SPIRV-Cross/spirv_cfg.cpp
Hans-Kristian Arntzen 5d97dae1eb Move branchless analysis to CFG.
Traverse backwards instead, far more robust. Should elide basically all
redundant continue; statements now.
2019-08-27 10:19:19 +02:00

398 lines
13 KiB
C++

/*
* Copyright 2016-2019 Arm Limited
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#include "spirv_cfg.hpp"
#include "spirv_cross.hpp"
#include <algorithm>
#include <assert.h>
using namespace std;
namespace SPIRV_CROSS_NAMESPACE
{
CFG::CFG(Compiler &compiler_, const SPIRFunction &func_)
: compiler(compiler_)
, func(func_)
{
build_post_order_visit_order();
build_immediate_dominators();
}
uint32_t CFG::find_common_dominator(uint32_t a, uint32_t b) const
{
while (a != b)
{
if (get_visit_order(a) < get_visit_order(b))
a = get_immediate_dominator(a);
else
b = get_immediate_dominator(b);
}
return a;
}
void CFG::build_immediate_dominators()
{
// Traverse the post-order in reverse and build up the immediate dominator tree.
immediate_dominators.clear();
immediate_dominators[func.entry_block] = func.entry_block;
for (auto i = post_order.size(); i; i--)
{
uint32_t block = post_order[i - 1];
auto &pred = preceding_edges[block];
if (pred.empty()) // This is for the entry block, but we've already set up the dominators.
continue;
for (auto &edge : pred)
{
if (immediate_dominators[block])
{
assert(immediate_dominators[edge]);
immediate_dominators[block] = find_common_dominator(block, edge);
}
else
immediate_dominators[block] = edge;
}
}
}
bool CFG::is_back_edge(uint32_t to) const
{
// We have a back edge if the visit order is set with the temporary magic value 0.
// Crossing edges will have already been recorded with a visit order.
auto itr = visit_order.find(to);
return itr != end(visit_order) && itr->second.get() == 0;
}
bool CFG::has_visited_forward_edge(uint32_t to) const
{
// If > 0, we have visited the edge already, and this is not a back edge branch.
auto itr = visit_order.find(to);
return itr != end(visit_order) && itr->second.get() > 0;
}
bool CFG::post_order_visit(uint32_t block_id)
{
// If we have already branched to this block (back edge), stop recursion.
// If our branches are back-edges, we do not record them.
// We have to record crossing edges however.
if (has_visited_forward_edge(block_id))
return true;
else if (is_back_edge(block_id))
return false;
// Block back-edges from recursively revisiting ourselves.
visit_order[block_id].get() = 0;
auto &block = compiler.get<SPIRBlock>(block_id);
// If this is a loop header, add an implied branch to the merge target.
// This is needed to avoid annoying cases with do { ... } while(false) loops often generated by inliners.
// To the CFG, this is linear control flow, but we risk picking the do/while scope as our dominating block.
// This makes sure that if we are accessing a variable outside the do/while, we choose the loop header as dominator.
// We could use has_visited_forward_edge, but this break code-gen where the merge block is unreachable in the CFG.
// Make a point out of visiting merge target first. This is to make sure that post visit order outside the loop
// is lower than inside the loop, which is going to be key for some traversal algorithms like post-dominance analysis.
// For selection constructs true/false blocks will end up visiting the merge block directly and it works out fine,
// but for loops, only the header might end up actually branching to merge block.
if (block.merge == SPIRBlock::MergeLoop && post_order_visit(block.merge_block))
add_branch(block_id, block.merge_block);
// First visit our branch targets.
switch (block.terminator)
{
case SPIRBlock::Direct:
if (post_order_visit(block.next_block))
add_branch(block_id, block.next_block);
break;
case SPIRBlock::Select:
if (post_order_visit(block.true_block))
add_branch(block_id, block.true_block);
if (post_order_visit(block.false_block))
add_branch(block_id, block.false_block);
break;
case SPIRBlock::MultiSelect:
for (auto &target : block.cases)
{
if (post_order_visit(target.block))
add_branch(block_id, target.block);
}
if (block.default_block && post_order_visit(block.default_block))
add_branch(block_id, block.default_block);
break;
default:
break;
}
// If this is a selection merge, add an implied branch to the merge target.
// This is needed to avoid cases where an inner branch dominates the outer branch.
// This can happen if one of the branches exit early, e.g.:
// if (cond) { ...; break; } else { var = 100 } use_var(var);
// We can use the variable without a Phi since there is only one possible parent here.
// However, in this case, we need to hoist out the inner variable to outside the branch.
// Use same strategy as loops.
if (block.merge == SPIRBlock::MergeSelection && post_order_visit(block.next_block))
{
// If there is only one preceding edge to the merge block and it's not ourselves, we need a fixup.
// Add a fake branch so any dominator in either the if (), or else () block, or a lone case statement
// will be hoisted out to outside the selection merge.
// If size > 1, the variable will be automatically hoisted, so we should not mess with it.
// The exception here is switch blocks, where we can have multiple edges to merge block,
// all coming from same scope, so be more conservative in this case.
// Adding fake branches unconditionally breaks parameter preservation analysis,
// which looks at how variables are accessed through the CFG.
auto pred_itr = preceding_edges.find(block.next_block);
if (pred_itr != end(preceding_edges))
{
auto &pred = pred_itr->second;
auto succ_itr = succeeding_edges.find(block_id);
uint32_t num_succeeding_edges = 0;
if (succ_itr != end(succeeding_edges))
num_succeeding_edges = succ_itr->second.size();
if (block.terminator == SPIRBlock::MultiSelect && num_succeeding_edges == 1)
{
// Multiple branches can come from the same scope due to "break;", so we need to assume that all branches
// come from same case scope in worst case, even if there are multiple preceding edges.
// If we have more than one succeeding edge from the block header, it should be impossible
// to have a dominator be inside the block.
// Only case this can go wrong is if we have 2 or more edges from block header and
// 2 or more edges to merge block, and still have dominator be inside a case label.
if (!pred.empty())
add_branch(block_id, block.next_block);
}
else
{
if (pred.size() == 1 && *pred.begin() != block_id)
add_branch(block_id, block.next_block);
}
}
else
{
// If the merge block does not have any preceding edges, i.e. unreachable, hallucinate it.
// We're going to do code-gen for it, and domination analysis requires that we have at least one preceding edge.
add_branch(block_id, block.next_block);
}
}
// Then visit ourselves. Start counting at one, to let 0 be a magic value for testing back vs. crossing edges.
visit_order[block_id].get() = ++visit_count;
post_order.push_back(block_id);
return true;
}
void CFG::build_post_order_visit_order()
{
uint32_t block = func.entry_block;
visit_count = 0;
visit_order.clear();
post_order.clear();
post_order_visit(block);
}
void CFG::add_branch(uint32_t from, uint32_t to)
{
const auto add_unique = [](SmallVector<uint32_t> &l, uint32_t value) {
auto itr = find(begin(l), end(l), value);
if (itr == end(l))
l.push_back(value);
};
add_unique(preceding_edges[to], from);
add_unique(succeeding_edges[from], to);
}
uint32_t CFG::find_loop_dominator(uint32_t block_id) const
{
while (block_id != SPIRBlock::NoDominator)
{
auto itr = preceding_edges.find(block_id);
if (itr == end(preceding_edges))
return SPIRBlock::NoDominator;
if (itr->second.empty())
return SPIRBlock::NoDominator;
uint32_t pred_block_id = SPIRBlock::NoDominator;
bool ignore_loop_header = false;
// If we are a merge block, go directly to the header block.
// Only consider a loop dominator if we are branching from inside a block to a loop header.
// NOTE: In the CFG we forced an edge from header to merge block always to support variable scopes properly.
for (auto &pred : itr->second)
{
auto &pred_block = compiler.get<SPIRBlock>(pred);
if (pred_block.merge == SPIRBlock::MergeLoop && pred_block.merge_block == block_id)
{
pred_block_id = pred;
ignore_loop_header = true;
break;
}
else if (pred_block.merge == SPIRBlock::MergeSelection && pred_block.next_block == block_id)
{
pred_block_id = pred;
break;
}
}
// No merge block means we can just pick any edge. Loop headers dominate the inner loop, so any path we
// take will lead there.
if (pred_block_id == SPIRBlock::NoDominator)
pred_block_id = itr->second.front();
block_id = pred_block_id;
if (!ignore_loop_header && block_id)
{
auto &block = compiler.get<SPIRBlock>(block_id);
if (block.merge == SPIRBlock::MergeLoop)
return block_id;
}
}
return block_id;
}
bool CFG::node_terminates_control_flow_in_sub_graph(uint32_t from, uint32_t to) const
{
// Walk backwards, starting from "to" block.
// Only follow pred edges if they have a 1:1 relationship, or a merge relationship.
// If we cannot find a path to "from", we must assume that to is inside control flow in some way.
auto &from_block = compiler.get<SPIRBlock>(from);
uint32_t ignore_block_id = 0;
if (from_block.merge == SPIRBlock::MergeLoop)
ignore_block_id = from_block.merge_block;
while (to != from)
{
auto pred_itr = preceding_edges.find(to);
if (pred_itr == end(preceding_edges))
return false;
DominatorBuilder builder(*this);
for (auto &edge : pred_itr->second)
builder.add_block(edge);
uint32_t dominator = builder.get_dominator();
if (dominator == 0)
return false;
auto &dom = compiler.get<SPIRBlock>(dominator);
bool true_path_ignore = false;
bool false_path_ignore = false;
if (ignore_block_id && dom.terminator == SPIRBlock::Select)
{
auto &true_block = compiler.get<SPIRBlock>(dom.true_block);
auto &false_block = compiler.get<SPIRBlock>(dom.false_block);
auto &ignore_block = compiler.get<SPIRBlock>(ignore_block_id);
true_path_ignore = compiler.execution_is_branchless(true_block, ignore_block);
false_path_ignore = compiler.execution_is_branchless(false_block, ignore_block);
}
if ((dom.merge == SPIRBlock::MergeSelection && dom.next_block == to) ||
(dom.merge == SPIRBlock::MergeLoop && dom.merge_block == to) ||
(dom.terminator == SPIRBlock::Direct && dom.next_block == to) ||
(dom.terminator == SPIRBlock::Select && dom.true_block == to && false_path_ignore) ||
(dom.terminator == SPIRBlock::Select && dom.false_block == to && true_path_ignore))
{
// Allow walking selection constructs if the other branch reaches out of a loop construct.
// It cannot be in-scope anymore.
to = dominator;
}
else
return false;
}
return true;
}
DominatorBuilder::DominatorBuilder(const CFG &cfg_)
: cfg(cfg_)
{
}
void DominatorBuilder::add_block(uint32_t block)
{
if (!cfg.get_immediate_dominator(block))
{
// Unreachable block via the CFG, we will never emit this code anyways.
return;
}
if (!dominator)
{
dominator = block;
return;
}
if (block != dominator)
dominator = cfg.find_common_dominator(block, dominator);
}
void DominatorBuilder::lift_continue_block_dominator()
{
// It is possible for a continue block to be the dominator of a variable is only accessed inside the while block of a do-while loop.
// We cannot safely declare variables inside a continue block, so move any variable declared
// in a continue block to the entry block to simplify.
// It makes very little sense for a continue block to ever be a dominator, so fall back to the simplest
// solution.
if (!dominator)
return;
auto &block = cfg.get_compiler().get<SPIRBlock>(dominator);
auto post_order = cfg.get_visit_order(dominator);
// If we are branching to a block with a higher post-order traversal index (continue blocks), we have a problem
// since we cannot create sensible GLSL code for this, fallback to entry block.
bool back_edge_dominator = false;
switch (block.terminator)
{
case SPIRBlock::Direct:
if (cfg.get_visit_order(block.next_block) > post_order)
back_edge_dominator = true;
break;
case SPIRBlock::Select:
if (cfg.get_visit_order(block.true_block) > post_order)
back_edge_dominator = true;
if (cfg.get_visit_order(block.false_block) > post_order)
back_edge_dominator = true;
break;
case SPIRBlock::MultiSelect:
for (auto &target : block.cases)
{
if (cfg.get_visit_order(target.block) > post_order)
back_edge_dominator = true;
}
if (block.default_block && cfg.get_visit_order(block.default_block) > post_order)
back_edge_dominator = true;
break;
default:
break;
}
if (back_edge_dominator)
dominator = cfg.get_function().entry_block;
}
} // namespace SPIRV_CROSS_NAMESPACE