[PURPOSE]:
The current process design for Uniform / Block / Pipe IO symbols reflection (during program linking) is as following :
1.1 using a global mapper called 'TNameToIndex' to store all the relationship of name (of symbols) to their indexes (in their own MapIndexToReflection vectors).
1.2 TNameToIndex mapper will be used during program linking and helps to check and merge duplicate symbols within each stage ( Uniform, Block and Pipe IO)
1.3 Different types of symbols will have their own index mapping storage. All those symbols will share TNameToIndex as a general searching mapper.
1.4 Only IN in first stage and OUT in last stage will be dealed within traversing functions.
Now, here we meet those problems:
2.1 In and Out variables for pipelines are mapping to different MapIndexToReflection vector (ioItems), but they may still have same names within the general symbol search mapper : TNameToIndex.
2.2 Then, when there are same symbols of IN in VS and OUT in FS, TNameToIndex could not tell the difference because it only stores one local index for one symbol (1:1) as a pair of KeyValue.
[What fixed]:
Seperate I/O from other symbols like Uniform and Block (it is wrong to keep them all in TNameToIndex), and save in new searching mappers called pipeInNameToIndex and pipeOutNameToIndex.
Expose new top-level functions defined as getReflectionPipeIOIndex and getPipeIOIndex for users who need to query Pipe I/O information (As they may reach those things through getUniformIndex and getReflectionIndex now, which is a confused way.)
As there are 2 mappers for above symbols, users needs to input second argument when they wanna reach those pipe I/O parameters, that's also why we need to modify GET functions either.
[Test Case]:
The shader is as following:
######### VS ############
layout(location = 0) in vec4 g_position;
layout(location = 1) in vec4 g_color;
out StageData {
vec4 color;
} g_vs_out;
void main() {
gl_Position = g_position;
g_vs_out.color = g_color;
}
########### FS #############
in StageData {
vec4 color;
} g_fs_in;
layout(location = 0) out vec4 g_color;
void main() {
g_color = g_fs_in.color;
}
Purpose :
According to GLSL SPEC 4.6 ( 4.4.1.4 Compute Shader Inputs), for compute shader input qualifiers, we should declare such qualifiers with same values in the same shader (local_size_x, y and z).
"If such a layout qualifier is declared more than once in the same shader, all those declarations must set the same set of local work-group sizes and set them to the same values; otherwise a compile-time error results."
Why this fix:
If we manually set "local_size_x = 1" and directly following a declaration like "local_size_x = 2", this would not be detected. That is because currently we treat all the '1' as default value and could not restrictly detect whether those are default values.
Test case:
......
layout(local_size_x=1) in;
layout(local_size_x=2) in;
......
So I add test cases for this fix:
1. set local_size_y = 1 => success
2. set local_size_y = 2 => error
3. set local_size_y = 1 => success
According to the 32-bit counterparts, their forms should be
genI64Type findLSB(genI64Type value)
genI64Type findLSB(genU64Type value)
genI64Type findMSB(genI64Type value)
genI64Type findMSB(genU64Type value)
The order of error checking was not quite being correct (maybe there is no correct
ordering, when many checks must be done and they affect each other).
So, check for block-name reuse twice.
glslang/include/intermediate.h -> Add a new interface to set TIntermBranch's expression.
glslang/include/Types.h -> Add interface to set Type's basicType and add interface to get basicType form a TSampler.
glslang/MachineIndependent/intermediate.cpp -> Part of the code in createConversion been encapsulating as a new function called buildConvertOp
glslang/MachineIndependent/localintermediate.h -> Export createConversion and
buildConvertOp as a public function
glslang/Public/ShaderLang.h -> Add interface to get shader object and shader source.
Also fixes, in practice, https://github.com/KhronosGroup/GLSL/issues/83.
When the specification language is correctly created, glslang can be
revisited for correctness. In the meantime, this seems like the best
"bug" to have relative to the specification.
Memory qualifiers are only relevant to parameters when they apply
to what the argument points to, as otherwise the argument is copied.
This leaves the fix from #1870 in place, and then more correctly
ignores memory qualifiers when something will be passed by copy.
In the current version of the code on non-debug builds these cases
will fallthrough, since assert is a no-op, and eventually make a call
passing in |op| which hasn't been initialized. clang is currently
throwing a warning about this behaviour when integrating downstream.
This patch changes the behaviour, so that in any branch that has an
assert now has a return nullptr, to indicate failure after it and
avoid the uninitialized variable usage.
If we don't do this then we get reflection output like so:
ArrayedBind[0].a.a: offset 0, type 1406, size 1, index 4, binding -1, stages 0
ArrayedBind[0].a.b: offset 4, type 1406, size 1, index 4, binding -1, stages 0
ArrayedBind[0].b.a: offset 4, type 1406, size 1, index 4, binding -1, stages 0
ArrayedBind[0].b.b: offset 8, type 1406, size 1, index 4, binding -1, stages 0
ArrayedBind[0].b: offset 4, type 1406, size 1, index 4, binding -1, stages 1
When the outer reflection loop that calls blowUpActiveAggregate incorrectly iterates over the struct members.
Saved about 21K, size down to 380K of MSVC x86 code.
Fixed one bug that needs to be looked at on the master branch:
The test for needing a Vulkan binding has a bug in it, "!layoutAttachment"
which does not mean "no layoutAttachment", because that is non-zero.
This is why some test and test results changed.
About 60K smaller (20% the target size is these units of 300K).
Over 300 fewer #ifdef.
This adds a new glslang.m4 file that needs to be processed by m4 to get
the glslang.y file needed by bison. See comment in glslang.m4 for more
detail.
This updates the updateGrammar script to do the .m4 -> .y processing,
to conditionally exclude grammar not needed for the web build.
About 50 fewer #ifdefs.
About 14K smaller.
Note, the base size is ill-defined due to optimizer settings (size vs. performance),
compression, and target architecture. Some recent %'s are accidentally reported as
3X the real savings. Early %'s were accurate. What matters though is that each
step got worthwhile gains, and what the final size ends up being.
Focus was on the front end (not SPIR-V), minus the grammar.
Reduces #ifdef count by around 320 and makes the web build 270K smaller,
which is about 90% the target size.
The grammar and scanner will be another step, as will the SPIR-V backend.
This makes heavy use of methods #ifdef'd to return false as a global way
of turning off code, relying on C++ DCE to do the rest.
Save about 100K.
N.B.: This is done by eliminating a function call, at a high level,
not by #ifdef'ing a bunch of code.
Also, removed no longer needed *_EXTENSION #ifdef in the code not
needed by GLSLANG_WEB.
