[Mesa-dev] [PATCH 00/22] Send ir_assignment::condition to the Upside Down

[Ian Romanick]

From: Ian Romanick <ian.d.romanick at intel.com>

I have also pushed this series to

https://cgit.freedesktop.org/~idr/mesa/log/?h=remove-ir_assignment-condition

I think -201 lines speaks for itself, but...

The condition field pre-dates the ir_triop_csel operation by a few
years.  The idea was that any assignment could be made conditional so
that platforms that didn't have flow control could replace an
if-statement with a bunch of conditional assignments.  However, handling
this condition is really annoying in most optimization passes, so most
passes bail as soon as a condition is encountered.

Much later in development, ir_triop_csel was added.  For the most part,
a conditional assignment is equivalent to a conditional select where the
"false" value is the destination variable.

My main motivation for this series is that I didn't want to deal with
ir_assignment::condition in the SPIR-V generator work.  That I was able
to delete a pile of code is a happy bonus.

There are two odd bits that may warrant a bit closer examination:

- Patch 6 changes the code that is generated for non-constant vector
  indexing, and that results in some shader-db changes.  Most of the
  changes are small (+1 instruction in a program), but some are more
  substantial.

- Patch 18 moves the 'foo = foo;' optimization from one optimization
  pass to another.  If we just eliminate that optimization altogether,
  shader-db results, even on platforms that use NIR, are hurt quite
  substantially.  I have not investigated why NIR isn't picking up the
  slack here.

Ian Romanick (22):
  glsl: Fix coding standards issues in lower_if_to_cond_assign
  glsl: Lower ifs to conditional-select instead of conditional-assign
  glsl: Fix coding standards issues in lower_vec_index_to_cond_assign
  glsl: Return ir_variable from compare_index_block
  glsl: Convert lower_vec_index_to_cond_assign to using ir_builder
  glsl: Lower vector indexing to conditional-select instead of
    conditional-assign
  glsl: Fix coding standards issues in
    lower_variable_index_to_cond_assign
  glsl: Convert lower_variable_index_to_cond_assign to ir_builder
  glsl: Lower array indexing to conditional-select instead of
    conditional-assign
  glsl: Don't pass NULL to ir_assignment constructor when not necessary
  glsl/ast: Use ir_binop_equal instead of ir_binop_all_equal
  glsl/ast: Convert ast_case_label::hir to ir_builder
  glsl/ast: Explicitly track the set of case labels that occur after
    default
  glsl/ast: Generate a more compact expression to disable execution of
    default case
  glsl/ast: Use logical-or instead of conditional assignment to set
    fallthru_var
  glsl: Eliminate ir_builder assign overloads that have a condition
  glsl: Eliminate ir_assignment constructors that have a condition
  glsl: Move 'foo = foo;' optimization to opt_dead_code_local
  glsl: Kill ir_assignment::condition with fire
  glsl: Fix indentation left weird from the previous commit
  glsl: Remove spurious assertions
  glsl: Simplify ir_assignment::clone

 src/compiler/glsl/ast_function.cpp            |  37 ++-
 src/compiler/glsl/ast_to_hir.cpp              | 152 ++++++-----
 src/compiler/glsl/glsl_to_nir.cpp             |  16 +-
 src/compiler/glsl/ir.cpp                      |  18 +-
 src/compiler/glsl/ir.h                        |  11 +-
 src/compiler/glsl/ir_builder.cpp              |  15 +-
 src/compiler/glsl/ir_builder.h                |   2 -
 .../glsl/ir_builder_print_visitor.cpp         |   4 -
 src/compiler/glsl/ir_clone.cpp                |  14 +-
 src/compiler/glsl/ir_constant_expression.cpp  |  12 +-
 .../glsl/ir_expression_flattening.cpp         |   4 +-
 src/compiler/glsl/ir_hv_accept.cpp            |   3 -
 src/compiler/glsl/ir_optimization.h           |   8 +-
 src/compiler/glsl/ir_print_visitor.cpp        |   3 -
 src/compiler/glsl/ir_reader.cpp               |   3 +-
 src/compiler/glsl/ir_rvalue_visitor.cpp       |   2 -
 src/compiler/glsl/loop_analysis.cpp           |   3 +-
 src/compiler/glsl/loop_controls.cpp           |   2 +-
 src/compiler/glsl/lower_discard.cpp           |   4 +-
 src/compiler/glsl/lower_distance.cpp          |   2 +-
 src/compiler/glsl/lower_if_to_cond_assign.cpp |  91 +++----
 src/compiler/glsl/lower_instructions.cpp      |   4 +-
 src/compiler/glsl/lower_jumps.cpp             |  11 +-
 src/compiler/glsl/lower_mat_op_to_vec.cpp     |   2 +-
 .../glsl/lower_texture_projection.cpp         |   2 +-
 .../lower_variable_index_to_cond_assign.cpp   | 253 +++++++-----------
 .../glsl/lower_vec_index_to_cond_assign.cpp   | 105 ++++----
 src/compiler/glsl/lower_vector.cpp            |   4 +-
 src/compiler/glsl/opt_array_splitting.cpp     |  10 +-
 src/compiler/glsl/opt_constant_folding.cpp    |  17 --
 .../glsl/opt_constant_propagation.cpp         |   3 -
 src/compiler/glsl/opt_constant_variable.cpp   |   3 -
 src/compiler/glsl/opt_copy_propagation.cpp    |  24 +-
 .../glsl/opt_copy_propagation_elements.cpp    |   3 -
 .../glsl/opt_dead_builtin_varyings.cpp        |   1 -
 src/compiler/glsl/opt_dead_code_local.cpp     | 180 +++++++------
 src/compiler/glsl/opt_function_inlining.cpp   |  11 +-
 src/compiler/glsl/opt_structure_splitting.cpp |  11 +-
 src/compiler/glsl/opt_tree_grafting.cpp       |   3 +-
 src/compiler/glsl/opt_vectorize.cpp           |   3 +-
 src/mesa/program/ir_to_mesa.cpp               |  24 +-
 src/mesa/state_tracker/st_glsl_to_tgsi.cpp    |  17 +-
 42 files changed, 448 insertions(+), 649 deletions(-)

-- 
2.9.5