Re: [Mesa-dev] [PATCH 05/12] glsl: Combine nop-swizzle optimization with swizzle-swizzle optimization

[Thomas Helland]

Changes all look really good, and some performance numbers
I have for a similar patch I've written shows a marginal benefit
in compiler runtime performance (perf-stat -> cycles executed)
For patches 4 and 5.

Reviewed-by: 

2. nov. 2017 21.26 skrev "Ian Romanick" :

From: Ian Romanick 

Signed-off-by: Ian Romanick 
---
 src/compiler/Makefile.sources  |  3 +-
 src/compiler/glsl/glsl_parser_extras.cpp   |  3 +-
 src/compiler/glsl/ir_optimization.h|  3 +-
 src/compiler/glsl/meson.build  |  3 +-
 .../glsl/{opt_noop_swizzle.cpp => opt_swizzle.cpp} | 56 ++---
 src/compiler/glsl/opt_swizzle_swizzle.cpp  | 96
--
 src/compiler/glsl/test_optpass.cpp |  6 +-
 7 files changed, 52 insertions(+), 118 deletions(-)
 rename src/compiler/glsl/{opt_noop_swizzle.cpp => opt_swizzle.cpp} (56%)
 delete mode 100644 src/compiler/glsl/opt_swizzle_swizzle.cpp

diff --git a/src/compiler/Makefile.sources b/src/compiler/Makefile.sources
index b80468c..60814e2 100644
--- a/src/compiler/Makefile.sources
+++ b/src/compiler/Makefile.sources
@@ -130,11 +130,10 @@ LIBGLSL_FILES = \
glsl/opt_function_inlining.cpp \
glsl/opt_if_simplification.cpp \
glsl/opt_minmax.cpp \
-   glsl/opt_noop_swizzle.cpp \
glsl/opt_rebalance_tree.cpp \
glsl/opt_redundant_jumps.cpp \
glsl/opt_structure_splitting.cpp \
-   glsl/opt_swizzle_swizzle.cpp \
+   glsl/opt_swizzle.cpp \
glsl/opt_tree_grafting.cpp \
glsl/opt_vectorize.cpp \
glsl/program.h \
diff --git a/src/compiler/glsl/glsl_parser_extras.cpp
b/src/compiler/glsl/glsl_parser_extras.cpp
index 822301a..5982173 100644
--- a/src/compiler/glsl/glsl_parser_extras.cpp
+++ b/src/compiler/glsl/glsl_parser_extras.cpp
@@ -2226,8 +2226,7 @@ do_common_optimization(exec_list *ir, bool linked,
options->EmitNoCont, options->EmitNoLoops);
OPT(do_vec_index_to_swizzle, ir);
OPT(lower_vector_insert, ir, false);
-   OPT(do_swizzle_swizzle, ir);
-   OPT(do_noop_swizzle, ir);
+   OPT(optimize_swizzles, ir);

OPT(optimize_split_arrays, ir, linked);
OPT(optimize_redundant_jumps, ir);
diff --git a/src/compiler/glsl/ir_optimization.h b/src/compiler/glsl/ir_
optimization.h
index f44ddcb..2b8c195 100644
--- a/src/compiler/glsl/ir_optimization.h
+++ b/src/compiler/glsl/ir_optimization.h
@@ -123,9 +123,8 @@ bool lower_if_to_cond_assign(gl_shader_stage stage,
exec_list *instructions,
  unsigned max_depth = 0, unsigned
min_branch_cost = 0);
 bool do_mat_op_to_vec(exec_list *instructions);
 bool do_minmax_prune(exec_list *instructions);
-bool do_noop_swizzle(exec_list *instructions);
 bool do_structure_splitting(exec_list *instructions);
-bool do_swizzle_swizzle(exec_list *instructions);
+bool optimize_swizzles(exec_list *instructions);
 bool do_vectorize(exec_list *instructions);
 bool do_tree_grafting(exec_list *instructions);
 bool do_vec_index_to_cond_assign(exec_list *instructions);
diff --git a/src/compiler/glsl/meson.build b/src/compiler/glsl/meson.build
index 76fcafb..6284d0c 100644
--- a/src/compiler/glsl/meson.build
+++ b/src/compiler/glsl/meson.build
@@ -169,11 +169,10 @@ files_libglsl = files(
   'opt_function_inlining.cpp',
   'opt_if_simplification.cpp',
   'opt_minmax.cpp',
-  'opt_noop_swizzle.cpp',
   'opt_rebalance_tree.cpp',
   'opt_redundant_jumps.cpp',
   'opt_structure_splitting.cpp',
-  'opt_swizzle_swizzle.cpp',
+  'opt_swizzle.cpp',
   'opt_tree_grafting.cpp',
   'opt_vectorize.cpp',
   'program.h',
diff --git a/src/compiler/glsl/opt_noop_swizzle.cpp b/src/compiler/glsl/opt_
swizzle.cpp
similarity index 56%
rename from src/compiler/glsl/opt_noop_swizzle.cpp
rename to src/compiler/glsl/opt_swizzle.cpp
index 41890ab..2fbe362 100644
--- a/src/compiler/glsl/opt_noop_swizzle.cpp
+++ b/src/compiler/glsl/opt_swizzle.cpp
@@ -22,11 +22,14 @@
  */

 /**
- * \file opt_noop_swizzle.cpp
+ * \file opt_swizzle.cpp
+ * Optimize swizzle operations.
  *
- * If a swizzle doesn't change the order or count of components, then
- * remove the swizzle so that other optimization passes see the value
- * behind it.
+ * First, compact a sequence of swizzled swizzles into a single swizzle.
+ *
+ * If the final resulting swizzle doesn't change the order or count of
+ * components, then remove the swizzle so that other optimization passes
see
+ * the value behind it.
  */

 #include "ir.h"
@@ -36,9 +39,9 @@

 namespace {

-class ir_noop_swizzle_visitor : public ir_rvalue_visitor {
+class ir_opt_swizzle_visitor : public ir_rvalue_visitor {
 public:
-   ir_noop_swizzle_visitor()
+   ir_opt_swizzle_visitor()
{
   this->progress = false;
}
@@ -50,13 +53,46 @@ public:
 } /* unnamed namespace */

 void
-ir_noop_swizzle_visitor::handle_rvalue(ir_rvalue **rvalue)

[Mesa-dev] [PATCH 05/12] glsl: Combine nop-swizzle optimization with swizzle-swizzle optimization

[Ian Romanick]

From: Ian Romanick 

Signed-off-by: Ian Romanick 
---
 src/compiler/Makefile.sources  |  3 +-
 src/compiler/glsl/glsl_parser_extras.cpp   |  3 +-
 src/compiler/glsl/ir_optimization.h|  3 +-
 src/compiler/glsl/meson.build  |  3 +-
 .../glsl/{opt_noop_swizzle.cpp => opt_swizzle.cpp} | 56 ++---
 src/compiler/glsl/opt_swizzle_swizzle.cpp  | 96 --
 src/compiler/glsl/test_optpass.cpp |  6 +-
 7 files changed, 52 insertions(+), 118 deletions(-)
 rename src/compiler/glsl/{opt_noop_swizzle.cpp => opt_swizzle.cpp} (56%)
 delete mode 100644 src/compiler/glsl/opt_swizzle_swizzle.cpp

diff --git a/src/compiler/Makefile.sources b/src/compiler/Makefile.sources
index b80468c..60814e2 100644
--- a/src/compiler/Makefile.sources
+++ b/src/compiler/Makefile.sources
@@ -130,11 +130,10 @@ LIBGLSL_FILES = \
glsl/opt_function_inlining.cpp \
glsl/opt_if_simplification.cpp \
glsl/opt_minmax.cpp \
-   glsl/opt_noop_swizzle.cpp \
glsl/opt_rebalance_tree.cpp \
glsl/opt_redundant_jumps.cpp \
glsl/opt_structure_splitting.cpp \
-   glsl/opt_swizzle_swizzle.cpp \
+   glsl/opt_swizzle.cpp \
glsl/opt_tree_grafting.cpp \
glsl/opt_vectorize.cpp \
glsl/program.h \
diff --git a/src/compiler/glsl/glsl_parser_extras.cpp 
b/src/compiler/glsl/glsl_parser_extras.cpp
index 822301a..5982173 100644
--- a/src/compiler/glsl/glsl_parser_extras.cpp
+++ b/src/compiler/glsl/glsl_parser_extras.cpp
@@ -2226,8 +2226,7 @@ do_common_optimization(exec_list *ir, bool linked,
options->EmitNoCont, options->EmitNoLoops);
OPT(do_vec_index_to_swizzle, ir);
OPT(lower_vector_insert, ir, false);
-   OPT(do_swizzle_swizzle, ir);
-   OPT(do_noop_swizzle, ir);
+   OPT(optimize_swizzles, ir);
 
OPT(optimize_split_arrays, ir, linked);
OPT(optimize_redundant_jumps, ir);
diff --git a/src/compiler/glsl/ir_optimization.h 
b/src/compiler/glsl/ir_optimization.h
index f44ddcb..2b8c195 100644
--- a/src/compiler/glsl/ir_optimization.h
+++ b/src/compiler/glsl/ir_optimization.h
@@ -123,9 +123,8 @@ bool lower_if_to_cond_assign(gl_shader_stage stage, 
exec_list *instructions,
  unsigned max_depth = 0, unsigned min_branch_cost 
= 0);
 bool do_mat_op_to_vec(exec_list *instructions);
 bool do_minmax_prune(exec_list *instructions);
-bool do_noop_swizzle(exec_list *instructions);
 bool do_structure_splitting(exec_list *instructions);
-bool do_swizzle_swizzle(exec_list *instructions);
+bool optimize_swizzles(exec_list *instructions);
 bool do_vectorize(exec_list *instructions);
 bool do_tree_grafting(exec_list *instructions);
 bool do_vec_index_to_cond_assign(exec_list *instructions);
diff --git a/src/compiler/glsl/meson.build b/src/compiler/glsl/meson.build
index 76fcafb..6284d0c 100644
--- a/src/compiler/glsl/meson.build
+++ b/src/compiler/glsl/meson.build
@@ -169,11 +169,10 @@ files_libglsl = files(
   'opt_function_inlining.cpp',
   'opt_if_simplification.cpp',
   'opt_minmax.cpp',
-  'opt_noop_swizzle.cpp',
   'opt_rebalance_tree.cpp',
   'opt_redundant_jumps.cpp',
   'opt_structure_splitting.cpp',
-  'opt_swizzle_swizzle.cpp',
+  'opt_swizzle.cpp',
   'opt_tree_grafting.cpp',
   'opt_vectorize.cpp',
   'program.h',
diff --git a/src/compiler/glsl/opt_noop_swizzle.cpp 
b/src/compiler/glsl/opt_swizzle.cpp
similarity index 56%
rename from src/compiler/glsl/opt_noop_swizzle.cpp
rename to src/compiler/glsl/opt_swizzle.cpp
index 41890ab..2fbe362 100644
--- a/src/compiler/glsl/opt_noop_swizzle.cpp
+++ b/src/compiler/glsl/opt_swizzle.cpp
@@ -22,11 +22,14 @@
  */
 
 /**
- * \file opt_noop_swizzle.cpp
+ * \file opt_swizzle.cpp
+ * Optimize swizzle operations.
  *
- * If a swizzle doesn't change the order or count of components, then
- * remove the swizzle so that other optimization passes see the value
- * behind it.
+ * First, compact a sequence of swizzled swizzles into a single swizzle.
+ *
+ * If the final resulting swizzle doesn't change the order or count of
+ * components, then remove the swizzle so that other optimization passes see
+ * the value behind it.
  */
 
 #include "ir.h"
@@ -36,9 +39,9 @@
 
 namespace {
 
-class ir_noop_swizzle_visitor : public ir_rvalue_visitor {
+class ir_opt_swizzle_visitor : public ir_rvalue_visitor {
 public:
-   ir_noop_swizzle_visitor()
+   ir_opt_swizzle_visitor()
{
   this->progress = false;
}
@@ -50,13 +53,46 @@ public:
 } /* unnamed namespace */
 
 void
-ir_noop_swizzle_visitor::handle_rvalue(ir_rvalue **rvalue)
+ir_opt_swizzle_visitor::handle_rvalue(ir_rvalue **rvalue)
 {
if (!*rvalue)
   return;
 
ir_swizzle *swiz = (*rvalue)->as_swizzle();
-   if (!swiz || swiz->type != swiz->val->type)
+
+   if (!swiz)
+  return;
+
+   ir_swizzle *swiz2;
+
+   while ((swiz2 = swiz->val->as_swizzle()) != NULL) {
+  int mask2[4];
+
+