_RESTRICT src_v, int stride,
                       int b_limit_in, int limit_in, int thresh_in) {
   v16u8 p3, p2, p1, p0, q3, q2, q1, q0, mask, hev;
   v16u8 row0, row1, row2, row3, row4, row5, row6, row7, row8;
diff --git a/src/dsp/dec_neon.c b/src/dsp/dec_neon.c
index b4c46f70..f150692a 100644
--- a/src/dsp/dec_neon.c
+++ b/src/dsp/dec_neon.c
@@ -916,8 +916,8 @@ static void HFilter16i_NEON(uint8_t* p, int stride,
 #endif  // !WORK_AROUND_GCC
 
 // 8-pixels wide variant, for chroma filtering
-static void VFilter8_NEON(uint8_t* u, uint8_t* v, int stride,
-                          int thresh, int ithresh, int hev_thresh) {
+static void VFilter8_NEON(uint8_t* WEBP_RESTRICT u, uint8_t* WEBP_RESTRICT v,
+                          int stride, int thresh, int ithresh, int hev_thresh) {
   uint8x16_t p3, p2, p1, p0, q0, q1, q2, q3;
   Load8x8x2_NEON(u, v, stride, &p3, &p2, &p1, &p0, &q0, &q1, &q2, &q3);
   {
@@ -932,7 +932,8 @@ static void VFilter8_NEON(uint8_t* u, uint8_t* v, int stride,
     Store8x2x2_NEON(oq1, oq2, u + 2 * stride, v + 2 * stride, stride);
   }
 }
-static void VFilter8i_NEON(uint8_t* u, uint8_t* v, int stride,
+static void VFilter8i_NEON(uint8_t* WEBP_RESTRICT u, uint8_t* WEBP_RESTRICT v,
+                           int stride,
                            int thresh, int ithresh, int hev_thresh) {
   uint8x16_t p3, p2, p1, p0, q0, q1, q2, q3;
   u += 4 * stride;
@@ -949,8 +950,8 @@ static void VFilter8i_NEON(uint8_t* u, uint8_t* v, int stride,
 }
 
 #if !defined(WORK_AROUND_GCC)
-static void HFilter8_NEON(uint8_t* u, uint8_t* v, int stride,
-                          int thresh, int ithresh, int hev_thresh) {
+static void HFilter8_NEON(uint8_t* WEBP_RESTRICT u, uint8_t* WEBP_RESTRICT v,
+                          int stride, int thresh, int ithresh, int hev_thresh) {
   uint8x16_t p3, p2, p1, p0, q0, q1, q2, q3;
   Load8x8x2T_NEON(u, v, stride, &p3, &p2, &p1, &p0, &q0, &q1, &q2, &q3);
   {
@@ -964,7 +965,8 @@ static void HFilter8_NEON(uint8_t* u, uint8_t* v, int stride,
   }
 }
 
-static void HFilter8i_NEON(uint8_t* u, uint8_t* v, int stride,
+static void HFilter8i_NEON(uint8_t* WEBP_RESTRICT u, uint8_t* WEBP_RESTRICT v,
+                           int stride,
                            int thresh, int ithresh, int hev_thresh) {
   uint8x16_t p3, p2, p1, p0, q0, q1, q2, q3;
   u += 4;
@@ -1041,7 +1043,8 @@ static WEBP_INLINE void TransformPass_NEON(int16x8x2_t* const rows) {
   Transpose8x2_NEON(E0, E1, rows);
 }
 
-static void TransformOne_NEON(const int16_t* in, uint8_t* dst) {
+static void TransformOne_NEON(const int16_t* WEBP_RESTRICT in,
+                              uint8_t* WEBP_RESTRICT dst) {
   int16x8x2_t rows;
   INIT_VECTOR2(rows, vld1q_s16(in + 0), vld1q_s16(in + 8));
   TransformPass_NEON(&rows);
@@ -1051,7 +1054,8 @@ static void TransformOne_NEON(const int16_t* in, uint8_t* dst) {
 
 #else
 
-static void TransformOne_NEON(const int16_t* in, uint8_t* dst) {
+static void TransformOne_NEON(const int16_t* WEBP_RESTRICT in,
+                              uint8_t* WEBP_RESTRICT dst) {
   const int kBPS = BPS;
   // kC1, kC2. Padded because vld1.16 loads 8 bytes
   const int16_t constants[4] = { kC1, kC2, 0, 0 };
@@ -1184,14 +1188,16 @@ static void TransformOne_NEON(const int16_t* in, uint8_t* dst) {
 
 #endif    // WEBP_USE_INTRINSICS
 
-static void TransformTwo_NEON(const int16_t* in, uint8_t* dst, int do_two) {
+static void TransformTwo_NEON(const int16_t* WEBP_RESTRICT in,
+                              uint8_t* WEBP_RESTRICT dst, int do_two) {
   TransformOne_NEON(in, dst);
   if (do_two) {
     TransformOne_NEON(in + 16, dst + 4);
   }
 }
 
-static void TransformDC_NEON(const int16_t* in, uint8_t* dst) {
+static void TransformDC_NEON(const int16_t* WEBP_RESTRICT in,
+                             uint8_t* WEBP_RESTRICT dst) {
   const int16x8_t DC = vdupq_n_s16(in[0]);
   Add4x4_NEON(DC, DC, dst);
 }
@@ -1205,7 +1211,8 @@ static void TransformDC_NEON(const int16_t* in, uint8_t* dst) {
   *dst = vgetq_lane_s32(rows.val[3], col); (dst) += 16; \
 } while (0)
 
-static void TransformWHT_NEON(const int16_t* in, int16_t* out) {
+static void TransformWHT_NEON(const int16_t* WEBP_RESTRICT in,
+                              int16_t* WEBP_RESTRICT out) {
   int32x4x4_t tmp;
 
   {
@@ -1256,7 +1263,8 @@ static void TransformWHT_NEON(const int16_t* in, int16_t* out) {
 
 //------------------------------------------------------------------------------
 
-static void TransformAC3_NEON(const int16_t* in, uint8_t* dst) {
+static void TransformAC3_NEON(const int16_t* WEBP_RESTRICT in,
+                              uint8_t* WEBP_RESTRICT dst) {
   const int16x4_t A = vld1_dup_s16(in);
   const int16x4_t c4 = vdup_n_s16(WEBP_TRANSFORM_AC3_MUL2(in[4]));
   const int16x4_t d4 = vdup_n_s16(WEBP_TRANSFORM_AC3_MUL1(in[4]));
diff --git a/src/dsp/dec_sse2.c b/src/dsp/dec_sse2.c
index 5fd85e6f..b0faada8 100644
--- a/src/dsp/dec_sse2.c
+++ b/src/dsp/dec_sse2.c
@@ -30,7 +30,8 @@
 //------------------------------------------------------------------------------
 // Transforms (Paragraph 14.4)
 
-static void Transform_SSE2(const int16_t* in, uint8_t* dst, int do_two) {
+static void Transform_SSE2(const int16_t* WEBP_RESTRICT in,
+                           uint8_t* WEBP_RESTRICT dst, int do_two) {
   // This implementation makes use of 16-bit fixed point versions of two
   // multiply constants:
   //    K1 = sqrt(2) * cos (pi/8) ~= 85627 / 2^16
@@ -197,7 +198,8 @@ static void Transform_SSE2(const int16_t* in, uint8_t* dst, int do_two) {
 
 #if (USE_TRANSFORM_AC3 == 1)
 
-static void TransformAC3_SSE2(const int16_t* in, uint8_t* dst) {
+static void TransformAC3_SSE2(const int16_t* WEBP_RESTRICT in,
+                              uint8_t* WEBP_RESTRICT dst) {
   const __m128i A = _mm_set1_epi16(in[0] + 4);
   const __m128i c4 = _mm_set1_epi16(WEBP_TRANSFORM_AC3_MUL2(in[4]));
   const __m128i d4 = _mm_set1_epi16(WEBP_TRANSFORM_AC3_MUL1(in[4]));
@@ -792,8 +794,8 @@ static void HFilter16i_SSE2(uint8_t* p, int stride,
 }
 
 // 8-pixels wide variant, for chroma filtering
-static void VFilter8_SSE2(uint8_t* u, uint8_t* v, int stride,
-                          int thresh, int ithresh, int hev_thresh) {
+static void VFilter8_SSE2(uint8_t* WEBP_RESTRICT u, uint8_t* WEBP_RESTRICT v,
+                          int stride, int thresh, int ithresh, int hev_thresh) {
   __m128i mask;
   __m128i t1, p2, p1, p0, q0, q1, q2;
 
@@ -817,8 +819,8 @@ static void VFilter8_SSE2(uint8_t* u, uint8_t* v, int stride,
   STOREUV(q2, u, v, 2 * stride);
 }
 
-static void HFilter8_SSE2(uint8_t* u, uint8_t* v, int stride,
-                          int thresh, int ithresh, int hev_thresh) {
+static void HFilter8_SSE2(uint8_t* WEBP_RESTRICT u, uint8_t* WEBP_RESTRICT v,
+                          int stride, int thresh, int ithresh, int hev_thresh) {
   __m128i mask;
   __m128i p3, p2, p1, p0, q0, q1, q2, q3;
 
@@ -837,7 +839,8 @@ static void HFilter8_SSE2(uint8_t* u, uint8_t* v, int stride,
   Store16x4_SSE2(&q0, &q1, &q2, &q3, u, v, stride);
 }
 
-static void VFilter8i_SSE2(uint8_t* u, uint8_t* v, int stride,
+static void VFilter8i_SSE2(uint8_t* WEBP_RESTRICT u, uint8_t* WEBP_RESTRICT v,
+                           int stride,
                            int thresh, int ithresh, int hev_thresh) {
   __m128i mask;
   __m128i t1, t2, p1, p0, q0, q1;
@@ -863,7 +866,8 @@ static void VFilter8i_SSE2(uint8_t* u, uint8_t* v, int stride,
   STOREUV(q1, u, v, 1 * stride);
 }
 
-static void HFilter8i_SSE2(uint8_t* u, uint8_t* v, int stride,
+static void HFilter8i_SSE2(uint8_t* WEBP_RESTRICT u, uint8_t* WEBP_RESTRICT v,
+                           int stride,
                            int thresh, int ithresh, int hev_thresh) {
   __m128i mask;
   __m128i t1, t2, p1, p0, q0, q1;
diff --git a/src/dsp/dsp.h b/src/dsp/dsp.h
index 82029533..bd126dea 100644
--- a/src/dsp/dsp.h
+++ b/src/dsp/dsp.h
@@ -63,11 +63,15 @@ extern "C" {
 typedef void (*VP8Idct)(const uint8_t* ref, const int16_t* in, uint8_t* dst,
                         int do_two);
 typedef void (*VP8Fdct)(const uint8_t* src, const uint8_t* ref, int16_t* out);
-typedef void (*VP8WHT)(const int16_t* in, int16_t* out);
+// TODO(jzern): merge these two typedefs after the encoder functions are
+// updated to use WEBP_RESTRICT.
+typedef void (*VP8FWHT)(const int16_t* in, int16_t* out);
+typedef void (*VP8IWHT)(const int16_t* WEBP_RESTRICT in,
+                        int16_t* WEBP_RESTRICT out);
 extern VP8Idct VP8ITransform;
 extern VP8Fdct VP8FTransform;
 extern VP8Fdct VP8FTransform2;   // performs two transforms at a time
-extern VP8WHT VP8FTransformWHT;
+extern VP8FWHT VP8FTransformWHT;
 // Predictions
 // *dst is the destination block. *top and *left can be NULL.
 typedef void (*VP8IntraPreds)(uint8_t* dst, const uint8_t* left,
@@ -194,15 +198,17 @@ void VP8SSIMDspInit(void);
 //------------------------------------------------------------------------------
 // Decoding
 
-typedef void (*VP8DecIdct)(const int16_t* coeffs, uint8_t* dst);
+typedef void (*VP8DecIdct)(const int16_t* WEBP_RESTRICT coeffs,
+                           uint8_t* WEBP_RESTRICT dst);
 // when doing two transforms, coeffs is actually int16_t[2][16].
-typedef void (*VP8DecIdct2)(const int16_t* coeffs, uint8_t* dst, int do_two);
+typedef void (*VP8DecIdct2)(const int16_t* WEBP_RESTRICT coeffs,
+                            uint8_t* WEBP_RESTRICT dst, int do_two);
 extern VP8DecIdct2 VP8Transform;
 extern VP8DecIdct VP8TransformAC3;
 extern VP8DecIdct VP8TransformUV;
 extern VP8DecIdct VP8TransformDC;
 extern VP8DecIdct VP8TransformDCUV;
-extern VP8WHT VP8TransformWHT;
+extern VP8IWHT VP8TransformWHT;
 
 #define WEBP_TRANSFORM_AC3_C1 20091
 #define WEBP_TRANSFORM_AC3_C2 35468
@@ -234,7 +240,8 @@ extern VP8SimpleFilterFunc VP8SimpleHFilter16i;
 // regular filter (on both macroblock edges and inner edges)
 typedef void (*VP8LumaFilterFunc)(uint8_t* luma, int stride,
                                   int thresh, int ithresh, int hev_t);
-typedef void (*VP8ChromaFilterFunc)(uint8_t* u, uint8_t* v, int stride,
+typedef void (*VP8ChromaFilterFunc)(uint8_t* WEBP_RESTRICT u,
+                                    uint8_t* WEBP_RESTRICT v, int stride,
                                     int thresh, int ithresh, int hev_t);
 // on outer edge
 extern VP8LumaFilterFunc VP8VFilter16;
@@ -254,8 +261,8 @@ extern VP8ChromaFilterFunc VP8HFilter8i;
 #define VP8_DITHER_DESCALE_ROUNDER (1 << (VP8_DITHER_DESCALE - 1))
 #define VP8_DITHER_AMP_BITS 7
 #define VP8_DITHER_AMP_CENTER (1 << VP8_DITHER_AMP_BITS)
-extern void (*VP8DitherCombine8x8)(const uint8_t* dither, uint8_t* dst,
-                                   int dst_stride);
+extern void (*VP8DitherCombine8x8)(const uint8_t* WEBP_RESTRICT dither,
+                                   uint8_t* WEBP_RESTRICT dst, int dst_stride);
 
 // must be called before anything using the above
 void VP8DspInit(void);
diff --git a/src/dsp/enc.c b/src/dsp/enc.c
index cb44ba6f..95c623d9 100644
--- a/src/dsp/enc.c
+++ b/src/dsp/enc.c
@@ -720,7 +720,7 @@ VP8CHisto VP8CollectHistogram;
 VP8Idct VP8ITransform;
 VP8Fdct VP8FTransform;
 VP8Fdct VP8FTransform2;
-VP8WHT VP8FTransformWHT;
+VP8FWHT VP8FTransformWHT;
 VP8Intra4Preds VP8EncPredLuma4;
 VP8IntraPreds VP8EncPredLuma16;
 VP8IntraPreds VP8EncPredChroma8;

From b1cb37e65921649514b7a4fad46eb8d28dcbfb10 Mon Sep 17 00:00:00 2001
From: James Zern <jzern@google.com>
Date: Sat, 3 Jul 2021 17:59:44 -0700
Subject: [PATCH 3/8] dsp/enc*: use WEBP_RESTRICT qualifier

This allows for better vectorization of the C code, inlining of
TrueMotion_SSE2, better load usage in aarch64 and other minor
reordering with ndk r27/gcc-13/clang-16.

This only affects non-vector pointers; any vector pointers are left as a
follow up.

Change-Id: I07e9944d5c0aa5a079b22883ac5a2d649695e4a0
---
 src/dsp/dec.c             |   2 +-
 src/dsp/dsp.h             |  55 ++++++-----
 src/dsp/enc.c             | 126 +++++++++++++++---------
 src/dsp/enc_mips32.c      |  40 +++++---
 src/dsp/enc_mips_dsp_r2.c | 101 +++++++++++--------
 src/dsp/enc_msa.c         | 125 ++++++++++++++---------
 src/dsp/enc_neon.c        |  82 +++++++++-------
 src/dsp/enc_sse2.c        | 202 +++++++++++++++++++++++---------------
 src/dsp/enc_sse41.c       |  21 ++--
 9 files changed, 457 insertions(+), 297 deletions(-)

diff --git a/src/dsp/dec.c b/src/dsp/dec.c
index 51067f45..dc1a7625 100644
--- a/src/dsp/dec.c
+++ b/src/dsp/dec.c
@@ -168,7 +168,7 @@ static void TransformWHT_C(const int16_t* WEBP_RESTRICT in,
 }
 #endif  // !WEBP_NEON_OMIT_C_CODE
 
-VP8IWHT VP8TransformWHT;
+VP8WHT VP8TransformWHT;
 
 //------------------------------------------------------------------------------
 // Intra predictions
diff --git a/src/dsp/dsp.h b/src/dsp/dsp.h
index bd126dea..ec0302f2 100644
--- a/src/dsp/dsp.h
+++ b/src/dsp/dsp.h
@@ -60,57 +60,66 @@ extern "C" {
 // Transforms
 // VP8Idct: Does one of two inverse transforms. If do_two is set, the transforms
 //          will be done for (ref, in, dst) and (ref + 4, in + 16, dst + 4).
-typedef void (*VP8Idct)(const uint8_t* ref, const int16_t* in, uint8_t* dst,
-                        int do_two);
-typedef void (*VP8Fdct)(const uint8_t* src, const uint8_t* ref, int16_t* out);
-// TODO(jzern): merge these two typedefs after the encoder functions are
-// updated to use WEBP_RESTRICT.
-typedef void (*VP8FWHT)(const int16_t* in, int16_t* out);
-typedef void (*VP8IWHT)(const int16_t* WEBP_RESTRICT in,
+typedef void (*VP8Idct)(const uint8_t* WEBP_RESTRICT ref,
+                        const int16_t* WEBP_RESTRICT in,
+                        uint8_t* WEBP_RESTRICT dst, int do_two);
+typedef void (*VP8Fdct)(const uint8_t* WEBP_RESTRICT src,
+                        const uint8_t* WEBP_RESTRICT ref,
                         int16_t* WEBP_RESTRICT out);
+typedef void (*VP8WHT)(const int16_t* WEBP_RESTRICT in,
+                       int16_t* WEBP_RESTRICT out);
 extern VP8Idct VP8ITransform;
 extern VP8Fdct VP8FTransform;
 extern VP8Fdct VP8FTransform2;   // performs two transforms at a time
-extern VP8FWHT VP8FTransformWHT;
+extern VP8WHT VP8FTransformWHT;
 // Predictions
 // *dst is the destination block. *top and *left can be NULL.
-typedef void (*VP8IntraPreds)(uint8_t* dst, const uint8_t* left,
-                              const uint8_t* top);
-typedef void (*VP8Intra4Preds)(uint8_t* dst, const uint8_t* top);
+typedef void (*VP8IntraPreds)(uint8_t* WEBP_RESTRICT dst,
+                              const uint8_t* WEBP_RESTRICT left,
+                              const uint8_t* WEBP_RESTRICT top);
+typedef void (*VP8Intra4Preds)(uint8_t* WEBP_RESTRICT dst,
+                               const uint8_t* WEBP_RESTRICT top);
 extern VP8Intra4Preds VP8EncPredLuma4;
 extern VP8IntraPreds VP8EncPredLuma16;
 extern VP8IntraPreds VP8EncPredChroma8;
 
-typedef int (*VP8Metric)(const uint8_t* pix, const uint8_t* ref);
+typedef int (*VP8Metric)(const uint8_t* WEBP_RESTRICT pix,
+                         const uint8_t* WEBP_RESTRICT ref);
 extern VP8Metric VP8SSE16x16, VP8SSE16x8, VP8SSE8x8, VP8SSE4x4;
-typedef int (*VP8WMetric)(const uint8_t* pix, const uint8_t* ref,
-                          const uint16_t* const weights);
+typedef int (*VP8WMetric)(const uint8_t* WEBP_RESTRICT pix,
+                          const uint8_t* WEBP_RESTRICT ref,
+                          const uint16_t* WEBP_RESTRICT const weights);
 // The weights for VP8TDisto4x4 and VP8TDisto16x16 contain a row-major
 // 4 by 4 symmetric matrix.
 extern VP8WMetric VP8TDisto4x4, VP8TDisto16x16;
 
 // Compute the average (DC) of four 4x4 blocks.
 // Each sub-4x4 block #i sum is stored in dc[i].
-typedef void (*VP8MeanMetric)(const uint8_t* ref, uint32_t dc[4]);
+typedef void (*VP8MeanMetric)(const uint8_t* WEBP_RESTRICT ref,
+                              uint32_t dc[4]);
 extern VP8MeanMetric VP8Mean16x4;
 
-typedef void (*VP8BlockCopy)(const uint8_t* src, uint8_t* dst);
+typedef void (*VP8BlockCopy)(const uint8_t* WEBP_RESTRICT src,
+                             uint8_t* WEBP_RESTRICT dst);
 extern VP8BlockCopy VP8Copy4x4;
 extern VP8BlockCopy VP8Copy16x8;
 // Quantization
 struct VP8Matrix;   // forward declaration
-typedef int (*VP8QuantizeBlock)(int16_t in[16], int16_t out[16],
-                                const struct VP8Matrix* const mtx);
+typedef int (*VP8QuantizeBlock)(
+    int16_t in[16], int16_t out[16],
+    const struct VP8Matrix* WEBP_RESTRICT const mtx);
 // Same as VP8QuantizeBlock, but quantizes two consecutive blocks.
-typedef int (*VP8Quantize2Blocks)(int16_t in[32], int16_t out[32],
-                                  const struct VP8Matrix* const mtx);
+typedef int (*VP8Quantize2Blocks)(
+    int16_t in[32], int16_t out[32],
+    const struct VP8Matrix* WEBP_RESTRICT const mtx);
 
 extern VP8QuantizeBlock VP8EncQuantizeBlock;
 extern VP8Quantize2Blocks VP8EncQuantize2Blocks;
 
 // specific to 2nd transform:
-typedef int (*VP8QuantizeBlockWHT)(int16_t in[16], int16_t out[16],
-                                   const struct VP8Matrix* const mtx);
+typedef int (*VP8QuantizeBlockWHT)(
+    int16_t in[16], int16_t out[16],
+    const struct VP8Matrix* WEBP_RESTRICT const mtx);
 extern VP8QuantizeBlockWHT VP8EncQuantizeBlockWHT;
 
 extern const int VP8DspScan[16 + 4 + 4];
@@ -208,7 +217,7 @@ extern VP8DecIdct VP8TransformAC3;
 extern VP8DecIdct VP8TransformUV;
 extern VP8DecIdct VP8TransformDC;
 extern VP8DecIdct VP8TransformDCUV;
-extern VP8IWHT VP8TransformWHT;
+extern VP8WHT VP8TransformWHT;
 
 #define WEBP_TRANSFORM_AC3_C1 20091
 #define WEBP_TRANSFORM_AC3_C2 35468
diff --git a/src/dsp/enc.c b/src/dsp/enc.c
index 95c623d9..b177031d 100644
--- a/src/dsp/enc.c
+++ b/src/dsp/enc.c
@@ -59,9 +59,10 @@ void VP8SetHistogramData(const int distribution[MAX_COEFF_THRESH + 1],
 }
 
 #if !WEBP_NEON_OMIT_C_CODE
-static void CollectHistogram_C(const uint8_t* ref, const uint8_t* pred,
+static void CollectHistogram_C(const uint8_t* WEBP_RESTRICT ref,
+                               const uint8_t* WEBP_RESTRICT pred,
                                int start_block, int end_block,
-                               VP8Histogram* const histo) {
+                               VP8Histogram* WEBP_RESTRICT const histo) {
   int j;
   int distribution[MAX_COEFF_THRESH + 1] = { 0 };
   for (j = start_block; j < end_block; ++j) {
@@ -109,8 +110,9 @@ static WEBP_TSAN_IGNORE_FUNCTION void InitTables(void) {
 #define STORE(x, y, v) \
   dst[(x) + (y) * BPS] = clip_8b(ref[(x) + (y) * BPS] + ((v) >> 3))
 
-static WEBP_INLINE void ITransformOne(const uint8_t* ref, const int16_t* in,
-                                      uint8_t* dst) {
+static WEBP_INLINE void ITransformOne(const uint8_t* WEBP_RESTRICT ref,
+                                      const int16_t* WEBP_RESTRICT in,
+                                      uint8_t* WEBP_RESTRICT dst) {
   int C[4 * 4], *tmp;
   int i;
   tmp = C;
@@ -146,7 +148,9 @@ static WEBP_INLINE void ITransformOne(const uint8_t* ref, const int16_t* in,
   }
 }
 
-static void ITransform_C(const uint8_t* ref, const int16_t* in, uint8_t* dst,
+static void ITransform_C(const uint8_t* WEBP_RESTRICT ref,
+                         const int16_t* WEBP_RESTRICT in,
+                         uint8_t* WEBP_RESTRICT dst,
                          int do_two) {
   ITransformOne(ref, in, dst);
   if (do_two) {
@@ -154,7 +158,9 @@ static void ITransform_C(const uint8_t* ref, const int16_t* in, uint8_t* dst,
   }
 }
 
-static void FTransform_C(const uint8_t* src, const uint8_t* ref, int16_t* out) {
+static void FTransform_C(const uint8_t* WEBP_RESTRICT src,
+                         const uint8_t* WEBP_RESTRICT ref,
+                         int16_t* WEBP_RESTRICT out) {
   int i;
   int tmp[16];
   for (i = 0; i < 4; ++i, src += BPS, ref += BPS) {
@@ -184,14 +190,16 @@ static void FTransform_C(const uint8_t* src, const uint8_t* ref, int16_t* out) {
 }
 #endif  // !WEBP_NEON_OMIT_C_CODE
 
-static void FTransform2_C(const uint8_t* src, const uint8_t* ref,
-                          int16_t* out) {
+static void FTransform2_C(const uint8_t* WEBP_RESTRICT src,
+                          const uint8_t* WEBP_RESTRICT ref,
+                          int16_t* WEBP_RESTRICT out) {
   VP8FTransform(src, ref, out);
   VP8FTransform(src + 4, ref + 4, out + 16);
 }
 
 #if !WEBP_NEON_OMIT_C_CODE
-static void FTransformWHT_C(const int16_t* in, int16_t* out) {
+static void FTransformWHT_C(const int16_t* WEBP_RESTRICT in,
+                            int16_t* WEBP_RESTRICT out) {
   // input is 12b signed
   int32_t tmp[16];
   int i;
@@ -234,8 +242,9 @@ static WEBP_INLINE void Fill(uint8_t* dst, int value, int size) {
   }
 }
 
-static WEBP_INLINE void VerticalPred(uint8_t* dst,
-                                     const uint8_t* top, int size) {
+static WEBP_INLINE void VerticalPred(uint8_t* WEBP_RESTRICT dst,
+                                     const uint8_t* WEBP_RESTRICT top,
+                                     int size) {
   int j;
   if (top != NULL) {
     for (j = 0; j < size; ++j) memcpy(dst + j * BPS, top, size);
@@ -244,8 +253,9 @@ static WEBP_INLINE void VerticalPred(uint8_t* dst,
   }
 }
 
-static WEBP_INLINE void HorizontalPred(uint8_t* dst,
-                                       const uint8_t* left, int size) {
+static WEBP_INLINE void HorizontalPred(uint8_t* WEBP_RESTRICT dst,
+                                       const uint8_t* WEBP_RESTRICT left,
+                                       int size) {
   if (left != NULL) {
     int j;
     for (j = 0; j < size; ++j) {
@@ -256,8 +266,9 @@ static WEBP_INLINE void HorizontalPred(uint8_t* dst,
   }
 }
 
-static WEBP_INLINE void TrueMotion(uint8_t* dst, const uint8_t* left,
-                                   const uint8_t* top, int size) {
+static WEBP_INLINE void TrueMotion(uint8_t* WEBP_RESTRICT dst,
+                                   const uint8_t* WEBP_RESTRICT left,
+                                   const uint8_t* WEBP_RESTRICT top, int size) {
   int y;
   if (left != NULL) {
     if (top != NULL) {
@@ -286,8 +297,9 @@ static WEBP_INLINE void TrueMotion(uint8_t* dst, const uint8_t* left,
   }
 }
 
-static WEBP_INLINE void DCMode(uint8_t* dst, const uint8_t* left,
-                               const uint8_t* top,
+static WEBP_INLINE void DCMode(uint8_t* WEBP_RESTRICT dst,
+                               const uint8_t* WEBP_RESTRICT left,
+                               const uint8_t* WEBP_RESTRICT top,
                                int size, int round, int shift) {
   int DC = 0;
   int j;
@@ -312,8 +324,9 @@ static WEBP_INLINE void DCMode(uint8_t* dst, const uint8_t* left,
 //------------------------------------------------------------------------------
 // Chroma 8x8 prediction (paragraph 12.2)
 
-static void IntraChromaPreds_C(uint8_t* dst, const uint8_t* left,
-                               const uint8_t* top) {
+static void IntraChromaPreds_C(uint8_t* WEBP_RESTRICT dst,
+                               const uint8_t* WEBP_RESTRICT left,
+                               const uint8_t* WEBP_RESTRICT top) {
   // U block
   DCMode(C8DC8 + dst, left, top, 8, 8, 4);
   VerticalPred(C8VE8 + dst, top, 8);
@@ -333,8 +346,9 @@ static void IntraChromaPreds_C(uint8_t* dst, const uint8_t* left,
 // luma 16x16 prediction (paragraph 12.3)
 
 #if !WEBP_NEON_OMIT_C_CODE || !WEBP_AARCH64
-static void Intra16Preds_C(uint8_t* dst,
-                           const uint8_t* left, const uint8_t* top) {
+static void Intra16Preds_C(uint8_t* WEBP_RESTRICT dst,
+                           const uint8_t* WEBP_RESTRICT left,
+                           const uint8_t* WEBP_RESTRICT top) {
   DCMode(I16DC16 + dst, left, top, 16, 16, 5);
   VerticalPred(I16VE16 + dst, top, 16);
   HorizontalPred(I16HE16 + dst, left, 16);
@@ -351,7 +365,8 @@ static void Intra16Preds_C(uint8_t* dst,
 #define AVG3(a, b, c) ((uint8_t)(((a) + 2 * (b) + (c) + 2) >> 2))
 #define AVG2(a, b) (((a) + (b) + 1) >> 1)
 
-static void VE4(uint8_t* dst, const uint8_t* top) {    // vertical
+// vertical
+static void VE4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
   const uint8_t vals[4] = {
     AVG3(top[-1], top[0], top[1]),
     AVG3(top[ 0], top[1], top[2]),
@@ -364,7 +379,8 @@ static void VE4(uint8_t* dst, const uint8_t* top) {    // vertical
   }
 }
 
-static void HE4(uint8_t* dst, const uint8_t* top) {    // horizontal
+// horizontal
+static void HE4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
   const int X = top[-1];
   const int I = top[-2];
   const int J = top[-3];
@@ -376,14 +392,14 @@ static void HE4(uint8_t* dst, const uint8_t* top) {    // horizontal
   WebPUint32ToMem(dst + 3 * BPS, 0x01010101U * AVG3(K, L, L));
 }
 
-static void DC4(uint8_t* dst, const uint8_t* top) {
+static void DC4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
   uint32_t dc = 4;
   int i;
   for (i = 0; i < 4; ++i) dc += top[i] + top[-5 + i];
   Fill(dst, dc >> 3, 4);
 }
 
-static void RD4(uint8_t* dst, const uint8_t* top) {
+static void RD4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
   const int X = top[-1];
   const int I = top[-2];
   const int J = top[-3];
@@ -402,7 +418,7 @@ static void RD4(uint8_t* dst, const uint8_t* top) {
   DST(3, 0)                                     = AVG3(D, C, B);
 }
 
-static void LD4(uint8_t* dst, const uint8_t* top) {
+static void LD4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
   const int A = top[0];
   const int B = top[1];
   const int C = top[2];
@@ -420,7 +436,7 @@ static void LD4(uint8_t* dst, const uint8_t* top) {
   DST(3, 3)                                     = AVG3(G, H, H);
 }
 
-static void VR4(uint8_t* dst, const uint8_t* top) {
+static void VR4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
   const int X = top[-1];
   const int I = top[-2];
   const int J = top[-3];
@@ -442,7 +458,7 @@ static void VR4(uint8_t* dst, const uint8_t* top) {
   DST(3, 1) =             AVG3(B, C, D);
 }
 
-static void VL4(uint8_t* dst, const uint8_t* top) {
+static void VL4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
   const int A = top[0];
   const int B = top[1];
   const int C = top[2];
@@ -464,7 +480,7 @@ static void VL4(uint8_t* dst, const uint8_t* top) {
               DST(3, 3) = AVG3(F, G, H);
 }
 
-static void HU4(uint8_t* dst, const uint8_t* top) {
+static void HU4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
   const int I = top[-2];
   const int J = top[-3];
   const int K = top[-4];
@@ -479,7 +495,7 @@ static void HU4(uint8_t* dst, const uint8_t* top) {
   DST(0, 3) = DST(1, 3) = DST(2, 3) = DST(3, 3) = L;
 }
 
-static void HD4(uint8_t* dst, const uint8_t* top) {
+static void HD4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
   const int X = top[-1];
   const int I = top[-2];
   const int J = top[-3];
@@ -502,7 +518,7 @@ static void HD4(uint8_t* dst, const uint8_t* top) {
   DST(1, 3)             = AVG3(L, K, J);
 }
 
-static void TM4(uint8_t* dst, const uint8_t* top) {
+static void TM4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
   int x, y;
   const uint8_t* const clip = clip1 + 255 - top[-1];
   for (y = 0; y < 4; ++y) {
@@ -520,7 +536,8 @@ static void TM4(uint8_t* dst, const uint8_t* top) {
 
 // Left samples are top[-5 .. -2], top_left is top[-1], top are
 // located at top[0..3], and top right is top[4..7]
-static void Intra4Preds_C(uint8_t* dst, const uint8_t* top) {
+static void Intra4Preds_C(uint8_t* WEBP_RESTRICT dst,
+                          const uint8_t* WEBP_RESTRICT top) {
   DC4(I4DC4 + dst, top);
   TM4(I4TM4 + dst, top);
   VE4(I4VE4 + dst, top);
@@ -539,7 +556,8 @@ static void Intra4Preds_C(uint8_t* dst, const uint8_t* top) {
 // Metric
 
 #if !WEBP_NEON_OMIT_C_CODE
-static WEBP_INLINE int GetSSE(const uint8_t* a, const uint8_t* b,
+static WEBP_INLINE int GetSSE(const uint8_t* WEBP_RESTRICT a,
+                              const uint8_t* WEBP_RESTRICT b,
                               int w, int h) {
   int count = 0;
   int y, x;
@@ -554,21 +572,25 @@ static WEBP_INLINE int GetSSE(const uint8_t* a, const uint8_t* b,
   return count;
 }
 
-static int SSE16x16_C(const uint8_t* a, const uint8_t* b) {
+static int SSE16x16_C(const uint8_t* WEBP_RESTRICT a,
+                      const uint8_t* WEBP_RESTRICT b) {
   return GetSSE(a, b, 16, 16);
 }
-static int SSE16x8_C(const uint8_t* a, const uint8_t* b) {
+static int SSE16x8_C(const uint8_t* WEBP_RESTRICT a,
+                     const uint8_t* WEBP_RESTRICT b) {
   return GetSSE(a, b, 16, 8);
 }
-static int SSE8x8_C(const uint8_t* a, const uint8_t* b) {
+static int SSE8x8_C(const uint8_t* WEBP_RESTRICT a,
+                    const uint8_t* WEBP_RESTRICT b) {
   return GetSSE(a, b, 8, 8);
 }
-static int SSE4x4_C(const uint8_t* a, const uint8_t* b) {
+static int SSE4x4_C(const uint8_t* WEBP_RESTRICT a,
+                    const uint8_t* WEBP_RESTRICT b) {
   return GetSSE(a, b, 4, 4);
 }
 #endif  // !WEBP_NEON_OMIT_C_CODE
 
-static void Mean16x4_C(const uint8_t* ref, uint32_t dc[4]) {
+static void Mean16x4_C(const uint8_t* WEBP_RESTRICT ref, uint32_t dc[4]) {
   int k, x, y;
   for (k = 0; k < 4; ++k) {
     uint32_t avg = 0;
@@ -592,7 +614,8 @@ static void Mean16x4_C(const uint8_t* ref, uint32_t dc[4]) {
 // Hadamard transform
 // Returns the weighted sum of the absolute value of transformed coefficients.
 // w[] contains a row-major 4 by 4 symmetric matrix.
-static int TTransform(const uint8_t* in, const uint16_t* w) {
+static int TTransform(const uint8_t* WEBP_RESTRICT in,
+                      const uint16_t* WEBP_RESTRICT w) {
   int sum = 0;
   int tmp[16];
   int i;
@@ -626,15 +649,17 @@ static int TTransform(const uint8_t* in, const uint16_t* w) {
   return sum;
 }
 
-static int Disto4x4_C(const uint8_t* const a, const uint8_t* const b,
-                      const uint16_t* const w) {
+static int Disto4x4_C(const uint8_t* WEBP_RESTRICT const a,
+                      const uint8_t* WEBP_RESTRICT const b,
+                      const uint16_t* WEBP_RESTRICT const w) {
   const int sum1 = TTransform(a, w);
   const int sum2 = TTransform(b, w);
   return abs(sum2 - sum1) >> 5;
 }
 
-static int Disto16x16_C(const uint8_t* const a, const uint8_t* const b,
-                        const uint16_t* const w) {
+static int Disto16x16_C(const uint8_t* WEBP_RESTRICT const a,
+                        const uint8_t* WEBP_RESTRICT const b,
+                        const uint16_t* WEBP_RESTRICT const w) {
   int D = 0;
   int x, y;
   for (y = 0; y < 16 * BPS; y += 4 * BPS) {
@@ -657,7 +682,7 @@ static const uint8_t kZigzag[16] = {
 
 // Simple quantization
 static int QuantizeBlock_C(int16_t in[16], int16_t out[16],
-                           const VP8Matrix* const mtx) {
+                           const VP8Matrix* WEBP_RESTRICT const mtx) {
   int last = -1;
   int n;
   for (n = 0; n < 16; ++n) {
@@ -683,7 +708,7 @@ static int QuantizeBlock_C(int16_t in[16], int16_t out[16],
 }
 
 static int Quantize2Blocks_C(int16_t in[32], int16_t out[32],
-                             const VP8Matrix* const mtx) {
+                             const VP8Matrix* WEBP_RESTRICT const mtx) {
   int nz;
   nz  = VP8EncQuantizeBlock(in + 0 * 16, out + 0 * 16, mtx) << 0;
   nz |= VP8EncQuantizeBlock(in + 1 * 16, out + 1 * 16, mtx) << 1;
@@ -694,7 +719,8 @@ static int Quantize2Blocks_C(int16_t in[32], int16_t out[32],
 //------------------------------------------------------------------------------
 // Block copy
 
-static WEBP_INLINE void Copy(const uint8_t* src, uint8_t* dst, int w, int h) {
+static WEBP_INLINE void Copy(const uint8_t* WEBP_RESTRICT src,
+                             uint8_t* WEBP_RESTRICT dst, int w, int h) {
   int y;
   for (y = 0; y < h; ++y) {
     memcpy(dst, src, w);
@@ -703,11 +729,13 @@ static WEBP_INLINE void Copy(const uint8_t* src, uint8_t* dst, int w, int h) {
   }
 }
 
-static void Copy4x4_C(const uint8_t* src, uint8_t* dst) {
+static void Copy4x4_C(const uint8_t* WEBP_RESTRICT src,
+                      uint8_t* WEBP_RESTRICT dst) {
   Copy(src, dst, 4, 4);
 }
 
-static void Copy16x8_C(const uint8_t* src, uint8_t* dst) {
+static void Copy16x8_C(const uint8_t* WEBP_RESTRICT src,
+                       uint8_t* WEBP_RESTRICT dst) {
   Copy(src, dst, 16, 8);
 }
 
@@ -720,7 +748,7 @@ VP8CHisto VP8CollectHistogram;
 VP8Idct VP8ITransform;
 VP8Fdct VP8FTransform;
 VP8Fdct VP8FTransform2;
-VP8FWHT VP8FTransformWHT;
+VP8WHT VP8FTransformWHT;
 VP8Intra4Preds VP8EncPredLuma4;
 VP8IntraPreds VP8EncPredLuma16;
 VP8IntraPreds VP8EncPredChroma8;
diff --git a/src/dsp/enc_mips32.c b/src/dsp/enc_mips32.c
index 50518a5f..6cd8c93d 100644
--- a/src/dsp/enc_mips32.c
+++ b/src/dsp/enc_mips32.c
@@ -109,9 +109,9 @@ static const int kC2 = WEBP_TRANSFORM_AC3_C2;
   "sb      %[" #TEMP12 "],   3+" XSTR(BPS) "*" #A "(%[temp16]) \n\t"
 
 // Does one or two inverse transforms.
-static WEBP_INLINE void ITransformOne_MIPS32(const uint8_t* ref,
-                                             const int16_t* in,
-                                             uint8_t* dst) {
+static WEBP_INLINE void ITransformOne_MIPS32(const uint8_t* WEBP_RESTRICT ref,
+                                             const int16_t* WEBP_RESTRICT in,
+                                             uint8_t* WEBP_RESTRICT dst) {
   int temp0, temp1, temp2, temp3, temp4, temp5, temp6;
   int temp7, temp8, temp9, temp10, temp11, temp12, temp13;
   int temp14, temp15, temp16, temp17, temp18, temp19, temp20;
@@ -141,8 +141,9 @@ static WEBP_INLINE void ITransformOne_MIPS32(const uint8_t* ref,
   );
 }
 
-static void ITransform_MIPS32(const uint8_t* ref, const int16_t* in,
-                              uint8_t* dst, int do_two) {
+static void ITransform_MIPS32(const uint8_t* WEBP_RESTRICT ref,
+                              const int16_t* WEBP_RESTRICT in,
+                              uint8_t* WEBP_RESTRICT dst, int do_two) {
   ITransformOne_MIPS32(ref, in, dst);
   if (do_two) {
     ITransformOne_MIPS32(ref + 4, in + 16, dst + 4);
@@ -236,7 +237,7 @@ static int QuantizeBlock_MIPS32(int16_t in[16], int16_t out[16],
 }
 
 static int Quantize2Blocks_MIPS32(int16_t in[32], int16_t out[32],
-                                  const VP8Matrix* const mtx) {
+                                  const VP8Matrix* WEBP_RESTRICT const mtx) {
   int nz;
   nz  = QuantizeBlock_MIPS32(in + 0 * 16, out + 0 * 16, mtx) << 0;
   nz |= QuantizeBlock_MIPS32(in + 1 * 16, out + 1 * 16, mtx) << 1;
@@ -358,8 +359,9 @@ static int Quantize2Blocks_MIPS32(int16_t in[32], int16_t out[32],
   "msub   %[temp6],  %[temp0]                \n\t"                \
   "msub   %[temp7],  %[temp1]                \n\t"
 
-static int Disto4x4_MIPS32(const uint8_t* const a, const uint8_t* const b,
-                           const uint16_t* const w) {
+static int Disto4x4_MIPS32(const uint8_t* WEBP_RESTRICT const a,
+                           const uint8_t* WEBP_RESTRICT const b,
+                           const uint16_t* WEBP_RESTRICT const w) {
   int tmp[32];
   int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8;
 
@@ -393,8 +395,9 @@ static int Disto4x4_MIPS32(const uint8_t* const a, const uint8_t* const b,
 #undef VERTICAL_PASS
 #undef HORIZONTAL_PASS
 
-static int Disto16x16_MIPS32(const uint8_t* const a, const uint8_t* const b,
-                             const uint16_t* const w) {
+static int Disto16x16_MIPS32(const uint8_t* WEBP_RESTRICT const a,
+                             const uint8_t* WEBP_RESTRICT const b,
+                             const uint16_t* WEBP_RESTRICT const w) {
   int D = 0;
   int x, y;
   for (y = 0; y < 16 * BPS; y += 4 * BPS) {
@@ -475,8 +478,9 @@ static int Disto16x16_MIPS32(const uint8_t* const a, const uint8_t* const b,
   "sh     %[" #TEMP8 "],  " #D "(%[temp20])              \n\t"    \
   "sh     %[" #TEMP12 "], " #B "(%[temp20])              \n\t"
 
-static void FTransform_MIPS32(const uint8_t* src, const uint8_t* ref,
-                              int16_t* out) {
+static void FTransform_MIPS32(const uint8_t* WEBP_RESTRICT src,
+                              const uint8_t* WEBP_RESTRICT ref,
+                              int16_t* WEBP_RESTRICT out) {
   int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8;
   int temp9, temp10, temp11, temp12, temp13, temp14, temp15, temp16;
   int temp17, temp18, temp19, temp20;
@@ -537,7 +541,8 @@ static void FTransform_MIPS32(const uint8_t* src, const uint8_t* ref,
   GET_SSE_INNER(C, C + 1, C + 2, C + 3)   \
   GET_SSE_INNER(D, D + 1, D + 2, D + 3)
 
-static int SSE16x16_MIPS32(const uint8_t* a, const uint8_t* b) {
+static int SSE16x16_MIPS32(const uint8_t* WEBP_RESTRICT a,
+                           const uint8_t* WEBP_RESTRICT b) {
   int count;
   int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
 
@@ -571,7 +576,8 @@ static int SSE16x16_MIPS32(const uint8_t* a, const uint8_t* b) {
   return count;
 }
 
-static int SSE16x8_MIPS32(const uint8_t* a, const uint8_t* b) {
+static int SSE16x8_MIPS32(const uint8_t* WEBP_RESTRICT a,
+                          const uint8_t* WEBP_RESTRICT b) {
   int count;
   int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
 
@@ -597,7 +603,8 @@ static int SSE16x8_MIPS32(const uint8_t* a, const uint8_t* b) {
   return count;
 }
 
-static int SSE8x8_MIPS32(const uint8_t* a, const uint8_t* b) {
+static int SSE8x8_MIPS32(const uint8_t* WEBP_RESTRICT a,
+                         const uint8_t* WEBP_RESTRICT b) {
   int count;
   int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
 
@@ -619,7 +626,8 @@ static int SSE8x8_MIPS32(const uint8_t* a, const uint8_t* b) {
   return count;
 }
 
-static int SSE4x4_MIPS32(const uint8_t* a, const uint8_t* b) {
+static int SSE4x4_MIPS32(const uint8_t* WEBP_RESTRICT a,
+                         const uint8_t* WEBP_RESTRICT b) {
   int count;
   int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
 
diff --git a/src/dsp/enc_mips_dsp_r2.c b/src/dsp/enc_mips_dsp_r2.c
index e1431f3b..4d808960 100644
--- a/src/dsp/enc_mips_dsp_r2.c
+++ b/src/dsp/enc_mips_dsp_r2.c
@@ -141,8 +141,9 @@ static const int kC2 = WEBP_TRANSFORM_AC3_C2;
   "sh              %[" #TEMP8 "],   " #D "(%[temp20])               \n\t"      \
   "sh              %[" #TEMP12 "],  " #B "(%[temp20])               \n\t"
 
-static void FTransform_MIPSdspR2(const uint8_t* src, const uint8_t* ref,
-                                 int16_t* out) {
+static void FTransform_MIPSdspR2(const uint8_t* WEBP_RESTRICT src,
+                                 const uint8_t* WEBP_RESTRICT ref,
+                                 int16_t* WEBP_RESTRICT out) {
   const int c2217 = 2217;
   const int c5352 = 5352;
   int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8;
@@ -171,8 +172,9 @@ static void FTransform_MIPSdspR2(const uint8_t* src, const uint8_t* ref,
 #undef VERTICAL_PASS
 #undef HORIZONTAL_PASS
 
-static WEBP_INLINE void ITransformOne(const uint8_t* ref, const int16_t* in,
-                                      uint8_t* dst) {
+static WEBP_INLINE void ITransformOne(const uint8_t* WEBP_RESTRICT ref,
+                                      const int16_t* WEBP_RESTRICT in,
+                                      uint8_t* WEBP_RESTRICT dst) {
   int temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8, temp9;
   int temp10, temp11, temp12, temp13, temp14, temp15, temp16, temp17, temp18;
 
@@ -239,16 +241,18 @@ static WEBP_INLINE void ITransformOne(const uint8_t* ref, const int16_t* in,
   );
 }
 
-static void ITransform_MIPSdspR2(const uint8_t* ref, const int16_t* in,
-                                 uint8_t* dst, int do_two) {
+static void ITransform_MIPSdspR2(const uint8_t* WEBP_RESTRICT ref,
+                                 const int16_t* WEBP_RESTRICT in,
+                                 uint8_t* WEBP_RESTRICT dst, int do_two) {
   ITransformOne(ref, in, dst);
   if (do_two) {
     ITransformOne(ref + 4, in + 16, dst + 4);
   }
 }
 
-static int Disto4x4_MIPSdspR2(const uint8_t* const a, const uint8_t* const b,
-                              const uint16_t* const w) {
+static int Disto4x4_MIPSdspR2(const uint8_t* WEBP_RESTRICT const a,
+                              const uint8_t* WEBP_RESTRICT const b,
+                              const uint16_t* WEBP_RESTRICT const w) {
   int temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8, temp9;
   int temp10, temp11, temp12, temp13, temp14, temp15, temp16, temp17;
 
@@ -314,9 +318,9 @@ static int Disto4x4_MIPSdspR2(const uint8_t* const a, const uint8_t* const b,
   return abs(temp3 - temp17) >> 5;
 }
 
-static int Disto16x16_MIPSdspR2(const uint8_t* const a,
-                                const uint8_t* const b,
-                                const uint16_t* const w) {
+static int Disto16x16_MIPSdspR2(const uint8_t* WEBP_RESTRICT const a,
+                                const uint8_t* WEBP_RESTRICT const b,
+                                const uint16_t* WEBP_RESTRICT const w) {
   int D = 0;
   int x, y;
   for (y = 0; y < 16 * BPS; y += 4 * BPS) {
@@ -367,8 +371,8 @@ static int Disto16x16_MIPSdspR2(const uint8_t* const a,
 } while (0)
 
 #define VERTICAL_PRED(DST, TOP, SIZE)                                          \
-static WEBP_INLINE void VerticalPred##SIZE(uint8_t* (DST),                     \
-                                           const uint8_t* (TOP)) {             \
+static WEBP_INLINE void VerticalPred##SIZE(                                    \
+    uint8_t* WEBP_RESTRICT (DST), const uint8_t* WEBP_RESTRICT (TOP)) {        \
   int j;                                                                       \
   if ((TOP)) {                                                                 \
     for (j = 0; j < (SIZE); ++j) memcpy((DST) + j * BPS, (TOP), (SIZE));       \
@@ -383,8 +387,8 @@ VERTICAL_PRED(dst, top, 16)
 #undef VERTICAL_PRED
 
 #define HORIZONTAL_PRED(DST, LEFT, SIZE)                                       \
-static WEBP_INLINE void HorizontalPred##SIZE(uint8_t* (DST),                   \
-                                             const uint8_t* (LEFT)) {          \
+static WEBP_INLINE void HorizontalPred##SIZE(                                  \
+    uint8_t* WEBP_RESTRICT (DST), const uint8_t* WEBP_RESTRICT (LEFT)) {       \
   if (LEFT) {                                                                  \
     int j;                                                                     \
     for (j = 0; j < (SIZE); ++j) {                                             \
@@ -451,8 +455,9 @@ HORIZONTAL_PRED(dst, left, 16)
 } while (0)
 
 #define TRUE_MOTION(DST, LEFT, TOP, SIZE)                                      \
-static WEBP_INLINE void TrueMotion##SIZE(uint8_t* (DST), const uint8_t* (LEFT),\
-                                         const uint8_t* (TOP)) {               \
+static WEBP_INLINE void TrueMotion##SIZE(uint8_t* WEBP_RESTRICT (DST),         \
+                                         const uint8_t* WEBP_RESTRICT (LEFT),  \
+                                         const uint8_t* WEBP_RESTRICT (TOP)) { \
   if ((LEFT) != NULL) {                                                        \
     if ((TOP) != NULL) {                                                       \
       CLIP_TO_DST((DST), (LEFT), (TOP), (SIZE));                               \
@@ -480,8 +485,9 @@ TRUE_MOTION(dst, left, top, 16)
 #undef CLIP_8B_TO_DST
 #undef CLIPPING
 
-static WEBP_INLINE void DCMode16(uint8_t* dst, const uint8_t* left,
-                                 const uint8_t* top) {
+static WEBP_INLINE void DCMode16(uint8_t* WEBP_RESTRICT dst,
+                                 const uint8_t* WEBP_RESTRICT left,
+                                 const uint8_t* WEBP_RESTRICT top) {
   int DC, DC1;
   int temp0, temp1, temp2, temp3;
 
@@ -543,8 +549,9 @@ static WEBP_INLINE void DCMode16(uint8_t* dst, const uint8_t* left,
   FILL_8_OR_16(dst, DC, 16);
 }
 
-static WEBP_INLINE void DCMode8(uint8_t* dst, const uint8_t* left,
-                                const uint8_t* top) {
+static WEBP_INLINE void DCMode8(uint8_t* WEBP_RESTRICT dst,
+                                const uint8_t* WEBP_RESTRICT left,
+                                const uint8_t* WEBP_RESTRICT top) {
   int DC, DC1;
   int temp0, temp1, temp2, temp3;
 
@@ -588,7 +595,7 @@ static WEBP_INLINE void DCMode8(uint8_t* dst, const uint8_t* left,
   FILL_8_OR_16(dst, DC, 8);
 }
 
-static void DC4(uint8_t* dst, const uint8_t* top) {
+static void DC4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
   int temp0, temp1;
   __asm__ volatile(
     "ulw          %[temp0],   0(%[top])               \n\t"
@@ -609,7 +616,7 @@ static void DC4(uint8_t* dst, const uint8_t* top) {
   );
 }
 
-static void TM4(uint8_t* dst, const uint8_t* top) {
+static void TM4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
   int a10, a32, temp0, temp1, temp2, temp3, temp4, temp5;
   const int c35 = 0xff00ff;
   __asm__ volatile (
@@ -664,7 +671,7 @@ static void TM4(uint8_t* dst, const uint8_t* top) {
   );
 }
 
-static void VE4(uint8_t* dst, const uint8_t* top) {
+static void VE4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
   int temp0, temp1, temp2, temp3, temp4, temp5, temp6;
   __asm__ volatile(
     "ulw             %[temp0],   -1(%[top])              \n\t"
@@ -695,7 +702,7 @@ static void VE4(uint8_t* dst, const uint8_t* top) {
   );
 }
 
-static void HE4(uint8_t* dst, const uint8_t* top) {
+static void HE4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
   int temp0, temp1, temp2, temp3, temp4, temp5, temp6;
   __asm__ volatile(
     "ulw             %[temp0],   -4(%[top])              \n\t"
@@ -731,7 +738,7 @@ static void HE4(uint8_t* dst, const uint8_t* top) {
   );
 }
 
-static void RD4(uint8_t* dst, const uint8_t* top) {
+static void RD4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
   int temp0, temp1, temp2, temp3, temp4, temp5;
   int temp6, temp7, temp8, temp9, temp10, temp11;
   __asm__ volatile(
@@ -780,7 +787,7 @@ static void RD4(uint8_t* dst, const uint8_t* top) {
   );
 }
 
-static void VR4(uint8_t* dst, const uint8_t* top) {
+static void VR4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
   int temp0, temp1, temp2, temp3, temp4;
   int temp5, temp6, temp7, temp8, temp9;
   __asm__ volatile (
@@ -830,7 +837,7 @@ static void VR4(uint8_t* dst, const uint8_t* top) {
   );
 }
 
-static void LD4(uint8_t* dst, const uint8_t* top) {
+static void LD4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
   int temp0, temp1, temp2, temp3, temp4, temp5;
   int temp6, temp7, temp8, temp9, temp10, temp11;
   __asm__ volatile(
@@ -877,7 +884,7 @@ static void LD4(uint8_t* dst, const uint8_t* top) {
   );
 }
 
-static void VL4(uint8_t* dst, const uint8_t* top) {
+static void VL4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
   int temp0, temp1, temp2, temp3, temp4;
   int temp5, temp6, temp7, temp8, temp9;
   __asm__ volatile (
@@ -926,7 +933,7 @@ static void VL4(uint8_t* dst, const uint8_t* top) {
   );
 }
 
-static void HD4(uint8_t* dst, const uint8_t* top) {
+static void HD4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
   int temp0, temp1, temp2, temp3, temp4;
   int temp5, temp6, temp7, temp8, temp9;
   __asm__ volatile (
@@ -974,7 +981,7 @@ static void HD4(uint8_t* dst, const uint8_t* top) {
   );
 }
 
-static void HU4(uint8_t* dst, const uint8_t* top) {
+static void HU4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
   int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
   __asm__ volatile (
     "ulw             %[temp0],   -5(%[top])              \n\t"
@@ -1013,8 +1020,9 @@ static void HU4(uint8_t* dst, const uint8_t* top) {
 //------------------------------------------------------------------------------
 // Chroma 8x8 prediction (paragraph 12.2)
 
-static void IntraChromaPreds_MIPSdspR2(uint8_t* dst, const uint8_t* left,
-                                       const uint8_t* top) {
+static void IntraChromaPreds_MIPSdspR2(uint8_t* WEBP_RESTRICT dst,
+                                       const uint8_t* WEBP_RESTRICT left,
+                                       const uint8_t* WEBP_RESTRICT top) {
   // U block
   DCMode8(C8DC8 + dst, left, top);
   VerticalPred8(C8VE8 + dst, top);
@@ -1033,8 +1041,9 @@ static void IntraChromaPreds_MIPSdspR2(uint8_t* dst, const uint8_t* left,
 //------------------------------------------------------------------------------
 // luma 16x16 prediction (paragraph 12.3)
 
-static void Intra16Preds_MIPSdspR2(uint8_t* dst,
-                                   const uint8_t* left, const uint8_t* top) {
+static void Intra16Preds_MIPSdspR2(uint8_t* WEBP_RESTRICT dst,
+                                   const uint8_t* WEBP_RESTRICT left,
+                                   const uint8_t* WEBP_RESTRICT top) {
   DCMode16(I16DC16 + dst, left, top);
   VerticalPred16(I16VE16 + dst, top);
   HorizontalPred16(I16HE16 + dst, left);
@@ -1043,7 +1052,8 @@ static void Intra16Preds_MIPSdspR2(uint8_t* dst,
 
 // Left samples are top[-5 .. -2], top_left is top[-1], top are
 // located at top[0..3], and top right is top[4..7]
-static void Intra4Preds_MIPSdspR2(uint8_t* dst, const uint8_t* top) {
+static void Intra4Preds_MIPSdspR2(uint8_t* WEBP_RESTRICT dst,
+                                  const uint8_t* WEBP_RESTRICT top) {
   DC4(I4DC4 + dst, top);
   TM4(I4TM4 + dst, top);
   VE4(I4VE4 + dst, top);
@@ -1079,7 +1089,8 @@ static void Intra4Preds_MIPSdspR2(uint8_t* dst, const uint8_t* top) {
   GET_SSE_INNER(C)                        \
   GET_SSE_INNER(D)
 
-static int SSE16x16_MIPSdspR2(const uint8_t* a, const uint8_t* b) {
+static int SSE16x16_MIPSdspR2(const uint8_t* WEBP_RESTRICT a,
+                              const uint8_t* WEBP_RESTRICT b) {
   int count;
   int temp0, temp1, temp2, temp3;
   __asm__ volatile (
@@ -1109,7 +1120,8 @@ static int SSE16x16_MIPSdspR2(const uint8_t* a, const uint8_t* b) {
   return count;
 }
 
-static int SSE16x8_MIPSdspR2(const uint8_t* a, const uint8_t* b) {
+static int SSE16x8_MIPSdspR2(const uint8_t* WEBP_RESTRICT a,
+                             const uint8_t* WEBP_RESTRICT b) {
   int count;
   int temp0, temp1, temp2, temp3;
   __asm__ volatile (
@@ -1131,7 +1143,8 @@ static int SSE16x8_MIPSdspR2(const uint8_t* a, const uint8_t* b) {
   return count;
 }
 
-static int SSE8x8_MIPSdspR2(const uint8_t* a, const uint8_t* b) {
+static int SSE8x8_MIPSdspR2(const uint8_t* WEBP_RESTRICT a,
+                            const uint8_t* WEBP_RESTRICT b) {
   int count;
   int temp0, temp1, temp2, temp3;
   __asm__ volatile (
@@ -1149,7 +1162,8 @@ static int SSE8x8_MIPSdspR2(const uint8_t* a, const uint8_t* b) {
   return count;
 }
 
-static int SSE4x4_MIPSdspR2(const uint8_t* a, const uint8_t* b) {
+static int SSE4x4_MIPSdspR2(const uint8_t* WEBP_RESTRICT a,
+                            const uint8_t* WEBP_RESTRICT b) {
   int count;
   int temp0, temp1, temp2, temp3;
   __asm__ volatile (
@@ -1273,7 +1287,7 @@ static int SSE4x4_MIPSdspR2(const uint8_t* a, const uint8_t* b) {
 "3:                                                          \n\t"
 
 static int QuantizeBlock_MIPSdspR2(int16_t in[16], int16_t out[16],
-                                   const VP8Matrix* const mtx) {
+                                   const VP8Matrix* WEBP_RESTRICT const mtx) {
   int temp0, temp1, temp2, temp3, temp4, temp5,temp6;
   int sign, coeff, level;
   int max_level = MAX_LEVEL;
@@ -1314,7 +1328,7 @@ static int QuantizeBlock_MIPSdspR2(int16_t in[16], int16_t out[16],
 }
 
 static int Quantize2Blocks_MIPSdspR2(int16_t in[32], int16_t out[32],
-                                     const VP8Matrix* const mtx) {
+                                     const VP8Matrix* WEBP_RESTRICT const mtx) {
   int nz;
   nz  = QuantizeBlock_MIPSdspR2(in + 0 * 16, out + 0 * 16, mtx) << 0;
   nz |= QuantizeBlock_MIPSdspR2(in + 1 * 16, out + 1 * 16, mtx) << 1;
@@ -1360,7 +1374,8 @@ static int Quantize2Blocks_MIPSdspR2(int16_t in[32], int16_t out[32],
   "usw             %[" #TEMP4 "],  " #C "(%[out])                 \n\t"        \
   "usw             %[" #TEMP6 "],  " #D "(%[out])                 \n\t"
 
-static void FTransformWHT_MIPSdspR2(const int16_t* in, int16_t* out) {
+static void FTransformWHT_MIPSdspR2(const int16_t* WEBP_RESTRICT in,
+                                    int16_t* WEBP_RESTRICT out) {
   int temp0, temp1, temp2, temp3, temp4;
   int temp5, temp6, temp7, temp8, temp9;
 
diff --git a/src/dsp/enc_msa.c b/src/dsp/enc_msa.c
index 6f85add4..31ecb942 100644
--- a/src/dsp/enc_msa.c
+++ b/src/dsp/enc_msa.c
@@ -41,8 +41,9 @@
   BUTTERFLY_4(a1_m, b1_m, c1_m, d1_m, out0, out1, out2, out3);      \
 } while (0)
 
-static WEBP_INLINE void ITransformOne(const uint8_t* ref, const int16_t* in,
-                                      uint8_t* dst) {
+static WEBP_INLINE void ITransformOne(const uint8_t* WEBP_RESTRICT ref,
+                                      const int16_t* WEBP_RESTRICT in,
+                                      uint8_t* WEBP_RESTRICT dst) {
   v8i16 input0, input1;
   v4i32 in0, in1, in2, in3, hz0, hz1, hz2, hz3, vt0, vt1, vt2, vt3;
   v4i32 res0, res1, res2, res3;
@@ -69,16 +70,18 @@ static WEBP_INLINE void ITransformOne(const uint8_t* ref, const int16_t* in,
   ST4x4_UB(res0, res0, 3, 2, 1, 0, dst, BPS);
 }
 
-static void ITransform_MSA(const uint8_t* ref, const int16_t* in, uint8_t* dst,
-                           int do_two) {
+static void ITransform_MSA(const uint8_t* WEBP_RESTRICT ref,
+                           const int16_t* WEBP_RESTRICT in,
+                           uint8_t* WEBP_RESTRICT dst, int do_two) {
   ITransformOne(ref, in, dst);
   if (do_two) {
     ITransformOne(ref + 4, in + 16, dst + 4);
   }
 }
 
-static void FTransform_MSA(const uint8_t* src, const uint8_t* ref,
-                           int16_t* out) {
+static void FTransform_MSA(const uint8_t* WEBP_RESTRICT src,
+                           const uint8_t* WEBP_RESTRICT ref,
+                           int16_t* WEBP_RESTRICT out) {
   uint64_t out0, out1, out2, out3;
   uint32_t in0, in1, in2, in3;
   v4i32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
@@ -131,7 +134,8 @@ static void FTransform_MSA(const uint8_t* src, const uint8_t* ref,
   SD4(out0, out1, out2, out3, out, 8);
 }
 
-static void FTransformWHT_MSA(const int16_t* in, int16_t* out) {
+static void FTransformWHT_MSA(const int16_t* WEBP_RESTRICT in,
+                              int16_t* WEBP_RESTRICT out) {
   v8i16 in0 = { 0 };
   v8i16 in1 = { 0 };
   v8i16 tmp0, tmp1, tmp2, tmp3;
@@ -168,7 +172,8 @@ static void FTransformWHT_MSA(const int16_t* in, int16_t* out) {
   ST_SH2(out0, out1, out, 8);
 }
 
-static int TTransform_MSA(const uint8_t* in, const uint16_t* w) {
+static int TTransform_MSA(const uint8_t* WEBP_RESTRICT in,
+                          const uint16_t* WEBP_RESTRICT w) {
   int sum;
   uint32_t in0_m, in1_m, in2_m, in3_m;
   v16i8 src0 = { 0 };
@@ -200,15 +205,17 @@ static int TTransform_MSA(const uint8_t* in, const uint16_t* w) {
   return sum;
 }
 
-static int Disto4x4_MSA(const uint8_t* const a, const uint8_t* const b,
-                        const uint16_t* const w) {
+static int Disto4x4_MSA(const uint8_t* WEBP_RESTRICT const a,
+                        const uint8_t* WEBP_RESTRICT const b,
+                        const uint16_t* WEBP_RESTRICT const w) {
   const int sum1 = TTransform_MSA(a, w);
   const int sum2 = TTransform_MSA(b, w);
   return abs(sum2 - sum1) >> 5;
 }
 
-static int Disto16x16_MSA(const uint8_t* const a, const uint8_t* const b,
-                          const uint16_t* const w) {
+static int Disto16x16_MSA(const uint8_t* WEBP_RESTRICT const a,
+                          const uint8_t* WEBP_RESTRICT const b,
+                          const uint16_t* WEBP_RESTRICT const w) {
   int D = 0;
   int x, y;
   for (y = 0; y < 16 * BPS; y += 4 * BPS) {
@@ -259,7 +266,9 @@ static void CollectHistogram_MSA(const uint8_t* ref, const uint8_t* pred,
 #define AVG3(a, b, c) (((a) + 2 * (b) + (c) + 2) >> 2)
 #define AVG2(a, b) (((a) + (b) + 1) >> 1)
 
-static WEBP_INLINE void VE4(uint8_t* dst, const uint8_t* top) {    // vertical
+// vertical
+static WEBP_INLINE void VE4(uint8_t* WEBP_RESTRICT dst,
+                            const uint8_t* WEBP_RESTRICT top) {
   const v16u8 A1 = { 0 };
   const uint64_t val_m = LD(top - 1);
   const v16u8 A = (v16u8)__msa_insert_d((v2i64)A1, 0, val_m);
@@ -272,7 +281,9 @@ static WEBP_INLINE void VE4(uint8_t* dst, const uint8_t* top) {    // vertical
   SW4(out, out, out, out, dst, BPS);
 }
 
-static WEBP_INLINE void HE4(uint8_t* dst, const uint8_t* top) {    // horizontal
+// horizontal
+static WEBP_INLINE void HE4(uint8_t* WEBP_RESTRICT dst,
+                            const uint8_t* WEBP_RESTRICT top) {
   const int X = top[-1];
   const int I = top[-2];
   const int J = top[-3];
@@ -284,7 +295,8 @@ static WEBP_INLINE void HE4(uint8_t* dst, const uint8_t* top) {    // horizontal
   WebPUint32ToMem(dst + 3 * BPS, 0x01010101U * AVG3(K, L, L));
 }
 
-static WEBP_INLINE void DC4(uint8_t* dst, const uint8_t* top) {
+static WEBP_INLINE void DC4(uint8_t* WEBP_RESTRICT dst,
+                            const uint8_t* WEBP_RESTRICT top) {
   uint32_t dc = 4;
   int i;
   for (i = 0; i < 4; ++i) dc += top[i] + top[-5 + i];
@@ -293,7 +305,8 @@ static WEBP_INLINE void DC4(uint8_t* dst, const uint8_t* top) {
   SW4(dc, dc, dc, dc, dst, BPS);
 }
 
-static WEBP_INLINE void RD4(uint8_t* dst, const uint8_t* top) {
+static WEBP_INLINE void RD4(uint8_t* WEBP_RESTRICT dst,
+                            const uint8_t* WEBP_RESTRICT top) {
   const v16u8 A2 = { 0 };
   const uint64_t val_m = LD(top - 5);
   const v16u8 A1 = (v16u8)__msa_insert_d((v2i64)A2, 0, val_m);
@@ -313,7 +326,8 @@ static WEBP_INLINE void RD4(uint8_t* dst, const uint8_t* top) {
   SW4(val3, val2, val1, val0, dst, BPS);
 }
 
-static WEBP_INLINE void LD4(uint8_t* dst, const uint8_t* top) {
+static WEBP_INLINE void LD4(uint8_t* WEBP_RESTRICT dst,
+                            const uint8_t* WEBP_RESTRICT top) {
   const v16u8 A1 = { 0 };
   const uint64_t val_m = LD(top);
   const v16u8 A = (v16u8)__msa_insert_d((v2i64)A1, 0, val_m);
@@ -333,7 +347,8 @@ static WEBP_INLINE void LD4(uint8_t* dst, const uint8_t* top) {
   SW4(val0, val1, val2, val3, dst, BPS);
 }
 
-static WEBP_INLINE void VR4(uint8_t* dst, const uint8_t* top) {
+static WEBP_INLINE void VR4(uint8_t* WEBP_RESTRICT dst,
+                            const uint8_t* WEBP_RESTRICT top) {
   const int X = top[-1];
   const int I = top[-2];
   const int J = top[-3];
@@ -354,7 +369,8 @@ static WEBP_INLINE void VR4(uint8_t* dst, const uint8_t* top) {
   DST(3, 1) =             AVG3(B, C, D);
 }
 
-static WEBP_INLINE void VL4(uint8_t* dst, const uint8_t* top) {
+static WEBP_INLINE void VL4(uint8_t* WEBP_RESTRICT dst,
+                            const uint8_t* WEBP_RESTRICT top) {
   const int A = top[0];
   const int B = top[1];
   const int C = top[2];
@@ -375,7 +391,8 @@ static WEBP_INLINE void VL4(uint8_t* dst, const uint8_t* top) {
               DST(3, 3) = AVG3(F, G, H);
 }
 
-static WEBP_INLINE void HU4(uint8_t* dst, const uint8_t* top) {
+static WEBP_INLINE void HU4(uint8_t* WEBP_RESTRICT dst,
+                            const uint8_t* WEBP_RESTRICT top) {
   const int I = top[-2];
   const int J = top[-3];
   const int K = top[-4];
@@ -390,7 +407,8 @@ static WEBP_INLINE void HU4(uint8_t* dst, const uint8_t* top) {
   DST(0, 3) = DST(1, 3) = DST(2, 3) = DST(3, 3) = L;
 }
 
-static WEBP_INLINE void HD4(uint8_t* dst, const uint8_t* top) {
+static WEBP_INLINE void HD4(uint8_t* WEBP_RESTRICT dst,
+                            const uint8_t* WEBP_RESTRICT top) {
   const int X = top[-1];
   const int I = top[-2];
   const int J = top[-3];
@@ -411,7 +429,8 @@ static WEBP_INLINE void HD4(uint8_t* dst, const uint8_t* top) {
   DST(1, 3)             = AVG3(L, K, J);
 }
 
-static WEBP_INLINE void TM4(uint8_t* dst, const uint8_t* top) {
+static WEBP_INLINE void TM4(uint8_t* WEBP_RESTRICT dst,
+                            const uint8_t* WEBP_RESTRICT top) {
   const v16i8 zero = { 0 };
   const v8i16 TL = (v8i16)__msa_fill_h(top[-1]);
   const v8i16 L0 = (v8i16)__msa_fill_h(top[-2]);
@@ -431,7 +450,8 @@ static WEBP_INLINE void TM4(uint8_t* dst, const uint8_t* top) {
 #undef AVG3
 #undef AVG2
 
-static void Intra4Preds_MSA(uint8_t* dst, const uint8_t* top) {
+static void Intra4Preds_MSA(uint8_t* WEBP_RESTRICT dst,
+                            const uint8_t* WEBP_RESTRICT top) {
   DC4(I4DC4 + dst, top);
   TM4(I4TM4 + dst, top);
   VE4(I4VE4 + dst, top);
@@ -451,7 +471,8 @@ static void Intra4Preds_MSA(uint8_t* dst, const uint8_t* top) {
     ST_UB8(out, out, out, out, out, out, out, out, dst + 8 * BPS, BPS);  \
 } while (0)
 
-static WEBP_INLINE void VerticalPred16x16(uint8_t* dst, const uint8_t* top) {
+static WEBP_INLINE void VerticalPred16x16(uint8_t* WEBP_RESTRICT dst,
+                                          const uint8_t* WEBP_RESTRICT top) {
   if (top != NULL) {
     const v16u8 out = LD_UB(top);
     STORE16x16(out, dst);
@@ -461,8 +482,8 @@ static WEBP_INLINE void VerticalPred16x16(uint8_t* dst, const uint8_t* top) {
   }
 }
 
-static WEBP_INLINE void HorizontalPred16x16(uint8_t* dst,
-                                            const uint8_t* left) {
+static WEBP_INLINE void HorizontalPred16x16(uint8_t* WEBP_RESTRICT dst,
+                                            const uint8_t* WEBP_RESTRICT left) {
   if (left != NULL) {
     int j;
     for (j = 0; j < 16; j += 4) {
@@ -480,8 +501,9 @@ static WEBP_INLINE void HorizontalPred16x16(uint8_t* dst,
   }
 }
 
-static WEBP_INLINE void TrueMotion16x16(uint8_t* dst, const uint8_t* left,
-                                        const uint8_t* top) {
+static WEBP_INLINE void TrueMotion16x16(uint8_t* WEBP_RESTRICT dst,
+                                        const uint8_t* WEBP_RESTRICT left,
+                                        const uint8_t* WEBP_RESTRICT top) {
   if (left != NULL) {
     if (top != NULL) {
       int j;
@@ -519,8 +541,9 @@ static WEBP_INLINE void TrueMotion16x16(uint8_t* dst, const uint8_t* left,
   }
 }
 
-static WEBP_INLINE void DCMode16x16(uint8_t* dst, const uint8_t* left,
-                                    const uint8_t* top) {
+static WEBP_INLINE void DCMode16x16(uint8_t* WEBP_RESTRICT dst,
+                                    const uint8_t* WEBP_RESTRICT left,
+                                    const uint8_t* WEBP_RESTRICT top) {
   int DC;
   v16u8 out;
   if (top != NULL && left != NULL) {
@@ -548,8 +571,9 @@ static WEBP_INLINE void DCMode16x16(uint8_t* dst, const uint8_t* left,
   STORE16x16(out, dst);
 }
 
-static void Intra16Preds_MSA(uint8_t* dst,
-                             const uint8_t* left, const uint8_t* top) {
+static void Intra16Preds_MSA(uint8_t* WEBP_RESTRICT dst,
+                             const uint8_t* WEBP_RESTRICT left,
+                             const uint8_t* WEBP_RESTRICT top) {
   DCMode16x16(I16DC16 + dst, left, top);
   VerticalPred16x16(I16VE16 + dst, top);
   HorizontalPred16x16(I16HE16 + dst, left);
@@ -574,7 +598,8 @@ static void Intra16Preds_MSA(uint8_t* dst,
   SD4(out, out, out, out, dst + 4 * BPS, BPS);  \
 } while (0)
 
-static WEBP_INLINE void VerticalPred8x8(uint8_t* dst, const uint8_t* top) {
+static WEBP_INLINE void VerticalPred8x8(uint8_t* WEBP_RESTRICT dst,
+                                        const uint8_t* WEBP_RESTRICT top) {
   if (top != NULL) {
     const uint64_t out = LD(top);
     STORE8x8(out, dst);
@@ -584,7 +609,8 @@ static WEBP_INLINE void VerticalPred8x8(uint8_t* dst, const uint8_t* top) {
   }
 }
 
-static WEBP_INLINE void HorizontalPred8x8(uint8_t* dst, const uint8_t* left) {
+static WEBP_INLINE void HorizontalPred8x8(uint8_t* WEBP_RESTRICT dst,
+                                          const uint8_t* WEBP_RESTRICT left) {
   if (left != NULL) {
     int j;
     for (j = 0; j < 8; j += 4) {
@@ -606,8 +632,9 @@ static WEBP_INLINE void HorizontalPred8x8(uint8_t* dst, const uint8_t* left) {
   }
 }
 
-static WEBP_INLINE void TrueMotion8x8(uint8_t* dst, const uint8_t* left,
-                                      const uint8_t* top) {
+static WEBP_INLINE void TrueMotion8x8(uint8_t* WEBP_RESTRICT dst,
+                                      const uint8_t* WEBP_RESTRICT left,
+                                      const uint8_t* WEBP_RESTRICT top) {
   if (left != NULL) {
     if (top != NULL) {
       int j;
@@ -646,8 +673,9 @@ static WEBP_INLINE void TrueMotion8x8(uint8_t* dst, const uint8_t* left,
   }
 }
 
-static WEBP_INLINE void DCMode8x8(uint8_t* dst, const uint8_t* left,
-                                  const uint8_t* top) {
+static WEBP_INLINE void DCMode8x8(uint8_t* WEBP_RESTRICT dst,
+                                  const uint8_t* WEBP_RESTRICT left,
+                                  const uint8_t* WEBP_RESTRICT top) {
   uint64_t out;
   v16u8 src = { 0 };
   if (top != NULL && left != NULL) {
@@ -670,8 +698,9 @@ static WEBP_INLINE void DCMode8x8(uint8_t* dst, const uint8_t* left,
   STORE8x8(out, dst);
 }
 
-static void IntraChromaPreds_MSA(uint8_t* dst, const uint8_t* left,
-                                 const uint8_t* top) {
+static void IntraChromaPreds_MSA(uint8_t* WEBP_RESTRICT dst,
+                                 const uint8_t* WEBP_RESTRICT left,
+                                 const uint8_t* WEBP_RESTRICT top) {
   // U block
   DCMode8x8(C8DC8 + dst, left, top);
   VerticalPred8x8(C8VE8 + dst, top);
@@ -712,7 +741,8 @@ static void IntraChromaPreds_MSA(uint8_t* dst, const uint8_t* left,
   DPADD_SH2_SW(tmp2, tmp3, tmp2, tmp3, out2, out3);                         \
 } while (0)
 
-static int SSE16x16_MSA(const uint8_t* a, const uint8_t* b) {
+static int SSE16x16_MSA(const uint8_t* WEBP_RESTRICT a,
+                        const uint8_t* WEBP_RESTRICT b) {
   uint32_t sum;
   v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
   v16u8 ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7;
@@ -739,7 +769,8 @@ static int SSE16x16_MSA(const uint8_t* a, const uint8_t* b) {
   return sum;
 }
 
-static int SSE16x8_MSA(const uint8_t* a, const uint8_t* b) {
+static int SSE16x8_MSA(const uint8_t* WEBP_RESTRICT a,
+                       const uint8_t* WEBP_RESTRICT b) {
   uint32_t sum;
   v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
   v16u8 ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7;
@@ -758,7 +789,8 @@ static int SSE16x8_MSA(const uint8_t* a, const uint8_t* b) {
   return sum;
 }
 
-static int SSE8x8_MSA(const uint8_t* a, const uint8_t* b) {
+static int SSE8x8_MSA(const uint8_t* WEBP_RESTRICT a,
+                      const uint8_t* WEBP_RESTRICT b) {
   uint32_t sum;
   v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
   v16u8 ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7;
@@ -778,7 +810,8 @@ static int SSE8x8_MSA(const uint8_t* a, const uint8_t* b) {
   return sum;
 }
 
-static int SSE4x4_MSA(const uint8_t* a, const uint8_t* b) {
+static int SSE4x4_MSA(const uint8_t* WEBP_RESTRICT a,
+                      const uint8_t* WEBP_RESTRICT b) {
   uint32_t sum = 0;
   uint32_t src0, src1, src2, src3, ref0, ref1, ref2, ref3;
   v16u8 src = { 0 }, ref = { 0 }, tmp0, tmp1;
@@ -801,7 +834,7 @@ static int SSE4x4_MSA(const uint8_t* a, const uint8_t* b) {
 // Quantization
 
 static int QuantizeBlock_MSA(int16_t in[16], int16_t out[16],
-                             const VP8Matrix* const mtx) {
+                             const VP8Matrix* WEBP_RESTRICT const mtx) {
   int sum;
   v8i16 in0, in1, sh0, sh1, out0, out1;
   v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, sign0, sign1;
@@ -854,7 +887,7 @@ static int QuantizeBlock_MSA(int16_t in[16], int16_t out[16],
 }
 
 static int Quantize2Blocks_MSA(int16_t in[32], int16_t out[32],
-                               const VP8Matrix* const mtx) {
+                               const VP8Matrix* WEBP_RESTRICT const mtx) {
   int nz;
   nz  = VP8EncQuantizeBlock(in + 0 * 16, out + 0 * 16, mtx) << 0;
   nz |= VP8EncQuantizeBlock(in + 1 * 16, out + 1 * 16, mtx) << 1;
diff --git a/src/dsp/enc_neon.c b/src/dsp/enc_neon.c
index 7ba5b2d6..30a66fc5 100644
--- a/src/dsp/enc_neon.c
+++ b/src/dsp/enc_neon.c
@@ -60,8 +60,8 @@ static WEBP_INLINE void SaturateAndStore4x4_NEON(uint8_t* const dst,
 
 static WEBP_INLINE void Add4x4_NEON(const int16x8_t row01,
                                     const int16x8_t row23,
-                                    const uint8_t* const ref,
-                                    uint8_t* const dst) {
+                                    const uint8_t* WEBP_RESTRICT const ref,
+                                    uint8_t* WEBP_RESTRICT const dst) {
   uint32x2_t dst01 = vdup_n_u32(0);
   uint32x2_t dst23 = vdup_n_u32(0);
 
@@ -120,8 +120,9 @@ static WEBP_INLINE void TransformPass_NEON(int16x8x2_t* const rows) {
   Transpose8x2_NEON(E0, E1, rows);
 }
 
-static void ITransformOne_NEON(const uint8_t* ref,
-                               const int16_t* in, uint8_t* dst) {
+static void ITransformOne_NEON(const uint8_t* WEBP_RESTRICT ref,
+                               const int16_t* WEBP_RESTRICT in,
+                               uint8_t* WEBP_RESTRICT dst) {
   int16x8x2_t rows;
   INIT_VECTOR2(rows, vld1q_s16(in + 0), vld1q_s16(in + 8));
   TransformPass_NEON(&rows);
@@ -131,8 +132,9 @@ static void ITransformOne_NEON(const uint8_t* ref,
 
 #else
 
-static void ITransformOne_NEON(const uint8_t* ref,
-                               const int16_t* in, uint8_t* dst) {
+static void ITransformOne_NEON(const uint8_t* WEBP_RESTRICT ref,
+                               const int16_t* WEBP_RESTRICT in,
+                               uint8_t* WEBP_RESTRICT dst) {
   const int kBPS = BPS;
   const int16_t kC1C2[] = { kC1, kC2, 0, 0 };
 
@@ -247,8 +249,9 @@ static void ITransformOne_NEON(const uint8_t* ref,
 
 #endif    // WEBP_USE_INTRINSICS
 
-static void ITransform_NEON(const uint8_t* ref,
-                            const int16_t* in, uint8_t* dst, int do_two) {
+static void ITransform_NEON(const uint8_t* WEBP_RESTRICT ref,
+                            const int16_t* WEBP_RESTRICT in,
+                            uint8_t* WEBP_RESTRICT dst, int do_two) {
   ITransformOne_NEON(ref, in, dst);
   if (do_two) {
     ITransformOne_NEON(ref + 4, in + 16, dst + 4);
@@ -294,8 +297,9 @@ static WEBP_INLINE int16x8_t DiffU8ToS16_NEON(const uint8x8_t a,
   return vreinterpretq_s16_u16(vsubl_u8(a, b));
 }
 
-static void FTransform_NEON(const uint8_t* src, const uint8_t* ref,
-                            int16_t* out) {
+static void FTransform_NEON(const uint8_t* WEBP_RESTRICT src,
+                            const uint8_t* WEBP_RESTRICT ref,
+                            int16_t* WEBP_RESTRICT out) {
   int16x8_t d0d1, d3d2;   // working 4x4 int16 variables
   {
     const uint8x16_t S0 = Load4x4_NEON(src);
@@ -364,8 +368,9 @@ static const int32_t kCoeff32[] = {
   51000, 51000, 51000, 51000
 };
 
-static void FTransform_NEON(const uint8_t* src, const uint8_t* ref,
-                            int16_t* out) {
+static void FTransform_NEON(const uint8_t* WEBP_RESTRICT src,
+                            const uint8_t* WEBP_RESTRICT ref,
+                            int16_t* WEBP_RESTRICT out) {
   const int kBPS = BPS;
   const uint8_t* src_ptr = src;
   const uint8_t* ref_ptr = ref;
@@ -484,7 +489,8 @@ static void FTransform_NEON(const uint8_t* src, const uint8_t* ref,
   src += stride;                                    \
 } while (0)
 
-static void FTransformWHT_NEON(const int16_t* src, int16_t* out) {
+static void FTransformWHT_NEON(const int16_t* WEBP_RESTRICT src,
+                               int16_t* WEBP_RESTRICT out) {
   const int stride = 16;
   const int16x4_t zero = vdup_n_s16(0);
   int32x4x4_t tmp0;
@@ -659,8 +665,9 @@ static WEBP_INLINE int32x2_t DistoSum_NEON(const int16x8x4_t q4_in,
 // Hadamard transform
 // Returns the weighted sum of the absolute value of transformed coefficients.
 // w[] contains a row-major 4 by 4 symmetric matrix.
-static int Disto4x4_NEON(const uint8_t* const a, const uint8_t* const b,
-                         const uint16_t* const w) {
+static int Disto4x4_NEON(const uint8_t* WEBP_RESTRICT const a,
+                         const uint8_t* WEBP_RESTRICT const b,
+                         const uint16_t* WEBP_RESTRICT const w) {
   uint32x2_t d_in_ab_0123 = vdup_n_u32(0);
   uint32x2_t d_in_ab_4567 = vdup_n_u32(0);
   uint32x2_t d_in_ab_89ab = vdup_n_u32(0);
@@ -701,8 +708,9 @@ static int Disto4x4_NEON(const uint8_t* const a, const uint8_t* const b,
 }
 #undef LOAD_LANE_32b
 
-static int Disto16x16_NEON(const uint8_t* const a, const uint8_t* const b,
-                           const uint16_t* const w) {
+static int Disto16x16_NEON(const uint8_t* WEBP_RESTRICT const a,
+                           const uint8_t* WEBP_RESTRICT const b,
+                           const uint16_t* WEBP_RESTRICT const w) {
   int D = 0;
   int x, y;
   for (y = 0; y < 16 * BPS; y += 4 * BPS) {
@@ -715,9 +723,10 @@ static int Disto16x16_NEON(const uint8_t* const a, const uint8_t* const b,
 
 //------------------------------------------------------------------------------
 
-static void CollectHistogram_NEON(const uint8_t* ref, const uint8_t* pred,
+static void CollectHistogram_NEON(const uint8_t* WEBP_RESTRICT ref,
+                                  const uint8_t* WEBP_RESTRICT pred,
                                   int start_block, int end_block,
-                                  VP8Histogram* const histo) {
+                                  VP8Histogram* WEBP_RESTRICT const histo) {
   const uint16x8_t max_coeff_thresh = vdupq_n_u16(MAX_COEFF_THRESH);
   int j;
   int distribution[MAX_COEFF_THRESH + 1] = { 0 };
@@ -747,9 +756,9 @@ static void CollectHistogram_NEON(const uint8_t* ref, const uint8_t* pred,
 
 //------------------------------------------------------------------------------
 
-static WEBP_INLINE void AccumulateSSE16_NEON(const uint8_t* const a,
-                                             const uint8_t* const b,
-                                             uint32x4_t* const sum) {
+static WEBP_INLINE void AccumulateSSE16_NEON(
+    const uint8_t* WEBP_RESTRICT const a, const uint8_t* WEBP_RESTRICT const b,
+    uint32x4_t* const sum) {
   const uint8x16_t a0 = vld1q_u8(a);
   const uint8x16_t b0 = vld1q_u8(b);
   const uint8x16_t abs_diff = vabdq_u8(a0, b0);
@@ -775,7 +784,8 @@ static int SumToInt_NEON(uint32x4_t sum) {
 #endif
 }
 
-static int SSE16x16_NEON(const uint8_t* a, const uint8_t* b) {
+static int SSE16x16_NEON(const uint8_t* WEBP_RESTRICT a,
+                         const uint8_t* WEBP_RESTRICT b) {
   uint32x4_t sum = vdupq_n_u32(0);
   int y;
   for (y = 0; y < 16; ++y) {
@@ -784,7 +794,8 @@ static int SSE16x16_NEON(const uint8_t* a, const uint8_t* b) {
   return SumToInt_NEON(sum);
 }
 
-static int SSE16x8_NEON(const uint8_t* a, const uint8_t* b) {
+static int SSE16x8_NEON(const uint8_t* WEBP_RESTRICT a,
+                        const uint8_t* WEBP_RESTRICT b) {
   uint32x4_t sum = vdupq_n_u32(0);
   int y;
   for (y = 0; y < 8; ++y) {
@@ -793,7 +804,8 @@ static int SSE16x8_NEON(const uint8_t* a, const uint8_t* b) {
   return SumToInt_NEON(sum);
 }
 
-static int SSE8x8_NEON(const uint8_t* a, const uint8_t* b) {
+static int SSE8x8_NEON(const uint8_t* WEBP_RESTRICT a,
+                       const uint8_t* WEBP_RESTRICT b) {
   uint32x4_t sum = vdupq_n_u32(0);
   int y;
   for (y = 0; y < 8; ++y) {
@@ -806,7 +818,8 @@ static int SSE8x8_NEON(const uint8_t* a, const uint8_t* b) {
   return SumToInt_NEON(sum);
 }
 
-static int SSE4x4_NEON(const uint8_t* a, const uint8_t* b) {
+static int SSE4x4_NEON(const uint8_t* WEBP_RESTRICT a,
+                       const uint8_t* WEBP_RESTRICT b) {
   const uint8x16_t a0 = Load4x4_NEON(a);
   const uint8x16_t b0 = Load4x4_NEON(b);
   const uint8x16_t abs_diff = vabdq_u8(a0, b0);
@@ -825,8 +838,9 @@ static int SSE4x4_NEON(const uint8_t* a, const uint8_t* b) {
 // Compilation with gcc-4.6.x is problematic for now.
 #if !defined(WORK_AROUND_GCC)
 
-static int16x8_t Quantize_NEON(int16_t* const in,
-                               const VP8Matrix* const mtx, int offset) {
+static int16x8_t Quantize_NEON(int16_t* WEBP_RESTRICT const in,
+                               const VP8Matrix* WEBP_RESTRICT const mtx,
+                               int offset) {
   const uint16x8_t sharp = vld1q_u16(&mtx->sharpen_[offset]);
   const uint16x8_t q = vld1q_u16(&mtx->q_[offset]);
   const uint16x8_t iq = vld1q_u16(&mtx->iq_[offset]);
@@ -860,7 +874,7 @@ static const uint8_t kShuffles[4][8] = {
 };
 
 static int QuantizeBlock_NEON(int16_t in[16], int16_t out[16],
-                              const VP8Matrix* const mtx) {
+                              const VP8Matrix* WEBP_RESTRICT const mtx) {
   const int16x8_t out0 = Quantize_NEON(in, mtx, 0);
   const int16x8_t out1 = Quantize_NEON(in, mtx, 8);
   uint8x8x4_t shuffles;
@@ -902,7 +916,7 @@ static int QuantizeBlock_NEON(int16_t in[16], int16_t out[16],
 }
 
 static int Quantize2Blocks_NEON(int16_t in[32], int16_t out[32],
-                                const VP8Matrix* const mtx) {
+                                const VP8Matrix* WEBP_RESTRICT const mtx) {
   int nz;
   nz  = QuantizeBlock_NEON(in + 0 * 16, out + 0 * 16, mtx) << 0;
   nz |= QuantizeBlock_NEON(in + 1 * 16, out + 1 * 16, mtx) << 1;
@@ -930,7 +944,8 @@ static int Quantize2Blocks_NEON(int16_t in[32], int16_t out[32],
     vst1q_u8(dst, r);                                                          \
   } while (0)
 
-static void Intra4Preds_NEON(uint8_t* dst, const uint8_t* top) {
+static void Intra4Preds_NEON(uint8_t* WEBP_RESTRICT dst,
+                             const uint8_t* WEBP_RESTRICT top) {
   // 0   1   2   3   4   5   6   7   8   9  10  11  12  13
   //     L   K   J   I   X   A   B   C   D   E   F   G   H
   //    -5  -4  -3  -2  -1   0   1   2   3   4   5   6   7
@@ -1162,8 +1177,9 @@ static WEBP_INLINE void TrueMotion_NEON(uint8_t* dst, const uint8_t* left,
   }
 }
 
-static void Intra16Preds_NEON(uint8_t* dst, const uint8_t* left,
-                              const uint8_t* top) {
+static void Intra16Preds_NEON(uint8_t* WEBP_RESTRICT dst,
+                              const uint8_t* WEBP_RESTRICT left,
+                              const uint8_t* WEBP_RESTRICT top) {
   DCMode_NEON(I16DC16 + dst, left, top);
   VerticalPred16_NEON(I16VE16 + dst, top);
   HorizontalPred16_NEON(I16HE16 + dst, left);
diff --git a/src/dsp/enc_sse2.c b/src/dsp/enc_sse2.c
index 010624a2..588a6292 100644
--- a/src/dsp/enc_sse2.c
+++ b/src/dsp/enc_sse2.c
@@ -26,8 +26,9 @@
 // Transforms (Paragraph 14.4)
 
 // Does one inverse transform.
-static void ITransform_One_SSE2(const uint8_t* ref, const int16_t* in,
-                                uint8_t* dst) {
+static void ITransform_One_SSE2(const uint8_t* WEBP_RESTRICT ref,
+                                const int16_t* WEBP_RESTRICT in,
+                                uint8_t* WEBP_RESTRICT dst) {
   // This implementation makes use of 16-bit fixed point versions of two
   // multiply constants:
   //    K1 = sqrt(2) * cos (pi/8) ~= 85627 / 2^16
@@ -177,8 +178,9 @@ static void ITransform_One_SSE2(const uint8_t* ref, const int16_t* in,
 }
 
 // Does two inverse transforms.
-static void ITransform_Two_SSE2(const uint8_t* ref, const int16_t* in,
-                                uint8_t* dst) {
+static void ITransform_Two_SSE2(const uint8_t* WEBP_RESTRICT ref,
+                                const int16_t* WEBP_RESTRICT in,
+                                uint8_t* WEBP_RESTRICT dst) {
   // This implementation makes use of 16-bit fixed point versions of two
   // multiply constants:
   //    K1 = sqrt(2) * cos (pi/8) ~= 85627 / 2^16
@@ -316,7 +318,9 @@ static void ITransform_Two_SSE2(const uint8_t* ref, const int16_t* in,
 }
 
 // Does one or two inverse transforms.
-static void ITransform_SSE2(const uint8_t* ref, const int16_t* in, uint8_t* dst,
+static void ITransform_SSE2(const uint8_t* WEBP_RESTRICT ref,
+                            const int16_t* WEBP_RESTRICT in,
+                            uint8_t* WEBP_RESTRICT dst,
                             int do_two) {
   if (do_two) {
     ITransform_Two_SSE2(ref, in, dst);
@@ -373,7 +377,7 @@ static void FTransformPass1_SSE2(const __m128i* const in01,
 
 static void FTransformPass2_SSE2(const __m128i* const v01,
                                  const __m128i* const v32,
-                                 int16_t* out) {
+                                 int16_t* WEBP_RESTRICT out) {
   const __m128i zero = _mm_setzero_si128();
   const __m128i seven = _mm_set1_epi16(7);
   const __m128i k5352_2217 = _mm_set_epi16(5352,  2217, 5352,  2217,
@@ -424,8 +428,9 @@ static void FTransformPass2_SSE2(const __m128i* const v01,
   _mm_storeu_si128((__m128i*)&out[8], d2_f3);
 }
 
-static void FTransform_SSE2(const uint8_t* src, const uint8_t* ref,
-                            int16_t* out) {
+static void FTransform_SSE2(const uint8_t* WEBP_RESTRICT src,
+                            const uint8_t* WEBP_RESTRICT ref,
+                            int16_t* WEBP_RESTRICT out) {
   const __m128i zero = _mm_setzero_si128();
   // Load src.
   const __m128i src0 = _mm_loadl_epi64((const __m128i*)&src[0 * BPS]);
@@ -468,8 +473,9 @@ static void FTransform_SSE2(const uint8_t* src, const uint8_t* ref,
   FTransformPass2_SSE2(&v01, &v32, out);
 }
 
-static void FTransform2_SSE2(const uint8_t* src, const uint8_t* ref,
-                             int16_t* out) {
+static void FTransform2_SSE2(const uint8_t* WEBP_RESTRICT src,
+                             const uint8_t* WEBP_RESTRICT ref,
+                             int16_t* WEBP_RESTRICT out) {
   const __m128i zero = _mm_setzero_si128();
 
   // Load src and convert to 16b.
@@ -517,7 +523,8 @@ static void FTransform2_SSE2(const uint8_t* src, const uint8_t* ref,
   FTransformPass2_SSE2(&v01h, &v32h, out + 16);
 }
 
-static void FTransformWHTRow_SSE2(const int16_t* const in, __m128i* const out) {
+static void FTransformWHTRow_SSE2(const int16_t* WEBP_RESTRICT const in,
+                                  __m128i* const out) {
   const __m128i kMult = _mm_set_epi16(-1, 1, -1, 1, 1, 1, 1, 1);
   const __m128i src0 = _mm_loadl_epi64((__m128i*)&in[0 * 16]);
   const __m128i src1 = _mm_loadl_epi64((__m128i*)&in[1 * 16]);
@@ -533,7 +540,8 @@ static void FTransformWHTRow_SSE2(const int16_t* const in, __m128i* const out) {
   *out = _mm_madd_epi16(D, kMult);
 }
 
-static void FTransformWHT_SSE2(const int16_t* in, int16_t* out) {
+static void FTransformWHT_SSE2(const int16_t* WEBP_RESTRICT in,
+                               int16_t* WEBP_RESTRICT out) {
   // Input is 12b signed.
   __m128i row0, row1, row2, row3;
   // Rows are 14b signed.
@@ -566,9 +574,10 @@ static void FTransformWHT_SSE2(const int16_t* in, int16_t* out) {
 // Compute susceptibility based on DCT-coeff histograms:
 // the higher, the "easier" the macroblock is to compress.
 
-static void CollectHistogram_SSE2(const uint8_t* ref, const uint8_t* pred,
+static void CollectHistogram_SSE2(const uint8_t* WEBP_RESTRICT ref,
+                                  const uint8_t* WEBP_RESTRICT pred,
                                   int start_block, int end_block,
-                                  VP8Histogram* const histo) {
+                                  VP8Histogram* WEBP_RESTRICT const histo) {
   const __m128i zero = _mm_setzero_si128();
   const __m128i max_coeff_thresh = _mm_set1_epi16(MAX_COEFF_THRESH);
   int j;
@@ -640,7 +649,8 @@ static WEBP_INLINE void Fill_SSE2(uint8_t* dst, int value, int size) {
   }
 }
 
-static WEBP_INLINE void VE8uv_SSE2(uint8_t* dst, const uint8_t* top) {
+static WEBP_INLINE void VE8uv_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                   const uint8_t* WEBP_RESTRICT top) {
   int j;
   const __m128i top_values = _mm_loadl_epi64((const __m128i*)top);
   for (j = 0; j < 8; ++j) {
@@ -648,7 +658,8 @@ static WEBP_INLINE void VE8uv_SSE2(uint8_t* dst, const uint8_t* top) {
   }
 }
 
-static WEBP_INLINE void VE16_SSE2(uint8_t* dst, const uint8_t* top) {
+static WEBP_INLINE void VE16_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                  const uint8_t* WEBP_RESTRICT top) {
   const __m128i top_values = _mm_load_si128((const __m128i*)top);
   int j;
   for (j = 0; j < 16; ++j) {
@@ -656,8 +667,9 @@ static WEBP_INLINE void VE16_SSE2(uint8_t* dst, const uint8_t* top) {
   }
 }
 
-static WEBP_INLINE void VerticalPred_SSE2(uint8_t* dst,
-                                          const uint8_t* top, int size) {
+static WEBP_INLINE void VerticalPred_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                          const uint8_t* WEBP_RESTRICT top,
+                                          int size) {
   if (top != NULL) {
     if (size == 8) {
       VE8uv_SSE2(dst, top);
@@ -669,7 +681,8 @@ static WEBP_INLINE void VerticalPred_SSE2(uint8_t* dst,
   }
 }
 
-static WEBP_INLINE void HE8uv_SSE2(uint8_t* dst, const uint8_t* left) {
+static WEBP_INLINE void HE8uv_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                   const uint8_t* WEBP_RESTRICT left) {
   int j;
   for (j = 0; j < 8; ++j) {
     const __m128i values = _mm_set1_epi8((char)left[j]);
@@ -678,7 +691,8 @@ static WEBP_INLINE void HE8uv_SSE2(uint8_t* dst, const uint8_t* left) {
   }
 }
 
-static WEBP_INLINE void HE16_SSE2(uint8_t* dst, const uint8_t* left) {
+static WEBP_INLINE void HE16_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                  const uint8_t* WEBP_RESTRICT left) {
   int j;
   for (j = 0; j < 16; ++j) {
     const __m128i values = _mm_set1_epi8((char)left[j]);
@@ -687,8 +701,9 @@ static WEBP_INLINE void HE16_SSE2(uint8_t* dst, const uint8_t* left) {
   }
 }
 
-static WEBP_INLINE void HorizontalPred_SSE2(uint8_t* dst,
-                                            const uint8_t* left, int size) {
+static WEBP_INLINE void HorizontalPred_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                            const uint8_t* WEBP_RESTRICT left,
+                                            int size) {
   if (left != NULL) {
     if (size == 8) {
       HE8uv_SSE2(dst, left);
@@ -700,8 +715,9 @@ static WEBP_INLINE void HorizontalPred_SSE2(uint8_t* dst,
   }
 }
 
-static WEBP_INLINE void TM_SSE2(uint8_t* dst, const uint8_t* left,
-                                const uint8_t* top, int size) {
+static WEBP_INLINE void TM_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                const uint8_t* WEBP_RESTRICT left,
+                                const uint8_t* WEBP_RESTRICT top, int size) {
   const __m128i zero = _mm_setzero_si128();
   int y;
   if (size == 8) {
@@ -728,8 +744,10 @@ static WEBP_INLINE void TM_SSE2(uint8_t* dst, const uint8_t* left,
   }
 }
 
-static WEBP_INLINE void TrueMotion_SSE2(uint8_t* dst, const uint8_t* left,
-                                        const uint8_t* top, int size) {
+static WEBP_INLINE void TrueMotion_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                        const uint8_t* WEBP_RESTRICT left,
+                                        const uint8_t* WEBP_RESTRICT top,
+                                        int size) {
   if (left != NULL) {
     if (top != NULL) {
       TM_SSE2(dst, left, top, size);
@@ -749,8 +767,9 @@ static WEBP_INLINE void TrueMotion_SSE2(uint8_t* dst, const uint8_t* left,
   }
 }
 
-static WEBP_INLINE void DC8uv_SSE2(uint8_t* dst, const uint8_t* left,
-                                   const uint8_t* top) {
+static WEBP_INLINE void DC8uv_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                   const uint8_t* WEBP_RESTRICT left,
+                                   const uint8_t* WEBP_RESTRICT top) {
   const __m128i top_values = _mm_loadl_epi64((const __m128i*)top);
   const __m128i left_values = _mm_loadl_epi64((const __m128i*)left);
   const __m128i combined = _mm_unpacklo_epi64(top_values, left_values);
@@ -758,7 +777,8 @@ static WEBP_INLINE void DC8uv_SSE2(uint8_t* dst, const uint8_t* left,
   Put8x8uv_SSE2(DC >> 4, dst);
 }
 
-static WEBP_INLINE void DC8uvNoLeft_SSE2(uint8_t* dst, const uint8_t* top) {
+static WEBP_INLINE void DC8uvNoLeft_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                         const uint8_t* WEBP_RESTRICT top) {
   const __m128i zero = _mm_setzero_si128();
   const __m128i top_values = _mm_loadl_epi64((const __m128i*)top);
   const __m128i sum = _mm_sad_epu8(top_values, zero);
@@ -766,7 +786,8 @@ static WEBP_INLINE void DC8uvNoLeft_SSE2(uint8_t* dst, const uint8_t* top) {
   Put8x8uv_SSE2(DC >> 3, dst);
 }
 
-static WEBP_INLINE void DC8uvNoTop_SSE2(uint8_t* dst, const uint8_t* left) {
+static WEBP_INLINE void DC8uvNoTop_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                        const uint8_t* WEBP_RESTRICT left) {
   // 'left' is contiguous so we can reuse the top summation.
   DC8uvNoLeft_SSE2(dst, left);
 }
@@ -775,8 +796,9 @@ static WEBP_INLINE void DC8uvNoTopLeft_SSE2(uint8_t* dst) {
   Put8x8uv_SSE2(0x80, dst);
 }
 
-static WEBP_INLINE void DC8uvMode_SSE2(uint8_t* dst, const uint8_t* left,
-                                       const uint8_t* top) {
+static WEBP_INLINE void DC8uvMode_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                       const uint8_t* WEBP_RESTRICT left,
+                                       const uint8_t* WEBP_RESTRICT top) {
   if (top != NULL) {
     if (left != NULL) {  // top and left present
       DC8uv_SSE2(dst, left, top);
@@ -790,8 +812,9 @@ static WEBP_INLINE void DC8uvMode_SSE2(uint8_t* dst, const uint8_t* left,
   }
 }
 
-static WEBP_INLINE void DC16_SSE2(uint8_t* dst, const uint8_t* left,
-                                  const uint8_t* top) {
+static WEBP_INLINE void DC16_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                  const uint8_t* WEBP_RESTRICT left,
+                                  const uint8_t* WEBP_RESTRICT top) {
   const __m128i top_row = _mm_load_si128((const __m128i*)top);
   const __m128i left_row = _mm_load_si128((const __m128i*)left);
   const int DC =
@@ -799,13 +822,15 @@ static WEBP_INLINE void DC16_SSE2(uint8_t* dst, const uint8_t* left,
   Put16_SSE2(DC >> 5, dst);
 }
 
-static WEBP_INLINE void DC16NoLeft_SSE2(uint8_t* dst, const uint8_t* top) {
+static WEBP_INLINE void DC16NoLeft_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                        const uint8_t* WEBP_RESTRICT top) {
   const __m128i top_row = _mm_load_si128((const __m128i*)top);
   const int DC = VP8HorizontalAdd8b(&top_row) + 8;
   Put16_SSE2(DC >> 4, dst);
 }
 
-static WEBP_INLINE void DC16NoTop_SSE2(uint8_t* dst, const uint8_t* left) {
+static WEBP_INLINE void DC16NoTop_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                       const uint8_t* WEBP_RESTRICT left) {
   // 'left' is contiguous so we can reuse the top summation.
   DC16NoLeft_SSE2(dst, left);
 }
@@ -814,8 +839,9 @@ static WEBP_INLINE void DC16NoTopLeft_SSE2(uint8_t* dst) {
   Put16_SSE2(0x80, dst);
 }
 
-static WEBP_INLINE void DC16Mode_SSE2(uint8_t* dst, const uint8_t* left,
-                                      const uint8_t* top) {
+static WEBP_INLINE void DC16Mode_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                      const uint8_t* WEBP_RESTRICT left,
+                                      const uint8_t* WEBP_RESTRICT top) {
   if (top != NULL) {
     if (left != NULL) {  // top and left present
       DC16_SSE2(dst, left, top);
@@ -844,8 +870,9 @@ static WEBP_INLINE void DC16Mode_SSE2(uint8_t* dst, const uint8_t* left,
 //   where: AC = (a + b + 1) >> 1,   BC = (b + c + 1) >> 1
 //   and ab = a ^ b, bc = b ^ c, lsb = (AC^BC)&1
 
-static WEBP_INLINE void VE4_SSE2(uint8_t* dst,
-                                 const uint8_t* top) {  // vertical
+// vertical
+static WEBP_INLINE void VE4_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                 const uint8_t* WEBP_RESTRICT top) {
   const __m128i one = _mm_set1_epi8(1);
   const __m128i ABCDEFGH = _mm_loadl_epi64((__m128i*)(top - 1));
   const __m128i BCDEFGH0 = _mm_srli_si128(ABCDEFGH, 1);
@@ -861,8 +888,9 @@ static WEBP_INLINE void VE4_SSE2(uint8_t* dst,
   }
 }
 
-static WEBP_INLINE void HE4_SSE2(uint8_t* dst,
-                                 const uint8_t* top) {  // horizontal
+// horizontal
+static WEBP_INLINE void HE4_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                 const uint8_t* WEBP_RESTRICT top) {
   const int X = top[-1];
   const int I = top[-2];
   const int J = top[-3];
@@ -874,15 +902,17 @@ static WEBP_INLINE void HE4_SSE2(uint8_t* dst,
   WebPUint32ToMem(dst + 3 * BPS, 0x01010101U * AVG3(K, L, L));
 }
 
-static WEBP_INLINE void DC4_SSE2(uint8_t* dst, const uint8_t* top) {
+static WEBP_INLINE void DC4_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                 const uint8_t* WEBP_RESTRICT top) {
   uint32_t dc = 4;
   int i;
   for (i = 0; i < 4; ++i) dc += top[i] + top[-5 + i];
   Fill_SSE2(dst, dc >> 3, 4);
 }
 
-static WEBP_INLINE void LD4_SSE2(uint8_t* dst,
-                                 const uint8_t* top) {  // Down-Left
+// Down-Left
+static WEBP_INLINE void LD4_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                 const uint8_t* WEBP_RESTRICT top) {
   const __m128i one = _mm_set1_epi8(1);
   const __m128i ABCDEFGH = _mm_loadl_epi64((const __m128i*)top);
   const __m128i BCDEFGH0 = _mm_srli_si128(ABCDEFGH, 1);
@@ -898,8 +928,9 @@ static WEBP_INLINE void LD4_SSE2(uint8_t* dst,
   WebPInt32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3)));
 }
 
-static WEBP_INLINE void VR4_SSE2(uint8_t* dst,
-                                 const uint8_t* top) {  // Vertical-Right
+// Vertical-Right
+static WEBP_INLINE void VR4_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                 const uint8_t* WEBP_RESTRICT top) {
   const __m128i one = _mm_set1_epi8(1);
   const int I = top[-2];
   const int J = top[-3];
@@ -924,8 +955,9 @@ static WEBP_INLINE void VR4_SSE2(uint8_t* dst,
   DST(0, 3) = AVG3(K, J, I);
 }
 
-static WEBP_INLINE void VL4_SSE2(uint8_t* dst,
-                                 const uint8_t* top) {  // Vertical-Left
+// Vertical-Left
+static WEBP_INLINE void VL4_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                 const uint8_t* WEBP_RESTRICT top) {
   const __m128i one = _mm_set1_epi8(1);
   const __m128i ABCDEFGH = _mm_loadl_epi64((const __m128i*)top);
   const __m128i BCDEFGH_ = _mm_srli_si128(ABCDEFGH, 1);
@@ -951,8 +983,9 @@ static WEBP_INLINE void VL4_SSE2(uint8_t* dst,
   DST(3, 3) = (extra_out >> 8) & 0xff;
 }
 
-static WEBP_INLINE void RD4_SSE2(uint8_t* dst,
-                                 const uint8_t* top) {  // Down-right
+// Down-right
+static WEBP_INLINE void RD4_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                 const uint8_t* WEBP_RESTRICT top) {
   const __m128i one = _mm_set1_epi8(1);
   const __m128i LKJIXABC = _mm_loadl_epi64((const __m128i*)(top - 5));
   const __m128i LKJIXABCD = _mm_insert_epi16(LKJIXABC, top[3], 4);
@@ -968,7 +1001,8 @@ static WEBP_INLINE void RD4_SSE2(uint8_t* dst,
   WebPInt32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3)));
 }
 
-static WEBP_INLINE void HU4_SSE2(uint8_t* dst, const uint8_t* top) {
+static WEBP_INLINE void HU4_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                 const uint8_t* WEBP_RESTRICT top) {
   const int I = top[-2];
   const int J = top[-3];
   const int K = top[-4];
@@ -983,7 +1017,8 @@ static WEBP_INLINE void HU4_SSE2(uint8_t* dst, const uint8_t* top) {
   DST(0, 3) = DST(1, 3) = DST(2, 3) = DST(3, 3) = L;
 }
 
-static WEBP_INLINE void HD4_SSE2(uint8_t* dst, const uint8_t* top) {
+static WEBP_INLINE void HD4_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                 const uint8_t* WEBP_RESTRICT top) {
   const int X = top[-1];
   const int I = top[-2];
   const int J = top[-3];
@@ -1006,7 +1041,8 @@ static WEBP_INLINE void HD4_SSE2(uint8_t* dst, const uint8_t* top) {
   DST(1, 3)             = AVG3(L, K, J);
 }
 
-static WEBP_INLINE void TM4_SSE2(uint8_t* dst, const uint8_t* top) {
+static WEBP_INLINE void TM4_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                 const uint8_t* WEBP_RESTRICT top) {
   const __m128i zero = _mm_setzero_si128();
   const __m128i top_values = _mm_cvtsi32_si128(WebPMemToInt32(top));
   const __m128i top_base = _mm_unpacklo_epi8(top_values, zero);
@@ -1028,7 +1064,8 @@ static WEBP_INLINE void TM4_SSE2(uint8_t* dst, const uint8_t* top) {
 
 // Left samples are top[-5 .. -2], top_left is top[-1], top are
 // located at top[0..3], and top right is top[4..7]
-static void Intra4Preds_SSE2(uint8_t* dst, const uint8_t* top) {
+static void Intra4Preds_SSE2(uint8_t* WEBP_RESTRICT dst,
+                             const uint8_t* WEBP_RESTRICT top) {
   DC4_SSE2(I4DC4 + dst, top);
   TM4_SSE2(I4TM4 + dst, top);
   VE4_SSE2(I4VE4 + dst, top);
@@ -1044,8 +1081,9 @@ static void Intra4Preds_SSE2(uint8_t* dst, const uint8_t* top) {
 //------------------------------------------------------------------------------
 // Chroma 8x8 prediction (paragraph 12.2)
 
-static void IntraChromaPreds_SSE2(uint8_t* dst, const uint8_t* left,
-                                  const uint8_t* top) {
+static void IntraChromaPreds_SSE2(uint8_t* WEBP_RESTRICT dst,
+                                  const uint8_t* WEBP_RESTRICT left,
+                                  const uint8_t* WEBP_RESTRICT top) {
   // U block
   DC8uvMode_SSE2(C8DC8 + dst, left, top);
   VerticalPred_SSE2(C8VE8 + dst, top, 8);
@@ -1064,8 +1102,9 @@ static void IntraChromaPreds_SSE2(uint8_t* dst, const uint8_t* left,
 //------------------------------------------------------------------------------
 // luma 16x16 prediction (paragraph 12.3)
 
-static void Intra16Preds_SSE2(uint8_t* dst,
-                              const uint8_t* left, const uint8_t* top) {
+static void Intra16Preds_SSE2(uint8_t* WEBP_RESTRICT dst,
+                              const uint8_t* WEBP_RESTRICT left,
+                              const uint8_t* WEBP_RESTRICT top) {
   DC16Mode_SSE2(I16DC16 + dst, left, top);
   VerticalPred_SSE2(I16VE16 + dst, top, 16);
   HorizontalPred_SSE2(I16HE16 + dst, left, 16);
@@ -1092,7 +1131,8 @@ static WEBP_INLINE void SubtractAndAccumulate_SSE2(const __m128i a,
   *sum = _mm_add_epi32(sum1, sum2);
 }
 
-static WEBP_INLINE int SSE_16xN_SSE2(const uint8_t* a, const uint8_t* b,
+static WEBP_INLINE int SSE_16xN_SSE2(const uint8_t* WEBP_RESTRICT a,
+                                     const uint8_t* WEBP_RESTRICT b,
                                      int num_pairs) {
   __m128i sum = _mm_setzero_si128();
   int32_t tmp[4];
@@ -1114,18 +1154,21 @@ static WEBP_INLINE int SSE_16xN_SSE2(const uint8_t* a, const uint8_t* b,
   return (tmp[3] + tmp[2] + tmp[1] + tmp[0]);
 }
 
-static int SSE16x16_SSE2(const uint8_t* a, const uint8_t* b) {
+static int SSE16x16_SSE2(const uint8_t* WEBP_RESTRICT a,
+                         const uint8_t* WEBP_RESTRICT b) {
   return SSE_16xN_SSE2(a, b, 8);
 }
 
-static int SSE16x8_SSE2(const uint8_t* a, const uint8_t* b) {
+static int SSE16x8_SSE2(const uint8_t* WEBP_RESTRICT a,
+                        const uint8_t* WEBP_RESTRICT b) {
   return SSE_16xN_SSE2(a, b, 4);
 }
 
 #define LOAD_8x16b(ptr) \
   _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i*)(ptr)), zero)
 
-static int SSE8x8_SSE2(const uint8_t* a, const uint8_t* b) {
+static int SSE8x8_SSE2(const uint8_t* WEBP_RESTRICT a,
+                       const uint8_t* WEBP_RESTRICT b) {
   const __m128i zero = _mm_setzero_si128();
   int num_pairs = 4;
   __m128i sum = zero;
@@ -1152,7 +1195,8 @@ static int SSE8x8_SSE2(const uint8_t* a, const uint8_t* b) {
 }
 #undef LOAD_8x16b
 
-static int SSE4x4_SSE2(const uint8_t* a, const uint8_t* b) {
+static int SSE4x4_SSE2(const uint8_t* WEBP_RESTRICT a,
+                       const uint8_t* WEBP_RESTRICT b) {
   const __m128i zero = _mm_setzero_si128();
 
   // Load values. Note that we read 8 pixels instead of 4,
@@ -1189,7 +1233,7 @@ static int SSE4x4_SSE2(const uint8_t* a, const uint8_t* b) {
 
 //------------------------------------------------------------------------------
 
-static void Mean16x4_SSE2(const uint8_t* ref, uint32_t dc[4]) {
+static void Mean16x4_SSE2(const uint8_t* WEBP_RESTRICT ref, uint32_t dc[4]) {
   const __m128i mask = _mm_set1_epi16(0x00ff);
   const __m128i a0 = _mm_loadu_si128((const __m128i*)&ref[BPS * 0]);
   const __m128i a1 = _mm_loadu_si128((const __m128i*)&ref[BPS * 1]);
@@ -1227,8 +1271,9 @@ static void Mean16x4_SSE2(const uint8_t* ref, uint32_t dc[4]) {
 // Hadamard transform
 // Returns the weighted sum of the absolute value of transformed coefficients.
 // w[] contains a row-major 4 by 4 symmetric matrix.
-static int TTransform_SSE2(const uint8_t* inA, const uint8_t* inB,
-                           const uint16_t* const w) {
+static int TTransform_SSE2(const uint8_t* WEBP_RESTRICT inA,
+                           const uint8_t* WEBP_RESTRICT inB,
+                           const uint16_t* WEBP_RESTRICT const w) {
   int32_t sum[4];
   __m128i tmp_0, tmp_1, tmp_2, tmp_3;
   const __m128i zero = _mm_setzero_si128();
@@ -1328,14 +1373,16 @@ static int TTransform_SSE2(const uint8_t* inA, const uint8_t* inB,
   return sum[0] + sum[1] + sum[2] + sum[3];
 }
 
-static int Disto4x4_SSE2(const uint8_t* const a, const uint8_t* const b,
-                         const uint16_t* const w) {
+static int Disto4x4_SSE2(const uint8_t* WEBP_RESTRICT const a,
+                         const uint8_t* WEBP_RESTRICT const b,
+                         const uint16_t* WEBP_RESTRICT const w) {
   const int diff_sum = TTransform_SSE2(a, b, w);
   return abs(diff_sum) >> 5;
 }
 
-static int Disto16x16_SSE2(const uint8_t* const a, const uint8_t* const b,
-                           const uint16_t* const w) {
+static int Disto16x16_SSE2(const uint8_t* WEBP_RESTRICT const a,
+                           const uint8_t* WEBP_RESTRICT const b,
+                           const uint16_t* WEBP_RESTRICT const w) {
   int D = 0;
   int x, y;
   for (y = 0; y < 16 * BPS; y += 4 * BPS) {
@@ -1350,9 +1397,10 @@ static int Disto16x16_SSE2(const uint8_t* const a, const uint8_t* const b,
 // Quantization
 //
 
-static WEBP_INLINE int DoQuantizeBlock_SSE2(int16_t in[16], int16_t out[16],
-                                            const uint16_t* const sharpen,
-                                            const VP8Matrix* const mtx) {
+static WEBP_INLINE int DoQuantizeBlock_SSE2(
+    int16_t in[16], int16_t out[16],
+    const uint16_t* WEBP_RESTRICT const sharpen,
+    const VP8Matrix* WEBP_RESTRICT const mtx) {
   const __m128i max_coeff_2047 = _mm_set1_epi16(MAX_LEVEL);
   const __m128i zero = _mm_setzero_si128();
   __m128i coeff0, coeff8;
@@ -1463,17 +1511,17 @@ static WEBP_INLINE int DoQuantizeBlock_SSE2(int16_t in[16], int16_t out[16],
 }
 
 static int QuantizeBlock_SSE2(int16_t in[16], int16_t out[16],
-                              const VP8Matrix* const mtx) {
+                              const VP8Matrix* WEBP_RESTRICT const mtx) {
   return DoQuantizeBlock_SSE2(in, out, &mtx->sharpen_[0], mtx);
 }
 
 static int QuantizeBlockWHT_SSE2(int16_t in[16], int16_t out[16],
-                                 const VP8Matrix* const mtx) {
+                                 const VP8Matrix* WEBP_RESTRICT const mtx) {
   return DoQuantizeBlock_SSE2(in, out, NULL, mtx);
 }
 
 static int Quantize2Blocks_SSE2(int16_t in[32], int16_t out[32],
-                                const VP8Matrix* const mtx) {
+                                const VP8Matrix* WEBP_RESTRICT const mtx) {
   int nz;
   const uint16_t* const sharpen = &mtx->sharpen_[0];
   nz  = DoQuantizeBlock_SSE2(in + 0 * 16, out + 0 * 16, sharpen, mtx) << 0;
diff --git a/src/dsp/enc_sse41.c b/src/dsp/enc_sse41.c
index 924035a6..613c44cf 100644
--- a/src/dsp/enc_sse41.c
+++ b/src/dsp/enc_sse41.c
@@ -23,9 +23,10 @@
 //------------------------------------------------------------------------------
 // Compute susceptibility based on DCT-coeff histograms.
 
-static void CollectHistogram_SSE41(const uint8_t* ref, const uint8_t* pred,
+static void CollectHistogram_SSE41(const uint8_t* WEBP_RESTRICT ref,
+                                   const uint8_t* WEBP_RESTRICT pred,
                                    int start_block, int end_block,
-                                   VP8Histogram* const histo) {
+                                   VP8Histogram* WEBP_RESTRICT const histo) {
   const __m128i max_coeff_thresh = _mm_set1_epi16(MAX_COEFF_THRESH);
   int j;
   int distribution[MAX_COEFF_THRESH + 1] = { 0 };
@@ -168,14 +169,16 @@ static int TTransform_SSE41(const uint8_t* inA, const uint8_t* inB,
   return sum[0] + sum[1] + sum[2] + sum[3];
 }
 
-static int Disto4x4_SSE41(const uint8_t* const a, const uint8_t* const b,
-                          const uint16_t* const w) {
+static int Disto4x4_SSE41(const uint8_t* WEBP_RESTRICT const a,
+                          const uint8_t* WEBP_RESTRICT const b,
+                          const uint16_t* WEBP_RESTRICT const w) {
   const int diff_sum = TTransform_SSE41(a, b, w);
   return abs(diff_sum) >> 5;
 }
 
-static int Disto16x16_SSE41(const uint8_t* const a, const uint8_t* const b,
-                            const uint16_t* const w) {
+static int Disto16x16_SSE41(const uint8_t* WEBP_RESTRICT const a,
+                            const uint8_t* WEBP_RESTRICT const b,
+                            const uint16_t* WEBP_RESTRICT const w) {
   int D = 0;
   int x, y;
   for (y = 0; y < 16 * BPS; y += 4 * BPS) {
@@ -301,17 +304,17 @@ static WEBP_INLINE int DoQuantizeBlock_SSE41(int16_t in[16], int16_t out[16],
 #undef PSHUFB_CST
 
 static int QuantizeBlock_SSE41(int16_t in[16], int16_t out[16],
-                               const VP8Matrix* const mtx) {
+                               const VP8Matrix* WEBP_RESTRICT const mtx) {
   return DoQuantizeBlock_SSE41(in, out, &mtx->sharpen_[0], mtx);
 }
 
 static int QuantizeBlockWHT_SSE41(int16_t in[16], int16_t out[16],
-                                  const VP8Matrix* const mtx) {
+                                  const VP8Matrix* WEBP_RESTRICT const mtx) {
   return DoQuantizeBlock_SSE41(in, out, NULL, mtx);
 }
 
 static int Quantize2Blocks_SSE41(int16_t in[32], int16_t out[32],
-                                 const VP8Matrix* const mtx) {
+                                 const VP8Matrix* WEBP_RESTRICT const mtx) {
   int nz;
   const uint16_t* const sharpen = &mtx->sharpen_[0];
   nz  = DoQuantizeBlock_SSE41(in + 0 * 16, out + 0 * 16, sharpen, mtx) << 0;

From 04d4b4f38781490fc1831a3876d3c3cb8a287e24 Mon Sep 17 00:00:00 2001
From: James Zern <jzern@google.com>
Date: Fri, 16 Aug 2024 13:52:32 -0700
Subject: [PATCH 4/8] dsp/filters*: use WEBP_RESTRICT qualifier

Better stack/register usage in SSE2/NEON code and improved vectorization
of the C code with ndk r27/gcc-13/clang-16.

This only affects non-vector pointers; any vector pointers are left as a
follow up.

Change-Id: I32b53dd38bfc7e2231d875409e7dfda7c513cfb6
---
 src/dsp/dsp.h                 |  5 +++--
 src/dsp/filters.c             | 33 +++++++++++++++------------
 src/dsp/filters_mips_dsp_r2.c | 40 ++++++++++++++++++---------------
 src/dsp/filters_msa.c         | 27 +++++++++++++---------
 src/dsp/filters_neon.c        | 42 ++++++++++++++++++++---------------
 src/dsp/filters_sse2.c        | 41 ++++++++++++++++++++--------------
 6 files changed, 108 insertions(+), 80 deletions(-)

diff --git a/src/dsp/dsp.h b/src/dsp/dsp.h
index ec0302f2..13df7dfa 100644
--- a/src/dsp/dsp.h
+++ b/src/dsp/dsp.h
@@ -497,8 +497,9 @@ typedef enum {     // Filter types.
   WEBP_FILTER_FAST
 } WEBP_FILTER_TYPE;
 
-typedef void (*WebPFilterFunc)(const uint8_t* in, int width, int height,
-                               int stride, uint8_t* out);
+typedef void (*WebPFilterFunc)(const uint8_t* WEBP_RESTRICT in,
+                               int width, int height, int stride,
+                               uint8_t* WEBP_RESTRICT out);
 // In-place un-filtering.
 // Warning! 'prev_line' pointer can be equal to 'cur_line' or 'preds'.
 typedef void (*WebPUnfilterFunc)(const uint8_t* prev_line, const uint8_t* preds,
diff --git a/src/dsp/filters.c b/src/dsp/filters.c
index 9c1fa919..f5e1e5f9 100644
--- a/src/dsp/filters.c
+++ b/src/dsp/filters.c
@@ -23,14 +23,16 @@
   do {                                                                         \
     assert((in) != NULL);                                                      \
     assert((out) != NULL);                                                     \
+    assert((in) != (out));                                                     \
     assert(width > 0);                                                         \
     assert(height > 0);                                                        \
     assert(stride >= width);                                                   \
   } while (0)
 
 #if !WEBP_NEON_OMIT_C_CODE
-static WEBP_INLINE void PredictLine_C(const uint8_t* src, const uint8_t* pred,
-                                      uint8_t* dst, int length) {
+static WEBP_INLINE void PredictLine_C(const uint8_t* WEBP_RESTRICT src,
+                                      const uint8_t* WEBP_RESTRICT pred,
+                                      uint8_t* WEBP_RESTRICT dst, int length) {
   int i;
   for (i = 0; i < length; ++i) dst[i] = (uint8_t)(src[i] - pred[i]);
 }
@@ -38,9 +40,9 @@ static WEBP_INLINE void PredictLine_C(const uint8_t* src, const uint8_t* pred,
 //------------------------------------------------------------------------------
 // Horizontal filter.
 
-static WEBP_INLINE void DoHorizontalFilter_C(const uint8_t* in,
+static WEBP_INLINE void DoHorizontalFilter_C(const uint8_t* WEBP_RESTRICT in,
                                              int width, int height, int stride,
-                                             uint8_t* out) {
+                                             uint8_t* WEBP_RESTRICT out) {
   const uint8_t* preds = in;
   int row;
   DCHECK(in, out);
@@ -66,9 +68,9 @@ static WEBP_INLINE void DoHorizontalFilter_C(const uint8_t* in,
 //------------------------------------------------------------------------------
 // Vertical filter.
 
-static WEBP_INLINE void DoVerticalFilter_C(const uint8_t* in,
+static WEBP_INLINE void DoVerticalFilter_C(const uint8_t* WEBP_RESTRICT in,
                                            int width, int height, int stride,
-                                           uint8_t* out) {
+                                           uint8_t* WEBP_RESTRICT out) {
   const uint8_t* preds = in;
   int row;
   DCHECK(in, out);
@@ -99,9 +101,9 @@ static WEBP_INLINE int GradientPredictor_C(uint8_t a, uint8_t b, uint8_t c) {
 }
 
 #if !WEBP_NEON_OMIT_C_CODE
-static WEBP_INLINE void DoGradientFilter_C(const uint8_t* in,
+static WEBP_INLINE void DoGradientFilter_C(const uint8_t* WEBP_RESTRICT in,
                                            int width, int height, int stride,
-                                           uint8_t* out) {
+                                           uint8_t* WEBP_RESTRICT out) {
   const uint8_t* preds = in;
   int row;
   DCHECK(in, out);
@@ -136,18 +138,21 @@ static WEBP_INLINE void DoGradientFilter_C(const uint8_t* in,
 //------------------------------------------------------------------------------
 
 #if !WEBP_NEON_OMIT_C_CODE
-static void HorizontalFilter_C(const uint8_t* data, int width, int height,
-                               int stride, uint8_t* filtered_data) {
+static void HorizontalFilter_C(const uint8_t* WEBP_RESTRICT data,
+                               int width, int height, int stride,
+                               uint8_t* WEBP_RESTRICT filtered_data) {
   DoHorizontalFilter_C(data, width, height, stride, filtered_data);
 }
 
-static void VerticalFilter_C(const uint8_t* data, int width, int height,
-                             int stride, uint8_t* filtered_data) {
+static void VerticalFilter_C(const uint8_t* WEBP_RESTRICT data,
+                             int width, int height, int stride,
+                             uint8_t* WEBP_RESTRICT filtered_data) {
   DoVerticalFilter_C(data, width, height, stride, filtered_data);
 }
 
-static void GradientFilter_C(const uint8_t* data, int width, int height,
-                             int stride, uint8_t* filtered_data) {
+static void GradientFilter_C(const uint8_t* WEBP_RESTRICT data,
+                             int width, int height, int stride,
+                             uint8_t* WEBP_RESTRICT filtered_data) {
   DoGradientFilter_C(data, width, height, stride, filtered_data);
 }
 #endif  // !WEBP_NEON_OMIT_C_CODE
diff --git a/src/dsp/filters_mips_dsp_r2.c b/src/dsp/filters_mips_dsp_r2.c
index 2c2c63c0..c62bb872 100644
--- a/src/dsp/filters_mips_dsp_r2.c
+++ b/src/dsp/filters_mips_dsp_r2.c
@@ -26,8 +26,9 @@
 
 #define DCHECK(in, out)                                                        \
   do {                                                                         \
-    assert(in != NULL);                                                        \
-    assert(out != NULL);                                                       \
+    assert((in) != NULL);                                                      \
+    assert((out) != NULL);                                                     \
+    assert((in) != (out));                                                     \
     assert(width > 0);                                                         \
     assert(height > 0);                                                        \
     assert(stride >= width);                                                   \
@@ -101,7 +102,8 @@
     );                                                                         \
   } while (0)
 
-static WEBP_INLINE void PredictLine_MIPSdspR2(const uint8_t* src, uint8_t* dst,
+static WEBP_INLINE void PredictLine_MIPSdspR2(const uint8_t* WEBP_RESTRICT src,
+                                              uint8_t* WEBP_RESTRICT dst,
                                               int length) {
   DO_PREDICT_LINE(src, dst, length, 0);
 }
@@ -191,9 +193,9 @@ static WEBP_INLINE void PredictLine_MIPSdspR2(const uint8_t* src, uint8_t* dst,
     }                                                                          \
   } while (0)
 
-static WEBP_INLINE void DoHorizontalFilter_MIPSdspR2(const uint8_t* in,
-                                                     int width, int height,
-                                                     int stride, uint8_t* out) {
+static WEBP_INLINE void DoHorizontalFilter_MIPSdspR2(
+    const uint8_t* WEBP_RESTRICT in, int width, int height, int stride,
+    uint8_t* WEBP_RESTRICT out) {
   const uint8_t* preds = in;
   int row;
   DCHECK(in, out);
@@ -210,9 +212,9 @@ static WEBP_INLINE void DoHorizontalFilter_MIPSdspR2(const uint8_t* in,
 }
 #undef FILTER_LINE_BY_LINE
 
-static void HorizontalFilter_MIPSdspR2(const uint8_t* data,
-                                       int width, int height,
-                                       int stride, uint8_t* filtered_data) {
+static void HorizontalFilter_MIPSdspR2(const uint8_t* WEBP_RESTRICT data,
+                                       int width, int height, int stride,
+                                       uint8_t* WEBP_RESTRICT filtered_data) {
   DoHorizontalFilter_MIPSdspR2(data, width, height, stride, filtered_data);
 }
 
@@ -228,9 +230,9 @@ static void HorizontalFilter_MIPSdspR2(const uint8_t* data,
     }                                                                          \
   } while (0)
 
-static WEBP_INLINE void DoVerticalFilter_MIPSdspR2(const uint8_t* in,
-                                                   int width, int height,
-                                                   int stride, uint8_t* out) {
+static WEBP_INLINE void DoVerticalFilter_MIPSdspR2(
+    const uint8_t* WEBP_RESTRICT in, int width, int height, int stride,
+    uint8_t* WEBP_RESTRICT out) {
   const uint8_t* preds = in;
   int row;
   DCHECK(in, out);
@@ -247,8 +249,9 @@ static WEBP_INLINE void DoVerticalFilter_MIPSdspR2(const uint8_t* in,
 }
 #undef FILTER_LINE_BY_LINE
 
-static void VerticalFilter_MIPSdspR2(const uint8_t* data, int width, int height,
-                                     int stride, uint8_t* filtered_data) {
+static void VerticalFilter_MIPSdspR2(const uint8_t* WEBP_RESTRICT data,
+                                     int width, int height, int stride,
+                                     uint8_t* WEBP_RESTRICT filtered_data) {
   DoVerticalFilter_MIPSdspR2(data, width, height, stride, filtered_data);
 }
 
@@ -284,9 +287,9 @@ static int GradientPredictor_MIPSdspR2(uint8_t a, uint8_t b, uint8_t c) {
     }                                                                          \
   } while (0)
 
-static void DoGradientFilter_MIPSdspR2(const uint8_t* in,
+static void DoGradientFilter_MIPSdspR2(const uint8_t* WEBP_RESTRICT in,
                                        int width, int height, int stride,
-                                       uint8_t* out) {
+                                       uint8_t* WEBP_RESTRICT out) {
   const uint8_t* preds = in;
   int row;
   DCHECK(in, out);
@@ -303,8 +306,9 @@ static void DoGradientFilter_MIPSdspR2(const uint8_t* in,
 }
 #undef FILTER_LINE_BY_LINE
 
-static void GradientFilter_MIPSdspR2(const uint8_t* data, int width, int height,
-                                     int stride, uint8_t* filtered_data) {
+static void GradientFilter_MIPSdspR2(const uint8_t* WEBP_RESTRICT data,
+                                     int width, int height, int stride,
+                                     uint8_t* WEBP_RESTRICT filtered_data) {
   DoGradientFilter_MIPSdspR2(data, width, height, stride, filtered_data);
 }
 
diff --git a/src/dsp/filters_msa.c b/src/dsp/filters_msa.c
index 33a1b20b..ae3d3699 100644
--- a/src/dsp/filters_msa.c
+++ b/src/dsp/filters_msa.c
@@ -21,7 +21,8 @@
 
 static WEBP_INLINE void PredictLineInverse0(const uint8_t* src,
                                             const uint8_t* pred,
-                                            uint8_t* dst, int length) {
+                                            uint8_t* WEBP_RESTRICT dst,
+                                            int length) {
   v16u8 src0, pred0, dst0;
   assert(length >= 0);
   while (length >= 32) {
@@ -58,8 +59,9 @@ static WEBP_INLINE void PredictLineInverse0(const uint8_t* src,
 
 #define DCHECK(in, out)        \
   do {                         \
-    assert(in != NULL);        \
-    assert(out != NULL);       \
+    assert((in) != NULL);      \
+    assert((out) != NULL);     \
+    assert((in) != (out));     \
     assert(width > 0);         \
     assert(height > 0);        \
     assert(stride >= width);   \
@@ -68,8 +70,9 @@ static WEBP_INLINE void PredictLineInverse0(const uint8_t* src,
 //------------------------------------------------------------------------------
 // Horrizontal filter
 
-static void HorizontalFilter_MSA(const uint8_t* data, int width, int height,
-                                 int stride, uint8_t* filtered_data) {
+static void HorizontalFilter_MSA(const uint8_t* WEBP_RESTRICT data,
+                                 int width, int height, int stride,
+                                 uint8_t* WEBP_RESTRICT filtered_data) {
   const uint8_t* preds = data;
   const uint8_t* in = data;
   uint8_t* out = filtered_data;
@@ -99,8 +102,8 @@ static void HorizontalFilter_MSA(const uint8_t* data, int width, int height,
 
 static WEBP_INLINE void PredictLineGradient(const uint8_t* pinput,
                                             const uint8_t* ppred,
-                                            uint8_t* poutput, int stride,
-                                            int size) {
+                                            uint8_t* WEBP_RESTRICT poutput,
+                                            int stride, int size) {
   int w;
   const v16i8 zero = { 0 };
   while (size >= 16) {
@@ -131,8 +134,9 @@ static WEBP_INLINE void PredictLineGradient(const uint8_t* pinput,
 }
 
 
-static void GradientFilter_MSA(const uint8_t* data, int width, int height,
-                               int stride, uint8_t* filtered_data) {
+static void GradientFilter_MSA(const uint8_t* WEBP_RESTRICT data,
+                               int width, int height, int stride,
+                               uint8_t* WEBP_RESTRICT filtered_data) {
   const uint8_t* in = data;
   const uint8_t* preds = data;
   uint8_t* out = filtered_data;
@@ -159,8 +163,9 @@ static void GradientFilter_MSA(const uint8_t* data, int width, int height,
 //------------------------------------------------------------------------------
 // Vertical filter
 
-static void VerticalFilter_MSA(const uint8_t* data, int width, int height,
-                               int stride, uint8_t* filtered_data) {
+static void VerticalFilter_MSA(const uint8_t* WEBP_RESTRICT data,
+                               int width, int height, int stride,
+                               uint8_t* WEBP_RESTRICT filtered_data) {
   const uint8_t* in = data;
   const uint8_t* preds = data;
   uint8_t* out = filtered_data;
diff --git a/src/dsp/filters_neon.c b/src/dsp/filters_neon.c
index 0b0a8421..4df10172 100644
--- a/src/dsp/filters_neon.c
+++ b/src/dsp/filters_neon.c
@@ -23,8 +23,9 @@
 
 #define DCHECK(in, out)                                                        \
   do {                                                                         \
-    assert(in != NULL);                                                        \
-    assert(out != NULL);                                                       \
+    assert((in) != NULL);                                                      \
+    assert((out) != NULL);                                                     \
+    assert((in) != (out));                                                     \
     assert(width > 0);                                                         \
     assert(height > 0);                                                        \
     assert(stride >= width);                                                   \
@@ -44,7 +45,7 @@
 #define ROTATE_RIGHT_N(A, N)   vext_u8((A), (A), (8 - (N)) % 8)
 
 static void PredictLine_NEON(const uint8_t* src, const uint8_t* pred,
-                             uint8_t* dst, int length) {
+                             uint8_t* WEBP_RESTRICT dst, int length) {
   int i;
   assert(length >= 0);
   for (i = 0; i + 16 <= length; i += 16) {
@@ -57,16 +58,17 @@ static void PredictLine_NEON(const uint8_t* src, const uint8_t* pred,
 }
 
 // Special case for left-based prediction (when preds==dst-1 or preds==src-1).
-static void PredictLineLeft_NEON(const uint8_t* src, uint8_t* dst, int length) {
+static void PredictLineLeft_NEON(const uint8_t* WEBP_RESTRICT src,
+                                 uint8_t* WEBP_RESTRICT dst, int length) {
   PredictLine_NEON(src, src - 1, dst, length);
 }
 
 //------------------------------------------------------------------------------
 // Horizontal filter.
 
-static WEBP_INLINE void DoHorizontalFilter_NEON(const uint8_t* in,
-                                                int width, int height,
-                                                int stride, uint8_t* out) {
+static WEBP_INLINE void DoHorizontalFilter_NEON(
+    const uint8_t* WEBP_RESTRICT in, int width, int height, int stride,
+    uint8_t* WEBP_RESTRICT out) {
   int row;
   DCHECK(in, out);
 
@@ -86,17 +88,18 @@ static WEBP_INLINE void DoHorizontalFilter_NEON(const uint8_t* in,
   }
 }
 
-static void HorizontalFilter_NEON(const uint8_t* data, int width, int height,
-                                  int stride, uint8_t* filtered_data) {
+static void HorizontalFilter_NEON(const uint8_t* WEBP_RESTRICT data,
+                                  int width, int height, int stride,
+                                  uint8_t* WEBP_RESTRICT filtered_data) {
   DoHorizontalFilter_NEON(data, width, height, stride, filtered_data);
 }
 
 //------------------------------------------------------------------------------
 // Vertical filter.
 
-static WEBP_INLINE void DoVerticalFilter_NEON(const uint8_t* in,
+static WEBP_INLINE void DoVerticalFilter_NEON(const uint8_t* WEBP_RESTRICT in,
                                               int width, int height, int stride,
-                                              uint8_t* out) {
+                                              uint8_t* WEBP_RESTRICT out) {
   int row;
   DCHECK(in, out);
 
@@ -115,8 +118,9 @@ static WEBP_INLINE void DoVerticalFilter_NEON(const uint8_t* in,
   }
 }
 
-static void VerticalFilter_NEON(const uint8_t* data, int width, int height,
-                                int stride, uint8_t* filtered_data) {
+static void VerticalFilter_NEON(const uint8_t* WEBP_RESTRICT data,
+                                int width, int height, int stride,
+                                uint8_t* WEBP_RESTRICT filtered_data) {
   DoVerticalFilter_NEON(data, width, height, stride, filtered_data);
 }
 
@@ -130,7 +134,8 @@ static WEBP_INLINE int GradientPredictor_C(uint8_t a, uint8_t b, uint8_t c) {
 
 static void GradientPredictDirect_NEON(const uint8_t* const row,
                                        const uint8_t* const top,
-                                       uint8_t* const out, int length) {
+                                       uint8_t* WEBP_RESTRICT const out,
+                                       int length) {
   int i;
   for (i = 0; i + 8 <= length; i += 8) {
     const uint8x8_t A = vld1_u8(&row[i - 1]);
@@ -146,9 +151,9 @@ static void GradientPredictDirect_NEON(const uint8_t* const row,
   }
 }
 
-static WEBP_INLINE void DoGradientFilter_NEON(const uint8_t* in,
+static WEBP_INLINE void DoGradientFilter_NEON(const uint8_t* WEBP_RESTRICT in,
                                               int width, int height, int stride,
-                                              uint8_t* out) {
+                                              uint8_t* WEBP_RESTRICT out) {
   int row;
   DCHECK(in, out);
 
@@ -167,8 +172,9 @@ static WEBP_INLINE void DoGradientFilter_NEON(const uint8_t* in,
   }
 }
 
-static void GradientFilter_NEON(const uint8_t* data, int width, int height,
-                                int stride, uint8_t* filtered_data) {
+static void GradientFilter_NEON(const uint8_t* WEBP_RESTRICT data,
+                                int width, int height, int stride,
+                                uint8_t* WEBP_RESTRICT filtered_data) {
   DoGradientFilter_NEON(data, width, height, stride, filtered_data);
 }
 
diff --git a/src/dsp/filters_sse2.c b/src/dsp/filters_sse2.c
index add4f493..d2ba7894 100644
--- a/src/dsp/filters_sse2.c
+++ b/src/dsp/filters_sse2.c
@@ -27,13 +27,15 @@
   do {                                                                         \
     assert((in) != NULL);                                                      \
     assert((out) != NULL);                                                     \
+    assert((in) != (out));                                                     \
     assert(width > 0);                                                         \
     assert(height > 0);                                                        \
     assert(stride >= width);                                                   \
   } while (0)
 
-static void PredictLineTop_SSE2(const uint8_t* src, const uint8_t* pred,
-                                uint8_t* dst, int length) {
+static void PredictLineTop_SSE2(const uint8_t* WEBP_RESTRICT src,
+                                const uint8_t* WEBP_RESTRICT pred,
+                                uint8_t* WEBP_RESTRICT dst, int length) {
   int i;
   const int max_pos = length & ~31;
   assert(length >= 0);
@@ -51,7 +53,8 @@ static void PredictLineTop_SSE2(const uint8_t* src, const uint8_t* pred,
 }
 
 // Special case for left-based prediction (when preds==dst-1 or preds==src-1).
-static void PredictLineLeft_SSE2(const uint8_t* src, uint8_t* dst, int length) {
+static void PredictLineLeft_SSE2(const uint8_t* WEBP_RESTRICT src,
+                                 uint8_t* WEBP_RESTRICT dst, int length) {
   int i;
   const int max_pos = length & ~31;
   assert(length >= 0);
@@ -71,9 +74,9 @@ static void PredictLineLeft_SSE2(const uint8_t* src, uint8_t* dst, int length) {
 //------------------------------------------------------------------------------
 // Horizontal filter.
 
-static WEBP_INLINE void DoHorizontalFilter_SSE2(const uint8_t* in,
-                                                int width, int height,
-                                                int stride, uint8_t* out) {
+static WEBP_INLINE void DoHorizontalFilter_SSE2(
+    const uint8_t* WEBP_RESTRICT in, int width, int height, int stride,
+    uint8_t* WEBP_RESTRICT out) {
   int row;
   DCHECK(in, out);
 
@@ -96,9 +99,9 @@ static WEBP_INLINE void DoHorizontalFilter_SSE2(const uint8_t* in,
 //------------------------------------------------------------------------------
 // Vertical filter.
 
-static WEBP_INLINE void DoVerticalFilter_SSE2(const uint8_t* in,
+static WEBP_INLINE void DoVerticalFilter_SSE2(const uint8_t* WEBP_RESTRICT in,
                                               int width, int height, int stride,
-                                              uint8_t* out) {
+                                              uint8_t* WEBP_RESTRICT out) {
   int row;
   DCHECK(in, out);
 
@@ -127,7 +130,8 @@ static WEBP_INLINE int GradientPredictor_SSE2(uint8_t a, uint8_t b, uint8_t c) {
 
 static void GradientPredictDirect_SSE2(const uint8_t* const row,
                                        const uint8_t* const top,
-                                       uint8_t* const out, int length) {
+                                       uint8_t* WEBP_RESTRICT const out,
+                                       int length) {
   const int max_pos = length & ~7;
   int i;
   const __m128i zero = _mm_setzero_si128();
@@ -151,9 +155,9 @@ static void GradientPredictDirect_SSE2(const uint8_t* const row,
   }
 }
 
-static WEBP_INLINE void DoGradientFilter_SSE2(const uint8_t* in,
+static WEBP_INLINE void DoGradientFilter_SSE2(const uint8_t* WEBP_RESTRICT in,
                                               int width, int height, int stride,
-                                              uint8_t* out) {
+                                              uint8_t* WEBP_RESTRICT out) {
   int row;
   DCHECK(in, out);
 
@@ -176,18 +180,21 @@ static WEBP_INLINE void DoGradientFilter_SSE2(const uint8_t* in,
 
 //------------------------------------------------------------------------------
 
-static void HorizontalFilter_SSE2(const uint8_t* data, int width, int height,
-                                  int stride, uint8_t* filtered_data) {
+static void HorizontalFilter_SSE2(const uint8_t* WEBP_RESTRICT data,
+                                  int width, int height, int stride,
+                                  uint8_t* WEBP_RESTRICT filtered_data) {
   DoHorizontalFilter_SSE2(data, width, height, stride, filtered_data);
 }
 
-static void VerticalFilter_SSE2(const uint8_t* data, int width, int height,
-                                int stride, uint8_t* filtered_data) {
+static void VerticalFilter_SSE2(const uint8_t* WEBP_RESTRICT data,
+                                int width, int height, int stride,
+                                uint8_t* WEBP_RESTRICT filtered_data) {
   DoVerticalFilter_SSE2(data, width, height, stride, filtered_data);
 }
 
-static void GradientFilter_SSE2(const uint8_t* data, int width, int height,
-                                int stride, uint8_t* filtered_data) {
+static void GradientFilter_SSE2(const uint8_t* WEBP_RESTRICT data,
+                                int width, int height, int stride,
+                                uint8_t* WEBP_RESTRICT filtered_data) {
   DoGradientFilter_SSE2(data, width, height, stride, filtered_data);
 }
 

From a32b436bd5878265286c3eb9c3284715c60913b2 Mon Sep 17 00:00:00 2001
From: James Zern <jzern@google.com>
Date: Fri, 16 Aug 2024 16:51:11 -0700
Subject: [PATCH 5/8] dsp/lossless*: use WEBP_RESTRICT qualifier

lossless_enc: better vectorization, most benefits seen in AddVector/Eq
              w/ndk r27/gcc-13/clang-16
lossless: minor reordering and some improvement to PredictorAdd5_SSE2
          w/gcc-13

This only affects non-vector pointers; any vector pointers are left as a
follow up.

Change-Id: I2356e314f391ee2f2c71f00bc6ee10097d3881e7
---
 src/dsp/dsp.h                      |  5 ++-
 src/dsp/lossless.c                 | 28 ++++++-------
 src/dsp/lossless.h                 | 65 +++++++++++++++++-------------
 src/dsp/lossless_common.h          | 18 ++++-----
 src/dsp/lossless_enc.c             | 63 ++++++++++++++++-------------
 src/dsp/lossless_enc_mips32.c      | 35 +++++++++-------
 src/dsp/lossless_enc_mips_dsp_r2.c | 16 ++++----
 src/dsp/lossless_enc_msa.c         |  4 +-
 src/dsp/lossless_enc_neon.c        |  5 ++-
 src/dsp/lossless_enc_sse2.c        | 42 +++++++++++--------
 src/dsp/lossless_enc_sse41.c       | 11 +++--
 src/dsp/lossless_neon.c            | 46 ++++++++++-----------
 src/dsp/lossless_sse2.c            | 38 +++++++++--------
 src/dsp/lossless_sse41.c           |  8 ++--
 14 files changed, 213 insertions(+), 171 deletions(-)

diff --git a/src/dsp/dsp.h b/src/dsp/dsp.h
index 13df7dfa..30a1d43f 100644
--- a/src/dsp/dsp.h
+++ b/src/dsp/dsp.h
@@ -131,9 +131,10 @@ typedef struct {
   int max_value;
   int last_non_zero;
 } VP8Histogram;
-typedef void (*VP8CHisto)(const uint8_t* ref, const uint8_t* pred,
+typedef void (*VP8CHisto)(const uint8_t* WEBP_RESTRICT ref,
+                          const uint8_t* WEBP_RESTRICT pred,
                           int start_block, int end_block,
-                          VP8Histogram* const histo);
+                          VP8Histogram* WEBP_RESTRICT const histo);
 extern VP8CHisto VP8CollectHistogram;
 // General-purpose util function to help VP8CollectHistogram().
 void VP8SetHistogramData(const int distribution[MAX_COEFF_THRESH + 1],
diff --git a/src/dsp/lossless.c b/src/dsp/lossless.c
index e7cbfd9e..a02443f1 100644
--- a/src/dsp/lossless.c
+++ b/src/dsp/lossless.c
@@ -182,13 +182,13 @@ uint32_t VP8LPredictor13_C(const uint32_t* const left,
 }
 
 static void PredictorAdd0_C(const uint32_t* in, const uint32_t* upper,
-                            int num_pixels, uint32_t* out) {
+                            int num_pixels, uint32_t* WEBP_RESTRICT out) {
   int x;
   (void)upper;
   for (x = 0; x < num_pixels; ++x) out[x] = VP8LAddPixels(in[x], ARGB_BLACK);
 }
 static void PredictorAdd1_C(const uint32_t* in, const uint32_t* upper,
-                            int num_pixels, uint32_t* out) {
+                            int num_pixels, uint32_t* WEBP_RESTRICT out) {
   int i;
   uint32_t left = out[-1];
   (void)upper;
@@ -441,8 +441,8 @@ static int is_big_endian(void) {
   return (tmp.b[0] != 1);
 }
 
-void VP8LConvertBGRAToRGB_C(const uint32_t* src,
-                            int num_pixels, uint8_t* dst) {
+void VP8LConvertBGRAToRGB_C(const uint32_t* WEBP_RESTRICT src,
+                            int num_pixels, uint8_t* WEBP_RESTRICT dst) {
   const uint32_t* const src_end = src + num_pixels;
   while (src < src_end) {
     const uint32_t argb = *src++;
@@ -452,8 +452,8 @@ void VP8LConvertBGRAToRGB_C(const uint32_t* src,
   }
 }
 
-void VP8LConvertBGRAToRGBA_C(const uint32_t* src,
-                             int num_pixels, uint8_t* dst) {
+void VP8LConvertBGRAToRGBA_C(const uint32_t* WEBP_RESTRICT src,
+                             int num_pixels, uint8_t* WEBP_RESTRICT dst) {
   const uint32_t* const src_end = src + num_pixels;
   while (src < src_end) {
     const uint32_t argb = *src++;
@@ -464,8 +464,8 @@ void VP8LConvertBGRAToRGBA_C(const uint32_t* src,
   }
 }
 
-void VP8LConvertBGRAToRGBA4444_C(const uint32_t* src,
-                                 int num_pixels, uint8_t* dst) {
+void VP8LConvertBGRAToRGBA4444_C(const uint32_t* WEBP_RESTRICT src,
+                                 int num_pixels, uint8_t* WEBP_RESTRICT dst) {
   const uint32_t* const src_end = src + num_pixels;
   while (src < src_end) {
     const uint32_t argb = *src++;
@@ -481,8 +481,8 @@ void VP8LConvertBGRAToRGBA4444_C(const uint32_t* src,
   }
 }
 
-void VP8LConvertBGRAToRGB565_C(const uint32_t* src,
-                               int num_pixels, uint8_t* dst) {
+void VP8LConvertBGRAToRGB565_C(const uint32_t* WEBP_RESTRICT src,
+                               int num_pixels, uint8_t* WEBP_RESTRICT dst) {
   const uint32_t* const src_end = src + num_pixels;
   while (src < src_end) {
     const uint32_t argb = *src++;
@@ -498,8 +498,8 @@ void VP8LConvertBGRAToRGB565_C(const uint32_t* src,
   }
 }
 
-void VP8LConvertBGRAToBGR_C(const uint32_t* src,
-                            int num_pixels, uint8_t* dst) {
+void VP8LConvertBGRAToBGR_C(const uint32_t* WEBP_RESTRICT src,
+                            int num_pixels, uint8_t* WEBP_RESTRICT dst) {
   const uint32_t* const src_end = src + num_pixels;
   while (src < src_end) {
     const uint32_t argb = *src++;
@@ -509,8 +509,8 @@ void VP8LConvertBGRAToBGR_C(const uint32_t* src,
   }
 }
 
-static void CopyOrSwap(const uint32_t* src, int num_pixels, uint8_t* dst,
-                       int swap_on_big_endian) {
+static void CopyOrSwap(const uint32_t* WEBP_RESTRICT src, int num_pixels,
+                       uint8_t* WEBP_RESTRICT dst, int swap_on_big_endian) {
   if (is_big_endian() == swap_on_big_endian) {
     const uint32_t* const src_end = src + num_pixels;
     while (src < src_end) {
diff --git a/src/dsp/lossless.h b/src/dsp/lossless.h
index 53278cda..a72e3b2a 100644
--- a/src/dsp/lossless.h
+++ b/src/dsp/lossless.h
@@ -18,6 +18,7 @@
 #include "src/webp/types.h"
 #include "src/webp/decode.h"
 
+#include "src/dsp/dsp.h"
 #include "src/enc/histogram_enc.h"
 #include "src/utils/utils.h"
 
@@ -60,7 +61,7 @@ uint32_t VP8LPredictor13_C(const uint32_t* const left,
 // These Add/Sub function expects upper[-1] and out[-1] to be readable.
 typedef void (*VP8LPredictorAddSubFunc)(const uint32_t* in,
                                         const uint32_t* upper, int num_pixels,
-                                        uint32_t* out);
+                                        uint32_t* WEBP_RESTRICT out);
 extern VP8LPredictorAddSubFunc VP8LPredictorsAdd[16];
 extern VP8LPredictorAddSubFunc VP8LPredictorsAdd_C[16];
 
@@ -91,8 +92,8 @@ void VP8LInverseTransform(const struct VP8LTransform* const transform,
                           const uint32_t* const in, uint32_t* const out);
 
 // Color space conversion.
-typedef void (*VP8LConvertFunc)(const uint32_t* src, int num_pixels,
-                                uint8_t* dst);
+typedef void (*VP8LConvertFunc)(const uint32_t* WEBP_RESTRICT src,
+                                int num_pixels, uint8_t* WEBP_RESTRICT dst);
 extern VP8LConvertFunc VP8LConvertBGRAToRGB;
 extern VP8LConvertFunc VP8LConvertBGRAToRGBA;
 extern VP8LConvertFunc VP8LConvertBGRAToRGBA4444;
@@ -145,29 +146,33 @@ void VP8LDspInit(void);
 
 typedef void (*VP8LProcessEncBlueAndRedFunc)(uint32_t* dst, int num_pixels);
 extern VP8LProcessEncBlueAndRedFunc VP8LSubtractGreenFromBlueAndRed;
-typedef void (*VP8LTransformColorFunc)(const VP8LMultipliers* const m,
-                                       uint32_t* dst, int num_pixels);
+typedef void (*VP8LTransformColorFunc)(
+    const VP8LMultipliers* WEBP_RESTRICT const m, uint32_t* WEBP_RESTRICT dst,
+    int num_pixels);
 extern VP8LTransformColorFunc VP8LTransformColor;
 typedef void (*VP8LCollectColorBlueTransformsFunc)(
-    const uint32_t* argb, int stride,
+    const uint32_t* WEBP_RESTRICT argb, int stride,
     int tile_width, int tile_height,
     int green_to_blue, int red_to_blue, uint32_t histo[]);
 extern VP8LCollectColorBlueTransformsFunc VP8LCollectColorBlueTransforms;
 
 typedef void (*VP8LCollectColorRedTransformsFunc)(
-    const uint32_t* argb, int stride,
+    const uint32_t* WEBP_RESTRICT argb, int stride,
     int tile_width, int tile_height,
     int green_to_red, uint32_t histo[]);
 extern VP8LCollectColorRedTransformsFunc VP8LCollectColorRedTransforms;
 
 // Expose some C-only fallback functions
-void VP8LTransformColor_C(const VP8LMultipliers* const m,
-                          uint32_t* data, int num_pixels);
-void VP8LSubtractGreenFromBlueAndRed_C(uint32_t* argb_data, int num_pixels);
-void VP8LCollectColorRedTransforms_C(const uint32_t* argb, int stride,
+void VP8LTransformColor_C(const VP8LMultipliers* WEBP_RESTRICT const m,
+                          uint32_t* WEBP_RESTRICT data, int num_pixels);
+void VP8LSubtractGreenFromBlueAndRed_C(uint32_t* WEBP_RESTRICT argb_data,
+                                       int num_pixels);
+void VP8LCollectColorRedTransforms_C(const uint32_t* WEBP_RESTRICT argb,
+                                     int stride,
                                      int tile_width, int tile_height,
                                      int green_to_red, uint32_t histo[]);
-void VP8LCollectColorBlueTransforms_C(const uint32_t* argb, int stride,
+void VP8LCollectColorBlueTransforms_C(const uint32_t* WEBP_RESTRICT argb,
+                                      int stride,
                                       int tile_width, int tile_height,
                                       int green_to_blue, int red_to_blue,
                                       uint32_t histo[]);
@@ -179,7 +184,8 @@ extern VP8LPredictorAddSubFunc VP8LPredictorsSub_C[16];
 // Huffman-cost related functions.
 
 typedef uint32_t (*VP8LCostFunc)(const uint32_t* population, int length);
-typedef uint32_t (*VP8LCostCombinedFunc)(const uint32_t* X, const uint32_t* Y,
+typedef uint32_t (*VP8LCostCombinedFunc)(const uint32_t* WEBP_RESTRICT X,
+                                         const uint32_t* WEBP_RESTRICT Y,
                                          int length);
 typedef uint64_t (*VP8LCombinedShannonEntropyFunc)(const uint32_t X[256],
                                                    const uint32_t Y[256]);
@@ -210,26 +216,30 @@ void VP8LBitEntropyInit(VP8LBitEntropy* const entropy);
 // codec specific heuristics.
 typedef void (*VP8LGetCombinedEntropyUnrefinedFunc)(
     const uint32_t X[], const uint32_t Y[], int length,
-    VP8LBitEntropy* const bit_entropy, VP8LStreaks* const stats);
+    VP8LBitEntropy* WEBP_RESTRICT const bit_entropy,
+    VP8LStreaks* WEBP_RESTRICT const stats);
 extern VP8LGetCombinedEntropyUnrefinedFunc VP8LGetCombinedEntropyUnrefined;
 
 // Get the entropy for the distribution 'X'.
-typedef void (*VP8LGetEntropyUnrefinedFunc)(const uint32_t X[], int length,
-                                            VP8LBitEntropy* const bit_entropy,
-                                            VP8LStreaks* const stats);
+typedef void (*VP8LGetEntropyUnrefinedFunc)(
+    const uint32_t X[], int length,
+    VP8LBitEntropy* WEBP_RESTRICT const bit_entropy,
+    VP8LStreaks* WEBP_RESTRICT const stats);
 extern VP8LGetEntropyUnrefinedFunc VP8LGetEntropyUnrefined;
 
 void VP8LBitsEntropyUnrefined(const uint32_t* const array, int n,
                               VP8LBitEntropy* const entropy);
 
-typedef void (*VP8LAddVectorFunc)(const uint32_t* a, const uint32_t* b,
-                                  uint32_t* out, int size);
+typedef void (*VP8LAddVectorFunc)(const uint32_t* WEBP_RESTRICT a,
+                                  const uint32_t* WEBP_RESTRICT b,
+                                  uint32_t* WEBP_RESTRICT out, int size);
 extern VP8LAddVectorFunc VP8LAddVector;
-typedef void (*VP8LAddVectorEqFunc)(const uint32_t* a, uint32_t* out, int size);
+typedef void (*VP8LAddVectorEqFunc)(const uint32_t* WEBP_RESTRICT a,
+                                    uint32_t* WEBP_RESTRICT out, int size);
 extern VP8LAddVectorEqFunc VP8LAddVectorEq;
-void VP8LHistogramAdd(const VP8LHistogram* const a,
-                      const VP8LHistogram* const b,
-                      VP8LHistogram* const out);
+void VP8LHistogramAdd(const VP8LHistogram* WEBP_RESTRICT const a,
+                      const VP8LHistogram* WEBP_RESTRICT const b,
+                      VP8LHistogram* WEBP_RESTRICT const out);
 
 // -----------------------------------------------------------------------------
 // PrefixEncode()
@@ -239,11 +249,12 @@ typedef int (*VP8LVectorMismatchFunc)(const uint32_t* const array1,
 // Returns the first index where array1 and array2 are different.
 extern VP8LVectorMismatchFunc VP8LVectorMismatch;
 
-typedef void (*VP8LBundleColorMapFunc)(const uint8_t* const row, int width,
-                                       int xbits, uint32_t* dst);
+typedef void (*VP8LBundleColorMapFunc)(const uint8_t* WEBP_RESTRICT const row,
+                                       int width, int xbits,
+                                       uint32_t* WEBP_RESTRICT dst);
 extern VP8LBundleColorMapFunc VP8LBundleColorMap;
-void VP8LBundleColorMap_C(const uint8_t* const row, int width, int xbits,
-                          uint32_t* dst);
+void VP8LBundleColorMap_C(const uint8_t* WEBP_RESTRICT const row,
+                          int width, int xbits, uint32_t* WEBP_RESTRICT dst);
 
 // Must be called before calling any of the above methods.
 void VP8LEncDspInit(void);
diff --git a/src/dsp/lossless_common.h b/src/dsp/lossless_common.h
index 33f2c4dc..66eadf1f 100644
--- a/src/dsp/lossless_common.h
+++ b/src/dsp/lossless_common.h
@@ -194,15 +194,15 @@ uint32_t VP8LSubPixels(uint32_t a, uint32_t b) {
 
 // The predictor is added to the output pixel (which
 // is therefore considered as a residual) to get the final prediction.
-#define GENERATE_PREDICTOR_ADD(PREDICTOR, PREDICTOR_ADD)             \
-static void PREDICTOR_ADD(const uint32_t* in, const uint32_t* upper, \
-                          int num_pixels, uint32_t* out) {           \
-  int x;                                                             \
-  assert(upper != NULL);                                             \
-  for (x = 0; x < num_pixels; ++x) {                                 \
-    const uint32_t pred = (PREDICTOR)(&out[x - 1], upper + x);       \
-    out[x] = VP8LAddPixels(in[x], pred);                             \
-  }                                                                  \
+#define GENERATE_PREDICTOR_ADD(PREDICTOR, PREDICTOR_ADD)                 \
+static void PREDICTOR_ADD(const uint32_t* in, const uint32_t* upper,     \
+                          int num_pixels, uint32_t* WEBP_RESTRICT out) { \
+  int x;                                                                 \
+  assert(upper != NULL);                                                 \
+  for (x = 0; x < num_pixels; ++x) {                                     \
+    const uint32_t pred = (PREDICTOR)(&out[x - 1], upper + x);           \
+    out[x] = VP8LAddPixels(in[x], pred);                                 \
+  }                                                                      \
 }
 
 #ifdef __cplusplus
diff --git a/src/dsp/lossless_enc.c b/src/dsp/lossless_enc.c
index 6ba1b373..7e621a71 100644
--- a/src/dsp/lossless_enc.c
+++ b/src/dsp/lossless_enc.c
@@ -359,8 +359,8 @@ void VP8LBitEntropyInit(VP8LBitEntropy* const entropy) {
   entropy->nonzero_code = VP8L_NON_TRIVIAL_SYM;
 }
 
-void VP8LBitsEntropyUnrefined(const uint32_t* const array, int n,
-                              VP8LBitEntropy* const entropy) {
+void VP8LBitsEntropyUnrefined(const uint32_t* WEBP_RESTRICT const array, int n,
+                              VP8LBitEntropy* WEBP_RESTRICT const entropy) {
   int i;
 
   VP8LBitEntropyInit(entropy);
@@ -380,8 +380,10 @@ void VP8LBitsEntropyUnrefined(const uint32_t* const array, int n,
 }
 
 static WEBP_INLINE void GetEntropyUnrefinedHelper(
-    uint32_t val, int i, uint32_t* const val_prev, int* const i_prev,
-    VP8LBitEntropy* const bit_entropy, VP8LStreaks* const stats) {
+    uint32_t val, int i, uint32_t* WEBP_RESTRICT const val_prev,
+    int* WEBP_RESTRICT const i_prev,
+    VP8LBitEntropy* WEBP_RESTRICT const bit_entropy,
+    VP8LStreaks* WEBP_RESTRICT const stats) {
   const int streak = i - *i_prev;
 
   // Gather info for the bit entropy.
@@ -403,9 +405,10 @@ static WEBP_INLINE void GetEntropyUnrefinedHelper(
   *i_prev = i;
 }
 
-static void GetEntropyUnrefined_C(const uint32_t X[], int length,
-                                  VP8LBitEntropy* const bit_entropy,
-                                  VP8LStreaks* const stats) {
+static void GetEntropyUnrefined_C(
+    const uint32_t X[], int length,
+    VP8LBitEntropy* WEBP_RESTRICT const bit_entropy,
+    VP8LStreaks* WEBP_RESTRICT const stats) {
   int i;
   int i_prev = 0;
   uint32_t x_prev = X[0];
@@ -424,11 +427,10 @@ static void GetEntropyUnrefined_C(const uint32_t X[], int length,
   bit_entropy->entropy = VP8LFastSLog2(bit_entropy->sum) - bit_entropy->entropy;
 }
 
-static void GetCombinedEntropyUnrefined_C(const uint32_t X[],
-                                          const uint32_t Y[],
-                                          int length,
-                                          VP8LBitEntropy* const bit_entropy,
-                                          VP8LStreaks* const stats) {
+static void GetCombinedEntropyUnrefined_C(
+    const uint32_t X[], const uint32_t Y[], int length,
+    VP8LBitEntropy* WEBP_RESTRICT const bit_entropy,
+    VP8LStreaks* WEBP_RESTRICT const stats) {
   int i = 1;
   int i_prev = 0;
   uint32_t xy_prev = X[0] + Y[0];
@@ -468,8 +470,8 @@ static WEBP_INLINE int8_t U32ToS8(uint32_t v) {
   return (int8_t)(v & 0xff);
 }
 
-void VP8LTransformColor_C(const VP8LMultipliers* const m, uint32_t* data,
-                          int num_pixels) {
+void VP8LTransformColor_C(const VP8LMultipliers* WEBP_RESTRICT const m,
+                          uint32_t* WEBP_RESTRICT data, int num_pixels) {
   int i;
   for (i = 0; i < num_pixels; ++i) {
     const uint32_t argb = data[i];
@@ -505,7 +507,8 @@ static WEBP_INLINE uint8_t TransformColorBlue(uint8_t green_to_blue,
   return (new_blue & 0xff);
 }
 
-void VP8LCollectColorRedTransforms_C(const uint32_t* argb, int stride,
+void VP8LCollectColorRedTransforms_C(const uint32_t* WEBP_RESTRICT argb,
+                                     int stride,
                                      int tile_width, int tile_height,
                                      int green_to_red, uint32_t histo[]) {
   while (tile_height-- > 0) {
@@ -517,7 +520,8 @@ void VP8LCollectColorRedTransforms_C(const uint32_t* argb, int stride,
   }
 }
 
-void VP8LCollectColorBlueTransforms_C(const uint32_t* argb, int stride,
+void VP8LCollectColorBlueTransforms_C(const uint32_t* WEBP_RESTRICT argb,
+                                      int stride,
                                       int tile_width, int tile_height,
                                       int green_to_blue, int red_to_blue,
                                       uint32_t histo[]) {
@@ -544,8 +548,8 @@ static int VectorMismatch_C(const uint32_t* const array1,
 }
 
 // Bundles multiple (1, 2, 4 or 8) pixels into a single pixel.
-void VP8LBundleColorMap_C(const uint8_t* const row, int width, int xbits,
-                          uint32_t* dst) {
+void VP8LBundleColorMap_C(const uint8_t* WEBP_RESTRICT const row,
+                          int width, int xbits, uint32_t* WEBP_RESTRICT dst) {
   int x;
   if (xbits > 0) {
     const int bit_depth = 1 << (3 - xbits);
@@ -576,7 +580,8 @@ static uint32_t ExtraCost_C(const uint32_t* population, int length) {
   return cost;
 }
 
-static uint32_t ExtraCostCombined_C(const uint32_t* X, const uint32_t* Y,
+static uint32_t ExtraCostCombined_C(const uint32_t* WEBP_RESTRICT X,
+                                    const uint32_t* WEBP_RESTRICT Y,
                                     int length) {
   int i;
   uint32_t cost = X[4] + Y[4] + X[5] + Y[5];
@@ -591,13 +596,15 @@ static uint32_t ExtraCostCombined_C(const uint32_t* X, const uint32_t* Y,
 
 //------------------------------------------------------------------------------
 
-static void AddVector_C(const uint32_t* a, const uint32_t* b, uint32_t* out,
-                        int size) {
+static void AddVector_C(const uint32_t* WEBP_RESTRICT a,
+                        const uint32_t* WEBP_RESTRICT b,
+                        uint32_t* WEBP_RESTRICT out, int size) {
   int i;
   for (i = 0; i < size; ++i) out[i] = a[i] + b[i];
 }
 
-static void AddVectorEq_C(const uint32_t* a, uint32_t* out, int size) {
+static void AddVectorEq_C(const uint32_t* WEBP_RESTRICT a,
+                          uint32_t* WEBP_RESTRICT out, int size) {
   int i;
   for (i = 0; i < size; ++i) out[i] += a[i];
 }
@@ -626,8 +633,9 @@ static void AddVectorEq_C(const uint32_t* a, uint32_t* out, int size) {
   }                                                                            \
 } while (0)
 
-void VP8LHistogramAdd(const VP8LHistogram* const a,
-                      const VP8LHistogram* const b, VP8LHistogram* const out) {
+void VP8LHistogramAdd(const VP8LHistogram* WEBP_RESTRICT const a,
+                      const VP8LHistogram* WEBP_RESTRICT const b,
+                      VP8LHistogram* WEBP_RESTRICT const out) {
   int i;
   const int literal_size = VP8LHistogramNumCodes(a->palette_code_bits_);
   assert(a->palette_code_bits_ == b->palette_code_bits_);
@@ -657,14 +665,14 @@ void VP8LHistogramAdd(const VP8LHistogram* const a,
 // Image transforms.
 
 static void PredictorSub0_C(const uint32_t* in, const uint32_t* upper,
-                            int num_pixels, uint32_t* out) {
+                            int num_pixels, uint32_t* WEBP_RESTRICT out) {
   int i;
   for (i = 0; i < num_pixels; ++i) out[i] = VP8LSubPixels(in[i], ARGB_BLACK);
   (void)upper;
 }
 
 static void PredictorSub1_C(const uint32_t* in, const uint32_t* upper,
-                            int num_pixels, uint32_t* out) {
+                            int num_pixels, uint32_t* WEBP_RESTRICT out) {
   int i;
   for (i = 0; i < num_pixels; ++i) out[i] = VP8LSubPixels(in[i], in[i - 1]);
   (void)upper;
@@ -675,7 +683,8 @@ static void PredictorSub1_C(const uint32_t* in, const uint32_t* upper,
 #define GENERATE_PREDICTOR_SUB(PREDICTOR_I)                                \
 static void PredictorSub##PREDICTOR_I##_C(const uint32_t* in,              \
                                           const uint32_t* upper,           \
-                                          int num_pixels, uint32_t* out) { \
+                                          int num_pixels,                  \
+                                          uint32_t* WEBP_RESTRICT out) {   \
   int x;                                                                   \
   assert(upper != NULL);                                                   \
   for (x = 0; x < num_pixels; ++x) {                                       \
diff --git a/src/dsp/lossless_enc_mips32.c b/src/dsp/lossless_enc_mips32.c
index 58529f9a..8e9d7358 100644
--- a/src/dsp/lossless_enc_mips32.c
+++ b/src/dsp/lossless_enc_mips32.c
@@ -149,8 +149,9 @@ static uint32_t ExtraCost_MIPS32(const uint32_t* const population, int length) {
 //     pY += 2;
 //   }
 //   return cost;
-static uint32_t ExtraCostCombined_MIPS32(const uint32_t* const X,
-                                         const uint32_t* const Y, int length) {
+static uint32_t ExtraCostCombined_MIPS32(const uint32_t* WEBP_RESTRICT const X,
+                                         const uint32_t* WEBP_RESTRICT const Y,
+                                         int length) {
   int i, temp0, temp1, temp2, temp3;
   const uint32_t* pX = &X[4];
   const uint32_t* pY = &Y[4];
@@ -215,8 +216,10 @@ static uint32_t ExtraCostCombined_MIPS32(const uint32_t* const X,
 
 // Returns the various RLE counts
 static WEBP_INLINE void GetEntropyUnrefinedHelper(
-    uint32_t val, int i, uint32_t* const val_prev, int* const i_prev,
-    VP8LBitEntropy* const bit_entropy, VP8LStreaks* const stats) {
+    uint32_t val, int i, uint32_t* WEBP_RESTRICT const val_prev,
+    int* WEBP_RESTRICT const i_prev,
+    VP8LBitEntropy* WEBP_RESTRICT const bit_entropy,
+    VP8LStreaks* WEBP_RESTRICT const stats) {
   int* const pstreaks = &stats->streaks[0][0];
   int* const pcnts = &stats->counts[0];
   int temp0, temp1, temp2, temp3;
@@ -241,9 +244,10 @@ static WEBP_INLINE void GetEntropyUnrefinedHelper(
   *i_prev = i;
 }
 
-static void GetEntropyUnrefined_MIPS32(const uint32_t X[], int length,
-                                       VP8LBitEntropy* const bit_entropy,
-                                       VP8LStreaks* const stats) {
+static void GetEntropyUnrefined_MIPS32(
+    const uint32_t X[], int length,
+    VP8LBitEntropy* WEBP_RESTRICT const bit_entropy,
+    VP8LStreaks* WEBP_RESTRICT const stats) {
   int i;
   int i_prev = 0;
   uint32_t x_prev = X[0];
@@ -262,11 +266,10 @@ static void GetEntropyUnrefined_MIPS32(const uint32_t X[], int length,
   bit_entropy->entropy = VP8LFastSLog2(bit_entropy->sum) - bit_entropy->entropy;
 }
 
-static void GetCombinedEntropyUnrefined_MIPS32(const uint32_t X[],
-                                               const uint32_t Y[],
-                                               int length,
-                                               VP8LBitEntropy* const entropy,
-                                               VP8LStreaks* const stats) {
+static void GetCombinedEntropyUnrefined_MIPS32(
+    const uint32_t X[], const uint32_t Y[], int length,
+    VP8LBitEntropy* WEBP_RESTRICT const entropy,
+    VP8LStreaks* WEBP_RESTRICT const stats) {
   int i = 1;
   int i_prev = 0;
   uint32_t xy_prev = X[0] + Y[0];
@@ -344,8 +347,9 @@ static void GetCombinedEntropyUnrefined_MIPS32(const uint32_t X[],
     ASM_END_COMMON_0                                    \
     ASM_END_COMMON_1
 
-static void AddVector_MIPS32(const uint32_t* pa, const uint32_t* pb,
-                             uint32_t* pout, int size) {
+static void AddVector_MIPS32(const uint32_t* WEBP_RESTRICT pa,
+                             const uint32_t* WEBP_RESTRICT pb,
+                             uint32_t* WEBP_RESTRICT pout, int size) {
   uint32_t temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
   const int end = ((size) / 4) * 4;
   const uint32_t* const LoopEnd = pa + end;
@@ -356,7 +360,8 @@ static void AddVector_MIPS32(const uint32_t* pa, const uint32_t* pb,
   for (i = 0; i < size - end; ++i) pout[i] = pa[i] + pb[i];
 }
 
-static void AddVectorEq_MIPS32(const uint32_t* pa, uint32_t* pout, int size) {
+static void AddVectorEq_MIPS32(const uint32_t* WEBP_RESTRICT pa,
+                               uint32_t* WEBP_RESTRICT pout, int size) {
   uint32_t temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
   const int end = ((size) / 4) * 4;
   const uint32_t* const LoopEnd = pa + end;
diff --git a/src/dsp/lossless_enc_mips_dsp_r2.c b/src/dsp/lossless_enc_mips_dsp_r2.c
index 6eaab0af..e10b8f7e 100644
--- a/src/dsp/lossless_enc_mips_dsp_r2.c
+++ b/src/dsp/lossless_enc_mips_dsp_r2.c
@@ -78,8 +78,9 @@ static WEBP_INLINE uint32_t ColorTransformDelta(int8_t color_pred,
   return (uint32_t)((int)(color_pred) * color) >> 5;
 }
 
-static void TransformColor_MIPSdspR2(const VP8LMultipliers* const m,
-                                     uint32_t* data, int num_pixels) {
+static void TransformColor_MIPSdspR2(
+    const VP8LMultipliers* WEBP_RESTRICT const m, uint32_t* WEBP_RESTRICT data,
+    int num_pixels) {
   int temp0, temp1, temp2, temp3, temp4, temp5;
   uint32_t argb, argb1, new_red, new_red1;
   const uint32_t G_to_R = m->green_to_red_;
@@ -172,7 +173,8 @@ static WEBP_INLINE uint8_t TransformColorBlue(uint8_t green_to_blue,
 }
 
 static void CollectColorBlueTransforms_MIPSdspR2(
-    const uint32_t* argb, int stride, int tile_width, int tile_height,
+    const uint32_t* WEBP_RESTRICT argb, int stride,
+    int tile_width, int tile_height,
     int green_to_blue, int red_to_blue, uint32_t histo[]) {
   const int rtb = (red_to_blue << 16) | (red_to_blue & 0xffff);
   const int gtb = (green_to_blue << 16) | (green_to_blue & 0xffff);
@@ -221,11 +223,9 @@ static WEBP_INLINE uint8_t TransformColorRed(uint8_t green_to_red,
   return (new_red & 0xff);
 }
 
-static void CollectColorRedTransforms_MIPSdspR2(const uint32_t* argb,
-                                                int stride, int tile_width,
-                                                int tile_height,
-                                                int green_to_red,
-                                                uint32_t histo[]) {
+static void CollectColorRedTransforms_MIPSdspR2(
+    const uint32_t* WEBP_RESTRICT argb, int stride,
+    int tile_width, int tile_height, int green_to_red, uint32_t histo[]) {
   const int gtr = (green_to_red << 16) | (green_to_red & 0xffff);
   while (tile_height-- > 0) {
     int x;
diff --git a/src/dsp/lossless_enc_msa.c b/src/dsp/lossless_enc_msa.c
index 600dddfb..6d835ab7 100644
--- a/src/dsp/lossless_enc_msa.c
+++ b/src/dsp/lossless_enc_msa.c
@@ -48,8 +48,8 @@
   dst = VSHF_UB(src, t0, mask1);                                \
 } while (0)
 
-static void TransformColor_MSA(const VP8LMultipliers* const m, uint32_t* data,
-                               int num_pixels) {
+static void TransformColor_MSA(const VP8LMultipliers* WEBP_RESTRICT const m,
+                               uint32_t* WEBP_RESTRICT data, int num_pixels) {
   v16u8 src0, dst0;
   const v16i8 g2br = (v16i8)__msa_fill_w(m->green_to_blue_ |
                                          (m->green_to_red_ << 16));
diff --git a/src/dsp/lossless_enc_neon.c b/src/dsp/lossless_enc_neon.c
index e32c7961..838204a7 100644
--- a/src/dsp/lossless_enc_neon.c
+++ b/src/dsp/lossless_enc_neon.c
@@ -72,8 +72,9 @@ static void SubtractGreenFromBlueAndRed_NEON(uint32_t* argb_data,
 //------------------------------------------------------------------------------
 // Color Transform
 
-static void TransformColor_NEON(const VP8LMultipliers* const m,
-                                uint32_t* argb_data, int num_pixels) {
+static void TransformColor_NEON(const VP8LMultipliers* WEBP_RESTRICT const m,
+                                uint32_t* WEBP_RESTRICT argb_data,
+                                int num_pixels) {
   // sign-extended multiplying constants, pre-shifted by 6.
 #define CST(X)  (((int16_t)(m->X << 8)) >> 6)
   const int16_t rb[8] = {
diff --git a/src/dsp/lossless_enc_sse2.c b/src/dsp/lossless_enc_sse2.c
index 530acc37..f6706dd5 100644
--- a/src/dsp/lossless_enc_sse2.c
+++ b/src/dsp/lossless_enc_sse2.c
@@ -49,8 +49,9 @@ static void SubtractGreenFromBlueAndRed_SSE2(uint32_t* argb_data,
 #define MK_CST_16(HI, LO) \
   _mm_set1_epi32((int)(((uint32_t)(HI) << 16) | ((LO) & 0xffff)))
 
-static void TransformColor_SSE2(const VP8LMultipliers* const m,
-                                uint32_t* argb_data, int num_pixels) {
+static void TransformColor_SSE2(const VP8LMultipliers* WEBP_RESTRICT const m,
+                                uint32_t* WEBP_RESTRICT argb_data,
+                                int num_pixels) {
   const __m128i mults_rb = MK_CST_16(CST_5b(m->green_to_red_),
                                      CST_5b(m->green_to_blue_));
   const __m128i mults_b2 = MK_CST_16(CST_5b(m->red_to_blue_), 0);
@@ -79,7 +80,8 @@ static void TransformColor_SSE2(const VP8LMultipliers* const m,
 
 //------------------------------------------------------------------------------
 #define SPAN 8
-static void CollectColorBlueTransforms_SSE2(const uint32_t* argb, int stride,
+static void CollectColorBlueTransforms_SSE2(const uint32_t* WEBP_RESTRICT argb,
+                                            int stride,
                                             int tile_width, int tile_height,
                                             int green_to_blue, int red_to_blue,
                                             uint32_t histo[]) {
@@ -126,7 +128,8 @@ static void CollectColorBlueTransforms_SSE2(const uint32_t* argb, int stride,
   }
 }
 
-static void CollectColorRedTransforms_SSE2(const uint32_t* argb, int stride,
+static void CollectColorRedTransforms_SSE2(const uint32_t* WEBP_RESTRICT argb,
+                                           int stride,
                                            int tile_width, int tile_height,
                                            int green_to_red, uint32_t histo[]) {
   const __m128i mults_g = MK_CST_16(0, CST_5b(green_to_red));
@@ -173,8 +176,9 @@ static void CollectColorRedTransforms_SSE2(const uint32_t* argb, int stride,
 // Note we are adding uint32_t's as *signed* int32's (using _mm_add_epi32). But
 // that's ok since the histogram values are less than 1<<28 (max picture size).
 #define LINE_SIZE 16    // 8 or 16
-static void AddVector_SSE2(const uint32_t* a, const uint32_t* b, uint32_t* out,
-                           int size) {
+static void AddVector_SSE2(const uint32_t* WEBP_RESTRICT a,
+                           const uint32_t* WEBP_RESTRICT b,
+                           uint32_t* WEBP_RESTRICT out, int size) {
   int i;
   for (i = 0; i + LINE_SIZE <= size; i += LINE_SIZE) {
     const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i +  0]);
@@ -201,7 +205,8 @@ static void AddVector_SSE2(const uint32_t* a, const uint32_t* b, uint32_t* out,
   }
 }
 
-static void AddVectorEq_SSE2(const uint32_t* a, uint32_t* out, int size) {
+static void AddVectorEq_SSE2(const uint32_t* WEBP_RESTRICT a,
+                             uint32_t* WEBP_RESTRICT out, int size) {
   int i;
   for (i = 0; i + LINE_SIZE <= size; i += LINE_SIZE) {
     const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i +  0]);
@@ -333,8 +338,9 @@ static int VectorMismatch_SSE2(const uint32_t* const array1,
 }
 
 // Bundles multiple (1, 2, 4 or 8) pixels into a single pixel.
-static void BundleColorMap_SSE2(const uint8_t* const row, int width, int xbits,
-                                uint32_t* dst) {
+static void BundleColorMap_SSE2(const uint8_t* WEBP_RESTRICT const row,
+                                int width, int xbits,
+                                uint32_t* WEBP_RESTRICT dst) {
   int x;
   assert(xbits >= 0);
   assert(xbits <= 3);
@@ -423,7 +429,7 @@ static WEBP_INLINE void Average2_m128i(const __m128i* const a0,
 
 // Predictor0: ARGB_BLACK.
 static void PredictorSub0_SSE2(const uint32_t* in, const uint32_t* upper,
-                               int num_pixels, uint32_t* out) {
+                               int num_pixels, uint32_t* WEBP_RESTRICT out) {
   int i;
   const __m128i black = _mm_set1_epi32((int)ARGB_BLACK);
   for (i = 0; i + 4 <= num_pixels; i += 4) {
@@ -440,7 +446,8 @@ static void PredictorSub0_SSE2(const uint32_t* in, const uint32_t* upper,
 #define GENERATE_PREDICTOR_1(X, IN)                                         \
   static void PredictorSub##X##_SSE2(const uint32_t* const in,              \
                                      const uint32_t* const upper,           \
-                                     int num_pixels, uint32_t* const out) { \
+                                     int num_pixels,                        \
+                                     uint32_t* WEBP_RESTRICT const out) {   \
     int i;                                                                  \
     for (i = 0; i + 4 <= num_pixels; i += 4) {                              \
       const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);          \
@@ -462,7 +469,7 @@ GENERATE_PREDICTOR_1(4, upper[i - 1])    // Predictor4: TL
 
 // Predictor5: avg2(avg2(L, TR), T)
 static void PredictorSub5_SSE2(const uint32_t* in, const uint32_t* upper,
-                               int num_pixels, uint32_t* out) {
+                               int num_pixels, uint32_t* WEBP_RESTRICT out) {
   int i;
   for (i = 0; i + 4 <= num_pixels; i += 4) {
     const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]);
@@ -482,7 +489,8 @@ static void PredictorSub5_SSE2(const uint32_t* in, const uint32_t* upper,
 
 #define GENERATE_PREDICTOR_2(X, A, B)                                         \
 static void PredictorSub##X##_SSE2(const uint32_t* in, const uint32_t* upper, \
-                                   int num_pixels, uint32_t* out) {           \
+                                   int num_pixels,                            \
+                                   uint32_t* WEBP_RESTRICT out) {             \
   int i;                                                                      \
   for (i = 0; i + 4 <= num_pixels; i += 4) {                                  \
     const __m128i tA = _mm_loadu_si128((const __m128i*)&(A));                 \
@@ -506,7 +514,7 @@ GENERATE_PREDICTOR_2(9, upper[i], upper[i + 1])    // Predictor9: average(T, TR)
 
 // Predictor10: avg(avg(L,TL), avg(T, TR)).
 static void PredictorSub10_SSE2(const uint32_t* in, const uint32_t* upper,
-                                int num_pixels, uint32_t* out) {
+                                int num_pixels, uint32_t* WEBP_RESTRICT out) {
   int i;
   for (i = 0; i + 4 <= num_pixels; i += 4) {
     const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]);
@@ -541,7 +549,7 @@ static void GetSumAbsDiff32_SSE2(const __m128i* const A, const __m128i* const B,
 }
 
 static void PredictorSub11_SSE2(const uint32_t* in, const uint32_t* upper,
-                                int num_pixels, uint32_t* out) {
+                                int num_pixels, uint32_t* WEBP_RESTRICT out) {
   int i;
   for (i = 0; i + 4 <= num_pixels; i += 4) {
     const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]);
@@ -567,7 +575,7 @@ static void PredictorSub11_SSE2(const uint32_t* in, const uint32_t* upper,
 
 // Predictor12: ClampedSubSubtractFull.
 static void PredictorSub12_SSE2(const uint32_t* in, const uint32_t* upper,
-                                int num_pixels, uint32_t* out) {
+                                int num_pixels, uint32_t* WEBP_RESTRICT out) {
   int i;
   const __m128i zero = _mm_setzero_si128();
   for (i = 0; i + 4 <= num_pixels; i += 4) {
@@ -596,7 +604,7 @@ static void PredictorSub12_SSE2(const uint32_t* in, const uint32_t* upper,
 
 // Predictors13: ClampedAddSubtractHalf
 static void PredictorSub13_SSE2(const uint32_t* in, const uint32_t* upper,
-                                int num_pixels, uint32_t* out) {
+                                int num_pixels, uint32_t* WEBP_RESTRICT out) {
   int i;
   const __m128i zero = _mm_setzero_si128();
   for (i = 0; i + 2 <= num_pixels; i += 2) {
diff --git a/src/dsp/lossless_enc_sse41.c b/src/dsp/lossless_enc_sse41.c
index 9a0dcf9b..87ed056f 100644
--- a/src/dsp/lossless_enc_sse41.c
+++ b/src/dsp/lossless_enc_sse41.c
@@ -44,8 +44,9 @@ static uint32_t ExtraCost_SSE41(const uint32_t* const a, int length) {
   return HorizontalSum_SSE41(cost);
 }
 
-static uint32_t ExtraCostCombined_SSE41(const uint32_t* const a,
-                                        const uint32_t* const b, int length) {
+static uint32_t ExtraCostCombined_SSE41(const uint32_t* WEBP_RESTRICT const a,
+                                        const uint32_t* WEBP_RESTRICT const b,
+                                        int length) {
   int i;
   __m128i cost = _mm_add_epi32(_mm_set_epi32(2 * a[7], 2 * a[6], a[5], a[4]),
                                _mm_set_epi32(2 * b[7], 2 * b[6], b[5], b[4]));
@@ -95,7 +96,8 @@ static void SubtractGreenFromBlueAndRed_SSE41(uint32_t* argb_data,
 #define MK_CST_16(HI, LO) \
   _mm_set1_epi32((int)(((uint32_t)(HI) << 16) | ((LO) & 0xffff)))
 
-static void CollectColorBlueTransforms_SSE41(const uint32_t* argb, int stride,
+static void CollectColorBlueTransforms_SSE41(const uint32_t* WEBP_RESTRICT argb,
+                                             int stride,
                                              int tile_width, int tile_height,
                                              int green_to_blue, int red_to_blue,
                                              uint32_t histo[]) {
@@ -141,7 +143,8 @@ static void CollectColorBlueTransforms_SSE41(const uint32_t* argb, int stride,
   }
 }
 
-static void CollectColorRedTransforms_SSE41(const uint32_t* argb, int stride,
+static void CollectColorRedTransforms_SSE41(const uint32_t* WEBP_RESTRICT argb,
+                                            int stride,
                                             int tile_width, int tile_height,
                                             int green_to_red,
                                             uint32_t histo[]) {
diff --git a/src/dsp/lossless_neon.c b/src/dsp/lossless_neon.c
index e9960db3..93f41cef 100644
--- a/src/dsp/lossless_neon.c
+++ b/src/dsp/lossless_neon.c
@@ -26,8 +26,8 @@
 #if !defined(WORK_AROUND_GCC)
 // gcc 4.6.0 had some trouble (NDK-r9) with this code. We only use it for
 // gcc-4.8.x at least.
-static void ConvertBGRAToRGBA_NEON(const uint32_t* src,
-                                   int num_pixels, uint8_t* dst) {
+static void ConvertBGRAToRGBA_NEON(const uint32_t* WEBP_RESTRICT src,
+                                   int num_pixels, uint8_t* WEBP_RESTRICT dst) {
   const uint32_t* const end = src + (num_pixels & ~15);
   for (; src < end; src += 16) {
     uint8x16x4_t pixel = vld4q_u8((uint8_t*)src);
@@ -41,8 +41,8 @@ static void ConvertBGRAToRGBA_NEON(const uint32_t* src,
   VP8LConvertBGRAToRGBA_C(src, num_pixels & 15, dst);  // left-overs
 }
 
-static void ConvertBGRAToBGR_NEON(const uint32_t* src,
-                                  int num_pixels, uint8_t* dst) {
+static void ConvertBGRAToBGR_NEON(const uint32_t* WEBP_RESTRICT src,
+                                  int num_pixels, uint8_t* WEBP_RESTRICT dst) {
   const uint32_t* const end = src + (num_pixels & ~15);
   for (; src < end; src += 16) {
     const uint8x16x4_t pixel = vld4q_u8((uint8_t*)src);
@@ -53,8 +53,8 @@ static void ConvertBGRAToBGR_NEON(const uint32_t* src,
   VP8LConvertBGRAToBGR_C(src, num_pixels & 15, dst);  // left-overs
 }
 
-static void ConvertBGRAToRGB_NEON(const uint32_t* src,
-                                  int num_pixels, uint8_t* dst) {
+static void ConvertBGRAToRGB_NEON(const uint32_t* WEBP_RESTRICT src,
+                                  int num_pixels, uint8_t* WEBP_RESTRICT dst) {
   const uint32_t* const end = src + (num_pixels & ~15);
   for (; src < end; src += 16) {
     const uint8x16x4_t pixel = vld4q_u8((uint8_t*)src);
@@ -71,8 +71,8 @@ static void ConvertBGRAToRGB_NEON(const uint32_t* src,
 
 static const uint8_t kRGBAShuffle[8] = { 2, 1, 0, 3, 6, 5, 4, 7 };
 
-static void ConvertBGRAToRGBA_NEON(const uint32_t* src,
-                                   int num_pixels, uint8_t* dst) {
+static void ConvertBGRAToRGBA_NEON(const uint32_t* WEBP_RESTRICT src,
+                                   int num_pixels, uint8_t* WEBP_RESTRICT dst) {
   const uint32_t* const end = src + (num_pixels & ~1);
   const uint8x8_t shuffle = vld1_u8(kRGBAShuffle);
   for (; src < end; src += 2) {
@@ -89,8 +89,8 @@ static const uint8_t kBGRShuffle[3][8] = {
   { 21, 22, 24, 25, 26, 28, 29, 30 }
 };
 
-static void ConvertBGRAToBGR_NEON(const uint32_t* src,
-                                  int num_pixels, uint8_t* dst) {
+static void ConvertBGRAToBGR_NEON(const uint32_t* WEBP_RESTRICT src,
+                                  int num_pixels, uint8_t* WEBP_RESTRICT dst) {
   const uint32_t* const end = src + (num_pixels & ~7);
   const uint8x8_t shuffle0 = vld1_u8(kBGRShuffle[0]);
   const uint8x8_t shuffle1 = vld1_u8(kBGRShuffle[1]);
@@ -116,8 +116,8 @@ static const uint8_t kRGBShuffle[3][8] = {
   { 21, 20, 26, 25, 24, 30, 29, 28 }
 };
 
-static void ConvertBGRAToRGB_NEON(const uint32_t* src,
-                                  int num_pixels, uint8_t* dst) {
+static void ConvertBGRAToRGB_NEON(const uint32_t* WEBP_RESTRICT src,
+                                  int num_pixels, uint8_t* WEBP_RESTRICT dst) {
   const uint32_t* const end = src + (num_pixels & ~7);
   const uint8x8_t shuffle0 = vld1_u8(kRGBShuffle[0]);
   const uint8x8_t shuffle1 = vld1_u8(kRGBShuffle[1]);
@@ -209,7 +209,7 @@ static uint32_t Predictor13_NEON(const uint32_t* const left,
 
 // Predictor0: ARGB_BLACK.
 static void PredictorAdd0_NEON(const uint32_t* in, const uint32_t* upper,
-                               int num_pixels, uint32_t* out) {
+                               int num_pixels, uint32_t* WEBP_RESTRICT out) {
   int i;
   const uint8x16_t black = vreinterpretq_u8_u32(vdupq_n_u32(ARGB_BLACK));
   for (i = 0; i + 4 <= num_pixels; i += 4) {
@@ -222,7 +222,7 @@ static void PredictorAdd0_NEON(const uint32_t* in, const uint32_t* upper,
 
 // Predictor1: left.
 static void PredictorAdd1_NEON(const uint32_t* in, const uint32_t* upper,
-                               int num_pixels, uint32_t* out) {
+                               int num_pixels, uint32_t* WEBP_RESTRICT out) {
   int i;
   const uint8x16_t zero = LOADQ_U32_AS_U8(0);
   for (i = 0; i + 4 <= num_pixels; i += 4) {
@@ -248,7 +248,7 @@ static void PredictorAdd1_NEON(const uint32_t* in, const uint32_t* upper,
 #define GENERATE_PREDICTOR_1(X, IN)                                       \
 static void PredictorAdd##X##_NEON(const uint32_t* in,                    \
                                    const uint32_t* upper, int num_pixels, \
-                                   uint32_t* out) {                       \
+                                   uint32_t* WEBP_RESTRICT out) {         \
   int i;                                                                  \
   for (i = 0; i + 4 <= num_pixels; i += 4) {                              \
     const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);                      \
@@ -276,7 +276,7 @@ GENERATE_PREDICTOR_1(4, upper[i - 1])
 } while (0)
 
 static void PredictorAdd5_NEON(const uint32_t* in, const uint32_t* upper,
-                               int num_pixels, uint32_t* out) {
+                               int num_pixels, uint32_t* WEBP_RESTRICT out) {
   int i;
   uint8x16_t L = LOADQ_U32_AS_U8(out[-1]);
   for (i = 0; i + 4 <= num_pixels; i += 4) {
@@ -301,7 +301,7 @@ static void PredictorAdd5_NEON(const uint32_t* in, const uint32_t* upper,
 
 // Predictor6: average(left, TL)
 static void PredictorAdd6_NEON(const uint32_t* in, const uint32_t* upper,
-                               int num_pixels, uint32_t* out) {
+                               int num_pixels, uint32_t* WEBP_RESTRICT out) {
   int i;
   uint8x16_t L = LOADQ_U32_AS_U8(out[-1]);
   for (i = 0; i + 4 <= num_pixels; i += 4) {
@@ -317,7 +317,7 @@ static void PredictorAdd6_NEON(const uint32_t* in, const uint32_t* upper,
 
 // Predictor7: average(left, T)
 static void PredictorAdd7_NEON(const uint32_t* in, const uint32_t* upper,
-                               int num_pixels, uint32_t* out) {
+                               int num_pixels, uint32_t* WEBP_RESTRICT out) {
   int i;
   uint8x16_t L = LOADQ_U32_AS_U8(out[-1]);
   for (i = 0; i + 4 <= num_pixels; i += 4) {
@@ -335,7 +335,7 @@ static void PredictorAdd7_NEON(const uint32_t* in, const uint32_t* upper,
 #define GENERATE_PREDICTOR_2(X, IN)                                       \
 static void PredictorAdd##X##_NEON(const uint32_t* in,                    \
                                    const uint32_t* upper, int num_pixels, \
-                                   uint32_t* out) {                       \
+                                   uint32_t* WEBP_RESTRICT out) {         \
   int i;                                                                  \
   for (i = 0; i + 4 <= num_pixels; i += 4) {                              \
     const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);                      \
@@ -363,7 +363,7 @@ GENERATE_PREDICTOR_2(9, upper[i + 1])
 } while (0)
 
 static void PredictorAdd10_NEON(const uint32_t* in, const uint32_t* upper,
-                                int num_pixels, uint32_t* out) {
+                                int num_pixels, uint32_t* WEBP_RESTRICT out) {
   int i;
   uint8x16_t L = LOADQ_U32_AS_U8(out[-1]);
   for (i = 0; i + 4 <= num_pixels; i += 4) {
@@ -394,7 +394,7 @@ static void PredictorAdd10_NEON(const uint32_t* in, const uint32_t* upper,
 } while (0)
 
 static void PredictorAdd11_NEON(const uint32_t* in, const uint32_t* upper,
-                                int num_pixels, uint32_t* out) {
+                                int num_pixels, uint32_t* WEBP_RESTRICT out) {
   int i;
   uint8x16_t L = LOADQ_U32_AS_U8(out[-1]);
   for (i = 0; i + 4 <= num_pixels; i += 4) {
@@ -427,7 +427,7 @@ static void PredictorAdd11_NEON(const uint32_t* in, const uint32_t* upper,
 } while (0)
 
 static void PredictorAdd12_NEON(const uint32_t* in, const uint32_t* upper,
-                                int num_pixels, uint32_t* out) {
+                                int num_pixels, uint32_t* WEBP_RESTRICT out) {
   int i;
   uint16x8_t L = vmovl_u8(LOAD_U32_AS_U8(out[-1]));
   for (i = 0; i + 4 <= num_pixels; i += 4) {
@@ -468,7 +468,7 @@ static void PredictorAdd12_NEON(const uint32_t* in, const uint32_t* upper,
 } while (0)
 
 static void PredictorAdd13_NEON(const uint32_t* in, const uint32_t* upper,
-                                int num_pixels, uint32_t* out) {
+                                int num_pixels, uint32_t* WEBP_RESTRICT out) {
   int i;
   uint8x16_t L = LOADQ_U32_AS_U8(out[-1]);
   for (i = 0; i + 4 <= num_pixels; i += 4) {
diff --git a/src/dsp/lossless_sse2.c b/src/dsp/lossless_sse2.c
index 4b6a532c..5b68d1cf 100644
--- a/src/dsp/lossless_sse2.c
+++ b/src/dsp/lossless_sse2.c
@@ -186,7 +186,7 @@ static uint32_t Predictor13_SSE2(const uint32_t* const left,
 
 // Predictor0: ARGB_BLACK.
 static void PredictorAdd0_SSE2(const uint32_t* in, const uint32_t* upper,
-                               int num_pixels, uint32_t* out) {
+                               int num_pixels, uint32_t* WEBP_RESTRICT out) {
   int i;
   const __m128i black = _mm_set1_epi32((int)ARGB_BLACK);
   for (i = 0; i + 4 <= num_pixels; i += 4) {
@@ -202,7 +202,7 @@ static void PredictorAdd0_SSE2(const uint32_t* in, const uint32_t* upper,
 
 // Predictor1: left.
 static void PredictorAdd1_SSE2(const uint32_t* in, const uint32_t* upper,
-                               int num_pixels, uint32_t* out) {
+                               int num_pixels, uint32_t* WEBP_RESTRICT out) {
   int i;
   __m128i prev = _mm_set1_epi32((int)out[-1]);
   for (i = 0; i + 4 <= num_pixels; i += 4) {
@@ -230,7 +230,8 @@ static void PredictorAdd1_SSE2(const uint32_t* in, const uint32_t* upper,
 // per 8 bit channel.
 #define GENERATE_PREDICTOR_1(X, IN)                                           \
 static void PredictorAdd##X##_SSE2(const uint32_t* in, const uint32_t* upper, \
-                                  int num_pixels, uint32_t* out) {            \
+                                   int num_pixels,                            \
+                                   uint32_t* WEBP_RESTRICT out) {             \
   int i;                                                                      \
   for (i = 0; i + 4 <= num_pixels; i += 4) {                                  \
     const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);              \
@@ -259,7 +260,8 @@ GENERATE_PREDICTOR_ADD(Predictor7_SSE2, PredictorAdd7_SSE2)
 
 #define GENERATE_PREDICTOR_2(X, IN)                                           \
 static void PredictorAdd##X##_SSE2(const uint32_t* in, const uint32_t* upper, \
-                                   int num_pixels, uint32_t* out) {           \
+                                   int num_pixels,                            \
+                                   uint32_t* WEBP_RESTRICT out) {             \
   int i;                                                                      \
   for (i = 0; i + 4 <= num_pixels; i += 4) {                                  \
     const __m128i Tother = _mm_loadu_si128((const __m128i*)&(IN));            \
@@ -297,7 +299,7 @@ GENERATE_PREDICTOR_2(9, upper[i + 1])
 } while (0)
 
 static void PredictorAdd10_SSE2(const uint32_t* in, const uint32_t* upper,
-                                int num_pixels, uint32_t* out) {
+                                int num_pixels, uint32_t* WEBP_RESTRICT out) {
   int i;
   __m128i L = _mm_cvtsi32_si128((int)out[-1]);
   for (i = 0; i + 4 <= num_pixels; i += 4) {
@@ -344,7 +346,7 @@ static void PredictorAdd10_SSE2(const uint32_t* in, const uint32_t* upper,
 } while (0)
 
 static void PredictorAdd11_SSE2(const uint32_t* in, const uint32_t* upper,
-                                int num_pixels, uint32_t* out) {
+                                int num_pixels, uint32_t* WEBP_RESTRICT out) {
   int i;
   __m128i pa;
   __m128i L = _mm_cvtsi32_si128((int)out[-1]);
@@ -395,7 +397,7 @@ static void PredictorAdd11_SSE2(const uint32_t* in, const uint32_t* upper,
 } while (0)
 
 static void PredictorAdd12_SSE2(const uint32_t* in, const uint32_t* upper,
-                                int num_pixels, uint32_t* out) {
+                                int num_pixels, uint32_t* WEBP_RESTRICT out) {
   int i;
   const __m128i zero = _mm_setzero_si128();
   const __m128i L8 = _mm_cvtsi32_si128((int)out[-1]);
@@ -490,8 +492,8 @@ static void TransformColorInverse_SSE2(const VP8LMultipliers* const m,
 //------------------------------------------------------------------------------
 // Color-space conversion functions
 
-static void ConvertBGRAToRGB_SSE2(const uint32_t* src, int num_pixels,
-                                  uint8_t* dst) {
+static void ConvertBGRAToRGB_SSE2(const uint32_t* WEBP_RESTRICT src,
+                                  int num_pixels, uint8_t* WEBP_RESTRICT dst) {
   const __m128i* in = (const __m128i*)src;
   __m128i* out = (__m128i*)dst;
 
@@ -526,8 +528,8 @@ static void ConvertBGRAToRGB_SSE2(const uint32_t* src, int num_pixels,
   }
 }
 
-static void ConvertBGRAToRGBA_SSE2(const uint32_t* src,
-                                   int num_pixels, uint8_t* dst) {
+static void ConvertBGRAToRGBA_SSE2(const uint32_t* WEBP_RESTRICT src,
+                                   int num_pixels, uint8_t* WEBP_RESTRICT dst) {
   const __m128i red_blue_mask = _mm_set1_epi32(0x00ff00ff);
   const __m128i* in = (const __m128i*)src;
   __m128i* out = (__m128i*)dst;
@@ -554,8 +556,9 @@ static void ConvertBGRAToRGBA_SSE2(const uint32_t* src,
   }
 }
 
-static void ConvertBGRAToRGBA4444_SSE2(const uint32_t* src,
-                                       int num_pixels, uint8_t* dst) {
+static void ConvertBGRAToRGBA4444_SSE2(const uint32_t* WEBP_RESTRICT src,
+                                       int num_pixels,
+                                       uint8_t* WEBP_RESTRICT dst) {
   const __m128i mask_0x0f = _mm_set1_epi8(0x0f);
   const __m128i mask_0xf0 = _mm_set1_epi8((char)0xf0);
   const __m128i* in = (const __m128i*)src;
@@ -590,8 +593,9 @@ static void ConvertBGRAToRGBA4444_SSE2(const uint32_t* src,
   }
 }
 
-static void ConvertBGRAToRGB565_SSE2(const uint32_t* src,
-                                     int num_pixels, uint8_t* dst) {
+static void ConvertBGRAToRGB565_SSE2(const uint32_t* WEBP_RESTRICT src,
+                                     int num_pixels,
+                                     uint8_t* WEBP_RESTRICT dst) {
   const __m128i mask_0xe0 = _mm_set1_epi8((char)0xe0);
   const __m128i mask_0xf8 = _mm_set1_epi8((char)0xf8);
   const __m128i mask_0x07 = _mm_set1_epi8(0x07);
@@ -631,8 +635,8 @@ static void ConvertBGRAToRGB565_SSE2(const uint32_t* src,
   }
 }
 
-static void ConvertBGRAToBGR_SSE2(const uint32_t* src,
-                                  int num_pixels, uint8_t* dst) {
+static void ConvertBGRAToBGR_SSE2(const uint32_t* WEBP_RESTRICT src,
+                                  int num_pixels, uint8_t* WEBP_RESTRICT dst) {
   const __m128i mask_l = _mm_set_epi32(0, 0x00ffffff, 0, 0x00ffffff);
   const __m128i mask_h = _mm_set_epi32(0x00ffffff, 0, 0x00ffffff, 0);
   const __m128i* in = (const __m128i*)src;
diff --git a/src/dsp/lossless_sse41.c b/src/dsp/lossless_sse41.c
index bb7ce761..a2d19144 100644
--- a/src/dsp/lossless_sse41.c
+++ b/src/dsp/lossless_sse41.c
@@ -77,8 +77,8 @@ static void TransformColorInverse_SSE41(const VP8LMultipliers* const m,
   }                                                   \
 } while (0)
 
-static void ConvertBGRAToRGB_SSE41(const uint32_t* src, int num_pixels,
-                                   uint8_t* dst) {
+static void ConvertBGRAToRGB_SSE41(const uint32_t* WEBP_RESTRICT src,
+                                   int num_pixels, uint8_t* WEBP_RESTRICT dst) {
   const __m128i* in = (const __m128i*)src;
   __m128i* out = (__m128i*)dst;
   const __m128i perm0 = _mm_setr_epi8(2, 1, 0, 6, 5, 4, 10, 9,
@@ -95,8 +95,8 @@ static void ConvertBGRAToRGB_SSE41(const uint32_t* src, int num_pixels,
   }
 }
 
-static void ConvertBGRAToBGR_SSE41(const uint32_t* src,
-                                   int num_pixels, uint8_t* dst) {
+static void ConvertBGRAToBGR_SSE41(const uint32_t* WEBP_RESTRICT src,
+                                   int num_pixels, uint8_t* WEBP_RESTRICT dst) {
   const __m128i* in = (const __m128i*)src;
   __m128i* out = (__m128i*)dst;
   const __m128i perm0 = _mm_setr_epi8(0, 1, 2, 4, 5, 6, 8, 9, 10,

From 35915b389eab7fa08c6db4feeb52aa0ff2dbde03 Mon Sep 17 00:00:00 2001
From: James Zern <jzern@google.com>
Date: Fri, 16 Aug 2024 16:55:00 -0700
Subject: [PATCH 6/8] dsp/rescaler*: use WEBP_RESTRICT qualifier

Some improvement in the C code. No changes in NEON or SSE2 w/ndk
r27/gcc-13/clang-16.

This only affects non-vector pointers; any vector pointers are left as a
follow up.

Change-Id: I2316122db893f48f0afda90a147c83cac7f07526
---
 src/dsp/dsp.h             | 20 ++++++++++++--------
 src/dsp/rescaler.c        | 11 ++++++-----
 src/dsp/rescaler_mips32.c |  8 ++++----
 src/dsp/rescaler_msa.c    | 27 ++++++++++++++-------------
 src/dsp/rescaler_neon.c   |  4 ++--
 src/dsp/rescaler_sse2.c   | 18 ++++++++----------
 6 files changed, 46 insertions(+), 42 deletions(-)

diff --git a/src/dsp/dsp.h b/src/dsp/dsp.h
index 30a1d43f..b7c2acf0 100644
--- a/src/dsp/dsp.h
+++ b/src/dsp/dsp.h
@@ -366,8 +366,9 @@ struct WebPRescaler;
 // Import a row of data and save its contribution in the rescaler.
 // 'channel' denotes the channel number to be imported. 'Expand' corresponds to
 // the wrk->x_expand case. Otherwise, 'Shrink' is to be used.
-typedef void (*WebPRescalerImportRowFunc)(struct WebPRescaler* const wrk,
-                                          const uint8_t* src);
+typedef void (*WebPRescalerImportRowFunc)(
+    struct WebPRescaler* WEBP_RESTRICT const wrk,
+    const uint8_t* WEBP_RESTRICT src);
 
 extern WebPRescalerImportRowFunc WebPRescalerImportRowExpand;
 extern WebPRescalerImportRowFunc WebPRescalerImportRowShrink;
@@ -380,16 +381,19 @@ extern WebPRescalerExportRowFunc WebPRescalerExportRowExpand;
 extern WebPRescalerExportRowFunc WebPRescalerExportRowShrink;
 
 // Plain-C implementation, as fall-back.
-extern void WebPRescalerImportRowExpand_C(struct WebPRescaler* const wrk,
-                                          const uint8_t* src);
-extern void WebPRescalerImportRowShrink_C(struct WebPRescaler* const wrk,
-                                          const uint8_t* src);
+extern void WebPRescalerImportRowExpand_C(
+    struct WebPRescaler* WEBP_RESTRICT const wrk,
+    const uint8_t* WEBP_RESTRICT src);
+extern void WebPRescalerImportRowShrink_C(
+    struct WebPRescaler* WEBP_RESTRICT const wrk,
+    const uint8_t* WEBP_RESTRICT src);
 extern void WebPRescalerExportRowExpand_C(struct WebPRescaler* const wrk);
 extern void WebPRescalerExportRowShrink_C(struct WebPRescaler* const wrk);
 
 // Main entry calls:
-extern void WebPRescalerImportRow(struct WebPRescaler* const wrk,
-                                  const uint8_t* src);
+extern void WebPRescalerImportRow(
+    struct WebPRescaler* WEBP_RESTRICT const wrk,
+    const uint8_t* WEBP_RESTRICT src);
 // Export one row (starting at x_out position) from rescaler.
 extern void WebPRescalerExportRow(struct WebPRescaler* const wrk);
 
diff --git a/src/dsp/rescaler.c b/src/dsp/rescaler.c
index 325d8be1..a96ca669 100644
--- a/src/dsp/rescaler.c
+++ b/src/dsp/rescaler.c
@@ -26,8 +26,8 @@
 //------------------------------------------------------------------------------
 // Row import
 
-void WebPRescalerImportRowExpand_C(WebPRescaler* const wrk,
-                                   const uint8_t* src) {
+void WebPRescalerImportRowExpand_C(WebPRescaler* WEBP_RESTRICT const wrk,
+                                   const uint8_t* WEBP_RESTRICT src) {
   const int x_stride = wrk->num_channels;
   const int x_out_max = wrk->dst_width * wrk->num_channels;
   int channel;
@@ -59,8 +59,8 @@ void WebPRescalerImportRowExpand_C(WebPRescaler* const wrk,
   }
 }
 
-void WebPRescalerImportRowShrink_C(WebPRescaler* const wrk,
-                                   const uint8_t* src) {
+void WebPRescalerImportRowShrink_C(WebPRescaler* WEBP_RESTRICT const wrk,
+                                   const uint8_t* WEBP_RESTRICT src) {
   const int x_stride = wrk->num_channels;
   const int x_out_max = wrk->dst_width * wrk->num_channels;
   int channel;
@@ -158,7 +158,8 @@ void WebPRescalerExportRowShrink_C(WebPRescaler* const wrk) {
 //------------------------------------------------------------------------------
 // Main entry calls
 
-void WebPRescalerImportRow(WebPRescaler* const wrk, const uint8_t* src) {
+void WebPRescalerImportRow(WebPRescaler* WEBP_RESTRICT const wrk,
+                           const uint8_t* WEBP_RESTRICT src) {
   assert(!WebPRescalerInputDone(wrk));
   if (!wrk->x_expand) {
     WebPRescalerImportRowShrink(wrk, src);
diff --git a/src/dsp/rescaler_mips32.c b/src/dsp/rescaler_mips32.c
index 61f63c61..b5168caa 100644
--- a/src/dsp/rescaler_mips32.c
+++ b/src/dsp/rescaler_mips32.c
@@ -21,8 +21,8 @@
 //------------------------------------------------------------------------------
 // Row import
 
-static void ImportRowShrink_MIPS32(WebPRescaler* const wrk,
-                                   const uint8_t* src) {
+static void ImportRowShrink_MIPS32(WebPRescaler* WEBP_RESTRICT const wrk,
+                                   const uint8_t* WEBP_RESTRICT src) {
   const int x_stride = wrk->num_channels;
   const int x_out_max = wrk->dst_width * wrk->num_channels;
   const int fx_scale = wrk->fx_scale;
@@ -81,8 +81,8 @@ static void ImportRowShrink_MIPS32(WebPRescaler* const wrk,
   }
 }
 
-static void ImportRowExpand_MIPS32(WebPRescaler* const wrk,
-                                   const uint8_t* src) {
+static void ImportRowExpand_MIPS32(WebPRescaler* WEBP_RESTRICT const wrk,
+                                   const uint8_t* WEBP_RESTRICT src) {
   const int x_stride = wrk->num_channels;
   const int x_out_max = wrk->dst_width * wrk->num_channels;
   const int x_add = wrk->x_add;
diff --git a/src/dsp/rescaler_msa.c b/src/dsp/rescaler_msa.c
index 256dbdd4..954d0fdf 100644
--- a/src/dsp/rescaler_msa.c
+++ b/src/dsp/rescaler_msa.c
@@ -114,9 +114,9 @@
   dst = __msa_copy_s_w((v4i32)t0, 0);                             \
 } while (0)
 
-static WEBP_INLINE void ExportRowExpand_0(const uint32_t* frow, uint8_t* dst,
-                                          int length,
-                                          WebPRescaler* const wrk) {
+static WEBP_INLINE void ExportRowExpand_0(
+    const uint32_t* WEBP_RESTRICT frow, uint8_t* WEBP_RESTRICT dst, int length,
+    WebPRescaler* WEBP_RESTRICT const wrk) {
   const v4u32 scale = (v4u32)__msa_fill_w(wrk->fy_scale);
   const v4u32 shift = (v4u32)__msa_fill_w(WEBP_RESCALER_RFIX);
   const v4i32 zero = { 0 };
@@ -171,9 +171,10 @@ static WEBP_INLINE void ExportRowExpand_0(const uint32_t* frow, uint8_t* dst,
   }
 }
 
-static WEBP_INLINE void ExportRowExpand_1(const uint32_t* frow, uint32_t* irow,
-                                          uint8_t* dst, int length,
-                                          WebPRescaler* const wrk) {
+static WEBP_INLINE void ExportRowExpand_1(
+    const uint32_t* WEBP_RESTRICT frow, uint32_t* WEBP_RESTRICT irow,
+    uint8_t* WEBP_RESTRICT dst, int length,
+    WebPRescaler* WEBP_RESTRICT const wrk) {
   const uint32_t B = WEBP_RESCALER_FRAC(-wrk->y_accum, wrk->y_sub);
   const uint32_t A = (uint32_t)(WEBP_RESCALER_ONE - B);
   const v4i32 B1 = __msa_fill_w(B);
@@ -262,10 +263,10 @@ static void RescalerExportRowExpand_MIPSdspR2(WebPRescaler* const wrk) {
 }
 
 #if 0  // disabled for now. TODO(skal): make match the C-code
-static WEBP_INLINE void ExportRowShrink_0(const uint32_t* frow, uint32_t* irow,
-                                          uint8_t* dst, int length,
-                                          const uint32_t yscale,
-                                          WebPRescaler* const wrk) {
+static WEBP_INLINE void ExportRowShrink_0(
+    const uint32_t* WEBP_RESTRICT frow, uint32_t* WEBP_RESTRICT irow,
+    uint8_t* WEBP_RESTRICT dst, int length, const uint32_t yscale,
+    WebPRescaler* WEBP_RESTRICT const wrk) {
   const v4u32 y_scale = (v4u32)__msa_fill_w(yscale);
   const v4u32 fxyscale = (v4u32)__msa_fill_w(wrk->fxy_scale);
   const v4u32 shiftval = (v4u32)__msa_fill_w(WEBP_RESCALER_RFIX);
@@ -348,9 +349,9 @@ static WEBP_INLINE void ExportRowShrink_0(const uint32_t* frow, uint32_t* irow,
   }
 }
 
-static WEBP_INLINE void ExportRowShrink_1(uint32_t* irow, uint8_t* dst,
-                                          int length,
-                                          WebPRescaler* const wrk) {
+static WEBP_INLINE void ExportRowShrink_1(
+    uint32_t* WEBP_RESTRICT irow, uint8_t* WEBP_RESTRICT dst, int length,
+    WebPRescaler* WEBP_RESTRICT const wrk) {
   const v4u32 scale = (v4u32)__msa_fill_w(wrk->fxy_scale);
   const v4u32 shift = (v4u32)__msa_fill_w(WEBP_RESCALER_RFIX);
   const v4i32 zero = { 0 };
diff --git a/src/dsp/rescaler_neon.c b/src/dsp/rescaler_neon.c
index 957a92db..ab4ddc00 100644
--- a/src/dsp/rescaler_neon.c
+++ b/src/dsp/rescaler_neon.c
@@ -45,8 +45,8 @@
 #error "MULT_FIX/WEBP_RESCALER_RFIX need some more work"
 #endif
 
-static uint32x4_t Interpolate_NEON(const rescaler_t* const frow,
-                                   const rescaler_t* const irow,
+static uint32x4_t Interpolate_NEON(const rescaler_t* WEBP_RESTRICT const frow,
+                                   const rescaler_t* WEBP_RESTRICT const irow,
                                    uint32_t A, uint32_t B) {
   LOAD_32x4(frow, A0);
   LOAD_32x4(irow, B0);
diff --git a/src/dsp/rescaler_sse2.c b/src/dsp/rescaler_sse2.c
index 3f18e94e..e898e2ac 100644
--- a/src/dsp/rescaler_sse2.c
+++ b/src/dsp/rescaler_sse2.c
@@ -43,8 +43,8 @@ static void LoadEightPixels_SSE2(const uint8_t* const src, __m128i* out) {
   *out = _mm_unpacklo_epi8(A, zero);
 }
 
-static void RescalerImportRowExpand_SSE2(WebPRescaler* const wrk,
-                                         const uint8_t* src) {
+static void RescalerImportRowExpand_SSE2(WebPRescaler* WEBP_RESTRICT const wrk,
+                                         const uint8_t* WEBP_RESTRICT src) {
   rescaler_t* frow = wrk->frow;
   const rescaler_t* const frow_end = frow + wrk->dst_width * wrk->num_channels;
   const int x_add = wrk->x_add;
@@ -109,8 +109,8 @@ static void RescalerImportRowExpand_SSE2(WebPRescaler* const wrk,
   assert(accum == 0);
 }
 
-static void RescalerImportRowShrink_SSE2(WebPRescaler* const wrk,
-                                         const uint8_t* src) {
+static void RescalerImportRowShrink_SSE2(WebPRescaler* WEBP_RESTRICT const wrk,
+                                         const uint8_t* WEBP_RESTRICT src) {
   const int x_sub = wrk->x_sub;
   int accum = 0;
   const __m128i zero = _mm_setzero_si128();
@@ -168,12 +168,10 @@ static void RescalerImportRowShrink_SSE2(WebPRescaler* const wrk,
 // Row export
 
 // load *src as epi64, multiply by mult and store result in [out0 ... out3]
-static WEBP_INLINE void LoadDispatchAndMult_SSE2(const rescaler_t* const src,
-                                                 const __m128i* const mult,
-                                                 __m128i* const out0,
-                                                 __m128i* const out1,
-                                                 __m128i* const out2,
-                                                 __m128i* const out3) {
+static WEBP_INLINE void LoadDispatchAndMult_SSE2(
+    const rescaler_t* WEBP_RESTRICT const src, const __m128i* const mult,
+    __m128i* const out0, __m128i* const out1, __m128i* const out2,
+    __m128i* const out3) {
   const __m128i A0 = _mm_loadu_si128((const __m128i*)(src + 0));
   const __m128i A1 = _mm_loadu_si128((const __m128i*)(src + 4));
   const __m128i A2 = _mm_srli_epi64(A0, 32);

From 23bbafbeb8ea2bbd701738d712855a5bb7330838 Mon Sep 17 00:00:00 2001
From: James Zern <jzern@google.com>
Date: Fri, 16 Aug 2024 19:02:19 -0700
Subject: [PATCH 7/8] dsp/upsampling*: use WEBP_RESTRICT qualifier

Better vectorization in the C code, fewer instructions in NEON, and some
code reordering / better register usage in SSE2/SSE4 w/ndk
r27/gcc-13/clang-16.

This only affects non-vector pointers; any vector pointers are left as a
follow up.

Change-Id: Ib29980f778ad3dbb952178ad8dee39b8673c4ff8
---
 src/dsp/dsp.h                    | 29 +++++++++--------
 src/dsp/upsampling.c             | 36 ++++++++++++++-------
 src/dsp/upsampling_mips_dsp_r2.c | 18 +++++++----
 src/dsp/upsampling_msa.c         | 55 +++++++++++++++++++++-----------
 src/dsp/upsampling_neon.c        | 17 ++++++----
 src/dsp/upsampling_sse2.c        | 29 +++++++++++------
 src/dsp/upsampling_sse41.c       | 29 +++++++++++------
 7 files changed, 137 insertions(+), 76 deletions(-)

diff --git a/src/dsp/dsp.h b/src/dsp/dsp.h
index b7c2acf0..8e0b4143 100644
--- a/src/dsp/dsp.h
+++ b/src/dsp/dsp.h
@@ -285,10 +285,10 @@ void VP8DspInit(void);
 // Convert a pair of y/u/v lines together to the output rgb/a colorspace.
 // bottom_y can be NULL if only one line of output is needed (at top/bottom).
 typedef void (*WebPUpsampleLinePairFunc)(
-    const uint8_t* top_y, const uint8_t* bottom_y,
-    const uint8_t* top_u, const uint8_t* top_v,
-    const uint8_t* cur_u, const uint8_t* cur_v,
-    uint8_t* top_dst, uint8_t* bottom_dst, int len);
+    const uint8_t* WEBP_RESTRICT top_y, const uint8_t* WEBP_RESTRICT bottom_y,
+    const uint8_t* WEBP_RESTRICT top_u, const uint8_t* WEBP_RESTRICT top_v,
+    const uint8_t* WEBP_RESTRICT cur_u, const uint8_t* WEBP_RESTRICT cur_v,
+    uint8_t* WEBP_RESTRICT top_dst, uint8_t* WEBP_RESTRICT bottom_dst, int len);
 
 #ifdef FANCY_UPSAMPLING
 
@@ -298,13 +298,15 @@ extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];
 #endif    // FANCY_UPSAMPLING
 
 // Per-row point-sampling methods.
-typedef void (*WebPSamplerRowFunc)(const uint8_t* y,
-                                   const uint8_t* u, const uint8_t* v,
-                                   uint8_t* dst, int len);
+typedef void (*WebPSamplerRowFunc)(const uint8_t* WEBP_RESTRICT y,
+                                   const uint8_t* WEBP_RESTRICT u,
+                                   const uint8_t* WEBP_RESTRICT v,
+                                   uint8_t* WEBP_RESTRICT dst, int len);
 // Generic function to apply 'WebPSamplerRowFunc' to the whole plane:
-void WebPSamplerProcessPlane(const uint8_t* y, int y_stride,
-                             const uint8_t* u, const uint8_t* v, int uv_stride,
-                             uint8_t* dst, int dst_stride,
+void WebPSamplerProcessPlane(const uint8_t* WEBP_RESTRICT y, int y_stride,
+                             const uint8_t* WEBP_RESTRICT u,
+                             const uint8_t* WEBP_RESTRICT v, int uv_stride,
+                             uint8_t* WEBP_RESTRICT dst, int dst_stride,
                              int width, int height, WebPSamplerRowFunc func);
 
 // Sampling functions to convert rows of YUV to RGB(A)
@@ -316,9 +318,10 @@ extern WebPSamplerRowFunc WebPSamplers[/* MODE_LAST */];
 WebPUpsampleLinePairFunc WebPGetLinePairConverter(int alpha_is_last);
 
 // YUV444->RGB converters
-typedef void (*WebPYUV444Converter)(const uint8_t* y,
-                                    const uint8_t* u, const uint8_t* v,
-                                    uint8_t* dst, int len);
+typedef void (*WebPYUV444Converter)(const uint8_t* WEBP_RESTRICT y,
+                                    const uint8_t* WEBP_RESTRICT u,
+                                    const uint8_t* WEBP_RESTRICT v,
+                                    uint8_t* WEBP_RESTRICT dst, int len);
 
 extern WebPYUV444Converter WebPYUV444Converters[/* MODE_LAST */];
 
diff --git a/src/dsp/upsampling.c b/src/dsp/upsampling.c
index 983b9c42..5953fe48 100644
--- a/src/dsp/upsampling.c
+++ b/src/dsp/upsampling.c
@@ -35,10 +35,14 @@ WebPUpsampleLinePairFunc WebPUpsamplers[MODE_LAST];
 #define LOAD_UV(u, v) ((u) | ((v) << 16))
 
 #define UPSAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP)                                  \
-static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y,           \
-                      const uint8_t* top_u, const uint8_t* top_v,              \
-                      const uint8_t* cur_u, const uint8_t* cur_v,              \
-                      uint8_t* top_dst, uint8_t* bottom_dst, int len) {        \
+static void FUNC_NAME(const uint8_t* WEBP_RESTRICT top_y,                      \
+                      const uint8_t* WEBP_RESTRICT bottom_y,                   \
+                      const uint8_t* WEBP_RESTRICT top_u,                      \
+                      const uint8_t* WEBP_RESTRICT top_v,                      \
+                      const uint8_t* WEBP_RESTRICT cur_u,                      \
+                      const uint8_t* WEBP_RESTRICT cur_v,                      \
+                      uint8_t* WEBP_RESTRICT top_dst,                          \
+                      uint8_t* WEBP_RESTRICT bottom_dst, int len) {            \
   int x;                                                                       \
   const int last_pixel_pair = (len - 1) >> 1;                                  \
   uint32_t tl_uv = LOAD_UV(top_u[0], top_v[0]);   /* top-left sample */        \
@@ -136,10 +140,14 @@ static void EmptyUpsampleFunc(const uint8_t* top_y, const uint8_t* bottom_y,
 
 #if !defined(FANCY_UPSAMPLING)
 #define DUAL_SAMPLE_FUNC(FUNC_NAME, FUNC)                                      \
-static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bot_y,              \
-                      const uint8_t* top_u, const uint8_t* top_v,              \
-                      const uint8_t* bot_u, const uint8_t* bot_v,              \
-                      uint8_t* top_dst, uint8_t* bot_dst, int len) {           \
+static void FUNC_NAME(const uint8_t* WEBP_RESTRICT top_y,                      \
+                      const uint8_t* WEBP_RESTRICT bot_y,                      \
+                      const uint8_t* WEBP_RESTRICT top_u,                      \
+                      const uint8_t* WEBP_RESTRICT top_v,                      \
+                      const uint8_t* WEBP_RESTRICT bot_u,                      \
+                      const uint8_t* WEBP_RESTRICT bot_v,                      \
+                      uint8_t* WEBP_RESTRICT top_dst,                          \
+                      uint8_t* WEBP_RESTRICT bot_dst, int len) {               \
   const int half_len = len >> 1;                                               \
   int x;                                                                       \
   assert(top_dst != NULL);                                                     \
@@ -178,10 +186,14 @@ WebPUpsampleLinePairFunc WebPGetLinePairConverter(int alpha_is_last) {
 // YUV444 converter
 
 #define YUV444_FUNC(FUNC_NAME, FUNC, XSTEP)                                    \
-extern void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v,    \
-                      uint8_t* dst, int len);                                  \
-void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v,           \
-               uint8_t* dst, int len) {                                        \
+extern void FUNC_NAME(const uint8_t* WEBP_RESTRICT y,                          \
+                      const uint8_t* WEBP_RESTRICT u,                          \
+                      const uint8_t* WEBP_RESTRICT v,                          \
+                      uint8_t* WEBP_RESTRICT dst, int len);                    \
+void FUNC_NAME(const uint8_t* WEBP_RESTRICT y,                                 \
+               const uint8_t* WEBP_RESTRICT u,                                 \
+               const uint8_t* WEBP_RESTRICT v,                                 \
+               uint8_t* WEBP_RESTRICT dst, int len) {                          \
   int i;                                                                       \
   for (i = 0; i < len; ++i) FUNC(y[i], u[i], v[i], &dst[i * (XSTEP)]);         \
 }
diff --git a/src/dsp/upsampling_mips_dsp_r2.c b/src/dsp/upsampling_mips_dsp_r2.c
index 10d499d7..cbe8e71d 100644
--- a/src/dsp/upsampling_mips_dsp_r2.c
+++ b/src/dsp/upsampling_mips_dsp_r2.c
@@ -143,10 +143,14 @@ static WEBP_INLINE void YuvToRgba(uint8_t y, uint8_t u, uint8_t v,
 #define LOAD_UV(u, v) ((u) | ((v) << 16))
 
 #define UPSAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP)                                  \
-static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y,           \
-                      const uint8_t* top_u, const uint8_t* top_v,              \
-                      const uint8_t* cur_u, const uint8_t* cur_v,              \
-                      uint8_t* top_dst, uint8_t* bottom_dst, int len) {        \
+static void FUNC_NAME(const uint8_t* WEBP_RESTRICT top_y,                      \
+                      const uint8_t* WEBP_RESTRICT bottom_y,                   \
+                      const uint8_t* WEBP_RESTRICT top_u,                      \
+                      const uint8_t* WEBP_RESTRICT top_v,                      \
+                      const uint8_t* WEBP_RESTRICT cur_u,                      \
+                      const uint8_t* WEBP_RESTRICT cur_v,                      \
+                      uint8_t* WEBP_RESTRICT top_dst,                          \
+                      uint8_t* WEBP_RESTRICT bottom_dst, int len) {            \
   int x;                                                                       \
   const int last_pixel_pair = (len - 1) >> 1;                                  \
   uint32_t tl_uv = LOAD_UV(top_u[0], top_v[0]);   /* top-left sample */        \
@@ -241,8 +245,10 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplersMIPSdspR2(void) {
 // YUV444 converter
 
 #define YUV444_FUNC(FUNC_NAME, FUNC, XSTEP)                                    \
-static void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v,    \
-                      uint8_t* dst, int len) {                                 \
+static void FUNC_NAME(const uint8_t* WEBP_RESTRICT y,                          \
+                      const uint8_t* WEBP_RESTRICT u,                          \
+                      const uint8_t* WEBP_RESTRICT v,                          \
+                      uint8_t* WEBP_RESTRICT dst, int len) {                   \
   int i;                                                                       \
   for (i = 0; i < len; ++i) FUNC(y[i], u[i], v[i], &dst[i * XSTEP]);           \
 }
diff --git a/src/dsp/upsampling_msa.c b/src/dsp/upsampling_msa.c
index f2e03e85..72a526bc 100644
--- a/src/dsp/upsampling_msa.c
+++ b/src/dsp/upsampling_msa.c
@@ -320,8 +320,10 @@ static void YuvToRgba(uint8_t y, uint8_t u, uint8_t v, uint8_t* const rgba) {
 }
 
 #if !defined(WEBP_REDUCE_CSP)
-static void YuvToRgbLine(const uint8_t* y, const uint8_t* u,
-                         const uint8_t* v, uint8_t* dst, int length) {
+static void YuvToRgbLine(const uint8_t* WEBP_RESTRICT y,
+                         const uint8_t* WEBP_RESTRICT u,
+                         const uint8_t* WEBP_RESTRICT v,
+                         uint8_t* WEBP_RESTRICT dst, int length) {
   v16u8 R, G, B;
   while (length >= 16) {
     CALC_RGB16(y, u, v, R, G, B);
@@ -347,8 +349,10 @@ static void YuvToRgbLine(const uint8_t* y, const uint8_t* u,
   }
 }
 
-static void YuvToBgrLine(const uint8_t* y, const uint8_t* u,
-                         const uint8_t* v, uint8_t* dst, int length) {
+static void YuvToBgrLine(const uint8_t* WEBP_RESTRICT y,
+                         const uint8_t* WEBP_RESTRICT u,
+                         const uint8_t* WEBP_RESTRICT v,
+                         uint8_t* WEBP_RESTRICT dst, int length) {
   v16u8 R, G, B;
   while (length >= 16) {
     CALC_RGB16(y, u, v, R, G, B);
@@ -375,8 +379,10 @@ static void YuvToBgrLine(const uint8_t* y, const uint8_t* u,
 }
 #endif  // WEBP_REDUCE_CSP
 
-static void YuvToRgbaLine(const uint8_t* y, const uint8_t* u,
-                          const uint8_t* v, uint8_t* dst, int length) {
+static void YuvToRgbaLine(const uint8_t* WEBP_RESTRICT y,
+                          const uint8_t* WEBP_RESTRICT u,
+                          const uint8_t* WEBP_RESTRICT v,
+                          uint8_t* WEBP_RESTRICT dst, int length) {
   v16u8 R, G, B;
   const v16u8 A = (v16u8)__msa_ldi_b(ALPHAVAL);
   while (length >= 16) {
@@ -403,8 +409,10 @@ static void YuvToRgbaLine(const uint8_t* y, const uint8_t* u,
   }
 }
 
-static void YuvToBgraLine(const uint8_t* y, const uint8_t* u,
-                          const uint8_t* v, uint8_t* dst, int length) {
+static void YuvToBgraLine(const uint8_t* WEBP_RESTRICT y,
+                          const uint8_t* WEBP_RESTRICT u,
+                          const uint8_t* WEBP_RESTRICT v,
+                          uint8_t* WEBP_RESTRICT dst, int length) {
   v16u8 R, G, B;
   const v16u8 A = (v16u8)__msa_ldi_b(ALPHAVAL);
   while (length >= 16) {
@@ -432,8 +440,10 @@ static void YuvToBgraLine(const uint8_t* y, const uint8_t* u,
 }
 
 #if !defined(WEBP_REDUCE_CSP)
-static void YuvToArgbLine(const uint8_t* y, const uint8_t* u,
-                          const uint8_t* v, uint8_t* dst, int length) {
+static void YuvToArgbLine(const uint8_t* WEBP_RESTRICT y,
+                          const uint8_t* WEBP_RESTRICT u,
+                          const uint8_t* WEBP_RESTRICT v,
+                          uint8_t* WEBP_RESTRICT dst, int length) {
   v16u8 R, G, B;
   const v16u8 A = (v16u8)__msa_ldi_b(ALPHAVAL);
   while (length >= 16) {
@@ -460,8 +470,10 @@ static void YuvToArgbLine(const uint8_t* y, const uint8_t* u,
   }
 }
 
-static void YuvToRgba4444Line(const uint8_t* y, const uint8_t* u,
-                              const uint8_t* v, uint8_t* dst, int length) {
+static void YuvToRgba4444Line(const uint8_t* WEBP_RESTRICT y,
+                              const uint8_t* WEBP_RESTRICT u,
+                              const uint8_t* WEBP_RESTRICT v,
+                              uint8_t* WEBP_RESTRICT dst, int length) {
   v16u8 R, G, B, RG, BA, tmp0, tmp1;
   while (length >= 16) {
 #if (WEBP_SWAP_16BIT_CSP == 1)
@@ -496,8 +508,10 @@ static void YuvToRgba4444Line(const uint8_t* y, const uint8_t* u,
   }
 }
 
-static void YuvToRgb565Line(const uint8_t* y, const uint8_t* u,
-                            const uint8_t* v, uint8_t* dst, int length) {
+static void YuvToRgb565Line(const uint8_t* WEBP_RESTRICT y,
+                            const uint8_t* WEBP_RESTRICT u,
+                            const uint8_t* WEBP_RESTRICT v,
+                            uint8_t* WEBP_RESTRICT dst, int length) {
   v16u8 R, G, B, RG, GB, tmp0, tmp1;
   while (length >= 16) {
 #if (WEBP_SWAP_16BIT_CSP == 1)
@@ -564,11 +578,14 @@ static void YuvToRgb565Line(const uint8_t* y, const uint8_t* u,
 } while (0)
 
 #define UPSAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP)                            \
-static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bot_y,        \
-                      const uint8_t* top_u, const uint8_t* top_v,        \
-                      const uint8_t* cur_u, const uint8_t* cur_v,        \
-                      uint8_t* top_dst, uint8_t* bot_dst, int len)       \
-{                                                                        \
+static void FUNC_NAME(const uint8_t* WEBP_RESTRICT top_y,                \
+                      const uint8_t* WEBP_RESTRICT bot_y,                \
+                      const uint8_t* WEBP_RESTRICT top_u,                \
+                      const uint8_t* WEBP_RESTRICT top_v,                \
+                      const uint8_t* WEBP_RESTRICT cur_u,                \
+                      const uint8_t* WEBP_RESTRICT cur_v,                \
+                      uint8_t* WEBP_RESTRICT top_dst,                    \
+                      uint8_t* WEBP_RESTRICT bot_dst, int len) {         \
   int size = (len - 1) >> 1;                                             \
   uint8_t temp_u[64];                                                    \
   uint8_t temp_v[64];                                                    \
diff --git a/src/dsp/upsampling_neon.c b/src/dsp/upsampling_neon.c
index f39d75e0..2bd3e931 100644
--- a/src/dsp/upsampling_neon.c
+++ b/src/dsp/upsampling_neon.c
@@ -58,8 +58,9 @@
 } while (0)
 
 // Turn the macro into a function for reducing code-size when non-critical
-static void Upsample16Pixels_NEON(const uint8_t* r1, const uint8_t* r2,
-                                  uint8_t* out) {
+static void Upsample16Pixels_NEON(const uint8_t* WEBP_RESTRICT const r1,
+                                  const uint8_t* WEBP_RESTRICT const r2,
+                                  uint8_t* WEBP_RESTRICT const out) {
   UPSAMPLE_16PIXELS(r1, r2, out);
 }
 
@@ -190,10 +191,14 @@ static const int16_t kCoeffs1[4] = { 19077, 26149, 6419, 13320 };
 }
 
 #define NEON_UPSAMPLE_FUNC(FUNC_NAME, FMT, XSTEP)                              \
-static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y,           \
-                      const uint8_t* top_u, const uint8_t* top_v,              \
-                      const uint8_t* cur_u, const uint8_t* cur_v,              \
-                      uint8_t* top_dst, uint8_t* bottom_dst, int len) {        \
+static void FUNC_NAME(const uint8_t* WEBP_RESTRICT top_y,                      \
+                      const uint8_t* WEBP_RESTRICT bottom_y,                   \
+                      const uint8_t* WEBP_RESTRICT top_u,                      \
+                      const uint8_t* WEBP_RESTRICT top_v,                      \
+                      const uint8_t* WEBP_RESTRICT cur_u,                      \
+                      const uint8_t* WEBP_RESTRICT cur_v,                      \
+                      uint8_t* WEBP_RESTRICT top_dst,                          \
+                      uint8_t* WEBP_RESTRICT bottom_dst, int len) {            \
   int block;                                                                   \
   /* 16 byte aligned array to cache reconstructed u and v */                   \
   uint8_t uv_buf[2 * 32 + 15];                                                 \
diff --git a/src/dsp/upsampling_sse2.c b/src/dsp/upsampling_sse2.c
index 77b4f722..36226fb1 100644
--- a/src/dsp/upsampling_sse2.c
+++ b/src/dsp/upsampling_sse2.c
@@ -88,8 +88,9 @@
 } while (0)
 
 // Turn the macro into a function for reducing code-size when non-critical
-static void Upsample32Pixels_SSE2(const uint8_t r1[], const uint8_t r2[],
-                                  uint8_t* const out) {
+static void Upsample32Pixels_SSE2(const uint8_t* WEBP_RESTRICT const r1,
+                                  const uint8_t* WEBP_RESTRICT const r2,
+                                  uint8_t* WEBP_RESTRICT const out) {
   UPSAMPLE_32PIXELS(r1, r2, out);
 }
 
@@ -114,10 +115,14 @@ static void Upsample32Pixels_SSE2(const uint8_t r1[], const uint8_t r2[],
 } while (0)
 
 #define SSE2_UPSAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP)                             \
-static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y,           \
-                      const uint8_t* top_u, const uint8_t* top_v,              \
-                      const uint8_t* cur_u, const uint8_t* cur_v,              \
-                      uint8_t* top_dst, uint8_t* bottom_dst, int len) {        \
+static void FUNC_NAME(const uint8_t* WEBP_RESTRICT top_y,                      \
+                      const uint8_t* WEBP_RESTRICT bottom_y,                   \
+                      const uint8_t* WEBP_RESTRICT top_u,                      \
+                      const uint8_t* WEBP_RESTRICT top_v,                      \
+                      const uint8_t* WEBP_RESTRICT cur_u,                      \
+                      const uint8_t* WEBP_RESTRICT cur_v,                      \
+                      uint8_t* WEBP_RESTRICT top_dst,                          \
+                      uint8_t* WEBP_RESTRICT bottom_dst, int len) {            \
   int uv_pos, pos;                                                             \
   /* 16byte-aligned array to cache reconstructed u and v */                    \
   uint8_t uv_buf[14 * 32 + 15] = { 0 };                                        \
@@ -215,10 +220,14 @@ extern WebPYUV444Converter WebPYUV444Converters[/* MODE_LAST */];
 extern void WebPInitYUV444ConvertersSSE2(void);
 
 #define YUV444_FUNC(FUNC_NAME, CALL, CALL_C, XSTEP)                            \
-extern void CALL_C(const uint8_t* y, const uint8_t* u, const uint8_t* v,       \
-                   uint8_t* dst, int len);                                     \
-static void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v,    \
-                      uint8_t* dst, int len) {                                 \
+extern void CALL_C(const uint8_t* WEBP_RESTRICT y,                             \
+                   const uint8_t* WEBP_RESTRICT u,                             \
+                   const uint8_t* WEBP_RESTRICT v,                             \
+                   uint8_t* WEBP_RESTRICT dst, int len);                       \
+static void FUNC_NAME(const uint8_t* WEBP_RESTRICT y,                          \
+                      const uint8_t* WEBP_RESTRICT u,                          \
+                      const uint8_t* WEBP_RESTRICT v,                          \
+                      uint8_t* WEBP_RESTRICT dst, int len) {                   \
   int i;                                                                       \
   const int max_len = len & ~31;                                               \
   for (i = 0; i < max_len; i += 32) {                                          \
diff --git a/src/dsp/upsampling_sse41.c b/src/dsp/upsampling_sse41.c
index a880974a..823633c4 100644
--- a/src/dsp/upsampling_sse41.c
+++ b/src/dsp/upsampling_sse41.c
@@ -90,8 +90,9 @@
 } while (0)
 
 // Turn the macro into a function for reducing code-size when non-critical
-static void Upsample32Pixels_SSE41(const uint8_t r1[], const uint8_t r2[],
-                                  uint8_t* const out) {
+static void Upsample32Pixels_SSE41(const uint8_t* WEBP_RESTRICT const r1,
+                                   const uint8_t* WEBP_RESTRICT const r2,
+                                   uint8_t* WEBP_RESTRICT const out) {
   UPSAMPLE_32PIXELS(r1, r2, out);
 }
 
@@ -116,10 +117,14 @@ static void Upsample32Pixels_SSE41(const uint8_t r1[], const uint8_t r2[],
 } while (0)
 
 #define SSE4_UPSAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP)                             \
-static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y,           \
-                      const uint8_t* top_u, const uint8_t* top_v,              \
-                      const uint8_t* cur_u, const uint8_t* cur_v,              \
-                      uint8_t* top_dst, uint8_t* bottom_dst, int len) {        \
+static void FUNC_NAME(const uint8_t* WEBP_RESTRICT top_y,                      \
+                      const uint8_t* WEBP_RESTRICT bottom_y,                   \
+                      const uint8_t* WEBP_RESTRICT top_u,                      \
+                      const uint8_t* WEBP_RESTRICT top_v,                      \
+                      const uint8_t* WEBP_RESTRICT cur_u,                      \
+                      const uint8_t* WEBP_RESTRICT cur_v,                      \
+                      uint8_t* WEBP_RESTRICT top_dst,                          \
+                      uint8_t* WEBP_RESTRICT bottom_dst, int len) {            \
   int uv_pos, pos;                                                             \
   /* 16byte-aligned array to cache reconstructed u and v */                    \
   uint8_t uv_buf[14 * 32 + 15] = { 0 };                                        \
@@ -202,10 +207,14 @@ extern WebPYUV444Converter WebPYUV444Converters[/* MODE_LAST */];
 extern void WebPInitYUV444ConvertersSSE41(void);
 
 #define YUV444_FUNC(FUNC_NAME, CALL, CALL_C, XSTEP)                            \
-extern void CALL_C(const uint8_t* y, const uint8_t* u, const uint8_t* v,       \
-                   uint8_t* dst, int len);                                     \
-static void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v,    \
-                      uint8_t* dst, int len) {                                 \
+extern void CALL_C(const uint8_t* WEBP_RESTRICT y,                             \
+                   const uint8_t* WEBP_RESTRICT u,                             \
+                   const uint8_t* WEBP_RESTRICT v,                             \
+                   uint8_t* WEBP_RESTRICT dst, int len);                       \
+static void FUNC_NAME(const uint8_t* WEBP_RESTRICT y,                          \
+                      const uint8_t* WEBP_RESTRICT u,                          \
+                      const uint8_t* WEBP_RESTRICT v,                          \
+                      uint8_t* WEBP_RESTRICT dst, int len) {                   \
   int i;                                                                       \
   const int max_len = len & ~31;                                               \
   for (i = 0; i < max_len; i += 32) {                                          \

From 2dd5eb9862d4822d2746cfb55f4b59b53bda99eb Mon Sep 17 00:00:00 2001
From: James Zern <jzern@google.com>
Date: Fri, 16 Aug 2024 19:26:36 -0700
Subject: [PATCH 8/8] dsp/yuv*: use WEBP_RESTRICT qualifier

Better vectorization in the C code, fewer instructions / comparisons in
NEON, and fewer reloads in SSE2/SSE4 w/ndk r27/gcc-13/clang-16.

This only affects non-vector pointers; any vector pointers are left as a
follow up.

Change-Id: I07a7e36a2dce8632c71c0fbbeef94dc51453eaf7
---
 src/dsp/dsp.h             |  27 +++++---
 src/dsp/yuv.c             |  48 ++++++++------
 src/dsp/yuv.h             |  54 ++++++++++------
 src/dsp/yuv_mips32.c      |   7 +-
 src/dsp/yuv_mips_dsp_r2.c |   7 +-
 src/dsp/yuv_neon.c        |  18 ++++--
 src/dsp/yuv_sse2.c        | 130 ++++++++++++++++++++++----------------
 src/dsp/yuv_sse41.c       |  65 +++++++++++--------
 8 files changed, 218 insertions(+), 138 deletions(-)

diff --git a/src/dsp/dsp.h b/src/dsp/dsp.h
index 8e0b4143..1b37ef4b 100644
--- a/src/dsp/dsp.h
+++ b/src/dsp/dsp.h
@@ -337,26 +337,35 @@ void WebPInitYUV444Converters(void);
 // ARGB -> YUV converters
 
 // Convert ARGB samples to luma Y.
-extern void (*WebPConvertARGBToY)(const uint32_t* argb, uint8_t* y, int width);
+extern void (*WebPConvertARGBToY)(const uint32_t* WEBP_RESTRICT argb,
+                                  uint8_t* WEBP_RESTRICT y, int width);
 // Convert ARGB samples to U/V with downsampling. do_store should be '1' for
 // even lines and '0' for odd ones. 'src_width' is the original width, not
 // the U/V one.
-extern void (*WebPConvertARGBToUV)(const uint32_t* argb, uint8_t* u, uint8_t* v,
+extern void (*WebPConvertARGBToUV)(const uint32_t* WEBP_RESTRICT argb,
+                                   uint8_t* WEBP_RESTRICT u,
+                                   uint8_t* WEBP_RESTRICT v,
                                    int src_width, int do_store);
 
 // Convert a row of accumulated (four-values) of rgba32 toward U/V
-extern void (*WebPConvertRGBA32ToUV)(const uint16_t* rgb,
-                                     uint8_t* u, uint8_t* v, int width);
+extern void (*WebPConvertRGBA32ToUV)(const uint16_t* WEBP_RESTRICT rgb,
+                                     uint8_t* WEBP_RESTRICT u,
+                                     uint8_t* WEBP_RESTRICT v, int width);
 
 // Convert RGB or BGR to Y
-extern void (*WebPConvertRGB24ToY)(const uint8_t* rgb, uint8_t* y, int width);
-extern void (*WebPConvertBGR24ToY)(const uint8_t* bgr, uint8_t* y, int width);
+extern void (*WebPConvertRGB24ToY)(const uint8_t* WEBP_RESTRICT rgb,
+                                   uint8_t* WEBP_RESTRICT y, int width);
+extern void (*WebPConvertBGR24ToY)(const uint8_t* WEBP_RESTRICT bgr,
+                                   uint8_t* WEBP_RESTRICT y, int width);
 
 // used for plain-C fallback.
-extern void WebPConvertARGBToUV_C(const uint32_t* argb, uint8_t* u, uint8_t* v,
+extern void WebPConvertARGBToUV_C(const uint32_t* WEBP_RESTRICT argb,
+                                  uint8_t* WEBP_RESTRICT u,
+                                  uint8_t* WEBP_RESTRICT v,
                                   int src_width, int do_store);
-extern void WebPConvertRGBA32ToUV_C(const uint16_t* rgb,
-                                    uint8_t* u, uint8_t* v, int width);
+extern void WebPConvertRGBA32ToUV_C(const uint16_t* WEBP_RESTRICT rgb,
+                                    uint8_t* WEBP_RESTRICT u,
+                                    uint8_t* WEBP_RESTRICT v, int width);
 
 // Must be called before using the above.
 void WebPInitConvertARGBToYUV(void);
diff --git a/src/dsp/yuv.c b/src/dsp/yuv.c
index 8a04b85d..c1320f28 100644
--- a/src/dsp/yuv.c
+++ b/src/dsp/yuv.c
@@ -20,9 +20,10 @@
 // Plain-C version
 
 #define ROW_FUNC(FUNC_NAME, FUNC, XSTEP)                                       \
-static void FUNC_NAME(const uint8_t* y,                                        \
-                      const uint8_t* u, const uint8_t* v,                      \
-                      uint8_t* dst, int len) {                                 \
+static void FUNC_NAME(const uint8_t* WEBP_RESTRICT y,                          \
+                      const uint8_t* WEBP_RESTRICT u,                          \
+                      const uint8_t* WEBP_RESTRICT v,                          \
+                      uint8_t* WEBP_RESTRICT dst, int len) {                   \
   const uint8_t* const end = dst + (len & ~1) * (XSTEP);                       \
   while (dst != end) {                                                         \
     FUNC(y[0], u[0], v[0], dst);                                               \
@@ -49,9 +50,10 @@ ROW_FUNC(YuvToRgb565Row,   VP8YuvToRgb565, 2)
 #undef ROW_FUNC
 
 // Main call for processing a plane with a WebPSamplerRowFunc function:
-void WebPSamplerProcessPlane(const uint8_t* y, int y_stride,
-                             const uint8_t* u, const uint8_t* v, int uv_stride,
-                             uint8_t* dst, int dst_stride,
+void WebPSamplerProcessPlane(const uint8_t* WEBP_RESTRICT y, int y_stride,
+                             const uint8_t* WEBP_RESTRICT u,
+                             const uint8_t* WEBP_RESTRICT v, int uv_stride,
+                             uint8_t* WEBP_RESTRICT dst, int dst_stride,
                              int width, int height, WebPSamplerRowFunc func) {
   int j;
   for (j = 0; j < height; ++j) {
@@ -117,7 +119,8 @@ WEBP_DSP_INIT_FUNC(WebPInitSamplers) {
 //-----------------------------------------------------------------------------
 // ARGB -> YUV converters
 
-static void ConvertARGBToY_C(const uint32_t* argb, uint8_t* y, int width) {
+static void ConvertARGBToY_C(const uint32_t* WEBP_RESTRICT argb,
+                             uint8_t* WEBP_RESTRICT y, int width) {
   int i;
   for (i = 0; i < width; ++i) {
     const uint32_t p = argb[i];
@@ -126,7 +129,8 @@ static void ConvertARGBToY_C(const uint32_t* argb, uint8_t* y, int width) {
   }
 }
 
-void WebPConvertARGBToUV_C(const uint32_t* argb, uint8_t* u, uint8_t* v,
+void WebPConvertARGBToUV_C(const uint32_t* WEBP_RESTRICT argb,
+                           uint8_t* WEBP_RESTRICT u, uint8_t* WEBP_RESTRICT v,
                            int src_width, int do_store) {
   // No rounding. Last pixel is dealt with separately.
   const int uv_width = src_width >> 1;
@@ -169,22 +173,25 @@ void WebPConvertARGBToUV_C(const uint32_t* argb, uint8_t* u, uint8_t* v,
 
 //-----------------------------------------------------------------------------
 
-static void ConvertRGB24ToY_C(const uint8_t* rgb, uint8_t* y, int width) {
+static void ConvertRGB24ToY_C(const uint8_t* WEBP_RESTRICT rgb,
+                              uint8_t* WEBP_RESTRICT y, int width) {
   int i;
   for (i = 0; i < width; ++i, rgb += 3) {
     y[i] = VP8RGBToY(rgb[0], rgb[1], rgb[2], YUV_HALF);
   }
 }
 
-static void ConvertBGR24ToY_C(const uint8_t* bgr, uint8_t* y, int width) {
+static void ConvertBGR24ToY_C(const uint8_t* WEBP_RESTRICT bgr,
+                              uint8_t* WEBP_RESTRICT y, int width) {
   int i;
   for (i = 0; i < width; ++i, bgr += 3) {
     y[i] = VP8RGBToY(bgr[2], bgr[1], bgr[0], YUV_HALF);
   }
 }
 
-void WebPConvertRGBA32ToUV_C(const uint16_t* rgb,
-                             uint8_t* u, uint8_t* v, int width) {
+void WebPConvertRGBA32ToUV_C(const uint16_t* WEBP_RESTRICT rgb,
+                             uint8_t* WEBP_RESTRICT u, uint8_t* WEBP_RESTRICT v,
+                             int width) {
   int i;
   for (i = 0; i < width; i += 1, rgb += 4) {
     const int r = rgb[0], g = rgb[1], b = rgb[2];
@@ -195,13 +202,18 @@ void WebPConvertRGBA32ToUV_C(const uint16_t* rgb,
 
 //-----------------------------------------------------------------------------
 
-void (*WebPConvertRGB24ToY)(const uint8_t* rgb, uint8_t* y, int width);
-void (*WebPConvertBGR24ToY)(const uint8_t* bgr, uint8_t* y, int width);
-void (*WebPConvertRGBA32ToUV)(const uint16_t* rgb,
-                              uint8_t* u, uint8_t* v, int width);
+void (*WebPConvertRGB24ToY)(const uint8_t* WEBP_RESTRICT rgb,
+                            uint8_t* WEBP_RESTRICT y, int width);
+void (*WebPConvertBGR24ToY)(const uint8_t* WEBP_RESTRICT bgr,
+                            uint8_t* WEBP_RESTRICT y, int width);
+void (*WebPConvertRGBA32ToUV)(const uint16_t* WEBP_RESTRICT rgb,
+                              uint8_t* WEBP_RESTRICT u,
+                              uint8_t* WEBP_RESTRICT v, int width);
 
-void (*WebPConvertARGBToY)(const uint32_t* argb, uint8_t* y, int width);
-void (*WebPConvertARGBToUV)(const uint32_t* argb, uint8_t* u, uint8_t* v,
+void (*WebPConvertARGBToY)(const uint32_t* WEBP_RESTRICT argb,
+                           uint8_t* WEBP_RESTRICT y, int width);
+void (*WebPConvertARGBToUV)(const uint32_t* WEBP_RESTRICT argb,
+                            uint8_t* WEBP_RESTRICT u, uint8_t* WEBP_RESTRICT v,
                             int src_width, int do_store);
 
 extern void WebPInitConvertARGBToYUVSSE2(void);
diff --git a/src/dsp/yuv.h b/src/dsp/yuv.h
index 66a397d1..91fdba12 100644
--- a/src/dsp/yuv.h
+++ b/src/dsp/yuv.h
@@ -149,20 +149,34 @@ static WEBP_INLINE void VP8YuvToRgba(uint8_t y, uint8_t u, uint8_t v,
 #if defined(WEBP_USE_SSE2)
 
 // Process 32 pixels and store the result (16b, 24b or 32b per pixel) in *dst.
-void VP8YuvToRgba32_SSE2(const uint8_t* y, const uint8_t* u, const uint8_t* v,
-                         uint8_t* dst);
-void VP8YuvToRgb32_SSE2(const uint8_t* y, const uint8_t* u, const uint8_t* v,
-                        uint8_t* dst);
-void VP8YuvToBgra32_SSE2(const uint8_t* y, const uint8_t* u, const uint8_t* v,
-                         uint8_t* dst);
-void VP8YuvToBgr32_SSE2(const uint8_t* y, const uint8_t* u, const uint8_t* v,
-                        uint8_t* dst);
-void VP8YuvToArgb32_SSE2(const uint8_t* y, const uint8_t* u, const uint8_t* v,
-                         uint8_t* dst);
-void VP8YuvToRgba444432_SSE2(const uint8_t* y, const uint8_t* u,
-                             const uint8_t* v, uint8_t* dst);
-void VP8YuvToRgb56532_SSE2(const uint8_t* y, const uint8_t* u, const uint8_t* v,
-                           uint8_t* dst);
+void VP8YuvToRgba32_SSE2(const uint8_t* WEBP_RESTRICT y,
+                         const uint8_t* WEBP_RESTRICT u,
+                         const uint8_t* WEBP_RESTRICT v,
+                         uint8_t* WEBP_RESTRICT dst);
+void VP8YuvToRgb32_SSE2(const uint8_t* WEBP_RESTRICT y,
+                        const uint8_t* WEBP_RESTRICT u,
+                        const uint8_t* WEBP_RESTRICT v,
+                        uint8_t* WEBP_RESTRICT dst);
+void VP8YuvToBgra32_SSE2(const uint8_t* WEBP_RESTRICT y,
+                         const uint8_t* WEBP_RESTRICT u,
+                         const uint8_t* WEBP_RESTRICT v,
+                         uint8_t* WEBP_RESTRICT dst);
+void VP8YuvToBgr32_SSE2(const uint8_t* WEBP_RESTRICT y,
+                        const uint8_t* WEBP_RESTRICT u,
+                        const uint8_t* WEBP_RESTRICT v,
+                        uint8_t* WEBP_RESTRICT dst);
+void VP8YuvToArgb32_SSE2(const uint8_t* WEBP_RESTRICT y,
+                         const uint8_t* WEBP_RESTRICT u,
+                         const uint8_t* WEBP_RESTRICT v,
+                         uint8_t* WEBP_RESTRICT dst);
+void VP8YuvToRgba444432_SSE2(const uint8_t* WEBP_RESTRICT y,
+                             const uint8_t* WEBP_RESTRICT u,
+                             const uint8_t* WEBP_RESTRICT v,
+                             uint8_t* WEBP_RESTRICT dst);
+void VP8YuvToRgb56532_SSE2(const uint8_t* WEBP_RESTRICT y,
+                           const uint8_t* WEBP_RESTRICT u,
+                           const uint8_t* WEBP_RESTRICT v,
+                           uint8_t* WEBP_RESTRICT dst);
 
 #endif    // WEBP_USE_SSE2
 
@@ -172,10 +186,14 @@ void VP8YuvToRgb56532_SSE2(const uint8_t* y, const uint8_t* u, const uint8_t* v,
 #if defined(WEBP_USE_SSE41)
 
 // Process 32 pixels and store the result (16b, 24b or 32b per pixel) in *dst.
-void VP8YuvToRgb32_SSE41(const uint8_t* y, const uint8_t* u, const uint8_t* v,
-                         uint8_t* dst);
-void VP8YuvToBgr32_SSE41(const uint8_t* y, const uint8_t* u, const uint8_t* v,
-                         uint8_t* dst);
+void VP8YuvToRgb32_SSE41(const uint8_t* WEBP_RESTRICT y,
+                         const uint8_t* WEBP_RESTRICT u,
+                         const uint8_t* WEBP_RESTRICT v,
+                         uint8_t* WEBP_RESTRICT dst);
+void VP8YuvToBgr32_SSE41(const uint8_t* WEBP_RESTRICT y,
+                         const uint8_t* WEBP_RESTRICT u,
+                         const uint8_t* WEBP_RESTRICT v,
+                         uint8_t* WEBP_RESTRICT dst);
 
 #endif    // WEBP_USE_SSE41
 
diff --git a/src/dsp/yuv_mips32.c b/src/dsp/yuv_mips32.c
index 9d0a8878..1f634858 100644
--- a/src/dsp/yuv_mips32.c
+++ b/src/dsp/yuv_mips32.c
@@ -22,9 +22,10 @@
 // simple point-sampling
 
 #define ROW_FUNC(FUNC_NAME, XSTEP, R, G, B, A)                                 \
-static void FUNC_NAME(const uint8_t* y,                                        \
-                      const uint8_t* u, const uint8_t* v,                      \
-                      uint8_t* dst, int len) {                                 \
+static void FUNC_NAME(const uint8_t* WEBP_RESTRICT y,                          \
+                      const uint8_t* WEBP_RESTRICT u,                          \
+                      const uint8_t* WEBP_RESTRICT v,                          \
+                      uint8_t* WEBP_RESTRICT dst, int len) {                   \
   int i, r, g, b;                                                              \
   int temp0, temp1, temp2, temp3, temp4;                                       \
   for (i = 0; i < (len >> 1); i++) {                                           \
diff --git a/src/dsp/yuv_mips_dsp_r2.c b/src/dsp/yuv_mips_dsp_r2.c
index cc8afcc7..816340fe 100644
--- a/src/dsp/yuv_mips_dsp_r2.c
+++ b/src/dsp/yuv_mips_dsp_r2.c
@@ -69,9 +69,10 @@
   : "memory", "hi", "lo"                                                       \
 
 #define ROW_FUNC(FUNC_NAME, XSTEP, R, G, B, A)                                 \
-static void FUNC_NAME(const uint8_t* y,                                        \
-                      const uint8_t* u, const uint8_t* v,                      \
-                      uint8_t* dst, int len) {                                 \
+static void FUNC_NAME(const uint8_t* WEBP_RESTRICT y,                          \
+                      const uint8_t* WEBP_RESTRICT u,                          \
+                      const uint8_t* WEBP_RESTRICT v,                          \
+                      uint8_t* WEBP_RESTRICT dst, int len) {                   \
   int i;                                                                       \
   uint32_t temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;             \
   const int t_con_1 = 26149;                                                   \
diff --git a/src/dsp/yuv_neon.c b/src/dsp/yuv_neon.c
index ff77b009..b1b7c604 100644
--- a/src/dsp/yuv_neon.c
+++ b/src/dsp/yuv_neon.c
@@ -46,7 +46,8 @@ static uint8x8_t ConvertRGBToY_NEON(const uint8x8_t R,
   return vqmovn_u16(Y2);
 }
 
-static void ConvertRGB24ToY_NEON(const uint8_t* rgb, uint8_t* y, int width) {
+static void ConvertRGB24ToY_NEON(const uint8_t* WEBP_RESTRICT rgb,
+                                 uint8_t* WEBP_RESTRICT y, int width) {
   int i;
   for (i = 0; i + 8 <= width; i += 8, rgb += 3 * 8) {
     const uint8x8x3_t RGB = vld3_u8(rgb);
@@ -58,7 +59,8 @@ static void ConvertRGB24ToY_NEON(const uint8_t* rgb, uint8_t* y, int width) {
   }
 }
 
-static void ConvertBGR24ToY_NEON(const uint8_t* bgr, uint8_t* y, int width) {
+static void ConvertBGR24ToY_NEON(const uint8_t* WEBP_RESTRICT bgr,
+                                 uint8_t* WEBP_RESTRICT y, int width) {
   int i;
   for (i = 0; i + 8 <= width; i += 8, bgr += 3 * 8) {
     const uint8x8x3_t BGR = vld3_u8(bgr);
@@ -70,7 +72,8 @@ static void ConvertBGR24ToY_NEON(const uint8_t* bgr, uint8_t* y, int width) {
   }
 }
 
-static void ConvertARGBToY_NEON(const uint32_t* argb, uint8_t* y, int width) {
+static void ConvertARGBToY_NEON(const uint32_t* WEBP_RESTRICT argb,
+                                uint8_t* WEBP_RESTRICT y, int width) {
   int i;
   for (i = 0; i + 8 <= width; i += 8) {
     const uint8x8x4_t RGB = vld4_u8((const uint8_t*)&argb[i]);
@@ -114,8 +117,9 @@ static void ConvertARGBToY_NEON(const uint32_t* argb, uint8_t* y, int width) {
   MULTIPLY_16b(28800, -24116, -4684, 128 << SHIFT, V_DST);       \
 } while (0)
 
-static void ConvertRGBA32ToUV_NEON(const uint16_t* rgb,
-                                   uint8_t* u, uint8_t* v, int width) {
+static void ConvertRGBA32ToUV_NEON(const uint16_t* WEBP_RESTRICT rgb,
+                                   uint8_t* WEBP_RESTRICT u,
+                                   uint8_t* WEBP_RESTRICT v, int width) {
   int i;
   for (i = 0; i + 8 <= width; i += 8, rgb += 4 * 8) {
     const uint16x8x4_t RGB = vld4q_u16((const uint16_t*)rgb);
@@ -131,7 +135,9 @@ static void ConvertRGBA32ToUV_NEON(const uint16_t* rgb,
   }
 }
 
-static void ConvertARGBToUV_NEON(const uint32_t* argb, uint8_t* u, uint8_t* v,
+static void ConvertARGBToUV_NEON(const uint32_t* WEBP_RESTRICT argb,
+                                 uint8_t* WEBP_RESTRICT u,
+                                 uint8_t* WEBP_RESTRICT v,
                                  int src_width, int do_store) {
   int i;
   for (i = 0; i + 16 <= src_width; i += 16, u += 8, v += 8) {
diff --git a/src/dsp/yuv_sse2.c b/src/dsp/yuv_sse2.c
index 01a48f9a..a96b4522 100644
--- a/src/dsp/yuv_sse2.c
+++ b/src/dsp/yuv_sse2.c
@@ -82,9 +82,9 @@ static WEBP_INLINE __m128i Load_UV_HI_8_SSE2(const uint8_t* src) {
 }
 
 // Convert 32 samples of YUV444 to R/G/B
-static void YUV444ToRGB_SSE2(const uint8_t* const y,
-                             const uint8_t* const u,
-                             const uint8_t* const v,
+static void YUV444ToRGB_SSE2(const uint8_t* WEBP_RESTRICT const y,
+                             const uint8_t* WEBP_RESTRICT const u,
+                             const uint8_t* WEBP_RESTRICT const v,
                              __m128i* const R, __m128i* const G,
                              __m128i* const B) {
   const __m128i Y0 = Load_HI_16_SSE2(y), U0 = Load_HI_16_SSE2(u),
@@ -93,9 +93,9 @@ static void YUV444ToRGB_SSE2(const uint8_t* const y,
 }
 
 // Convert 32 samples of YUV420 to R/G/B
-static void YUV420ToRGB_SSE2(const uint8_t* const y,
-                             const uint8_t* const u,
-                             const uint8_t* const v,
+static void YUV420ToRGB_SSE2(const uint8_t* WEBP_RESTRICT const y,
+                             const uint8_t* WEBP_RESTRICT const u,
+                             const uint8_t* WEBP_RESTRICT const v,
                              __m128i* const R, __m128i* const G,
                              __m128i* const B) {
   const __m128i Y0 = Load_HI_16_SSE2(y), U0 = Load_UV_HI_8_SSE2(u),
@@ -108,7 +108,7 @@ static WEBP_INLINE void PackAndStore4_SSE2(const __m128i* const R,
                                            const __m128i* const G,
                                            const __m128i* const B,
                                            const __m128i* const A,
-                                           uint8_t* const dst) {
+                                           uint8_t* WEBP_RESTRICT const dst) {
   const __m128i rb = _mm_packus_epi16(*R, *B);
   const __m128i ga = _mm_packus_epi16(*G, *A);
   const __m128i rg = _mm_unpacklo_epi8(rb, ga);
@@ -120,11 +120,9 @@ static WEBP_INLINE void PackAndStore4_SSE2(const __m128i* const R,
 }
 
 // Pack R/G/B/A results into 16b output.
-static WEBP_INLINE void PackAndStore4444_SSE2(const __m128i* const R,
-                                              const __m128i* const G,
-                                              const __m128i* const B,
-                                              const __m128i* const A,
-                                              uint8_t* const dst) {
+static WEBP_INLINE void PackAndStore4444_SSE2(
+     const __m128i* const R, const __m128i* const G, const __m128i* const B,
+     const __m128i* const A, uint8_t* WEBP_RESTRICT const dst) {
 #if (WEBP_SWAP_16BIT_CSP == 0)
   const __m128i rg0 = _mm_packus_epi16(*R, *G);
   const __m128i ba0 = _mm_packus_epi16(*B, *A);
@@ -145,7 +143,7 @@ static WEBP_INLINE void PackAndStore4444_SSE2(const __m128i* const R,
 static WEBP_INLINE void PackAndStore565_SSE2(const __m128i* const R,
                                              const __m128i* const G,
                                              const __m128i* const B,
-                                             uint8_t* const dst) {
+                                             uint8_t* WEBP_RESTRICT const dst) {
   const __m128i r0 = _mm_packus_epi16(*R, *R);
   const __m128i g0 = _mm_packus_epi16(*G, *G);
   const __m128i b0 = _mm_packus_epi16(*B, *B);
@@ -170,7 +168,7 @@ static WEBP_INLINE void PackAndStore565_SSE2(const __m128i* const R,
 static WEBP_INLINE void PlanarTo24b_SSE2(__m128i* const in0, __m128i* const in1,
                                          __m128i* const in2, __m128i* const in3,
                                          __m128i* const in4, __m128i* const in5,
-                                         uint8_t* const rgb) {
+                                         uint8_t* WEBP_RESTRICT const rgb) {
   // The input is 6 registers of sixteen 8b but for the sake of explanation,
   // let's take 6 registers of four 8b values.
   // To pack, we will keep taking one every two 8b integer and move it
@@ -193,8 +191,10 @@ static WEBP_INLINE void PlanarTo24b_SSE2(__m128i* const in0, __m128i* const in1,
   _mm_storeu_si128((__m128i*)(rgb + 80), *in5);
 }
 
-void VP8YuvToRgba32_SSE2(const uint8_t* y, const uint8_t* u, const uint8_t* v,
-                         uint8_t* dst) {
+void VP8YuvToRgba32_SSE2(const uint8_t* WEBP_RESTRICT y,
+                         const uint8_t* WEBP_RESTRICT u,
+                         const uint8_t* WEBP_RESTRICT v,
+                         uint8_t* WEBP_RESTRICT dst) {
   const __m128i kAlpha = _mm_set1_epi16(255);
   int n;
   for (n = 0; n < 32; n += 8, dst += 32) {
@@ -204,8 +204,10 @@ void VP8YuvToRgba32_SSE2(const uint8_t* y, const uint8_t* u, const uint8_t* v,
   }
 }
 
-void VP8YuvToBgra32_SSE2(const uint8_t* y, const uint8_t* u, const uint8_t* v,
-                         uint8_t* dst) {
+void VP8YuvToBgra32_SSE2(const uint8_t* WEBP_RESTRICT y,
+                         const uint8_t* WEBP_RESTRICT u,
+                         const uint8_t* WEBP_RESTRICT v,
+                         uint8_t* WEBP_RESTRICT dst) {
   const __m128i kAlpha = _mm_set1_epi16(255);
   int n;
   for (n = 0; n < 32; n += 8, dst += 32) {
@@ -215,8 +217,10 @@ void VP8YuvToBgra32_SSE2(const uint8_t* y, const uint8_t* u, const uint8_t* v,
   }
 }
 
-void VP8YuvToArgb32_SSE2(const uint8_t* y, const uint8_t* u, const uint8_t* v,
-                         uint8_t* dst) {
+void VP8YuvToArgb32_SSE2(const uint8_t* WEBP_RESTRICT y,
+                         const uint8_t* WEBP_RESTRICT u,
+                         const uint8_t* WEBP_RESTRICT v,
+                         uint8_t* WEBP_RESTRICT dst) {
   const __m128i kAlpha = _mm_set1_epi16(255);
   int n;
   for (n = 0; n < 32; n += 8, dst += 32) {
@@ -226,8 +230,10 @@ void VP8YuvToArgb32_SSE2(const uint8_t* y, const uint8_t* u, const uint8_t* v,
   }
 }
 
-void VP8YuvToRgba444432_SSE2(const uint8_t* y, const uint8_t* u,
-                             const uint8_t* v, uint8_t* dst) {
+void VP8YuvToRgba444432_SSE2(const uint8_t* WEBP_RESTRICT y,
+                             const uint8_t* WEBP_RESTRICT u,
+                             const uint8_t* WEBP_RESTRICT v,
+                             uint8_t* WEBP_RESTRICT dst) {
   const __m128i kAlpha = _mm_set1_epi16(255);
   int n;
   for (n = 0; n < 32; n += 8, dst += 16) {
@@ -237,8 +243,10 @@ void VP8YuvToRgba444432_SSE2(const uint8_t* y, const uint8_t* u,
   }
 }
 
-void VP8YuvToRgb56532_SSE2(const uint8_t* y, const uint8_t* u, const uint8_t* v,
-                           uint8_t* dst) {
+void VP8YuvToRgb56532_SSE2(const uint8_t* WEBP_RESTRICT y,
+                           const uint8_t* WEBP_RESTRICT u,
+                           const uint8_t* WEBP_RESTRICT v,
+                           uint8_t* WEBP_RESTRICT dst) {
   int n;
   for (n = 0; n < 32; n += 8, dst += 16) {
     __m128i R, G, B;
@@ -247,8 +255,10 @@ void VP8YuvToRgb56532_SSE2(const uint8_t* y, const uint8_t* u, const uint8_t* v,
   }
 }
 
-void VP8YuvToRgb32_SSE2(const uint8_t* y, const uint8_t* u, const uint8_t* v,
-                        uint8_t* dst) {
+void VP8YuvToRgb32_SSE2(const uint8_t* WEBP_RESTRICT y,
+                        const uint8_t* WEBP_RESTRICT u,
+                        const uint8_t* WEBP_RESTRICT v,
+                        uint8_t* WEBP_RESTRICT dst) {
   __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3;
   __m128i rgb0, rgb1, rgb2, rgb3, rgb4, rgb5;
 
@@ -269,8 +279,10 @@ void VP8YuvToRgb32_SSE2(const uint8_t* y, const uint8_t* u, const uint8_t* v,
   PlanarTo24b_SSE2(&rgb0, &rgb1, &rgb2, &rgb3, &rgb4, &rgb5, dst);
 }
 
-void VP8YuvToBgr32_SSE2(const uint8_t* y, const uint8_t* u, const uint8_t* v,
-                        uint8_t* dst) {
+void VP8YuvToBgr32_SSE2(const uint8_t* WEBP_RESTRICT y,
+                        const uint8_t* WEBP_RESTRICT u,
+                        const uint8_t* WEBP_RESTRICT v,
+                        uint8_t* WEBP_RESTRICT dst) {
   __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3;
   __m128i bgr0, bgr1, bgr2, bgr3, bgr4, bgr5;
 
@@ -294,9 +306,10 @@ void VP8YuvToBgr32_SSE2(const uint8_t* y, const uint8_t* u, const uint8_t* v,
 //-----------------------------------------------------------------------------
 // Arbitrary-length row conversion functions
 
-static void YuvToRgbaRow_SSE2(const uint8_t* y,
-                              const uint8_t* u, const uint8_t* v,
-                              uint8_t* dst, int len) {
+static void YuvToRgbaRow_SSE2(const uint8_t* WEBP_RESTRICT y,
+                              const uint8_t* WEBP_RESTRICT u,
+                              const uint8_t* WEBP_RESTRICT v,
+                              uint8_t* WEBP_RESTRICT dst, int len) {
   const __m128i kAlpha = _mm_set1_epi16(255);
   int n;
   for (n = 0; n + 8 <= len; n += 8, dst += 32) {
@@ -316,9 +329,10 @@ static void YuvToRgbaRow_SSE2(const uint8_t* y,
   }
 }
 
-static void YuvToBgraRow_SSE2(const uint8_t* y,
-                              const uint8_t* u, const uint8_t* v,
-                              uint8_t* dst, int len) {
+static void YuvToBgraRow_SSE2(const uint8_t* WEBP_RESTRICT y,
+                              const uint8_t* WEBP_RESTRICT u,
+                              const uint8_t* WEBP_RESTRICT v,
+                              uint8_t* WEBP_RESTRICT dst, int len) {
   const __m128i kAlpha = _mm_set1_epi16(255);
   int n;
   for (n = 0; n + 8 <= len; n += 8, dst += 32) {
@@ -338,9 +352,10 @@ static void YuvToBgraRow_SSE2(const uint8_t* y,
   }
 }
 
-static void YuvToArgbRow_SSE2(const uint8_t* y,
-                              const uint8_t* u, const uint8_t* v,
-                              uint8_t* dst, int len) {
+static void YuvToArgbRow_SSE2(const uint8_t* WEBP_RESTRICT y,
+                              const uint8_t* WEBP_RESTRICT u,
+                              const uint8_t* WEBP_RESTRICT v,
+                              uint8_t* WEBP_RESTRICT dst, int len) {
   const __m128i kAlpha = _mm_set1_epi16(255);
   int n;
   for (n = 0; n + 8 <= len; n += 8, dst += 32) {
@@ -360,9 +375,10 @@ static void YuvToArgbRow_SSE2(const uint8_t* y,
   }
 }
 
-static void YuvToRgbRow_SSE2(const uint8_t* y,
-                             const uint8_t* u, const uint8_t* v,
-                             uint8_t* dst, int len) {
+static void YuvToRgbRow_SSE2(const uint8_t* WEBP_RESTRICT y,
+                             const uint8_t* WEBP_RESTRICT u,
+                             const uint8_t* WEBP_RESTRICT v,
+                             uint8_t* WEBP_RESTRICT dst, int len) {
   int n;
   for (n = 0; n + 32 <= len; n += 32, dst += 32 * 3) {
     __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3;
@@ -397,9 +413,10 @@ static void YuvToRgbRow_SSE2(const uint8_t* y,
   }
 }
 
-static void YuvToBgrRow_SSE2(const uint8_t* y,
-                             const uint8_t* u, const uint8_t* v,
-                             uint8_t* dst, int len) {
+static void YuvToBgrRow_SSE2(const uint8_t* WEBP_RESTRICT y,
+                             const uint8_t* WEBP_RESTRICT u,
+                             const uint8_t* WEBP_RESTRICT v,
+                             uint8_t* WEBP_RESTRICT dst, int len) {
   int n;
   for (n = 0; n + 32 <= len; n += 32, dst += 32 * 3) {
     __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3;
@@ -471,7 +488,7 @@ static WEBP_INLINE void RGB24PackedToPlanarHelper_SSE2(
 // rrrr... rrrr... gggg... gggg... bbbb... bbbb....
 // Similar to PlanarTo24bHelper(), but in reverse order.
 static WEBP_INLINE void RGB24PackedToPlanar_SSE2(
-    const uint8_t* const rgb, __m128i* const out /*out[6]*/) {
+    const uint8_t* WEBP_RESTRICT const rgb, __m128i* const out /*out[6]*/) {
   __m128i tmp[6];
   tmp[0] = _mm_loadu_si128((const __m128i*)(rgb +  0));
   tmp[1] = _mm_loadu_si128((const __m128i*)(rgb + 16));
@@ -488,8 +505,8 @@ static WEBP_INLINE void RGB24PackedToPlanar_SSE2(
 }
 
 // Convert 8 packed ARGB to r[], g[], b[]
-static WEBP_INLINE void RGB32PackedToPlanar_SSE2(const uint32_t* const argb,
-                                                 __m128i* const rgb /*in[6]*/) {
+static WEBP_INLINE void RGB32PackedToPlanar_SSE2(
+    const uint32_t* WEBP_RESTRICT const argb, __m128i* const rgb /*in[6]*/) {
   const __m128i zero = _mm_setzero_si128();
   __m128i a0 = LOAD_16(argb + 0);
   __m128i a1 = LOAD_16(argb + 4);
@@ -562,7 +579,8 @@ static WEBP_INLINE void ConvertRGBToUV_SSE2(const __m128i* const R,
 #undef MK_CST_16
 #undef TRANSFORM
 
-static void ConvertRGB24ToY_SSE2(const uint8_t* rgb, uint8_t* y, int width) {
+static void ConvertRGB24ToY_SSE2(const uint8_t* WEBP_RESTRICT rgb,
+                                 uint8_t* WEBP_RESTRICT y, int width) {
   const int max_width = width & ~31;
   int i;
   for (i = 0; i < max_width; rgb += 3 * 16 * 2) {
@@ -596,7 +614,8 @@ static void ConvertRGB24ToY_SSE2(const uint8_t* rgb, uint8_t* y, int width) {
   }
 }
 
-static void ConvertBGR24ToY_SSE2(const uint8_t* bgr, uint8_t* y, int width) {
+static void ConvertBGR24ToY_SSE2(const uint8_t* WEBP_RESTRICT bgr,
+                                 uint8_t* WEBP_RESTRICT y, int width) {
   const int max_width = width & ~31;
   int i;
   for (i = 0; i < max_width; bgr += 3 * 16 * 2) {
@@ -630,7 +649,8 @@ static void ConvertBGR24ToY_SSE2(const uint8_t* bgr, uint8_t* y, int width) {
   }
 }
 
-static void ConvertARGBToY_SSE2(const uint32_t* argb, uint8_t* y, int width) {
+static void ConvertARGBToY_SSE2(const uint32_t* WEBP_RESTRICT argb,
+                                uint8_t* WEBP_RESTRICT y, int width) {
   const int max_width = width & ~15;
   int i;
   for (i = 0; i < max_width; i += 16) {
@@ -658,8 +678,9 @@ static void HorizontalAddPack_SSE2(const __m128i* const A,
   *out = _mm_packs_epi32(C, D);
 }
 
-static void ConvertARGBToUV_SSE2(const uint32_t* argb,
-                                 uint8_t* u, uint8_t* v,
+static void ConvertARGBToUV_SSE2(const uint32_t* WEBP_RESTRICT argb,
+                                 uint8_t* WEBP_RESTRICT u,
+                                 uint8_t* WEBP_RESTRICT v,
                                  int src_width, int do_store) {
   const int max_width = src_width & ~31;
   int i;
@@ -695,7 +716,7 @@ static void ConvertARGBToUV_SSE2(const uint32_t* argb,
 
 // Convert 16 packed ARGB 16b-values to r[], g[], b[]
 static WEBP_INLINE void RGBA32PackedToPlanar_16b_SSE2(
-    const uint16_t* const rgbx,
+    const uint16_t* WEBP_RESTRICT const rgbx,
     __m128i* const r, __m128i* const g, __m128i* const b) {
   const __m128i in0 = LOAD_16(rgbx +  0);  // r0 | g0 | b0 |x| r1 | g1 | b1 |x
   const __m128i in1 = LOAD_16(rgbx +  8);  // r2 | g2 | b2 |x| r3 | g3 | b3 |x
@@ -715,8 +736,9 @@ static WEBP_INLINE void RGBA32PackedToPlanar_16b_SSE2(
   *b = _mm_unpacklo_epi64(B1, B3);
 }
 
-static void ConvertRGBA32ToUV_SSE2(const uint16_t* rgb,
-                                   uint8_t* u, uint8_t* v, int width) {
+static void ConvertRGBA32ToUV_SSE2(const uint16_t* WEBP_RESTRICT rgb,
+                                   uint8_t* WEBP_RESTRICT u,
+                                   uint8_t* WEBP_RESTRICT v, int width) {
   const int max_width = width & ~15;
   const uint16_t* const last_rgb = rgb + 4 * max_width;
   while (rgb < last_rgb) {
diff --git a/src/dsp/yuv_sse41.c b/src/dsp/yuv_sse41.c
index f79b802e..071e4908 100644
--- a/src/dsp/yuv_sse41.c
+++ b/src/dsp/yuv_sse41.c
@@ -82,9 +82,9 @@ static WEBP_INLINE __m128i Load_UV_HI_8_SSE41(const uint8_t* src) {
 }
 
 // Convert 32 samples of YUV444 to R/G/B
-static void YUV444ToRGB_SSE41(const uint8_t* const y,
-                              const uint8_t* const u,
-                              const uint8_t* const v,
+static void YUV444ToRGB_SSE41(const uint8_t* WEBP_RESTRICT const y,
+                              const uint8_t* WEBP_RESTRICT const u,
+                              const uint8_t* WEBP_RESTRICT const v,
                               __m128i* const R, __m128i* const G,
                               __m128i* const B) {
   const __m128i Y0 = Load_HI_16_SSE41(y), U0 = Load_HI_16_SSE41(u),
@@ -93,9 +93,9 @@ static void YUV444ToRGB_SSE41(const uint8_t* const y,
 }
 
 // Convert 32 samples of YUV420 to R/G/B
-static void YUV420ToRGB_SSE41(const uint8_t* const y,
-                              const uint8_t* const u,
-                              const uint8_t* const v,
+static void YUV420ToRGB_SSE41(const uint8_t* WEBP_RESTRICT const y,
+                              const uint8_t* WEBP_RESTRICT const u,
+                              const uint8_t* WEBP_RESTRICT const v,
                               __m128i* const R, __m128i* const G,
                               __m128i* const B) {
   const __m128i Y0 = Load_HI_16_SSE41(y), U0 = Load_UV_HI_8_SSE41(u),
@@ -109,7 +109,7 @@ static void YUV420ToRGB_SSE41(const uint8_t* const y,
 static WEBP_INLINE void PlanarTo24b_SSE41(
     __m128i* const in0, __m128i* const in1, __m128i* const in2,
     __m128i* const in3, __m128i* const in4, __m128i* const in5,
-    uint8_t* const rgb) {
+    uint8_t* WEBP_RESTRICT const rgb) {
   // The input is 6 registers of sixteen 8b but for the sake of explanation,
   // let's take 6 registers of four 8b values.
   // To pack, we will keep taking one every two 8b integer and move it
@@ -132,8 +132,10 @@ static WEBP_INLINE void PlanarTo24b_SSE41(
   _mm_storeu_si128((__m128i*)(rgb + 80), *in5);
 }
 
-void VP8YuvToRgb32_SSE41(const uint8_t* y, const uint8_t* u, const uint8_t* v,
-                         uint8_t* dst) {
+void VP8YuvToRgb32_SSE41(const uint8_t* WEBP_RESTRICT y,
+                         const uint8_t* WEBP_RESTRICT u,
+                         const uint8_t* WEBP_RESTRICT v,
+                         uint8_t* WEBP_RESTRICT dst) {
   __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3;
   __m128i rgb0, rgb1, rgb2, rgb3, rgb4, rgb5;
 
@@ -154,8 +156,10 @@ void VP8YuvToRgb32_SSE41(const uint8_t* y, const uint8_t* u, const uint8_t* v,
   PlanarTo24b_SSE41(&rgb0, &rgb1, &rgb2, &rgb3, &rgb4, &rgb5, dst);
 }
 
-void VP8YuvToBgr32_SSE41(const uint8_t* y, const uint8_t* u, const uint8_t* v,
-                         uint8_t* dst) {
+void VP8YuvToBgr32_SSE41(const uint8_t* WEBP_RESTRICT y,
+                         const uint8_t* WEBP_RESTRICT u,
+                         const uint8_t* WEBP_RESTRICT v,
+                         uint8_t* WEBP_RESTRICT dst) {
   __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3;
   __m128i bgr0, bgr1, bgr2, bgr3, bgr4, bgr5;
 
@@ -179,9 +183,10 @@ void VP8YuvToBgr32_SSE41(const uint8_t* y, const uint8_t* u, const uint8_t* v,
 //-----------------------------------------------------------------------------
 // Arbitrary-length row conversion functions
 
-static void YuvToRgbRow_SSE41(const uint8_t* y,
-                              const uint8_t* u, const uint8_t* v,
-                              uint8_t* dst, int len) {
+static void YuvToRgbRow_SSE41(const uint8_t* WEBP_RESTRICT y,
+                              const uint8_t* WEBP_RESTRICT u,
+                              const uint8_t* WEBP_RESTRICT v,
+                              uint8_t* WEBP_RESTRICT dst, int len) {
   int n;
   for (n = 0; n + 32 <= len; n += 32, dst += 32 * 3) {
     __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3;
@@ -216,9 +221,10 @@ static void YuvToRgbRow_SSE41(const uint8_t* y,
   }
 }
 
-static void YuvToBgrRow_SSE41(const uint8_t* y,
-                              const uint8_t* u, const uint8_t* v,
-                              uint8_t* dst, int len) {
+static void YuvToBgrRow_SSE41(const uint8_t* WEBP_RESTRICT y,
+                              const uint8_t* WEBP_RESTRICT u,
+                              const uint8_t* WEBP_RESTRICT v,
+                              uint8_t* WEBP_RESTRICT dst, int len) {
   int n;
   for (n = 0; n + 32 <= len; n += 32, dst += 32 * 3) {
     __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3;
@@ -290,7 +296,7 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitSamplersSSE41(void) {
 // rrrr... rrrr... gggg... gggg... bbbb... bbbb....
 // Similar to PlanarTo24bHelper(), but in reverse order.
 static WEBP_INLINE void RGB24PackedToPlanar_SSE41(
-    const uint8_t* const rgb, __m128i* const out /*out[6]*/) {
+    const uint8_t* WEBP_RESTRICT const rgb, __m128i* const out /*out[6]*/) {
   const __m128i A0 = _mm_loadu_si128((const __m128i*)(rgb +  0));
   const __m128i A1 = _mm_loadu_si128((const __m128i*)(rgb + 16));
   const __m128i A2 = _mm_loadu_si128((const __m128i*)(rgb + 32));
@@ -334,7 +340,7 @@ static WEBP_INLINE void RGB24PackedToPlanar_SSE41(
 
 // Convert 8 packed ARGB to r[], g[], b[]
 static WEBP_INLINE void RGB32PackedToPlanar_SSE41(
-    const uint32_t* const argb, __m128i* const rgb /*in[6]*/) {
+    const uint32_t* WEBP_RESTRICT const argb, __m128i* const rgb /*in[6]*/) {
   const __m128i zero = _mm_setzero_si128();
   __m128i a0 = LOAD_16(argb + 0);
   __m128i a1 = LOAD_16(argb + 4);
@@ -407,7 +413,8 @@ static WEBP_INLINE void ConvertRGBToUV_SSE41(const __m128i* const R,
 #undef MK_CST_16
 #undef TRANSFORM
 
-static void ConvertRGB24ToY_SSE41(const uint8_t* rgb, uint8_t* y, int width) {
+static void ConvertRGB24ToY_SSE41(const uint8_t* WEBP_RESTRICT rgb,
+                                  uint8_t* WEBP_RESTRICT y, int width) {
   const int max_width = width & ~31;
   int i;
   for (i = 0; i < max_width; rgb += 3 * 16 * 2) {
@@ -441,7 +448,8 @@ static void ConvertRGB24ToY_SSE41(const uint8_t* rgb, uint8_t* y, int width) {
   }
 }
 
-static void ConvertBGR24ToY_SSE41(const uint8_t* bgr, uint8_t* y, int width) {
+static void ConvertBGR24ToY_SSE41(const uint8_t* WEBP_RESTRICT bgr,
+                                  uint8_t* WEBP_RESTRICT y, int width) {
   const int max_width = width & ~31;
   int i;
   for (i = 0; i < max_width; bgr += 3 * 16 * 2) {
@@ -475,7 +483,8 @@ static void ConvertBGR24ToY_SSE41(const uint8_t* bgr, uint8_t* y, int width) {
   }
 }
 
-static void ConvertARGBToY_SSE41(const uint32_t* argb, uint8_t* y, int width) {
+static void ConvertARGBToY_SSE41(const uint32_t* WEBP_RESTRICT argb,
+                                 uint8_t* WEBP_RESTRICT y, int width) {
   const int max_width = width & ~15;
   int i;
   for (i = 0; i < max_width; i += 16) {
@@ -503,8 +512,9 @@ static void HorizontalAddPack_SSE41(const __m128i* const A,
   *out = _mm_packs_epi32(C, D);
 }
 
-static void ConvertARGBToUV_SSE41(const uint32_t* argb,
-                                  uint8_t* u, uint8_t* v,
+static void ConvertARGBToUV_SSE41(const uint32_t* WEBP_RESTRICT argb,
+                                  uint8_t* WEBP_RESTRICT u,
+                                  uint8_t* WEBP_RESTRICT v,
                                   int src_width, int do_store) {
   const int max_width = src_width & ~31;
   int i;
@@ -540,7 +550,7 @@ static void ConvertARGBToUV_SSE41(const uint32_t* argb,
 
 // Convert 16 packed ARGB 16b-values to r[], g[], b[]
 static WEBP_INLINE void RGBA32PackedToPlanar_16b_SSE41(
-    const uint16_t* const rgbx,
+    const uint16_t* WEBP_RESTRICT const rgbx,
     __m128i* const r, __m128i* const g, __m128i* const b) {
   const __m128i in0 = LOAD_16(rgbx +  0);  // r0 | g0 | b0 |x| r1 | g1 | b1 |x
   const __m128i in1 = LOAD_16(rgbx +  8);  // r2 | g2 | b2 |x| r3 | g3 | b3 |x
@@ -570,8 +580,9 @@ static WEBP_INLINE void RGBA32PackedToPlanar_16b_SSE41(
   *b = _mm_unpackhi_epi64(B1, B3);
 }
 
-static void ConvertRGBA32ToUV_SSE41(const uint16_t* rgb,
-                                    uint8_t* u, uint8_t* v, int width) {
+static void ConvertRGBA32ToUV_SSE41(const uint16_t* WEBP_RESTRICT rgb,
+                                    uint8_t* WEBP_RESTRICT u,
+                                    uint8_t* WEBP_RESTRICT v, int width) {
   const int max_width = width & ~15;
   const uint16_t* const last_rgb = rgb + 4 * max_width;
   while (rgb < last_rgb) {