CHANGES

Release 1.4.0 - in progress (compared to 1.3)
----------------------------------------------

Language, standard libary, and compiler changes (for shader writers):

* Simplex noise: new noise varieties "simplex" and "usimplex" are 
  faster than Perlin noise (and MUCH faster for the 4-D variety),
  but has a somewhat different appearance.  Judge for yourself, use
  whichever noise suits your purpose.
* oslc now catches mismatches between the format string and the types
  of arguments to format(), printf(), fprintf(), warning(), and error().
* oslc new gives a compile error when trying to use a function as
  if it were an ordinary variable. (1.3.2)
* Improvement to oslc function type checking -- allow more nuanced
  matching of polymorphism based on return types.  First we check for
  identical return types, then we test allowing non-identical spatial
  triples to match, then finally we allow any return type to match.
* New ShadingSystem attribute "debug_uninit", which when set to nonzero
  tries to detect use of uninitialized variables at runtime.  (Off by
  default because it's very expensive, just use if you are debugging.)
* Implemented the OSL aastep() function, which was always described in
  the docs but never implemented.

ShadingSystem API changes and new options (for renderer writers):

* New ShadingSystem attribute "renderer_outputs" enumerates the renderer
  outputs (AOVs) explicitly, so that elision of unused outputs can be done
  with full understanding of which params will or will not be sued.
  Beware: if you enable "opt_elide_unconnected_outputs", you MUST pass
  "renderer_outputs" or your AOVs may be eliminated!

Performance improvements:

* "Middleman" optimization notices when you have 3 layers A,B,C with
  connections A.Aout -> B.Bin and B.Bout -> C.Cin, an Bout is simply
  copied unconditionally from Bin.  In this case, Aout is connected
  directly to Cin, potentially leading to further optimizations, including
  layer B getting culled entirely.
* Layer-to-layer global propagation: Notice cases where an upstream
  output is guaranteed to be aassigned a global variable (such as N or u)
  and then alias the downstream parameter to that global (eliminating
  the connection and possibly culling the upstream layer entirely).
* Improve constant propagation from layer to layer (it used to do so only
  if the output of the earlier layer was never used in the upstream layer;
  but actually it's ok for it to be read, as long as it's not rewritten).
* Runtime constant-folding of cellnoise() calls.
* Local variables that are arrays initialized with float, int, or string
  constants have been greatly optimized.  If that's the only time anything
  is written into the array, it will be turned into a true static constant
  with no per-shader-invocation initialization or allocation costs.
  Thus, it is no longer expensive to use a large array of constants as a
  table... it may look like a per-shade cost to set up the array, but by
  the time the optimizer gets done with it, it is not.
* Improved runtime constant-folding of max/min (now works for the integer
  case, as well as previously working for the float-based types).
* Runtime constant-folding of cos, sin, acos, asin, normalize, inversesqrt,
  exp, exp2, expm1, log, log2, log10, logb, erf, erfc, radians, degrees.
* Optimization that speeds up certain closure manipulations and can
  cut overhead especially for relatively simple shaders.
* New peephole optimizations for two common patterns: 
  (1) add A B C ; sub D A C => [add A B C;] assign D B
  and the same if the order of add & sub are switched.
  (2) OP A B... ; assign A C => OP C B...  as long as A is not subsequently
  used and OP fully overwrites A and writes to no other variables.

Bug fixes:

* Fix 'debugnan' tendency to give false positives for certain assignments
  to arrays and multi-component variables (colors, points, etc.).
* The integer versions of min(), max(), and clamp() were not working
  properly. (1.3.1)
* fmod(triple,triple,float) was not working properly in Windows. (1.3.2)
* Avoid incorrect heap addressing when retrieving a symbol's address. (1.3.2)
* Fix for a bug where an early 'return' from a shader could lead to
  output parameters not being properly copied to the inputs of their
  downstream connections. (1.3.2)
* Fix bug where filterwidth() didn't properly mark itself as needing
  derivatives of its argument. (1.3.2)
* Bug fix: Was not properly initializing the default values of output
  params if they were written but never subsequently read.

Under the hood:

* Eliminate use of custom DEBUG preprocessor symbol in favor of the more
  standard NDEBUG (beware: its sense is reversed).
* All references to and vestiges of TBB have been removed.

Build & test system improvements and developer goodies:

* testshade: --options lets you set arbitrary options to the ShadingSystem
  for test runs.
* Added CMake option 'ENABLERTTI' that enables use of RTTI for sites that
  must link against an RTTI-enabled LLVM (the default is to assume LLVM
  is built without RTTI, so we don't use it either). (1.3.1)
* Fix some ambiguity about which variety of shared_ptr we are using -- 
  this could produce problems in apps that include the OSL headers but also
  use different shared_ptr varieties.
* Fix a variety of build issues and warnings under Windows.
* Work around CMake bug where on some platforms, CMake doesn't define the
  'NDEBUG' when making the 'RelWithDebInfo' target (which is sometimes
  used for profiling). (1.3.1)
* Add Make/CMake option 'USE_EXTERNAL_PUGIXML' which if set to nonzero
  will find and use a system-installed PugiXML rather than assuming that
  it's included in libOpenImageIO.  When used, it will also use the
  environment variable PUGIXML_HOME (if set) as a hint for where to find
  it, in case it's not in a usual system library directory. (1.3.1)
* Nicer debug printing (for developers) when an op doesn't have an
  associated source line. (1.3.2)
* Reorder include directories for building LLVM bitcode so that the
  openimageio, ilmbase and boost directories are ahead of any system
  directories given by llvm-config --cxxflags.  This avoids the LLVM
  bitcode getting compiled against different library versions than the
  rest of the code if multiple versions are installed.
* Have debug_groupname and debug_layername apply to LLVM debugging info
  as well.

Documentation:

* Eliminate references to OSL random() function; it was never
  implemented and nobody seemed to miss it or have a compelling use
  case. (1.3.1)



Release 1.3.0 - 14 Feb 2013 (compared to 1.2)
----------------------------------------------

Language, standard libary, and compiler changes (for shader writers):

* pointcloud_write() allows shaders to write data and save it as a point
  cloud.
* spline now accepts a "constant" interpolation type, which interpolates
  with discrete steps.
* isconnected(var) returns true if var is a shader parameter and is
  connected to an earlier layer. This is helpful in having shader logic
  that can discern between a connected parameter versus a default.
* Whole-array assignment is now supported.  That is, if A and B are arrays
  of the same type and length(A) >= length(B), it is legal to write
  A=B without needing to write a loop to copy each element separately.
* stoi() and stof() convert strings to int or float types, respectively
  (just like the identically-named functions in C++11).  These do properly
  constant-fold during runtime optimization if their inputs can be deduced.
* split() splits a string into tokens and stores them in the elements of an
  array of strings (and does constant-fold!).
* distance(point A, point B, point Q) gives the distance from Q to the
  line segment joining A and B.  This was described by the OSL spec all
  along, but somehow was not properly implemented.

ShadingSystem API changes and new options (for renderer writers):

* Default implementation of all the pointcloud_*() functions in
  RendererServices, if Partio is found at build time.
* ShadingSystem attribute "compile_report" controls whether information
  should be sent to the renderer for each shading group as it compiles.
  (The default is now to be more silent, set this to nonzero if you want
  to see the old per-group status messages.)
* ShadingSystem attribute "countlayerexecs" adds debugging code to count
  the total number of shader layer executions (off by default; it slows
  shader execution down slightly, is only meant for debugging).
* Add a parameter to OSLCompiler::compile to specify the path to the
  stdosl.h file.
* New call: ShadingSystem::LoadMemoryCompiledShader() loads a shader with
  OSO data from a memory buffer, rather than from a file.

Performance improvements:

* Reduce instance symbol memory use, lowering OSL-related memory by 1/2-2/3
  for large scenes.
* Identical shader instances within a shader group are automatically
  merged -- this can have substiantial performance and memory
  improvements for very complex shader groups that may (inadvertently,
  for convenience, or as an artifact of certain lookdev tools) have
  multiple instances with identical input parameters and connections.
  We have seen some shots that render up to 20% faster with this change.
* Better constant-folding for pointcloud_get (if indices are known).
* Speed up implementation of exp() for some Linux platform (works around
  a known glibc issue).
* Speed up of mix() by making it a built-in function (rather than defined
  in stdosl.h) and doing aggressive constant folding on it. Also some special
  cases to try to avoid having mix() unnecessarily trigger execution of
  earlier layers when one or both of its arguments is a connected shader
  param and the mix value turns out to be 0 or 1.
* Runtime optimization now includes constant-folding of getattribute()
  calls, for those attributes which are scene-wide (such as renderer options
  that apply to the whole frame).
* Improvements in the algorithm that tracks which symbols need to carry
  derivatives around results in MUCH faster compile times for extremely
  complex shaders.
* Improve bad performance for shaders that call warning() prolifically, by
  imposing a maximum number of shader warnings that are echoed back to the
  renderer (controlled by ShadingSystem attribute "max_warnings_per_thread",
  which defaults to 100; 0 means that there is no maximum).

Bug fixes and minor improvements:

* Fix incorrect oso output of matrix parameter initialization. (This did
  not affect shader execution, but could result in oslquery/oslinfo not
  reporting default matrix parameter values correctly.)
* Bug fix: oslinfo didn't print int metadata values properly.
* oslc better error checking when calling a variable as if it were a
  function.
* Parsing of oso files was broken for compiled shaders that had string
  metadata whose metadata value contained the '}' character.
* Pointcloud function bug fixes: now robust to being passed empty filenames,
  or calling pointcloud_get when count=0.
* Fix crash with C preprocessor when #include fails or no shader function
  is defined.
* Improve error reporting when parsing oso, especially in improving the
  error messages when an old renderer tries to load a .oso file created by
  a "newer" oslc, and the oso contains new instruction opcodes that the
  older renderer didn't know about.
* oslc now gives a helpful error message, rather than hitting an assertion,
  if you try to index into a non-array or non-component type.
* Broken "XYZ" color space transformations because of confusion between
  "xyz" and "XYZ" names.
* oslc: correctly catch errors with integer literals that overflow.
* oslc: proper error, rather than assertion, for 
  'closure color * closure color'.
* oslc: proper handling of preprocessor errors when stdosl.h is in a path 
  containing spaces.
* Fix bugs in the OSL exit() function.
* Don't error/warn if metadata name masks a global scope name.

Under the hood:

* Simplify code generation of binary ops involving closures.

Build & test system improvements and developer goodies:

* Many, many fixes to enable building and correct running of OSL on
  Windows.  Too many to list individually, let's just say that OSL 1.3
  builds and runs pretty robustly on Windows, older versions did not.
* Remove unused OpenGL and GLEW searching from CMake files.
* Fix a variety of compiler warnings, particularly on newer compilers,
  32 bit platforms, and Windows.
* Find LLVM correctly even when using "svn" LLVM versions.
* Adjust failure threshold on some tests to account for minor platform
  differences.
* New minimum OIIO version: 1.1.
* CMake fixes: Simplify searching for and using IlmBase, and remove
  unneeded search for OpenEXR.
* Find Partio (automatically, or using 'make PARTIO_HOME=...'), and if
  found, use it to provide default RendererServices implementations of
  all the pointcloud-related functions.
* Make/CMake STOP_ON_WARNING flag (set to 0 or 1) controls whether the
  build will stop upon any compiler warning (default) or if it should
  keep going (=0).
* Fixed bad interaction between weave preprocessor, OIIO::Strutil::format,
  and gcc 4.7.2.
* Fixes for static linkage of liboslcomp.
* Addressed compilation errors on g++ 4.7.2.
* Improved logic and error reporting for finding OpenImageIO at build time.
* Add support for linking to static LLVM libraries, and for specifying
  a particular LLVM version to be used.
* New BUILDSTATIC option for CMake/Make allows building of static OSL 
  libraries.
* Fixes to allow OSL to use LLVM 3.2.
* Fixes to allow OSL to be compiled by clang 3.2.
* Fixes to allow building on Windows with MinGW.
* Fix for building from a chroot where the host may be, e.g., 64 bit, but
  the target is 32 bit.
* Minimize cmake output clutter for things that aren't errors, unless
  VERBOSE=1.
* Various fixes to get OSL building and running on FreeBSD.
* Cmake/make option USE_LLVM_BITCODE to be able to force LLVM bitcode
  compilation on or off (defaults to on for Windows, off for Unix/Linux).

Documentation:
* Clarify that documentation is licensed under Creative Commons 3.0 BY.
* Clean up OSL Spec discussion of built-in material closures.
* Overhaul the suggested metadata conventions in the OSL spec (now 
  conforms to the conventions used by Katana).




Release 1.2.1 - 6 Nov, 2012 (compared to 1.2.0)
-----------------------------------------------
* Fix incorrect oso output of matrix parameter initialization. (This did
  not affect shader execution, but could result in oslquery/oslinfo not
  reporting default matrix parameter values correctly.)
* Build: remove unused OpenGL and GLEW searching from CMake files.
* Build: Fix a variety of compiler warnings, particularly on newer compilers,
  32 bit platforms, and Windows.
* Build: Find LLVM correctly even when using "svn" LLVM versions.
* Bug fix: oslinfo didn't print int metadata values properly.


Release 1.2.0 - Aug 30, 2012 (compared to 1.1)
----------------------------------------------
Note: this is more or less the production-hardened version of OSL that
was used to complete Men in Black 3, The Amazing Spider-Man, and Hotel
Transylvania.

New tools/utilities:
* New program "testrender" is a tiny ray-tracing renderer that uses OSL
  for shading.  Features are very minimal (only spheres are permitted at
  this time) and there has been no attention to performance, but it
  demonstrates how the OSL libraries may be integrated into a working
  renderer, what interfaces the renderer needs to supply, and how the
  BSDFs/radiance closures should be evaluated and integrated (including
  with multiple importance sampling).
* shaders/ubersurface.osl is an example of a general purpose surface
  shader.

Language, standard libary, and compiler changes:
* texture()/texture3d() support for subimage/face selection by name as
  well as by numeric index.
* getattribute() performs automatic type conversions that mostly
  correspond to the kind of automatic type casting you would get from
  ordinary assignments in OSL.  For example, getattribute("attrib",mycolor)
  used to fail if "attrib" turned out to be a float rather than a color;
  but now it succeeds and copies the float value to all three channels
  of mycolor.

ShadingSystem API changes and new options:
* Remove unused 'size' parameter from the register_closure API.

Optimization improvements:
* Constant-fold pointcloud_search calls when the position is a constant if
  the search returns few enough results, by doing the query at optimization
  time and putting the results into new constant arrays.
* Matrix parameters initialized by m=matrix(constA,constB,...) now 
  statically initialize, and no longer need to run initialization code
  each time the shader executes.

Bug fixes and minor improvements:
* Fix pointcloud_search to optionally take a 'sort' parameter, as
  originally documented.
* Unit tests weren't properly run as part of the testsuite.
* Track local+temp memory usage of optimized shaders and consider it an 
  error if a shader needs more than a maximum amount at runtime, set with
  the "max_local_mem_KB" attribute.
* Add pointcloud statistics.
* Fix derivative error for sincos() when the inputs have no derivatives but
  the outputs do.
* Bug fix to vector-returning Gabor noise (it could previously generate
  different values for different platforms).
* printf() of closures built from other closures allows for proper
  recursive printing of the closure tree.

Build & test system improvements and developer goodies:
* Simplify the namespace scheme.
* Remove support for certain old dependencies: OIIO versions < 0.10,
  LLVM < 3.0, and Boost < 1.40.
* Lots of little fixes to solve compiler warnings on various compilers.
* Support for newer OSX releases, particularly if /usr/bin/cpp-4.2 is not
  found.
* Better support for Boost::wave (portable C preprocessor replacement).
  Build with 'make USE_BOOST_WAVE=1' to force Wave use instead of system
  cpp.
* You can select a custom LLVM namespace with 'make LLVM_NAMESPACE=...'.
* Symbols that are not part of the OSL public APIs are now hidden from the
  linker in Linux/OSX if the CMake variable HIDE_SYMBOLS is on.
* New Makefile/CMake option LLVM_STATIC can be used to use static LLVM
  libraries rather than the default dynamic libraries.
* Support for LLVM 3.1.
* Support for building with Clang 3.1 (lots of warning fixes).
* Makefile/CMake variable EXTRA_CPP_DEFINITIONS allows you to inject
  additional compiler flags that you may need to customize the build for
  your site or a choice of unusual compiler.
* Add support for 'PROFILE=1' builds that are appropriate for use with
  a profile.



Release 1.1.0 - Mar 14, 2012 (compared to 1.0.0)
------------------------------------------------
Language, standard libary, and compiler changes:
* Allow closures as parameters to closures.
* New constants: M_2PI, M_4PI
* Generic noise: noise("noisetype",coords,...)
* Gabor noise (anisotropic, automatically antialiased) via noise("gabor").
* Fix mod/fmod discrepancy: fmod() now matches C, mod() always returns a
  positive result like in RSL.
* Allow "if (closure): and "if (!closure)" to test if a closure is empty
  or not.
* New optional parameter to trace(): "traceset" allows you to specify a
  named geometry set for tracing.

ShadingSystem API changes and new options:
* New "greedyjit" option will optimize & JIT all shader groups up front,
  concurrently, without locking.
* Add a way to name shader groups.
* attribute("options",...) lets you set a bunch of options at once.
* Options to enable/disable individual optimizations (mostly useful for
  debugging)

Optimization improvements:
* Allow block alias tracking on non-constants when it's safe.
* Track "stale" values to eliminate pointless assignments.
* Eliminate redundant "useparam" ops.
* Assignments to output parameters that are not connected to any
  downstream layers are now eliminated.
* More aggressive elision of ops that only write to symbols that won't
  be subsequently used.
* More careful identification and removal of parameters (input and output)
  that are both unused in the shader and not connected downstream.

Bug fixes and minor improvements:
* Minor blackbody fixes.
* Bug fix: don't mark constants as having their derivatives taken.
* Clamp splineinverse() for out-of-knot-range x input.
* Bug fix: the optimization of "a=b; a=c" was incorrect if c was an
  alias for a (it incorrectly eliminated the first assignment).
* Bug fix: work around LLVM thread safety issues during JIT.
* Bug fix: symbol_data() wasn't returning the right address for non-heap
  parameters.
* Bug fix: optimization errors related to break, continue, and return not
  properly marking the next instruction as a new basic block.
* Bug fix: luminance() with derivatives didn't work.
* Bug fix: in code generation of structure initializers.
* Improved error messages from ConnectShaders.
* Bug fix: type checking bug could case non-exactly-matching polymorphic 
  functions to coerce a return value even when that was not intended.
* Type checking improvements: Make sure point-point is a vector
  expression, and point+vector & point-vector are point expressions.

Build & test system improvements and developer goodies:
* testsuite overhauls:
    - run each test both optimized and not
    - generate all tests in build, not directly in ./testsuite
    - greatly simplify the run.py scripts
* Much more detailed debugging logs of the optimization process.
* Upgrade to clang/llvm 3.0.
* Lots of infrastructure to make debugging the optimizer easier.
  Including new options debug_groupname, debug_layername, only_groupname.
* Improved the build system's LLVM-finding logic.
* Fix warnings from gcc 4.6.



Release 1.0.0 - Oct 12, 2011
----------------------------
* Modified testshade (and the underlying SimpleRender class) to handle
  several standard named coordinate systems such as "camera", "screen",
  "NDC", "raster."
* blackbody() and wavelength_color().
* New ShadingSystem configuration attribute: "colorspace" lets you explain
  to OSL what RGB really means (e.g., "Rec709", "sRGB", "NTSC", etc.).
  The luminance(), blackbody(), wavelength_color, and conversion to/from
  XYZ now takes this into account correctly.
* rotate()  (always in spec, never implemented)



Release 0.6.2 - Sept 29, 2011
-----------------------------
* Statistics overhaul -- added optimization stats, eliminated unused ones.
* Allow a shader parameter to mask a global built-in function with
  only a warning, and improve scope conflict errors by pointing out the
  file and line of the previous definition.
* Altered the RendererServices API to add transform_points() method, which
  allows renderers to support nonlinear transformations (i.e., those that
  are not expressible as a 4x4 matrix).
* Issue a renderer error when unknown coordinate system names are used
  (can be turned of by setting the new ShadingSystem attribute 
  "unknown_coordsys_error" to false).
* New OSL built-in function: splineinverse().


Release 0.6.1 - Sept 20, 2011
-----------------------------
* Be more aggressive in freeing shader instance memory that's no longer
  needed after optimization and LLVM JIT.  This greatly reduces
  OSL-related memory consumption for scenes with large numbers of very
  complicated shading networks.
* Add Dz() which is helpful for apps that use OSL to shade volumes.  At
  present, we only correctly compute Dz(P), all other Dz() queries
  return 0.
* Additional statistics on how many instances and groups compile, and
  how many are empty after all optimizations are performed.
* Make sure all the relevant statistics can be queried via
  ShadingSystem::getattribute.


Release 0.6.0 - Sept 9, 2011
----------------------------
* ShadeExec API overhaul -- an app using it no longer needs
  ShadingSystemImpl internal knowledge.
* Thread-parallel runtime optimization and LLVM JIT of different shader
  groups.
* Optimizations: runtime constant folding of arraylength and regex_search,
  new instruction 'arraycopy' will copy an entire array at once.
* Renamed patterns.h to oslutil.h.
* Do not generate unnecessary code when optional texture parameters are set
  to their default values.
* Restore long-lost ability for layers to run unconditionally (not lazily)
  if they were marked as "non-lazy" (for example, if they write to globals.
* Make the "debugnan" attribute work again -- when turned on, code will
  be inserted after every op to be sure that no NaN or Inf values are
  generated, and also verify that shader inputs (globals) don't have NaN
  or Inf values passed in by the renderer.  A similar facility existed a
  long time ago, but we lost that functionality when we switched from
  the interpreter to LLVM.
* Looks for release versions of LLVM-2.9 (allows using a Macports LLVM 
  installation).


Release 0.5.4 - Jul 21, 2011
----------------------------
* Several fixes related to arrays of structs, and structs containing
  other structs.
* Fixed arrays of closures.
* Removed support for old LLVM 2.7, nobody seemed to be using it any more.
* Changed the definition of dict_find() to return -1 for an invalid
  dictionary, to distinguish it from 0 meaning that the query failed but
  the dictionary itself was valid.
* Make array parameters safe to convert to constants during runtime
  optimization.
* Support derivatives in pointcloud searches.
* Fixed several runtime optimizer bugs.
* Fixed a bug where environment() calls with an optional "alpha" parameter
  that has derivatives was overwriting memory.
* Fixed code generation bug for a*=b and a/=b.
* Automatically initialize all local string variables to NULL, to avoid
  bad pointers for uninitialized strings.
* Bug fix: dict_value() wasn't properly marking its argument as writable.
* Automatic range checking of array and component access.
* Fix uninitialized derivatives for pointcloud().
* Speed up getattribute() calls by caching information about failed
  getattribute queries in the ShadingContext.
* Fix to constant folding of gettextureinfo: a failed lookup should not
  fold, because we want the error to occur in shader execution, not during
  optimization.
* Refactor, clean up, and comment testshade.
* oslc now gives an error on non-void functions that fail to return a value.
* Fixed implementation of area() that could generate an assertion.
* Fix escape sequences in string literals: we were handling it correctly
  for printf-like format strings, but not other string literals.
* break and continue now work properly (just like in C/C++).
* You can now return from anywhere in a user function (multiple times if
  you want), just like C/C++, and are no longer restricted to the only
  return statement being the last statement of the function.
* New include file for shaders: patterns.h.  Now, it only includes a handy
  'wireframe()' function, but will expand for other useful things.
* New function: int getmatrix(from,to,M) is like the matrix(from,to)
  constructor, but returns a success value so a shader can tell if the
  named coordinate systems failed to be found by the renderer.


Release 0.5.3 - Apr 19, 2011
----------------------------
* Fix missing derivatives for sign() function.
* Fix closure color type size (crashes).
* Fix bug with environment() when passed "alpha" pointers with derivatives.
* Improve error messages for getmessage/setmessage to catch the most
  common sources of non-deterministic behavior.
* Bug fix when constant-folding gettextureinfo().
* Fix mismatched prototype for subsurface() closure.
* Texture errors encountered in shader constant folding are now properly
  reported to the renderer.
* Allow functions to have array parameters of unspecified length.
* Fix subtle bug related to lifetime analysis of variables in loops (led
  to incorrect optimizations).


Release 0.5.2 - Mar 14, 2011
----------------------------

* Windows: use boost::wave instead of external cpp; various other Windows
  compilation fixes.
* texture & environment now take an optional "interp" parameter that
  overrides the interpolation/filtering method (valid arguments:
  "smartcubic", "cubic", "linear", "closest").
* Bug fixes to getmessage() and its handling of derivatives, which includes
  a slight RendererServices API change.