Update libjpeg_turbo to use clz for bitcounting on ARM

Cherry-picked r1220 from upstream:
Use clz/bsr instructions on ARM for bit counting rather than the lookup table (reduces memory footprint and can improve performance in some cases.)

Upstream review:
http://sourceforge.net/p/libjpeg-turbo/patches/57/

Original review:
https://codereview.appspot.com/77480045/

Removing the lookup table saves 64k data for each process that uses jpeg encoding. Benchmarks on a few ARM devices shows encoding performance changes, from a slowdown of 3-4% on some devices, to a speedup of 10-20% on other devices. In average performance improves.

x86 will still use the lookup table because the bsr instruction showed to be slower on some chips.

BUG=
R=noel@chromium.org

Review URL: https://codereview.appspot.com/97690043

git-svn-id: http://src.chromium.org/svn/trunk/deps/third_party/libjpeg_turbo@272637 4ff67af0-8c30-449e-8e8b-ad334ec8d88c

3 files changed, 108 insertions, 2 deletions

diff --git a/README.chromium b/README.chromium
index a595954..b01cb0a 100644
--- a/README.chromium
+++ b/README.chromium
@@ -9,6 +9,7 @@ Description:
 This consists of the components:
 * A partial copy of libjpeg-turbo 1.3.1 (r1219);
 * Revision r1188 cherry-picked from upstream trunk into config.h;
+* Revision r1220 cherry-picked from upstream trunk into jchuff.c;
 * A build file (libjpeg.gyp), and;
 * Patched header files used by Chromium.
 
diff --git a/google.patch b/google.patch
index 9fcfe9b..8a9179d 100644
--- a/google.patch
+++ b/google.patch
@@ -1767,3 +1767,78 @@ Index: jdhuff.c
  
  METHODDEF(boolean)
  decode_mcu (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+Index: jchuff.c
+===================================================================
+--- jchuff.c	(revision 1219)
++++ jchuff.c	(revision 1220)
+@@ -22,8 +22,36 @@
+ #include "jchuff.h"		/* Declarations shared with jcphuff.c */
+ #include <limits.h>
+ 
++/*
++ * NOTE: If USE_CLZ_INTRINSIC is defined, then clz/bsr instructions will be
++ * used for bit counting rather than the lookup table.  This will reduce the
++ * memory footprint by 64k, which is important for some mobile applications
++ * that create many isolated instances of libjpeg-turbo (web browsers, for
++ * instance.)  This may improve performance on some mobile platforms as well.
++ * This feature is enabled by default only on ARM processors, because some x86
++ * chips have a slow implementation of bsr, and the use of clz/bsr cannot be
++ * shown to have a significant performance impact even on the x86 chips that
++ * have a fast implementation of it.  When building for ARMv6, you can
++ * explicitly disable the use of clz/bsr by adding -mthumb to the compiler
++ * flags (this defines __thumb__).
++ */
++
++/* NOTE: Both GCC and Clang define __GNUC__ */
++#if defined __GNUC__ && defined __arm__
++#if !defined __thumb__ || defined __thumb2__
++#define USE_CLZ_INTRINSIC
++#endif
++#endif
++
++#ifdef USE_CLZ_INTRINSIC
++#define JPEG_NBITS_NONZERO(x) (32 - __builtin_clz(x))
++#define JPEG_NBITS(x) (x ? JPEG_NBITS_NONZERO(x) : 0)
++#else
+ static unsigned char jpeg_nbits_table[65536];
+ static int jpeg_nbits_table_init = 0;
++#define JPEG_NBITS(x) (jpeg_nbits_table[x])
++#define JPEG_NBITS_NONZERO(x) JPEG_NBITS(x)
++#endif
+ 
+ #ifndef min
+  #define min(a,b) ((a)<(b)?(a):(b))
+@@ -272,6 +300,7 @@
+     dtbl->ehufsi[i] = huffsize[p];
+   }
+ 
++#ifndef USE_CLZ_INTRINSIC
+   if(!jpeg_nbits_table_init) {
+     for(i = 0; i < 65536; i++) {
+       int nbits = 0, temp = i;
+@@ -280,6 +309,7 @@
+     }
+     jpeg_nbits_table_init = 1;
+   }
++#endif
+ }
+ 
+ 
+@@ -482,7 +512,7 @@
+   temp2 += temp3;
+ 
+   /* Find the number of bits needed for the magnitude of the coefficient */
+-  nbits = jpeg_nbits_table[temp];
++  nbits = JPEG_NBITS(temp);
+ 
+   /* Emit the Huffman-coded symbol for the number of bits */
+   code = dctbl->ehufco[nbits];
+@@ -516,7 +546,7 @@
+     temp ^= temp3; \
+     temp -= temp3; \
+     temp2 += temp3; \
+-    nbits = jpeg_nbits_table[temp]; \
++    nbits = JPEG_NBITS_NONZERO(temp); \
+     /* if run length > 15, must emit special run-length-16 codes (0xF0) */ \
+     while (r > 15) { \
+       EMIT_BITS(code_0xf0, size_0xf0) \
diff --git a/jchuff.c b/jchuff.c
index 29bf389..68e4e0e 100644
--- a/jchuff.c
+++ b/jchuff.c
@@ -22,8 +22,36 @@
 #include "jchuff.h"		/* Declarations shared with jcphuff.c */
 #include <limits.h>
 
+/*
+ * NOTE: If USE_CLZ_INTRINSIC is defined, then clz/bsr instructions will be
+ * used for bit counting rather than the lookup table.  This will reduce the
+ * memory footprint by 64k, which is important for some mobile applications
+ * that create many isolated instances of libjpeg-turbo (web browsers, for
+ * instance.)  This may improve performance on some mobile platforms as well.
+ * This feature is enabled by default only on ARM processors, because some x86
+ * chips have a slow implementation of bsr, and the use of clz/bsr cannot be
+ * shown to have a significant performance impact even on the x86 chips that
+ * have a fast implementation of it.  When building for ARMv6, you can
+ * explicitly disable the use of clz/bsr by adding -mthumb to the compiler
+ * flags (this defines __thumb__).
+ */
+
+/* NOTE: Both GCC and Clang define __GNUC__ */
+#if defined __GNUC__ && defined __arm__
+#if !defined __thumb__ || defined __thumb2__
+#define USE_CLZ_INTRINSIC
+#endif
+#endif
+
+#ifdef USE_CLZ_INTRINSIC
+#define JPEG_NBITS_NONZERO(x) (32 - __builtin_clz(x))
+#define JPEG_NBITS(x) (x ? JPEG_NBITS_NONZERO(x) : 0)
+#else
 static unsigned char jpeg_nbits_table[65536];
 static int jpeg_nbits_table_init = 0;
+#define JPEG_NBITS(x) (jpeg_nbits_table[x])
+#define JPEG_NBITS_NONZERO(x) JPEG_NBITS(x)
+#endif
 
 #ifndef min
  #define min(a,b) ((a)<(b)?(a):(b))
@@ -272,6 +300,7 @@ jpeg_make_c_derived_tbl (j_compress_ptr cinfo, boolean isDC, int tblno,
     dtbl->ehufsi[i] = huffsize[p];
   }
 
+#ifndef USE_CLZ_INTRINSIC
   if(!jpeg_nbits_table_init) {
     for(i = 0; i < 65536; i++) {
       int nbits = 0, temp = i;
@@ -280,6 +309,7 @@ jpeg_make_c_derived_tbl (j_compress_ptr cinfo, boolean isDC, int tblno,
     }
     jpeg_nbits_table_init = 1;
   }
+#endif
 }
 
 
@@ -482,7 +512,7 @@ encode_one_block (working_state * state, JCOEFPTR block, int last_dc_val,
   temp2 += temp3;
 
   /* Find the number of bits needed for the magnitude of the coefficient */
-  nbits = jpeg_nbits_table[temp];
+  nbits = JPEG_NBITS(temp);
 
   /* Emit the Huffman-coded symbol for the number of bits */
   code = dctbl->ehufco[nbits];
@@ -516,7 +546,7 @@ encode_one_block (working_state * state, JCOEFPTR block, int last_dc_val,
     temp ^= temp3; \
     temp -= temp3; \
     temp2 += temp3; \
-    nbits = jpeg_nbits_table[temp]; \
+    nbits = JPEG_NBITS_NONZERO(temp); \
     /* if run length > 15, must emit special run-length-16 codes (0xF0) */ \
     while (r > 15) { \
       EMIT_BITS(code_0xf0, size_0xf0) \