1 /*
2  * Copyright 2016-2020 Arm Limited
3  *
4  * Licensed under the Apache License, Version 2.0 (the "License");
5  * you may not use this file except in compliance with the License.
6  * You may obtain a copy of the License at
7  *
8  *     http://www.apache.org/licenses/LICENSE-2.0
9  *
10  * Unless required by applicable law or agreed to in writing, software
11  * distributed under the License is distributed on an "AS IS" BASIS,
12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13  * See the License for the specific language governing permissions and
14  * limitations under the License.
15  */
16 
17 #include "spirv_cfg.hpp"
18 #include "spirv_cross.hpp"
19 #include <algorithm>
20 #include <assert.h>
21 
22 using namespace std;
23 
24 namespace SPIRV_CROSS_NAMESPACE
25 {
CFG(Compiler & compiler_,const SPIRFunction & func_)26 CFG::CFG(Compiler &compiler_, const SPIRFunction &func_)
27     : compiler(compiler_)
28     , func(func_)
29 {
30 	build_post_order_visit_order();
31 	build_immediate_dominators();
32 }
33 
find_common_dominator(uint32_t a,uint32_t b) const34 uint32_t CFG::find_common_dominator(uint32_t a, uint32_t b) const
35 {
36 	while (a != b)
37 	{
38 		if (get_visit_order(a) < get_visit_order(b))
39 			a = get_immediate_dominator(a);
40 		else
41 			b = get_immediate_dominator(b);
42 	}
43 	return a;
44 }
45 
build_immediate_dominators()46 void CFG::build_immediate_dominators()
47 {
48 	// Traverse the post-order in reverse and build up the immediate dominator tree.
49 	immediate_dominators.clear();
50 	immediate_dominators[func.entry_block] = func.entry_block;
51 
52 	for (auto i = post_order.size(); i; i--)
53 	{
54 		uint32_t block = post_order[i - 1];
55 		auto &pred = preceding_edges[block];
56 		if (pred.empty()) // This is for the entry block, but we've already set up the dominators.
57 			continue;
58 
59 		for (auto &edge : pred)
60 		{
61 			if (immediate_dominators[block])
62 			{
63 				assert(immediate_dominators[edge]);
64 				immediate_dominators[block] = find_common_dominator(immediate_dominators[block], edge);
65 			}
66 			else
67 				immediate_dominators[block] = edge;
68 		}
69 	}
70 }
71 
is_back_edge(uint32_t to) const72 bool CFG::is_back_edge(uint32_t to) const
73 {
74 	// We have a back edge if the visit order is set with the temporary magic value 0.
75 	// Crossing edges will have already been recorded with a visit order.
76 	auto itr = visit_order.find(to);
77 	return itr != end(visit_order) && itr->second.get() == 0;
78 }
79 
has_visited_forward_edge(uint32_t to) const80 bool CFG::has_visited_forward_edge(uint32_t to) const
81 {
82 	// If > 0, we have visited the edge already, and this is not a back edge branch.
83 	auto itr = visit_order.find(to);
84 	return itr != end(visit_order) && itr->second.get() > 0;
85 }
86 
post_order_visit(uint32_t block_id)87 bool CFG::post_order_visit(uint32_t block_id)
88 {
89 	// If we have already branched to this block (back edge), stop recursion.
90 	// If our branches are back-edges, we do not record them.
91 	// We have to record crossing edges however.
92 	if (has_visited_forward_edge(block_id))
93 		return true;
94 	else if (is_back_edge(block_id))
95 		return false;
96 
97 	// Block back-edges from recursively revisiting ourselves.
98 	visit_order[block_id].get() = 0;
99 
100 	auto &block = compiler.get<SPIRBlock>(block_id);
101 
102 	// If this is a loop header, add an implied branch to the merge target.
103 	// This is needed to avoid annoying cases with do { ... } while(false) loops often generated by inliners.
104 	// To the CFG, this is linear control flow, but we risk picking the do/while scope as our dominating block.
105 	// This makes sure that if we are accessing a variable outside the do/while, we choose the loop header as dominator.
106 	// We could use has_visited_forward_edge, but this break code-gen where the merge block is unreachable in the CFG.
107 
108 	// Make a point out of visiting merge target first. This is to make sure that post visit order outside the loop
109 	// is lower than inside the loop, which is going to be key for some traversal algorithms like post-dominance analysis.
110 	// For selection constructs true/false blocks will end up visiting the merge block directly and it works out fine,
111 	// but for loops, only the header might end up actually branching to merge block.
112 	if (block.merge == SPIRBlock::MergeLoop && post_order_visit(block.merge_block))
113 		add_branch(block_id, block.merge_block);
114 
115 	// First visit our branch targets.
116 	switch (block.terminator)
117 	{
118 	case SPIRBlock::Direct:
119 		if (post_order_visit(block.next_block))
120 			add_branch(block_id, block.next_block);
121 		break;
122 
123 	case SPIRBlock::Select:
124 		if (post_order_visit(block.true_block))
125 			add_branch(block_id, block.true_block);
126 		if (post_order_visit(block.false_block))
127 			add_branch(block_id, block.false_block);
128 		break;
129 
130 	case SPIRBlock::MultiSelect:
131 		for (auto &target : block.cases)
132 		{
133 			if (post_order_visit(target.block))
134 				add_branch(block_id, target.block);
135 		}
136 		if (block.default_block && post_order_visit(block.default_block))
137 			add_branch(block_id, block.default_block);
138 		break;
139 
140 	default:
141 		break;
142 	}
143 
144 	// If this is a selection merge, add an implied branch to the merge target.
145 	// This is needed to avoid cases where an inner branch dominates the outer branch.
146 	// This can happen if one of the branches exit early, e.g.:
147 	// if (cond) { ...; break; } else { var = 100 } use_var(var);
148 	// We can use the variable without a Phi since there is only one possible parent here.
149 	// However, in this case, we need to hoist out the inner variable to outside the branch.
150 	// Use same strategy as loops.
151 	if (block.merge == SPIRBlock::MergeSelection && post_order_visit(block.next_block))
152 	{
153 		// If there is only one preceding edge to the merge block and it's not ourselves, we need a fixup.
154 		// Add a fake branch so any dominator in either the if (), or else () block, or a lone case statement
155 		// will be hoisted out to outside the selection merge.
156 		// If size > 1, the variable will be automatically hoisted, so we should not mess with it.
157 		// The exception here is switch blocks, where we can have multiple edges to merge block,
158 		// all coming from same scope, so be more conservative in this case.
159 		// Adding fake branches unconditionally breaks parameter preservation analysis,
160 		// which looks at how variables are accessed through the CFG.
161 		auto pred_itr = preceding_edges.find(block.next_block);
162 		if (pred_itr != end(preceding_edges))
163 		{
164 			auto &pred = pred_itr->second;
165 			auto succ_itr = succeeding_edges.find(block_id);
166 			size_t num_succeeding_edges = 0;
167 			if (succ_itr != end(succeeding_edges))
168 				num_succeeding_edges = succ_itr->second.size();
169 
170 			if (block.terminator == SPIRBlock::MultiSelect && num_succeeding_edges == 1)
171 			{
172 				// Multiple branches can come from the same scope due to "break;", so we need to assume that all branches
173 				// come from same case scope in worst case, even if there are multiple preceding edges.
174 				// If we have more than one succeeding edge from the block header, it should be impossible
175 				// to have a dominator be inside the block.
176 				// Only case this can go wrong is if we have 2 or more edges from block header and
177 				// 2 or more edges to merge block, and still have dominator be inside a case label.
178 				if (!pred.empty())
179 					add_branch(block_id, block.next_block);
180 			}
181 			else
182 			{
183 				if (pred.size() == 1 && *pred.begin() != block_id)
184 					add_branch(block_id, block.next_block);
185 			}
186 		}
187 		else
188 		{
189 			// If the merge block does not have any preceding edges, i.e. unreachable, hallucinate it.
190 			// We're going to do code-gen for it, and domination analysis requires that we have at least one preceding edge.
191 			add_branch(block_id, block.next_block);
192 		}
193 	}
194 
195 	// Then visit ourselves. Start counting at one, to let 0 be a magic value for testing back vs. crossing edges.
196 	visit_order[block_id].get() = ++visit_count;
197 	post_order.push_back(block_id);
198 	return true;
199 }
200 
build_post_order_visit_order()201 void CFG::build_post_order_visit_order()
202 {
203 	uint32_t block = func.entry_block;
204 	visit_count = 0;
205 	visit_order.clear();
206 	post_order.clear();
207 	post_order_visit(block);
208 }
209 
add_branch(uint32_t from,uint32_t to)210 void CFG::add_branch(uint32_t from, uint32_t to)
211 {
212 	const auto add_unique = [](SmallVector<uint32_t> &l, uint32_t value) {
213 		auto itr = find(begin(l), end(l), value);
214 		if (itr == end(l))
215 			l.push_back(value);
216 	};
217 	add_unique(preceding_edges[to], from);
218 	add_unique(succeeding_edges[from], to);
219 }
220 
find_loop_dominator(uint32_t block_id) const221 uint32_t CFG::find_loop_dominator(uint32_t block_id) const
222 {
223 	while (block_id != SPIRBlock::NoDominator)
224 	{
225 		auto itr = preceding_edges.find(block_id);
226 		if (itr == end(preceding_edges))
227 			return SPIRBlock::NoDominator;
228 		if (itr->second.empty())
229 			return SPIRBlock::NoDominator;
230 
231 		uint32_t pred_block_id = SPIRBlock::NoDominator;
232 		bool ignore_loop_header = false;
233 
234 		// If we are a merge block, go directly to the header block.
235 		// Only consider a loop dominator if we are branching from inside a block to a loop header.
236 		// NOTE: In the CFG we forced an edge from header to merge block always to support variable scopes properly.
237 		for (auto &pred : itr->second)
238 		{
239 			auto &pred_block = compiler.get<SPIRBlock>(pred);
240 			if (pred_block.merge == SPIRBlock::MergeLoop && pred_block.merge_block == ID(block_id))
241 			{
242 				pred_block_id = pred;
243 				ignore_loop_header = true;
244 				break;
245 			}
246 			else if (pred_block.merge == SPIRBlock::MergeSelection && pred_block.next_block == ID(block_id))
247 			{
248 				pred_block_id = pred;
249 				break;
250 			}
251 		}
252 
253 		// No merge block means we can just pick any edge. Loop headers dominate the inner loop, so any path we
254 		// take will lead there.
255 		if (pred_block_id == SPIRBlock::NoDominator)
256 			pred_block_id = itr->second.front();
257 
258 		block_id = pred_block_id;
259 
260 		if (!ignore_loop_header && block_id)
261 		{
262 			auto &block = compiler.get<SPIRBlock>(block_id);
263 			if (block.merge == SPIRBlock::MergeLoop)
264 				return block_id;
265 		}
266 	}
267 
268 	return block_id;
269 }
270 
node_terminates_control_flow_in_sub_graph(BlockID from,BlockID to) const271 bool CFG::node_terminates_control_flow_in_sub_graph(BlockID from, BlockID to) const
272 {
273 	// Walk backwards, starting from "to" block.
274 	// Only follow pred edges if they have a 1:1 relationship, or a merge relationship.
275 	// If we cannot find a path to "from", we must assume that to is inside control flow in some way.
276 
277 	auto &from_block = compiler.get<SPIRBlock>(from);
278 	BlockID ignore_block_id = 0;
279 	if (from_block.merge == SPIRBlock::MergeLoop)
280 		ignore_block_id = from_block.merge_block;
281 
282 	while (to != from)
283 	{
284 		auto pred_itr = preceding_edges.find(to);
285 		if (pred_itr == end(preceding_edges))
286 			return false;
287 
288 		DominatorBuilder builder(*this);
289 		for (auto &edge : pred_itr->second)
290 			builder.add_block(edge);
291 
292 		uint32_t dominator = builder.get_dominator();
293 		if (dominator == 0)
294 			return false;
295 
296 		auto &dom = compiler.get<SPIRBlock>(dominator);
297 
298 		bool true_path_ignore = false;
299 		bool false_path_ignore = false;
300 		if (ignore_block_id && dom.terminator == SPIRBlock::Select)
301 		{
302 			auto &true_block = compiler.get<SPIRBlock>(dom.true_block);
303 			auto &false_block = compiler.get<SPIRBlock>(dom.false_block);
304 			auto &ignore_block = compiler.get<SPIRBlock>(ignore_block_id);
305 			true_path_ignore = compiler.execution_is_branchless(true_block, ignore_block);
306 			false_path_ignore = compiler.execution_is_branchless(false_block, ignore_block);
307 		}
308 
309 		if ((dom.merge == SPIRBlock::MergeSelection && dom.next_block == to) ||
310 		    (dom.merge == SPIRBlock::MergeLoop && dom.merge_block == to) ||
311 		    (dom.terminator == SPIRBlock::Direct && dom.next_block == to) ||
312 		    (dom.terminator == SPIRBlock::Select && dom.true_block == to && false_path_ignore) ||
313 		    (dom.terminator == SPIRBlock::Select && dom.false_block == to && true_path_ignore))
314 		{
315 			// Allow walking selection constructs if the other branch reaches out of a loop construct.
316 			// It cannot be in-scope anymore.
317 			to = dominator;
318 		}
319 		else
320 			return false;
321 	}
322 
323 	return true;
324 }
325 
DominatorBuilder(const CFG & cfg_)326 DominatorBuilder::DominatorBuilder(const CFG &cfg_)
327     : cfg(cfg_)
328 {
329 }
330 
add_block(uint32_t block)331 void DominatorBuilder::add_block(uint32_t block)
332 {
333 	if (!cfg.get_immediate_dominator(block))
334 	{
335 		// Unreachable block via the CFG, we will never emit this code anyways.
336 		return;
337 	}
338 
339 	if (!dominator)
340 	{
341 		dominator = block;
342 		return;
343 	}
344 
345 	if (block != dominator)
346 		dominator = cfg.find_common_dominator(block, dominator);
347 }
348 
lift_continue_block_dominator()349 void DominatorBuilder::lift_continue_block_dominator()
350 {
351 	// 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.
352 	// We cannot safely declare variables inside a continue block, so move any variable declared
353 	// in a continue block to the entry block to simplify.
354 	// It makes very little sense for a continue block to ever be a dominator, so fall back to the simplest
355 	// solution.
356 
357 	if (!dominator)
358 		return;
359 
360 	auto &block = cfg.get_compiler().get<SPIRBlock>(dominator);
361 	auto post_order = cfg.get_visit_order(dominator);
362 
363 	// If we are branching to a block with a higher post-order traversal index (continue blocks), we have a problem
364 	// since we cannot create sensible GLSL code for this, fallback to entry block.
365 	bool back_edge_dominator = false;
366 	switch (block.terminator)
367 	{
368 	case SPIRBlock::Direct:
369 		if (cfg.get_visit_order(block.next_block) > post_order)
370 			back_edge_dominator = true;
371 		break;
372 
373 	case SPIRBlock::Select:
374 		if (cfg.get_visit_order(block.true_block) > post_order)
375 			back_edge_dominator = true;
376 		if (cfg.get_visit_order(block.false_block) > post_order)
377 			back_edge_dominator = true;
378 		break;
379 
380 	case SPIRBlock::MultiSelect:
381 		for (auto &target : block.cases)
382 		{
383 			if (cfg.get_visit_order(target.block) > post_order)
384 				back_edge_dominator = true;
385 		}
386 		if (block.default_block && cfg.get_visit_order(block.default_block) > post_order)
387 			back_edge_dominator = true;
388 		break;
389 
390 	default:
391 		break;
392 	}
393 
394 	if (back_edge_dominator)
395 		dominator = cfg.get_function().entry_block;
396 }
397 } // namespace SPIRV_CROSS_NAMESPACE
398