[libpng16] Allow run-time ARM NEON checking to be disabled. A new configure

option:
    --enable-arm-neon=always will stop the run-time checks. New checks
    within arm/arm_init.c will cause the code not to be compiled unless
    __ARM_NEON__ is set. This should make it fail safe (if someone asks
    for it on then the build will fail if it can't be done.)
Fixed ALIGNED_MEMORY support.

ANNOUNCE

@@ -50,6 +50,12 @@ Version 1.6.1beta02 [February 19, 2013]
   Reenabled code to allow zero length PLTE chunks for MNG.
 
 Version 1.6.1beta03 [February 19, 2013]
+  Fixed ALIGNED_MEMORY support.
+  Allow run-time ARM NEON checking to be disabled. A new configure option:
+    --enable-arm-neon=always will stop the run-time checks. New checks
+    within arm/arm_init.c will cause the code not to be compiled unless
+    __ARM_NEON__ is set. This should make it fail safe (if someone asks
+    for it on then the build will fail if it can't be done.)
 
 Send comments/corrections/commendations to png-mng-implement at lists.sf.net
 (subscription required; visit

CHANGES

@@ -4407,6 +4407,12 @@ Version 1.6.1beta02 [February 19, 2013]
   Reenabled code to allow zero length PLTE chunks for MNG.
 
 Version 1.6.1beta03 [February 19, 2013]
+  Fixed ALIGNED_MEMORY support.
+  Allow run-time ARM NEON checking to be disabled. A new configure option:
+    --enable-arm-neon=always will stop the run-time checks. New checks
+    within arm/arm_init.c will cause the code not to be compiled unless
+    __ARM_NEON__ is set. This should make it fail safe (if someone asks
+    for it on then the build will fail if it can't be done.)
 
 Send comments/corrections/commendations to png-mng-implement at lists.sf.net
 (subscription required; visit

arm/arm_init.c

@@ -16,12 +16,17 @@
 
 #include "../pngpriv.h"
 
+#if defined(PNG_FILTER_OPTIMIZATIONS) && defined(__arm__) && \
+   defined(__ARM_NEON__)
 /* __arm__ is defined by GCC, MSVC defines _M_ARM to the ARM version number,
  * Andoid intends to define __ANDROID__, however there are bugs in their
  * toolchain; use -D__ANDROID__ to work round this.
+ *
+ * __ARM_NEON__ is used to ensure that the compiler has the appropriate ARM
+ * NEON support
  */
-#if defined(__linux__) && defined(__arm__)
+
-#define CHECK_NEON
+#ifdef PNG_ARM_NEON_CHECK_SUPPORTED /* Do run-time checks */
 #include <signal.h> /* for sig_atomic_t */
 
 #ifdef __ANDROID__
@@ -43,7 +48,7 @@ png_have_neon(png_structp png_ptr)
    return andoid_getCpuFamily() == ANDROID_CPU_FAMILY_ARM &&
       (android_getCpuFeatures() & ANDROID_CPU_ARM_FEATURE_NEON) != 0;
 }
-#else
+#elif defined(__linux__)
 /* The generic __linux__ implementation requires reading /proc/self/auxv and
  * looking at each element for one that records NEON capabilities.
  */
@@ -143,14 +148,20 @@ png_have_neon(png_structp png_ptr)
    close(fd);
    return 0;
 }
-#endif /* !__ANDROID__ */
+#else
-#endif /* __linux__ && __arm__ */
+   /* We don't know how to do a run-time check on this system */
+#  error "no support for run-time ARM NEON checks"
+#endif /* OS checks */
+#endif /* PNG_ARM_NEON_CHECK_SUPPORTED */
+
+#ifndef PNG_ALIGNED_MEMORY_SUPPORTED
+#  error "ALIGNED_MEMORY is required; set: -DPNG_ALIGNED_MEMORY_SUPPORTED"
+#endif
 
 void
 png_init_filter_functions_neon(png_structp pp, unsigned int bpp)
 {
-#ifdef __arm__
+#ifdef PNG_ARM_NEON_CHECK_SUPPORTED
-#ifdef CHECK_NEON
    static volatile sig_atomic_t no_neon = -1; /* not checked */
 
    if (no_neon < 0)
@@ -158,7 +169,7 @@ png_init_filter_functions_neon(png_structp pp, unsigned int bpp)
 
    if (no_neon)
       return;
-#endif
+#endif /* PNG_ARM_NEON_CHECK_SUPPORTED */
 
    /* IMPORTANT: any new external functions used here must be declared using
     * PNG_INTERNAL_FUNCTION in ../pngpriv.h.  This is required so that the
@@ -188,8 +199,5 @@ png_init_filter_functions_neon(png_structp pp, unsigned int bpp)
       pp->read_filter[PNG_FILTER_VALUE_PAETH-1] =
           png_read_filter_row_paeth4_neon;
    }
-#else
-   PNG_UNUSED(pp)
-   PNG_UNUSED(bpp)
-#endif
 }
+#endif /* FILTER_OPTIMIZATIONS && __arm__ && __ARM_NEON__ */

arm/filter_neon.S

@@ -10,11 +10,14 @@
  * and license in png.h
  */
 
-/* This is required to get the symbol renames, which are #defines */
+/* This is required to get the symbol renames, which are #defines, and also
+ * includes the value of PNG_FILTER_OPTIMIZATIONS.
+ */
 #define PNG_VERSION_INFO_ONLY
 #include "../pngpriv.h"
 
-#ifdef __arm__
+#if defined(PNG_FILTER_OPTIMIZATIONS) && defined(__arm__) && \
+   defined(__ARM_NEON__)
 #if defined(__linux__) && defined(__ELF__)
 .section .note.GNU-stack,"",%progbits /* mark stack as non-executable */
 #endif
@@ -229,4 +232,4 @@ func    png_read_filter_row_paeth3_neon, export=1
 
         pop             {r4,pc}
 endfunc
-#endif
+#endif /* FILTER_OPTIMIZATIONS && __arm__ && __ARM_NEON__ */

configure.ac

@@ -232,15 +232,20 @@ AM_CONDITIONAL([DO_PNG_PREFIX], [test "${with_libpng_prefix:-no}" != "no"])
 AC_SUBST([AM_CCASFLAGS], [-Wa,--noexecstack])
 
 AC_ARG_ENABLE([arm-neon],
-   AS_HELP_STRING([[[--enable-arm-neon]]], [Enable ARM NEON optimizations]),
+   AS_HELP_STRING([[[--enable-arm-neon]]],
-   [if test "${enableval}" = "yes"; then
+      [Enable ARM NEON optimizations: use 'always' to turn off run-time checks]),
+   [if test "${enableval}" = "yes" -o "${enableval}" = "always"; then
       AC_DEFINE([PNG_FILTER_OPTIMIZATIONS],
                 [png_init_filter_functions_neon],
                 [ARM NEON filter initialization function])
-      AC_DEFINE([PNG_ALIGNED_MEMORY_SUPPORTED], [1],
+      AC_DEFINE([PNG_ALIGNED_MEMORY_SUPPORTED], [],
                 [Align row buffers])
+      if test "${enableval}" = "always"; then
+         AC_DEFINE([PNG_NO_ARM_NEON_CHECK], [],
+                   [Turn off run-time checking for ARM NEON support])
+      fi
     fi])
-AM_CONDITIONAL([PNG_ARM_NEON], [test "${enable_arm_neon:-no}" = yes])
+AM_CONDITIONAL([PNG_ARM_NEON], [test "${enable_arm_neon:-no}" != "no"])
 
 AC_MSG_NOTICE([[Extra options for compiler: $PNG_COPTS]])
 

scripts/pnglibconf.dfa

@@ -136,6 +136,12 @@ logunsupported = 1
 # - PNG_USER_VERSIONINFO_COMPANYNAME
 # - PNG_USER_VERSIONINFO_LEGALTRADEMARKS
 
+# It is necessary to include configures definitions here so that AC_DEFINE
+# in configure.ac works in a comprehensible way
+@#if defined(HAVE_CONFIG_H) && !defined(PNG_NO_CONFIG_H)
+@#  include "config.h"
+@#endif
+
 @#ifdef PNG_USER_CONFIG
 @#  include "pngusr.h"
 @#endif
@@ -187,6 +193,18 @@ setting PREFIX
 
 setting DEFAULT_READ_MACROS default 1
 
+# This setting allows a hardware or configuration specific filter optimization
+# function to be specified, the argument is the name of the filter initializaion
+# function to use.
+
+setting FILTER_OPTIMIZATIONS
+
+# This option turns on runtime checks for ARM NEON support, it is irrelevant
+# on other platforms and it is irrelevant unless NEON code is turned on.  Checks
+# are on by default
+
+option ARM_NEON_CHECK
+
 # These settings configure the default compression level (0-9) and 'strategy';
 # strategy is as defined by the implementors of zlib, it describes the input
 # data and modifies the zlib parameters in an attempt to optimize the balance
@@ -577,7 +595,7 @@ setting sRGB_PROFILE_CHECKS default 2
 # but can help (in theory) on some architectures.  Only affects
 # internal structures.  Added at libpng 1.4.0
 
-option ALIGN_MEMORY
+option ALIGNED_MEMORY
 
 # Buggy compilers (e.g., gcc 2.7.2.2) need PNG_NO_POINTER_INDEXING
 # See png[wr]util.c, normally this should always be *on*

scripts/pnglibconf.h.prebuilt

@@ -44,7 +44,7 @@
 /* end of settings */
 /* options */
 #define PNG_16BIT_SUPPORTED
-#define PNG_ALIGN_MEMORY_SUPPORTED
+#define PNG_ALIGNED_MEMORY_SUPPORTED
 #define PNG_BENIGN_ERRORS_SUPPORTED
 #define PNG_BENIGN_READ_ERRORS_SUPPORTED
 /*#undef PNG_BENIGN_WRITE_ERRORS_SUPPORTED*/