diff --git a/babl/babl-icc.c b/babl/babl-icc.c
index 38e382a..25f36a9 100644
--- a/babl/babl-icc.c
+++ b/babl/babl-icc.c
@@ -474,10 +474,9 @@ babl_trc_from_icc (ICC         *state,
         }
 
         ret = babl_trc_lut_find (lut, count);
-        if (ret)
-          return ret;
+        if (ret == NULL)
+          ret = babl_trc_lut (NULL, count, lut);
 
-        ret = babl_trc_lut (NULL, count, lut);
         babl_free (lut);
         return ret;
       }
diff --git a/babl/base/meson.build b/babl/base/meson.build
index e59609c..1084d8e 100644
--- a/babl/base/meson.build
+++ b/babl/base/meson.build
@@ -21,7 +21,7 @@ babl_base_sources = [
 babl_base = static_library('babl_base',
   babl_base_sources,
   include_directories: [rootInclude, bablInclude],
-  dependencies: [math, lcms],
+  dependencies: [math, lcms, log],
    c_args: common_c_flags + [sse2_cflags]
 ) 
 
@@ -30,14 +30,14 @@ if host_cpu_family == 'x86_64'
   babl_base_x86_64_v2 = static_library('babl_base-x86-64-v2',
     babl_base_sources,
     include_directories: [rootInclude, bablInclude],
-    dependencies: [math, lcms],
+    dependencies: [math, lcms, log],
     c_args: common_c_flags + x86_64_v2_flags
   )
 
   babl_base_x86_64_v3 = static_library('babl_base-x86-64-v3',
     babl_base_sources,
     include_directories: [rootInclude, bablInclude],
-    dependencies: [math, lcms],
+    dependencies: [math, lcms, log],
     c_args: common_c_flags + x86_64_v3_flags
   )
 
@@ -49,7 +49,7 @@ if host_cpu_family == 'arm'
   babl_base_arm_neon = static_library('babl_base-arm-neon',
     babl_base_sources,
     include_directories: [rootInclude, bablInclude],
-    dependencies: [math, lcms],
+    dependencies: [math, lcms, log],
     c_args: common_c_flags + arm_neon_flags
   )
 
diff --git a/babl/meson.build b/babl/meson.build
index 013384e..b98df1a 100644
--- a/babl/meson.build
+++ b/babl/meson.build
@@ -118,7 +118,7 @@ if meson.is_subproject()
   subdir('babl')
 endif
 
-babl_deps = [math, thread, dl, lcms]
+babl_deps = [math, thread, dl, lcms, log]
 babl_includes = [rootInclude, bablBaseInclude]
 
 if host_cpu_family == 'x86_64'
diff --git a/docs/meson.build b/docs/meson.build
index bd006b7..97b76ac 100644
--- a/docs/meson.build
+++ b/docs/meson.build
@@ -40,7 +40,7 @@ endif
 
 TOC = files('toc')
 html_files = {
-  'index': [index_static_html, [
+  'index': ['index-static.html.in', [
     ['AUTHORS', authors_file],
     ['TODO',    todo_file],
     ['NEWS',    news_file],
diff --git a/extensions/CIE.c b/extensions/CIE.c
index 1d7b0ca..5b367d4 100644
--- a/extensions/CIE.c
+++ b/extensions/CIE.c
@@ -2441,10 +2441,10 @@ rgbaf_to_Lf_sse2 (const Babl  *conversion,
 }
 
 static void
-rgbaf_to_Labaf_sse2 (const Babl  *conversion, 
-                     const float *src, 
-                     float       *dst, 
-                     long         samples)
+rgbaf_to_Labaf_sse2_aligned_4mult (const Babl  *conversion,
+                                   const float *src,
+                                   float       *dst,
+                                   long         samples)
 {
   const Babl *space = babl_conversion_get_source_space (conversion);
   const float m_0_0 = space->space.RGBtoXYZf[0] / D50_WHITE_REF_Xf;
@@ -2456,100 +2456,124 @@ rgbaf_to_Labaf_sse2 (const Babl  *conversion,
   const float m_2_0 = space->space.RGBtoXYZf[6] / D50_WHITE_REF_Zf;
   const float m_2_1 = space->space.RGBtoXYZf[7] / D50_WHITE_REF_Zf;
   const float m_2_2 = space->space.RGBtoXYZf[8] / D50_WHITE_REF_Zf;
-  long i = 0;
-  long remainder;
+  long        i = 0;
+
+  const __m128 m_0_0_v = _mm_set1_ps (m_0_0);
+  const __m128 m_0_1_v = _mm_set1_ps (m_0_1);
+  const __m128 m_0_2_v = _mm_set1_ps (m_0_2);
+  const __m128 m_1_0_v = _mm_set1_ps (m_1_0);
+  const __m128 m_1_1_v = _mm_set1_ps (m_1_1);
+  const __m128 m_1_2_v = _mm_set1_ps (m_1_2);
+  const __m128 m_2_0_v = _mm_set1_ps (m_2_0);
+  const __m128 m_2_1_v = _mm_set1_ps (m_2_1);
+  const __m128 m_2_2_v = _mm_set1_ps (m_2_2);
+
+  assert (((uintptr_t) src % 16) + ((uintptr_t) dst % 16) == 0);
+  assert (samples % 4 == 0);
+
+  for ( ; i < samples; i += 4)
+    {
+      __m128 Laba0;
+      __m128 Laba1;
+      __m128 Laba2;
+      __m128 Laba3;
+
+      __m128 rgba0 = _mm_load_ps (src);
+      __m128 rgba1 = _mm_load_ps (src + 4);
+      __m128 rgba2 = _mm_load_ps (src + 8);
+      __m128 rgba3 = _mm_load_ps (src + 12);
+
+      __m128 r = rgba0;
+      __m128 g = rgba1;
+      __m128 b = rgba2;
+      __m128 a = rgba3;
+      _MM_TRANSPOSE4_PS (r, g, b, a);
 
-  if (((uintptr_t) src % 16) + ((uintptr_t) dst % 16) == 0)
-    {
-      const long    n = (samples / 4) * 4;
-      const __m128 m_0_0_v = _mm_set1_ps (m_0_0);
-      const __m128 m_0_1_v = _mm_set1_ps (m_0_1);
-      const __m128 m_0_2_v = _mm_set1_ps (m_0_2);
-      const __m128 m_1_0_v = _mm_set1_ps (m_1_0);
-      const __m128 m_1_1_v = _mm_set1_ps (m_1_1);
-      const __m128 m_1_2_v = _mm_set1_ps (m_1_2);
-      const __m128 m_2_0_v = _mm_set1_ps (m_2_0);
-      const __m128 m_2_1_v = _mm_set1_ps (m_2_1);
-      const __m128 m_2_2_v = _mm_set1_ps (m_2_2);
-
-      for ( ; i < n; i += 4)
         {
-          __m128 Laba0;
-          __m128 Laba1;
-          __m128 Laba2;
-          __m128 Laba3;
+          __m128 xr = _mm_add_ps (_mm_add_ps (_mm_mul_ps (m_0_0_v, r), _mm_mul_ps (m_0_1_v, g)),
+                                  _mm_mul_ps (m_0_2_v, b));
+          __m128 yr = _mm_add_ps (_mm_add_ps (_mm_mul_ps (m_1_0_v, r), _mm_mul_ps (m_1_1_v, g)),
+                                  _mm_mul_ps (m_1_2_v, b));
+          __m128 zr = _mm_add_ps (_mm_add_ps (_mm_mul_ps (m_2_0_v, r), _mm_mul_ps (m_2_1_v, g)),
+                                  _mm_mul_ps (m_2_2_v, b));
 
-          __m128 rgba0 = _mm_load_ps (src);
-          __m128 rgba1 = _mm_load_ps (src + 4);
-          __m128 rgba2 = _mm_load_ps (src + 8);
-          __m128 rgba3 = _mm_load_ps (src + 12);
+          __m128 fx = lab_r_to_f_sse2 (xr);
+          __m128 fy = lab_r_to_f_sse2 (yr);
+          __m128 fz = lab_r_to_f_sse2 (zr);
 
-          __m128 r = rgba0;
-          __m128 g = rgba1;
-          __m128 b = rgba2;
-          __m128 a = rgba3;
-          _MM_TRANSPOSE4_PS (r, g, b, a);
+          __m128 L = _mm_sub_ps (_mm_mul_ps (_mm_set1_ps (116.0f), fy), _mm_set1_ps (16.0f));
+          __m128 A = _mm_mul_ps (_mm_set1_ps (500.0f), _mm_sub_ps (fx, fy));
+          __m128 B = _mm_mul_ps (_mm_set1_ps (200.0f), _mm_sub_ps (fy, fz));
+
+          Laba0 = L;
+          Laba1 = A;
+          Laba2 = B;
+          Laba3 = a;
+          _MM_TRANSPOSE4_PS (Laba0, Laba1, Laba2, Laba3);
+        }
 
-          {
-            __m128 xr = _mm_add_ps (_mm_add_ps (_mm_mul_ps (m_0_0_v, r), _mm_mul_ps (m_0_1_v, g)),
-                                    _mm_mul_ps (m_0_2_v, b));
-            __m128 yr = _mm_add_ps (_mm_add_ps (_mm_mul_ps (m_1_0_v, r), _mm_mul_ps (m_1_1_v, g)),
-                                    _mm_mul_ps (m_1_2_v, b));
-            __m128 zr = _mm_add_ps (_mm_add_ps (_mm_mul_ps (m_2_0_v, r), _mm_mul_ps (m_2_1_v, g)),
-                                    _mm_mul_ps (m_2_2_v, b));
+      _mm_store_ps (dst, Laba0);
+      _mm_store_ps (dst + 4, Laba1);
+      _mm_store_ps (dst + 8, Laba2);
+      _mm_store_ps (dst + 12, Laba3);
 
-            __m128 fx = lab_r_to_f_sse2 (xr);
-            __m128 fy = lab_r_to_f_sse2 (yr);
-            __m128 fz = lab_r_to_f_sse2 (zr);
+      src += 16;
+      dst += 16;
+    }
+}
 
-            __m128 L = _mm_sub_ps (_mm_mul_ps (_mm_set1_ps (116.0f), fy), _mm_set1_ps (16.0f));
-            __m128 A = _mm_mul_ps (_mm_set1_ps (500.0f), _mm_sub_ps (fx, fy));
-            __m128 B = _mm_mul_ps (_mm_set1_ps (200.0f), _mm_sub_ps (fy, fz));
-
-            Laba0 = L;
-            Laba1 = A;
-            Laba2 = B;
-            Laba3 = a;
-            _MM_TRANSPOSE4_PS (Laba0, Laba1, Laba2, Laba3);
-          }
+static void
+rgbaf_to_Labaf_sse2 (const Babl  *conversion,
+                     const float *src,
+                     float       *dst,
+                     long         samples)
+{
+  if (((uintptr_t) src % 16) + ((uintptr_t) dst % 16) == 0 ||
+      samples < 4)
+    {
+      long first_samples = samples / 4 * 4;
+      long remainder;
 
-          _mm_store_ps (dst, Laba0);
-          _mm_store_ps (dst + 4, Laba1);
-          _mm_store_ps (dst + 8, Laba2);
-          _mm_store_ps (dst + 12, Laba3);
+      rgbaf_to_Labaf_sse2_aligned_4mult (conversion, src, dst, first_samples);
+      remainder = samples - first_samples;
 
-          src += 16;
-          dst += 16;
+      if (remainder)
+        {
+          float __attribute__ ((aligned (16))) aligned_src[16];
+          float __attribute__ ((aligned (16))) aligned_dest[16];
+
+          memcpy (aligned_src, src + first_samples * 4, remainder * 16);
+          memset (aligned_src + remainder * 4, 0, 4 * 16 - (remainder * 16));
+          rgbaf_to_Labaf_sse2_aligned_4mult (conversion, (const float *) aligned_src, aligned_dest, 4);
+          memcpy (dst + first_samples * 4, aligned_dest, remainder * 16);
         }
     }
-
-  remainder = samples - i;
-  while (remainder--)
+  else
     {
-      float r = src[0];
-      float g = src[1];
-      float b = src[2];
-      float a = src[3];
-
-      float xr = m_0_0 * r + m_0_1 * g + m_0_2 * b;
-      float yr = m_1_0 * r + m_1_1 * g + m_1_2 * b;
-      float zr = m_2_0 * r + m_2_1 * g + m_2_2 * b;
+      float __attribute__ ((aligned (16))) _aligned_src[4 * samples];
+      float __attribute__ ((aligned (16))) _aligned_dst[4 * samples];
+      float                               *aligned_src;
+      float                               *aligned_dst;
 
-      float fx = xr > LAB_EPSILONf ? _cbrtf (xr) : (LAB_KAPPAf * xr + 16.0f) / 116.0f;
-      float fy = yr > LAB_EPSILONf ? _cbrtf (yr) : (LAB_KAPPAf * yr + 16.0f) / 116.0f;
-      float fz = zr > LAB_EPSILONf ? _cbrtf (zr) : (LAB_KAPPAf * zr + 16.0f) / 116.0f;
+      if (((uintptr_t) src % 16) != 0)
+        {
+          aligned_src = _aligned_src;
+          memcpy (aligned_src, src, samples * 16);
+        }
+      else
+        {
+          aligned_src = (float *) src;
+        }
 
-      float L = 116.0f * fy - 16.0f;
-      float A = 500.0f * (fx - fy);
-      float B = 200.0f * (fy - fz);
+      if (((uintptr_t) dst % 16) != 0)
+        aligned_dst = _aligned_dst;
+      else
+        aligned_dst = dst;
 
-      dst[0] = L;
-      dst[1] = A;
-      dst[2] = B;
-      dst[3] = a;
+      rgbaf_to_Labaf_sse2 (conversion, aligned_src, aligned_dst, samples);
 
-      src += 4;
-      dst += 4;
+      if (((uintptr_t) dst % 16) != 0)
+        memcpy (dst, aligned_dst, samples * 16);
     }
 }
 
diff --git a/extensions/HCY.c b/extensions/HCY.c
index 31cb8b4..46d7a99 100644
--- a/extensions/HCY.c
+++ b/extensions/HCY.c
@@ -86,7 +86,7 @@ components (void)
   babl_component_new ("hue", NULL);
   babl_component_new ("HCY chroma", "chroma", NULL);
   babl_component_new ("HCY luma", "luma", NULL);
-  babl_component_new ("alpha", NULL);
+  babl_component_new ("alpha", "alpha", NULL);
 }
 
 static void
diff --git a/extensions/HSL.c b/extensions/HSL.c
index 73b2f03..2c8ae9f 100644
--- a/extensions/HSL.c
+++ b/extensions/HSL.c
@@ -69,7 +69,7 @@ init (void)
   babl_component_new ("hue", NULL);
   babl_component_new ("saturation", NULL);
   babl_component_new ("lightness", NULL);
-  babl_component_new ("alpha", NULL);
+  babl_component_new ("alpha", "alpha", NULL);
 
   babl_model_new ("name", "HSL",
                   "doc", "HSL - Hue Saturation Lightness, an improvement over HSV; which uses lightness; defined as (MAX(R,G,B) + MIN(R,G,B))/2 for the grayscale axis; better than HSV, but look into the CIE based spaces for better perceptual uniformity. The HSL space is relative to the RGB space associated with the format.",
diff --git a/extensions/HSV.c b/extensions/HSV.c
index ad2e002..100636c 100644
--- a/extensions/HSV.c
+++ b/extensions/HSV.c
@@ -89,7 +89,7 @@ components (void)
   babl_component_new ("hue", NULL);
   babl_component_new ("saturation", NULL);
   babl_component_new ("value", NULL);
-  babl_component_new ("alpha", NULL);
+  babl_component_new ("alpha", "alpha", NULL);
 }
 
 static void
diff --git a/extensions/ycbcr.c b/extensions/ycbcr.c
index 1e779d7..9fe1fa0 100644
--- a/extensions/ycbcr.c
+++ b/extensions/ycbcr.c
@@ -49,7 +49,7 @@ init (void)
 static void
 components (void)
 {
-  babl_component_new ("alpha", NULL);
+  babl_component_new ("alpha", "alpha", NULL);
 }
 
 
diff --git a/meson.build b/meson.build
index 2ce13ec..9a292d9 100644
--- a/meson.build
+++ b/meson.build
@@ -1,6 +1,6 @@
 project('babl', 'c',
   license: 'LGPL3+',
-  version: '0.1.108',
+  version: '0.1.109',
   meson_version: '>=0.55.0',
   default_options: [
     'buildtype=debugoptimized'
@@ -110,11 +110,6 @@ platform_win32 = (host_os.startswith('mingw') or
                   host_os.startswith('windows'))
 
 platform_osx = host_os.startswith('darwin')
-if platform_osx
-  if cc.get_id() != 'clang'
-    error('You should use Clang/Clang++ on OSX.')
-  endif
-endif
 
 platform_android = host_os.contains('android')
 
@@ -161,6 +156,7 @@ can_run_host_binaries = meson.can_run_host_binaries()
 # Compiler arguments
 
 common_c_flags = []
+common_l_flags = []
 
 if buildtype == 'debugoptimized' or buildtype == 'release'
   common_c_flags += cc.get_supported_arguments(['-Ofast'])
@@ -180,6 +176,22 @@ extra_warnings_list = [
 ]
 common_c_flags += cc.get_supported_arguments(extra_warnings_list)
 
+# Generate .pdb (CodeView) debug symbols (makes possible to debug with DIA SDK)
+#pdb_support = cc.has_argument('-gcodeview') and cc.has_link_argument('-Wl,--pdb=')
+#if platform_win32 and pdb_support
+#  common_c_flags += '-gcodeview'
+#  common_l_flags += '-Wl,--pdb='
+#endif
+
+if platform_win32 and cc.get_id() == 'clang'
+  # Optimize DWARF symbols to Dr. Mingw
+  # https://github.com/jrfonseca/drmingw/issues/42
+  common_c_flags += '-gdwarf-aranges'
+  # Workaround to get colored output
+  # https://github.com/msys2/MINGW-packages/issues/2988
+  common_c_flags += '-fansi-escape-codes'
+endif
+
 add_project_arguments(common_c_flags, language: 'c')
 
 # Linker arguments
@@ -200,9 +212,11 @@ elif host_cpu_family == 'arm'
   arm_neon_flags = cc.get_supported_arguments(['-mfpu=neon-vfpv4'])
   arm_neon_flags += '-DARM_NEON'
 elif host_cpu_family == 'aarch64'
-  common_c_flags += cc.get_supported_arguments(['-mfpu=neon-fp-armv8'])
+  common_c_flags += cc.get_supported_arguments(['-mfpu=neon-fp-armv8', '-ftree-vectorize'])
 endif
 
+add_project_link_arguments(common_l_flags, language: 'c')
+
 ################################################################################
 # Check for compiler CPU extensions
 
@@ -497,6 +511,15 @@ todo_file = files('TODO')
 export_symbols_file = files('export-symbols')
 gen_babl_map_file = files('gen_babl_map.py')
 
+################################################################################
+# Install debug data (.pdb) on Windows
+# Ideally meson should take care of it automatically.
+# See: https://github.com/mesonbuild/meson/issues/12977
+#if platform_win32 and pdb_support
+#  install_win_debug_script = find_program('./meson_install_win_debug.sh')
+#  meson.add_install_script(install_win_debug_script)
+#endif
+
 ################################################################################
 # Subdirs
 
diff --git a/meson_install_win_debug.sh b/meson_install_win_debug.sh
new file mode 100644
index 0000000..5d09524
--- /dev/null
+++ b/meson_install_win_debug.sh
@@ -0,0 +1,20 @@
+#!/bin/sh
+
+find . \( -iname '*.dll' -or -iname '*.exe' \) | \
+while IFS= read -r build_bin;
+do
+  build_bin_name="${build_bin##*/}"
+  installed_bin=$(find ${MESON_INSTALL_DESTDIR_PREFIX} -iname "$build_bin_name")
+  if [ x"$installed_bin" != "x" ]; then
+    install_dir=$(dirname ${installed_bin})
+    pdb_debug=$(echo $build_bin|sed 's/\.\(dll\|exe\)$/.pdb/')
+    if [ -f "$pdb_debug" ]; then
+      # Note: meson hides script outputs anyway on success. But this can be
+      # useful when debugging.
+      echo Installing $pdb_debug to $install_dir
+      if [ -z "$MESON_INSTALL_DRY_RUN" ]; then
+        cp -f $pdb_debug $install_dir
+      fi
+    fi
+  fi
+done;
diff --git a/tests/meson.build b/tests/meson.build
index 7c67e70..fd09fb9 100644
--- a/tests/meson.build
+++ b/tests/meson.build
@@ -42,7 +42,7 @@ foreach test_name : test_names
     test_name + '.c',
     include_directories: [rootInclude, bablInclude],
     link_with: babl,
-    dependencies: [thread, lcms],
+    dependencies: [thread, lcms, log],
     export_dynamic: true,
     install: false,
   )
diff --git a/tools/meson.build b/tools/meson.build
index 89ccf40..a2dc1cb 100644
--- a/tools/meson.build
+++ b/tools/meson.build
@@ -18,7 +18,7 @@ foreach tool_name : tool_names
     tool_name + '.c',
     include_directories: [rootInclude, bablInclude],
     link_with: babl,
-    dependencies: [math, thread, lcms],
+    dependencies: [math, thread, lcms, log],
     install: false,
   )