I'd missed putting in the actual "find structures to split" part, so
most of the code didn't do anything. I was running on too large of an
app and assuming the lack of progress was elsewhere.
ir_structure_splitting: New pass to chop structures into their components.
This doesn't do anything if your structure goes through an uninlined
function call or if whole-structure assignment occurs. As such, the
impact is limited, at least until we do some global copy propagation
to reduce whole-structure assignment.
glsl2: Fix the dereferences_variable check in ir_tree_grafting.
The HV doesn't descend into ir_variable, which is generally a good
thing (allowing one to distinguish between variable declarations and
refs), but here we never saw tree grafting opportunities killed
because we were looking for the ir_variable child of a dereference to
get visited.
Fixes:
glsl1-function call with inout params
ir_to_mesa: Don't do function inlining until linking.
Optimizations at compile time should generally be done with the goal
of reducing instruction count so that other work, particularly
linking, is less time-consuming if the shader is used multiple times.
However, function inlining increases instruction count for the inlined
function bodies without removing the original function body, since we
don't know if it will be used at link time or not.
Reduces the runtime of linking and executing a Yo Frankie fragment
shader from 0.9 seconds to 0.5 seconds (-45.9%, +/- 2.2%, n=5).
For a shader involving many small functions, this avoids running
optimization across all of them after they've been inlined
post-linking.
Reduces the runtime of linking and running a fragment shader from Yo
Frankie from 1.6 seconds to 0.9 seconds (-44.9%, +/- 3.3%).
glsl2: Catch pointless copies in copy propagation.
We wouldn't want to go rewriting dereferences to variables to point at
the same variable it did before. While I didn't find a way to trigger
that, a shader in Yo Frankie managed to produce a self-assignment by
passing a constant to a function doing self assignment like this.
Cleans up the IR for glsl-deadcode-self-assign.shader_test
glsl2: Remove the shader_in/shader_out tracking separate from var->mode.
I introduced this for ir_dead_code to distinguish function parameter
outvals from varying outputs. Only, since ast_to_hir's
current_function is unset when setting up function parameters (they're
needed for making the function signature in the first place), all
function parameter outvals were marked as shader outputs anyway. This
meant that an inlined function's cloned outval was marked as a shader
output and couldn't be dead-code eliminated. Instead, since
ir_dead_code doesn't even look at function parameters, just use
var->mode.
The longest Mesa IR coming out of ir_to_mesa for Yo Frankie drops from
725 instructions to 636.
i965: Settle on printing our program debug to stdout.
Mixing stderr (_mesa_print_program, _mesa_print_instruction,
_mesa_print_alu) with stdout means that when writing both to a file,
there isn't a consistent ordering between the two.
ir_to_mesa: Print shader source and compiled IR under MESA_GLSL=dump.
While the Mesa IR dumping includes some corresponding GLSL IR for
correlating Mesa IR to GLSL IR, it doesn't completely express it.
This printing includes things like variable declarations and control
flow structure that is hard to read otherwise.
glsl2: Generate masked assignments in vector and matrix constructors
Previously the in-line matrix and vector constructors would generate
swizzles in the LHS. The code is actually more clear if it just
generates the masked assignments instead of relying on the
ir_assignment constructor to convert the swizzles to write masks.
glsl2: Add ir_assignment::write_mask and associated methods
Replace swizzles on the LHS with additional swizzles on the RHS and a
write mask in the assignment instruction. As part of this add
ir_assignment::set_lhs. Ideally we'd make ir_assignment::lhs private
to prevent erroneous writes, but that would require a lot of code
butchery at this point.
Add ir_assignment constructor that takes an explicit write mask. This
is required for ir_assignment::clone, but it can also be used in other
places. Without this, ir_assignment clones lose their write masks,
and incorrect IR is generated in optimization passes.
Add ir_assignment::whole_variable_written method. This method gets
the variable on the LHS if the whole variable is written or NULL
otherwise. This is different from
ir->lhs->whole_variable_referenced() because the latter has no
knowledge of the write mask stored in the ir_assignment.
Gut all code from ir_to_mesa that handled swizzles on the LHS of
assignments. There is probably some other refactoring that could be
done here, but that can be left for another day.
glcpp: Remove xtalloc wrappers in favor of plain talloc.
Calling exit() on a memory failure probably made sense for the
standalone preprocessor, but doesn't seem too appealing as part of
the GL library. Also, we don't use it in the main compiler.
glsl2: Don't try to assign locations for samplers during linking.
Mesa will do the mapping at _mesa_add_sampler() time. Fixes assertion
failures in debug builds, which might have caught real problems with
multiple samplers linked in a row.
ir_to_mesa: Clean up the mapping of samplers to Mesa's sampler uniforms.
Instead of using a linker-assigned location (since samplers don't
actually take up uniform space, being a link-time choice), use the
sampler's varaible pointer as a hash key.
mesa: Don't null deref looking for Mesa IR code at compile time.
The new compiler doesn't generate Mesa IR at compile time, and that
compile time code previously wouldn't have reflected the link time
code that actually got used. But do dump the info log of the compile
regardless.
glsl2: Make the clone() method take a talloc context.
In most cases, we needed to be reparenting the cloned IR to a
different context (for example, to the linked shader instead of the
unlinked shader), or optimization before the reparent would cause
memory usage of the original object to grow and grow.
glsl2: Use talloc_zero_size instead of talloc_size to allocate ast_node objects.
This is a zero-ing function, (like calloc), to avoid bugs due to
accessing uninitialized values. Thanks to valgrind for noticing the
use of uninitialized values.