This was detected by valgrind. I think GCC still does the right
thing, but the C++ spec allows the compiler to do something
stupid... like crash or only delete the first entry in the array.
Every platform that supports GLSL sets GL_MAX_TEXTURE_COORDS to at
least 4, so hard-code 4 for now.
This causes the following tests to pass:
glslparsertest/glsl2/norsetto-bumptbn_sh_fp.vert
glslparsertest/glsl2/xreal-lighting-d-omni.vert
glslparsertest/glsl2/xreal-lighting-db-omni.vert
glslparsertest/glsl2/xreal-lighting-dbs-omni.vert
This causes the following tests to pass:
glslparsertest/glsl2/precision-02.vert
glslparsertest/glsl2/precision-04.vert
glslparsertest/glsl2/precision-06.vert
This causes the following test to fail. This shader was previously
failing to compile, but it was failing for the wrong reasons.
glslparsertest/glsl2/precision-03.vert
Some valid shaders, such as 'precision highp float;', evaluate to
empty sets of instructions. This causes some of the optimization
stages to enter infinite loops. Instead, don't bother processing the
empty ones.
ir_dereference_array::array is always an r-value. If the dereference
is of a varaible, that r-value will be an ir_dereference_variable.
This simplifies the code a bit.
We were nicely constructing a new expression for the implicit type
conversion, but then checking that the previous types matched instead
of the new expression's type. Fixes errors in Regnum Online shaders.
This removes a bunch of gratuitous moving around of constant values
from constructors. Makes a shader ir I was looking at for structure
handling almost readable.
Similar to other situations where the visitor pattern doesn't fit, in
this case we need the pointer to the base instruction in the
instruction stream for where to insert any new instructions we
generate (not the instruction in the tree we're looking at). By
removing the code for setting the base_ir, flattened expressions would
end up, for example, before the function definition where they had appeared.
Previously, the syntax was (array_ref <variable name> <index>), but the
subject is now a general rvalue (not a name). In particular, it might
be a (var_ref ...).
Also, remove "expected ... or (swiz)" from error messages; swiz is not
allowed inside a var_ref.
Array dereferences now point to variable dereferences instead of
pointing directly to variables. This necessitated some changes to the
way the variable is accessed when setting the maximum index array element.
Kenneth Graunke