Mesa (master): glsl: Skip making a temporary for assignments when we don' t need one.
Eric Anholt
anholt at kemper.freedesktop.org
Tue Apr 8 08:24:35 UTC 2014
Module: Mesa
Branch: master
Commit: e9822f77a9cc024f528d30382fd5ad21c73a173b
URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=e9822f77a9cc024f528d30382fd5ad21c73a173b
Author: Eric Anholt <eric at anholt.net>
Date: Wed Mar 5 17:05:54 2014 -0800
glsl: Skip making a temporary for assignments when we don't need one.
While we wish our optimization passes could identify all the cases where
we can coalesce our variables, we miss out on a lot of opportunities.
total instructions in shared programs: 1673849 -> 1673166 (-0.04%)
instructions in affected programs: 299521 -> 298838 (-0.23%)
GAINED: 7
LOST: 0
Note that many programs are "hurt". The notable ones are where we produce
unrolling in cases we didn't before (presumably just because of the lower
instruction count). But there are also some cases where pushing things
right into the variables prevents copy propagation and tree grafting,
since we don't split our variable usage webs apart.
Reviewed-by: Kenneth Graunke <kenneth at whitecape.org>
---
src/glsl/ast.h | 13 ++++
src/glsl/ast_to_hir.cpp | 171 ++++++++++++++++++++++++++++++-----------------
2 files changed, 123 insertions(+), 61 deletions(-)
diff --git a/src/glsl/ast.h b/src/glsl/ast.h
index ae70b00..6b136f5 100644
--- a/src/glsl/ast.h
+++ b/src/glsl/ast.h
@@ -217,6 +217,13 @@ public:
virtual ir_rvalue *hir(exec_list *instructions,
struct _mesa_glsl_parse_state *state);
+ virtual void hir_no_rvalue(exec_list *instructions,
+ struct _mesa_glsl_parse_state *state);
+
+ ir_rvalue *do_hir(exec_list *instructions,
+ struct _mesa_glsl_parse_state *state,
+ bool needs_rvalue);
+
virtual void print(void) const;
enum ast_operators oper;
@@ -287,6 +294,9 @@ public:
virtual ir_rvalue *hir(exec_list *instructions,
struct _mesa_glsl_parse_state *state);
+ virtual void hir_no_rvalue(exec_list *instructions,
+ struct _mesa_glsl_parse_state *state);
+
private:
/**
* Is this function call actually a constructor?
@@ -359,6 +369,9 @@ public:
virtual ir_rvalue *hir(exec_list *instructions,
struct _mesa_glsl_parse_state *state);
+
+ virtual void hir_no_rvalue(exec_list *instructions,
+ struct _mesa_glsl_parse_state *state);
};
/**
diff --git a/src/glsl/ast_to_hir.cpp b/src/glsl/ast_to_hir.cpp
index 8f6e901..8d55ee3 100644
--- a/src/glsl/ast_to_hir.cpp
+++ b/src/glsl/ast_to_hir.cpp
@@ -56,6 +56,9 @@
#include "glsl_types.h"
#include "program/hash_table.h"
#include "ir.h"
+#include "ir_builder.h"
+
+using namespace ir_builder;
static void
detect_conflicting_assignments(struct _mesa_glsl_parse_state *state,
@@ -733,10 +736,12 @@ mark_whole_array_access(ir_rvalue *access)
}
}
-ir_rvalue *
+static bool
do_assignment(exec_list *instructions, struct _mesa_glsl_parse_state *state,
const char *non_lvalue_description,
- ir_rvalue *lhs, ir_rvalue *rhs, bool is_initializer,
+ ir_rvalue *lhs, ir_rvalue *rhs,
+ ir_rvalue **out_rvalue, bool needs_rvalue,
+ bool is_initializer,
YYLTYPE lhs_loc)
{
void *ctx = state;
@@ -855,27 +860,33 @@ do_assignment(exec_list *instructions, struct _mesa_glsl_parse_state *state,
* to handle things like:
*
* i = j += 1;
- *
- * So we always just store the computed value being assigned to a
- * temporary and return a deref of that temporary. If the rvalue
- * ends up not being used, the temp will get copy-propagated out.
*/
- ir_variable *var = new(ctx) ir_variable(rhs->type, "assignment_tmp",
- ir_var_temporary);
- ir_dereference_variable *deref_var = new(ctx) ir_dereference_variable(var);
- instructions->push_tail(var);
- instructions->push_tail(new(ctx) ir_assignment(deref_var, rhs));
- deref_var = new(ctx) ir_dereference_variable(var);
+ if (needs_rvalue) {
+ ir_variable *var = new(ctx) ir_variable(rhs->type, "assignment_tmp",
+ ir_var_temporary);
+ instructions->push_tail(var);
+ instructions->push_tail(assign(var, rhs));
+
+ if (!error_emitted) {
+ ir_dereference_variable *deref_var = new(ctx) ir_dereference_variable(var);
+ instructions->push_tail(new(ctx) ir_assignment(lhs, deref_var));
+ }
+ ir_rvalue *rvalue = new(ctx) ir_dereference_variable(var);
- if (!error_emitted)
- instructions->push_tail(new(ctx) ir_assignment(lhs, deref_var));
+ if (extract_channel) {
+ rvalue = new(ctx) ir_expression(ir_binop_vector_extract,
+ rvalue,
+ extract_channel->clone(ctx, NULL));
+ }
- if (extract_channel) {
- return new(ctx) ir_expression(ir_binop_vector_extract,
- new(ctx) ir_dereference_variable(var),
- extract_channel->clone(ctx, NULL));
+ *out_rvalue = rvalue;
+ } else {
+ if (!error_emitted)
+ instructions->push_tail(new(ctx) ir_assignment(lhs, rhs));
+ *out_rvalue = NULL;
}
- return new(ctx) ir_dereference_variable(var);
+
+ return error_emitted;
}
static ir_rvalue *
@@ -906,6 +917,20 @@ ast_node::hir(exec_list *instructions,
return NULL;
}
+void
+ast_function_expression::hir_no_rvalue(exec_list *instructions,
+ struct _mesa_glsl_parse_state *state)
+{
+ (void)hir(instructions, state);
+}
+
+void
+ast_aggregate_initializer::hir_no_rvalue(exec_list *instructions,
+ struct _mesa_glsl_parse_state *state)
+{
+ (void)hir(instructions, state);
+}
+
static ir_rvalue *
do_comparison(void *mem_ctx, int operation, ir_rvalue *op0, ir_rvalue *op1)
{
@@ -1075,6 +1100,21 @@ ir_rvalue *
ast_expression::hir(exec_list *instructions,
struct _mesa_glsl_parse_state *state)
{
+ return do_hir(instructions, state, true);
+}
+
+void
+ast_expression::hir_no_rvalue(exec_list *instructions,
+ struct _mesa_glsl_parse_state *state)
+{
+ do_hir(instructions, state, false);
+}
+
+ir_rvalue *
+ast_expression::do_hir(exec_list *instructions,
+ struct _mesa_glsl_parse_state *state,
+ bool needs_rvalue)
+{
void *ctx = state;
static const int operations[AST_NUM_OPERATORS] = {
-1, /* ast_assign doesn't convert to ir_expression. */
@@ -1148,11 +1188,11 @@ ast_expression::hir(exec_list *instructions,
op[0] = this->subexpressions[0]->hir(instructions, state);
op[1] = this->subexpressions[1]->hir(instructions, state);
- result = do_assignment(instructions, state,
- this->subexpressions[0]->non_lvalue_description,
- op[0], op[1], false,
- this->subexpressions[0]->get_location());
- error_emitted = result->type->is_error();
+ error_emitted =
+ do_assignment(instructions, state,
+ this->subexpressions[0]->non_lvalue_description,
+ op[0], op[1], &result, needs_rvalue, false,
+ this->subexpressions[0]->get_location());
break;
}
@@ -1418,11 +1458,12 @@ ast_expression::hir(exec_list *instructions,
ir_rvalue *temp_rhs = new(ctx) ir_expression(operations[this->oper], type,
op[0], op[1]);
- result = do_assignment(instructions, state,
- this->subexpressions[0]->non_lvalue_description,
- op[0]->clone(ctx, NULL), temp_rhs, false,
- this->subexpressions[0]->get_location());
- error_emitted = (op[0]->type->is_error());
+ error_emitted =
+ do_assignment(instructions, state,
+ this->subexpressions[0]->non_lvalue_description,
+ op[0]->clone(ctx, NULL), temp_rhs,
+ &result, needs_rvalue, false,
+ this->subexpressions[0]->get_location());
/* GLSL 1.10 does not allow array assignment. However, we don't have to
* explicitly test for this because none of the binary expression
@@ -1444,11 +1485,12 @@ ast_expression::hir(exec_list *instructions,
temp_rhs = new(ctx) ir_expression(operations[this->oper], type,
op[0], op[1]);
- result = do_assignment(instructions, state,
- this->subexpressions[0]->non_lvalue_description,
- op[0]->clone(ctx, NULL), temp_rhs, false,
- this->subexpressions[0]->get_location());
- error_emitted = type->is_error();
+ error_emitted =
+ do_assignment(instructions, state,
+ this->subexpressions[0]->non_lvalue_description,
+ op[0]->clone(ctx, NULL), temp_rhs,
+ &result, needs_rvalue, false,
+ this->subexpressions[0]->get_location());
break;
}
@@ -1460,11 +1502,12 @@ ast_expression::hir(exec_list *instructions,
&loc);
ir_rvalue *temp_rhs = new(ctx) ir_expression(operations[this->oper],
type, op[0], op[1]);
- result = do_assignment(instructions, state,
- this->subexpressions[0]->non_lvalue_description,
- op[0]->clone(ctx, NULL), temp_rhs, false,
- this->subexpressions[0]->get_location());
- error_emitted = op[0]->type->is_error() || op[1]->type->is_error();
+ error_emitted =
+ do_assignment(instructions, state,
+ this->subexpressions[0]->non_lvalue_description,
+ op[0]->clone(ctx, NULL), temp_rhs,
+ &result, needs_rvalue, false,
+ this->subexpressions[0]->get_location());
break;
}
@@ -1477,11 +1520,12 @@ ast_expression::hir(exec_list *instructions,
state, &loc);
ir_rvalue *temp_rhs = new(ctx) ir_expression(operations[this->oper],
type, op[0], op[1]);
- result = do_assignment(instructions, state,
- this->subexpressions[0]->non_lvalue_description,
- op[0]->clone(ctx, NULL), temp_rhs, false,
- this->subexpressions[0]->get_location());
- error_emitted = op[0]->type->is_error() || op[1]->type->is_error();
+ error_emitted =
+ do_assignment(instructions, state,
+ this->subexpressions[0]->non_lvalue_description,
+ op[0]->clone(ctx, NULL), temp_rhs,
+ &result, needs_rvalue, false,
+ this->subexpressions[0]->get_location());
break;
}
@@ -1589,11 +1633,12 @@ ast_expression::hir(exec_list *instructions,
temp_rhs = new(ctx) ir_expression(operations[this->oper], type,
op[0], op[1]);
- result = do_assignment(instructions, state,
- this->subexpressions[0]->non_lvalue_description,
- op[0]->clone(ctx, NULL), temp_rhs, false,
- this->subexpressions[0]->get_location());
- error_emitted = op[0]->type->is_error();
+ error_emitted =
+ do_assignment(instructions, state,
+ this->subexpressions[0]->non_lvalue_description,
+ op[0]->clone(ctx, NULL), temp_rhs,
+ &result, needs_rvalue, false,
+ this->subexpressions[0]->get_location());
break;
}
@@ -1617,12 +1662,14 @@ ast_expression::hir(exec_list *instructions,
*/
result = get_lvalue_copy(instructions, op[0]->clone(ctx, NULL));
- (void)do_assignment(instructions, state,
- this->subexpressions[0]->non_lvalue_description,
- op[0]->clone(ctx, NULL), temp_rhs, false,
- this->subexpressions[0]->get_location());
+ ir_rvalue *junk_rvalue;
+ error_emitted =
+ do_assignment(instructions, state,
+ this->subexpressions[0]->non_lvalue_description,
+ op[0]->clone(ctx, NULL), temp_rhs,
+ &junk_rvalue, false, false,
+ this->subexpressions[0]->get_location());
- error_emitted = op[0]->type->is_error();
break;
}
@@ -1744,9 +1791,9 @@ ast_expression::hir(exec_list *instructions,
}
}
type = NULL; /* use result->type, not type. */
- assert(result != NULL);
+ assert(result != NULL || !needs_rvalue);
- if (result->type->is_error() && !error_emitted)
+ if (result && result->type->is_error() && !error_emitted)
_mesa_glsl_error(& loc, state, "type mismatch");
return result;
@@ -1767,7 +1814,7 @@ ast_expression_statement::hir(exec_list *instructions,
* anything in that case.
*/
if (expression != NULL)
- expression->hir(instructions, state);
+ expression->hir_no_rvalue(instructions, state);
/* Statements do not have r-values.
*/
@@ -2831,10 +2878,12 @@ process_initializer(ir_variable *var, ast_declaration *decl,
*/
const glsl_type *initializer_type;
if (!type->qualifier.flags.q.uniform) {
- result = do_assignment(initializer_instructions, state,
- NULL,
- lhs, rhs, true,
- type->get_location());
+ do_assignment(initializer_instructions, state,
+ NULL,
+ lhs, rhs,
+ &result, true,
+ true,
+ type->get_location());
initializer_type = result->type;
} else
initializer_type = rhs->type;
More information about the mesa-commit
mailing list