Files
third_party_mesa3d/src/compiler/glsl/opt_copy_propagation.cpp

368 lines
10 KiB
C++
Raw Normal View History

/*
* Copyright © 2010 Intel Corporation
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
* DEALINGS IN THE SOFTWARE.
*/
/**
* \file opt_copy_propagation.cpp
*
* Moves usage of recently-copied variables to the previous copy of
* the variable.
*
* This should reduce the number of MOV instructions in the generated
* programs unless copy propagation is also done on the LIR, and may
* help anyway by triggering other optimizations that live in the HIR.
*/
#include "ir.h"
#include "ir_visitor.h"
#include "ir_basic_block.h"
#include "ir_optimization.h"
#include "compiler/glsl_types.h"
#include "util/hash_table.h"
namespace {
class kill_entry : public exec_node
{
public:
/* override operator new from exec_node */
DECLARE_LINEAR_ZALLOC_CXX_OPERATORS(kill_entry)
kill_entry(ir_variable *var)
{
assert(var);
this->var = var;
}
ir_variable *var;
};
class ir_copy_propagation_visitor : public ir_hierarchical_visitor {
public:
ir_copy_propagation_visitor()
{
progress = false;
mem_ctx = ralloc_context(0);
lin_ctx = linear_alloc_parent(mem_ctx, 0);
acp = _mesa_hash_table_create(mem_ctx, _mesa_hash_pointer,
_mesa_key_pointer_equal);
this->kills = new(mem_ctx) exec_list;
killed_all = false;
}
~ir_copy_propagation_visitor()
{
ralloc_free(mem_ctx);
}
virtual ir_visitor_status visit(class ir_dereference_variable *);
glsl: Make opt_copy_propagation actually propagate into loops. We've had a FINISHME here since Eric originally wrote the code in 2010. This patch implements his suggested approach, which makes us actually able to copy propagate into the loops, at the unfortunate cost of making this pass even more expensive. The shader-db statistics are not terribly impressive: total instructions in shared programs: 9008589 -> 9008613 (0.00%) instructions in affected programs: 4293 -> 4317 (0.56%) helped: 0 HURT: 10 total cycles in shared programs: 78550978 -> 78575760 (0.03%) cycles in affected programs: 655426 -> 680208 (3.78%) helped: 75 HURT: 88 GAINED: 2 Most of the "regressions" appear to be us successfully copy propagating uniforms, which i965 is loading as pull constants instead of push, so we occasionally have two pulls instead of one. That doesn't seem like this pass's job - it's propagating correctly, and we should be smarter about pull loads in the backend. This patch is also useful for a couple of reasons: 1. It can clean up copies created by varying packing (previously, we couldn't if the uses were inside a loop). This fixes a bug when interpolateAt*() is used on a packed varying inside a loop: glsl_to_nir struggles to see through the extra copy and mistakenly believed the variable was not an input. 2. It will help propagate uniform array access created by lower_const_array_to_uniforms(). Signed-off-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Matt Turner <mattst88@gmail.com>
2016-04-29 22:06:37 -07:00
void handle_loop(class ir_loop *, bool keep_acp);
virtual ir_visitor_status visit_enter(class ir_loop *);
virtual ir_visitor_status visit_enter(class ir_function_signature *);
virtual ir_visitor_status visit_enter(class ir_function *);
virtual ir_visitor_status visit_leave(class ir_assignment *);
virtual ir_visitor_status visit_enter(class ir_call *);
virtual ir_visitor_status visit_enter(class ir_if *);
void add_copy(ir_assignment *ir);
void kill(ir_variable *ir);
void handle_if_block(exec_list *instructions);
/** Hash of lhs->rhs: The available copies to propagate */
hash_table *acp;
/**
* List of kill_entry: The variables whose values were killed in this
* block.
*/
exec_list *kills;
bool progress;
bool killed_all;
void *mem_ctx;
void *lin_ctx;
};
} /* unnamed namespace */
ir_visitor_status
ir_copy_propagation_visitor::visit_enter(ir_function_signature *ir)
{
/* Treat entry into a function signature as a completely separate
* block. Any instructions at global scope will be shuffled into
* main() at link time, so they're irrelevant to us.
*/
hash_table *orig_acp = this->acp;
exec_list *orig_kills = this->kills;
bool orig_killed_all = this->killed_all;
acp = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
_mesa_key_pointer_equal);
this->kills = new(mem_ctx) exec_list;
this->killed_all = false;
visit_list_elements(this, &ir->body);
_mesa_hash_table_destroy(acp, NULL);
ralloc_free(this->kills);
this->kills = orig_kills;
this->acp = orig_acp;
this->killed_all = orig_killed_all;
return visit_continue_with_parent;
}
ir_visitor_status
ir_copy_propagation_visitor::visit_leave(ir_assignment *ir)
{
kill(ir->lhs->variable_referenced());
add_copy(ir);
return visit_continue;
}
ir_visitor_status
ir_copy_propagation_visitor::visit_enter(ir_function *ir)
{
(void) ir;
return visit_continue;
}
/**
* Replaces dereferences of ACP RHS variables with ACP LHS variables.
*
* This is where the actual copy propagation occurs. Note that the
* rewriting of ir_dereference means that the ir_dereference instance
* must not be shared by multiple IR operations!
*/
ir_visitor_status
ir_copy_propagation_visitor::visit(ir_dereference_variable *ir)
{
if (this->in_assignee)
return visit_continue;
struct hash_entry *entry = _mesa_hash_table_search(acp, ir->var);
if (entry) {
ir->var = (ir_variable *) entry->data;
progress = true;
}
return visit_continue;
}
ir_visitor_status
ir_copy_propagation_visitor::visit_enter(ir_call *ir)
{
/* Do copy propagation on call parameters, but skip any out params */
foreach_two_lists(formal_node, &ir->callee->parameters,
actual_node, &ir->actual_parameters) {
ir_variable *sig_param = (ir_variable *) formal_node;
ir_rvalue *ir = (ir_rvalue *) actual_node;
if (sig_param->data.mode != ir_var_function_out
&& sig_param->data.mode != ir_var_function_inout) {
ir->accept(this);
}
}
/* Since we're unlinked, we don't (necessarily) know the side effects of
* this call. So kill all copies.
*/
_mesa_hash_table_clear(acp, NULL);
this->killed_all = true;
return visit_continue_with_parent;
}
void
ir_copy_propagation_visitor::handle_if_block(exec_list *instructions)
{
hash_table *orig_acp = this->acp;
exec_list *orig_kills = this->kills;
bool orig_killed_all = this->killed_all;
acp = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
_mesa_key_pointer_equal);
this->kills = new(mem_ctx) exec_list;
this->killed_all = false;
/* Populate the initial acp with a copy of the original */
struct hash_entry *entry;
hash_table_foreach(orig_acp, entry) {
_mesa_hash_table_insert(acp, entry->key, entry->data);
}
visit_list_elements(this, instructions);
if (this->killed_all) {
_mesa_hash_table_clear(orig_acp, NULL);
}
exec_list *new_kills = this->kills;
this->kills = orig_kills;
_mesa_hash_table_destroy(acp, NULL);
this->acp = orig_acp;
this->killed_all = this->killed_all || orig_killed_all;
foreach_in_list(kill_entry, k, new_kills) {
kill(k->var);
}
ralloc_free(new_kills);
}
ir_visitor_status
ir_copy_propagation_visitor::visit_enter(ir_if *ir)
{
ir->condition->accept(this);
handle_if_block(&ir->then_instructions);
handle_if_block(&ir->else_instructions);
/* handle_if_block() already descended into the children. */
return visit_continue_with_parent;
}
glsl: Make opt_copy_propagation actually propagate into loops. We've had a FINISHME here since Eric originally wrote the code in 2010. This patch implements his suggested approach, which makes us actually able to copy propagate into the loops, at the unfortunate cost of making this pass even more expensive. The shader-db statistics are not terribly impressive: total instructions in shared programs: 9008589 -> 9008613 (0.00%) instructions in affected programs: 4293 -> 4317 (0.56%) helped: 0 HURT: 10 total cycles in shared programs: 78550978 -> 78575760 (0.03%) cycles in affected programs: 655426 -> 680208 (3.78%) helped: 75 HURT: 88 GAINED: 2 Most of the "regressions" appear to be us successfully copy propagating uniforms, which i965 is loading as pull constants instead of push, so we occasionally have two pulls instead of one. That doesn't seem like this pass's job - it's propagating correctly, and we should be smarter about pull loads in the backend. This patch is also useful for a couple of reasons: 1. It can clean up copies created by varying packing (previously, we couldn't if the uses were inside a loop). This fixes a bug when interpolateAt*() is used on a packed varying inside a loop: glsl_to_nir struggles to see through the extra copy and mistakenly believed the variable was not an input. 2. It will help propagate uniform array access created by lower_const_array_to_uniforms(). Signed-off-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Matt Turner <mattst88@gmail.com>
2016-04-29 22:06:37 -07:00
void
ir_copy_propagation_visitor::handle_loop(ir_loop *ir, bool keep_acp)
{
hash_table *orig_acp = this->acp;
exec_list *orig_kills = this->kills;
bool orig_killed_all = this->killed_all;
acp = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
_mesa_key_pointer_equal);
this->kills = new(mem_ctx) exec_list;
this->killed_all = false;
glsl: Make opt_copy_propagation actually propagate into loops. We've had a FINISHME here since Eric originally wrote the code in 2010. This patch implements his suggested approach, which makes us actually able to copy propagate into the loops, at the unfortunate cost of making this pass even more expensive. The shader-db statistics are not terribly impressive: total instructions in shared programs: 9008589 -> 9008613 (0.00%) instructions in affected programs: 4293 -> 4317 (0.56%) helped: 0 HURT: 10 total cycles in shared programs: 78550978 -> 78575760 (0.03%) cycles in affected programs: 655426 -> 680208 (3.78%) helped: 75 HURT: 88 GAINED: 2 Most of the "regressions" appear to be us successfully copy propagating uniforms, which i965 is loading as pull constants instead of push, so we occasionally have two pulls instead of one. That doesn't seem like this pass's job - it's propagating correctly, and we should be smarter about pull loads in the backend. This patch is also useful for a couple of reasons: 1. It can clean up copies created by varying packing (previously, we couldn't if the uses were inside a loop). This fixes a bug when interpolateAt*() is used on a packed varying inside a loop: glsl_to_nir struggles to see through the extra copy and mistakenly believed the variable was not an input. 2. It will help propagate uniform array access created by lower_const_array_to_uniforms(). Signed-off-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Matt Turner <mattst88@gmail.com>
2016-04-29 22:06:37 -07:00
if (keep_acp) {
struct hash_entry *entry;
hash_table_foreach(orig_acp, entry) {
_mesa_hash_table_insert(acp, entry->key, entry->data);
glsl: Make opt_copy_propagation actually propagate into loops. We've had a FINISHME here since Eric originally wrote the code in 2010. This patch implements his suggested approach, which makes us actually able to copy propagate into the loops, at the unfortunate cost of making this pass even more expensive. The shader-db statistics are not terribly impressive: total instructions in shared programs: 9008589 -> 9008613 (0.00%) instructions in affected programs: 4293 -> 4317 (0.56%) helped: 0 HURT: 10 total cycles in shared programs: 78550978 -> 78575760 (0.03%) cycles in affected programs: 655426 -> 680208 (3.78%) helped: 75 HURT: 88 GAINED: 2 Most of the "regressions" appear to be us successfully copy propagating uniforms, which i965 is loading as pull constants instead of push, so we occasionally have two pulls instead of one. That doesn't seem like this pass's job - it's propagating correctly, and we should be smarter about pull loads in the backend. This patch is also useful for a couple of reasons: 1. It can clean up copies created by varying packing (previously, we couldn't if the uses were inside a loop). This fixes a bug when interpolateAt*() is used on a packed varying inside a loop: glsl_to_nir struggles to see through the extra copy and mistakenly believed the variable was not an input. 2. It will help propagate uniform array access created by lower_const_array_to_uniforms(). Signed-off-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Matt Turner <mattst88@gmail.com>
2016-04-29 22:06:37 -07:00
}
}
visit_list_elements(this, &ir->body_instructions);
if (this->killed_all) {
_mesa_hash_table_clear(orig_acp, NULL);
}
exec_list *new_kills = this->kills;
this->kills = orig_kills;
_mesa_hash_table_destroy(acp, NULL);
this->acp = orig_acp;
this->killed_all = this->killed_all || orig_killed_all;
foreach_in_list(kill_entry, k, new_kills) {
kill(k->var);
}
ralloc_free(new_kills);
glsl: Make opt_copy_propagation actually propagate into loops. We've had a FINISHME here since Eric originally wrote the code in 2010. This patch implements his suggested approach, which makes us actually able to copy propagate into the loops, at the unfortunate cost of making this pass even more expensive. The shader-db statistics are not terribly impressive: total instructions in shared programs: 9008589 -> 9008613 (0.00%) instructions in affected programs: 4293 -> 4317 (0.56%) helped: 0 HURT: 10 total cycles in shared programs: 78550978 -> 78575760 (0.03%) cycles in affected programs: 655426 -> 680208 (3.78%) helped: 75 HURT: 88 GAINED: 2 Most of the "regressions" appear to be us successfully copy propagating uniforms, which i965 is loading as pull constants instead of push, so we occasionally have two pulls instead of one. That doesn't seem like this pass's job - it's propagating correctly, and we should be smarter about pull loads in the backend. This patch is also useful for a couple of reasons: 1. It can clean up copies created by varying packing (previously, we couldn't if the uses were inside a loop). This fixes a bug when interpolateAt*() is used on a packed varying inside a loop: glsl_to_nir struggles to see through the extra copy and mistakenly believed the variable was not an input. 2. It will help propagate uniform array access created by lower_const_array_to_uniforms(). Signed-off-by: Kenneth Graunke <kenneth@whitecape.org> Reviewed-by: Matt Turner <mattst88@gmail.com>
2016-04-29 22:06:37 -07:00
}
ir_visitor_status
ir_copy_propagation_visitor::visit_enter(ir_loop *ir)
{
/* Make a conservative first pass over the loop with an empty ACP set.
* This also removes any killed entries from the original ACP set.
*/
handle_loop(ir, false);
/* Then, run it again with the real ACP set, minus any killed entries.
* This takes care of propagating values from before the loop into it.
*/
handle_loop(ir, true);
/* already descended into the children. */
return visit_continue_with_parent;
}
void
ir_copy_propagation_visitor::kill(ir_variable *var)
{
assert(var != NULL);
/* Remove any entries currently in the ACP for this kill. */
struct hash_entry *entry = _mesa_hash_table_search(acp, var);
if (entry) {
_mesa_hash_table_remove(acp, entry);
}
hash_table_foreach(acp, entry) {
if (var == (ir_variable *) entry->data) {
_mesa_hash_table_remove(acp, entry);
}
}
/* Add the LHS variable to the list of killed variables in this block.
*/
this->kills->push_tail(new(this->lin_ctx) kill_entry(var));
}
/**
* Adds an entry to the available copy list if it's a plain assignment
* of a variable to a variable.
*/
void
ir_copy_propagation_visitor::add_copy(ir_assignment *ir)
{
if (ir->condition)
return;
ir_variable *lhs_var = ir->whole_variable_written();
ir_variable *rhs_var = ir->rhs->whole_variable_referenced();
if ((lhs_var != NULL) && (rhs_var != NULL)) {
if (lhs_var == rhs_var) {
/* This is a dumb assignment, but we've conveniently noticed
* it here. Removing it now would mess up the loop iteration
* calling us. Just flag it to not execute, and someone else
* will clean up the mess.
*/
ir->condition = new(ralloc_parent(ir)) ir_constant(false);
this->progress = true;
} else if (lhs_var->data.mode != ir_var_shader_storage &&
lhs_var->data.mode != ir_var_shader_shared &&
lhs_var->data.precise == rhs_var->data.precise) {
assert(lhs_var);
assert(rhs_var);
_mesa_hash_table_insert(acp, lhs_var, rhs_var);
}
}
}
/**
* Does a copy propagation pass on the code present in the instruction stream.
*/
bool
do_copy_propagation(exec_list *instructions)
{
ir_copy_propagation_visitor v;
visit_list_elements(&v, instructions);
return v.progress;
}