The big change is to delay address reg setup until the instruction
that needs the deref. It was hard to use the deref chain support for
the LHS because it does the copy of the dereffed value to a temporary
(to avoid problems when two src regs are array derefs), so we wouldn't
haev a pointer to actual storage in the end.
Fixes glsl-vs-arrays on swrast.
Find instructions in all shaders that are not contained in a function
(i.e., initializers for global variables). "Move" these instructions
to the top of the main function in the linked shader. As a
side-effect, many global variables will also be copied into the linked
shader.
The linker needs to use this function to get specific function signatures, but
it also needs to modify the returned signature. Since this method isn't itself
const (i.e., const this pointer), there is no value in making a const and
non-const version.
This currently involves an ugly hack so that every link doesn't result
in all the built-in functions showing up as multiply defined. As soon
as the built-in functions are stored in a separate compilation unit,
ir_function_signature::is_built_in can be removed.
The instruction can be hung off of any other in the tree, even if the
other one will be deleted, since it'll get stolen to the shader's
context later if it's still live.
We would clear the in_lhs flag early, avoiding copy propagation on the
array index variable (oops) and then copy propagating on the array
variable (ouch). Just avoid all copy propagation on the LHS instead.
Add two invariant checks related to functions and function signatures:
1. Ensure that function definitions (ir_function) are not nested.
2. Ensure that the ir_function pointed to by an ir_function_signature
is the one that contains it in its signatures list.
There is no setter function, the getter returns a constant pointer,
and ir_function_signature::_function is private for a reason. The
only way to make a connection between a function and function
signature is via ir_function::add_signature. This helps ensure that
certain invariants (i.e., a function signature is in the list of
signatures for its _function) are met.
Temporary variables added for &&, ||, and ?: were not being added to
the instruction stream. This resulted in either test failures or
Valgrind being angry after the original IR tree was destroyed by
talloc_free. The talloc_free caused the ir_variables to be destroyed
even though they were still referenced.
This will be used by the Mesa IR and likely most HW backends, as it
allows other optimizations to occur that might not otherwise.
Fixes glsl-vs-mat-sub-1, glsl-vs-mat-div-1.
This feels a little odd, but it will be useful for ir_mat_to_vec,
where I want to see a plain assignment of the expression to a
variable, not to a writemasked array dereference with a call as the
array index.
It would be easy to miss an entry either of the two visitors involved
that would result in trying to ir->remove() the call to remove it from
the instruction stream when really it's part of an expression tree
that wasn't flattened.
Previously, code like ivec4(mat2(...)) would fail because the compiler
would naively try to convert a mat2 to an imat2...which doesn't exist.
Now, a separate pass breaks such matrices down to their columns, which
can be converted from vec2 to ivec2.
Fixes piglit tests constructor-11.vert, constructor-14.vert,
constructor-15.vert, and CorrectConstFolding2.frag.
In particular, with foreach_list_safe, one can remove and free the current
node without crashes; if new nodes are added after the current node,
they will be properly visited as well.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Most places in the code simply use a static name, which works because
names are never used to look up an ir_variable. generate_temporary is
simply unnecessary (and looks like it would leak memory, and isn't
thread safe...)
Previously, a register defined at main scope and used in a loop in a
loop could end up getting marked as needed only from the definition
outside of the loops to the end of the inner loop, and we would
cleverly slot in something else in its register in the end of the
outer loop.
Fixes glsl-vs-loop-nested and glsl-fs-loop-nested on glsl2. This
doesn't happen much on master because the original compiler does its
own register allocation, so we find little we can do with linear scan
register (re)allocation.
