When capabilities are needed for specific SPIR-V instructions, it is
fragile to do so based on GLSL/AST usage; it should be based on actual
instructions they got translated to.
1) On some old versions of MSVC:
glslang\MachineIndependent\Constant.cpp(187): warning C4056: overflow in floating-point constant arithmetic
On this platform the definition of INFINITY is as follows:
#ifndef _HUGE_ENUF
#define _HUGE_ENUF 1e+300 // _HUGE_ENUF*_HUGE_ENUF must overflow
#endif
#define INFINITY ((float)(_HUGE_ENUF * _HUGE_ENUF))
Moving the negation outside the cast seems to resolve that issue.
2) Some Linux compilers were unhappy with lines 226/227
glslang/MachineIndependent/Constant.cpp: In member function 'virtual glslang::TIntermTyped* glslang::TIntermConstantUnion::fold(glslang::TOperator, const glslang::TIntermTyped*) const':
glslang/MachineIndependent/Constant.cpp:226:99: error: integer overflow in expression [-Werror=overflow]
else if (rightUnionArray[i].getIConst() == -1 && leftUnionArray[i].getIConst() == -(int)0x80000000)
^~~~~~~~~~~~~~~~
glslang/MachineIndependent/Constant.cpp:227:48: error: integer overflow in expression [-Werror=overflow]
newConstArray[i].setIConst(-(int)0x80000000);
^~~~~~~~~~~~~~~~
Moving the negation to the right side of the cast made those happy, but then some Windows compilers were unhappy:
glslang\MachineIndependent\Constant.cpp(226): warning C4146: unary minus operator applied to unsigned type, result still unsigned
glslang\MachineIndependent\Constant.cpp(227): warning C4146: unary minus operator applied to unsigned type, result still unsigned
which required adding on the "ll" suffix.
3) Android builds where unhappy with line 242:
glslang/MachineIndependent/Constant.cpp:242:100: error: comparison of integers of different signs: 'long long' and 'unsigned long long' [-Werror,-Wsign-compare]
else if (rightUnionArray[i].getI64Const() == -1 && leftUnionArray[i].getI64Const() == -0x8000000000000000ll)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ^ ~~~~~~~~~~~~~~~~~~~~~
1 error generated.
Adding an explicit (long long) cast resolved this.
And the negation needs to be on the right side of the cast otherwise linux
builds are unhappy as in (2).
4) Android builds are unhappy with out of order initializers:
glslang/MachineIndependent/reflection.h:60:84: error: field 'type' will be initialized after field 'stages' [-Werror,-Wreorder]
glDefineType(pGLDefineType), size(pSize), index(pIndex), counterIndex(-1), type(pType.clone()), stages(EShLanguageMask(0)) { }
^
1 error generated.
Change-Id: Ic9a05fa7912498284885113d8b051f93f822f62b
Fixes#854. But, only good if we are not trying to use the same
texture for both shadow and non-shadow constructors.
Force the type of the texture to have 'shadow' set when it is
constructed with a samplerShadow.
Includes the following:
Add Vulkan 1.1 capability sets
Don't merge types of resources
Remove stores of undef.
Make sure the constant folder get the correct type.
This is one step in providing full linker functionality for creating
correct SPIR-V from multiple compilation units for the same stage.
(This was the only remaining "hard" part. The rest should be simple.)
More recent NDK releases have their own CMake toolchain file.
Use it.
Also, download the NDK from github.com:dneto0/android-ndk.
That is a fork of the repo we used to use, but we have more
control over how long it stays stable.
The Travis-CI bot downloads a copy of the Android NDK. The source
we get it from recently updated to Android NDK r17b. However,
the android.toolchain.cmake file does not know how to parse the
Android native API level from that version of the NDK. So check
out the NDK r13b version that we were using until yesterday.
Fixes#1439
This is totally optional, but lets people build this repo with
ccache to improve rebuild speeds. It also can help a great deal
on ccache-enabled CI systems like Travis-CI. We build fixed
revisions of glslang a lot on Travis, so this will be a big help
with CI machine loading.
When constructing a matrix from another matrix with smaller dimensions,
there's no need to extract the scalars out of columns and rebuild the
resulting matrix from scalars - instead, we can just construct shorter
vectors with OpShuffle and combine them to the final result.
This keeps the common casts such as mat3(mat4) in vector registers,
which may improve performance for some GPUs, and cleans up output of
translation tools like SPIRV-Cross.
Fixes#1412.
These introduce limited support for 8/16-bit types such that they can only be accessed in buffer memory and converted to/from 32-bit types.
Contributed from Khronos-internal work.
