shader: repair nested merge structurization

This commit is contained in:
nmzik
2026-08-18 03:53:44 +02:00
parent d80fbea51d
commit 1dbac4fd79
2 changed files with 96 additions and 28 deletions
@@ -1051,15 +1051,6 @@ bool IsolateSemanticLoopHeader(Graph& graph, uint32_t old_header) {
return true;
}
bool IsSyntheticMergeForwarder(const Graph& graph, uint32_t block_id, uint32_t merge) {
const auto* block = graph.FindBlock(block_id);
return block != nullptr && block->inst_begin == block->inst_end &&
block->successors.size() == 1u && block->successors.front() == merge &&
block->terminator.kind == TerminatorKind::Branch &&
block->terminator.condition == BranchCondition::Always &&
block->terminator.true_block == merge;
}
const NaturalLoop* FindInnermostContainingLoop(const Graph& graph, uint32_t block_id) {
const NaturalLoop* innermost = nullptr;
for (const auto& loop: graph.natural_loops) {
@@ -1232,24 +1223,14 @@ bool SplitSharedMergeBlock(Graph& graph, uint32_t merge,
if (construct_predecessors.empty() || (!force_split && external_predecessors.empty())) {
return false;
}
std::vector<uint32_t> predecessors_to_split;
for (auto pred: construct_predecessors) {
if (!IsSyntheticMergeForwarder(graph, pred, merge)) {
AddUnique(predecessors_to_split, pred);
}
}
if (predecessors_to_split.empty()) {
return false;
}
const auto synthetic_merge = AppendSyntheticBranchBlock(graph, merge);
auto* synthetic_block = graph.FindBlock(synthetic_merge);
if (synthetic_block != nullptr) {
synthetic_block->predecessors = predecessors_to_split;
synthetic_block->predecessors = construct_predecessors;
SortUnique(synthetic_block->predecessors);
}
for (auto pred: predecessors_to_split) {
for (auto pred: construct_predecessors) {
auto* block = graph.FindBlock(pred);
if (block == nullptr) {
continue;
@@ -1260,7 +1241,7 @@ bool SplitSharedMergeBlock(Graph& graph, uint32_t merge,
auto* old_merge = graph.FindBlock(merge);
if (old_merge != nullptr) {
for (auto pred: predecessors_to_split) {
for (auto pred: construct_predecessors) {
RemoveValue(old_merge->predecessors, pred);
}
AddUnique(old_merge->predecessors, synthetic_merge);
@@ -1313,11 +1294,24 @@ bool SplitOneSelectionMerge(Graph& graph, std::string* error) {
AddUnique(loop_headers, loop.header);
}
const auto original_block_count = static_cast<uint32_t>(graph.blocks.size());
for (uint32_t block_id = 0; block_id < original_block_count; block_id++) {
std::vector<uint32_t> selection_headers;
for (const auto& block: graph.blocks) {
if (block.terminator.kind == TerminatorKind::ConditionalBranch &&
!Contains(loop_headers, block.id)) {
selection_headers.push_back(block.id);
}
}
std::sort(selection_headers.begin(), selection_headers.end(), [&](uint32_t lhs, uint32_t rhs) {
const auto* lhs_block = graph.FindBlock(lhs);
const auto* rhs_block = graph.FindBlock(rhs);
const auto lhs_depth = lhs_block != nullptr ? lhs_block->dominators.size() : 0u;
const auto rhs_depth = rhs_block != nullptr ? rhs_block->dominators.size() : 0u;
return lhs_depth != rhs_depth ? lhs_depth > rhs_depth : lhs < rhs;
});
for (const auto block_id: selection_headers) {
const auto* block = graph.FindBlock(block_id);
if (block == nullptr || block->terminator.kind != TerminatorKind::ConditionalBranch ||
Contains(loop_headers, block_id)) {
if (block == nullptr) {
continue;
}
if (IsInnermostLoopControlConditional(graph, *block)) {
@@ -1352,8 +1346,7 @@ bool SplitOneSelectionMerge(Graph& graph, std::string* error) {
bool SplitSharedMergeBlocks(Graph& graph, std::string* error) {
const auto original_block_count = static_cast<uint32_t>(graph.blocks.size());
const auto split_budget =
std::max<uint32_t>(16u, std::min<uint32_t>(128u, original_block_count * 4u));
const auto split_budget = std::max<uint32_t>(16u, original_block_count * 4u);
for (uint32_t splits = 0; splits < split_budget; splits++) {
if (!SplitOneLoopMerge(graph) && !SplitOneSelectionMerge(graph, error)) {
return !graph.unsupported;
+75
View File
@@ -6943,6 +6943,80 @@ void TestNewShaderRecompilerCfgDuplicateMergeStructuredSplit() {
CheckSpirvBinaryValidates(result.spirv);
}
void TestNewShaderRecompilerCfgNestedEarlyExitLoopForwarders() {
const uint32_t shader[] = {
EncodeSopc(0x06, 0, 0), // preheader condition
EncodeSopp(0x05, 10), // preheader early exit -> end
EncodeSopc(0x06, 1, 1), // outer condition
EncodeSopp(0x05, 8), // outer early exit -> end
EncodeSopc(0x06, 2, 2), // inner condition
EncodeSopp(0x05, 2), // inner -> loop, else linear arm
EncodeSMovB32(3, 129), // linear arm work
EncodeSopp(0x02, 4), // linear arm -> end
EncodeSopc(0x06, 4, 4), // loop header condition
EncodeSopp(0x04, 2), // loop exit -> end
EncodeSop2(0x00, 5, 5, 129), // loop work
EncodeSopp(0x02, 0xfffcu), // backedge -> loop header
0xbf810000u,
};
ShaderRecompiler::Decoder::Program decoded;
std::string error;
Check(ShaderRecompiler::Decoder::DecodeProgram(std::span{shader}, decoded,
&error),
error.c_str());
ShaderRecompiler::CFG::Graph graph;
Check(ShaderRecompiler::CFG::BuildGraph(decoded, graph, &error),
error.c_str());
const auto original_block_count = graph.blocks.size();
const auto original_coverage =
CfgInstructionCoverage(graph, decoded.instructions.size());
const auto original_empty_blocks =
std::ranges::count_if(graph.blocks, [](const auto &block) {
return block.inst_begin == block.inst_end;
});
Check(original_block_count == 8u && graph.natural_loops.size() == 1u,
"nested early-exit fixture has the wrong native CFG");
const bool structured = ShaderRecompiler::CFG::Structurize(graph, &error);
Check(structured, error.c_str());
Check(graph.blocks.size() == original_block_count + 3u &&
CfgInstructionCoverage(graph, decoded.instructions.size()) ==
original_coverage &&
std::ranges::count_if(graph.blocks,
[](const auto &block) {
return block.inst_begin == block.inst_end;
}) == original_empty_blocks + 3,
"nested early-exit structurization changed semantic coverage");
const auto *preheader = graph.FindBlockByPc(0x00u);
const auto *outer = graph.FindBlockByPc(0x08u);
const auto *inner = graph.FindBlockByPc(0x10u);
const auto *loop = graph.FindBlockByPc(0x20u);
Check(
preheader != nullptr && outer != nullptr && inner != nullptr &&
loop != nullptr && loop->terminator.loop_header &&
preheader->terminator.merge_block != UINT32_MAX &&
outer->terminator.merge_block != UINT32_MAX &&
inner->terminator.merge_block != UINT32_MAX &&
preheader->terminator.merge_block != outer->terminator.merge_block &&
preheader->terminator.merge_block != inner->terminator.merge_block &&
outer->terminator.merge_block != inner->terminator.merge_block,
"nested early-exit constructs do not have distinct structured merges");
auto options = MakeCompileOptions(ShaderType::Compute);
options.dump_ir = true;
ShaderRecompiler::CompileResult result;
Check(ShaderRecompiler::TryRecompile(shader, options, result, &error),
error.c_str());
Check(!result.program.dispatcher_fallback &&
Common::ContainsStr(result.ir_dump, "mode=structured"),
"nested early-exit loop unexpectedly selected dispatcher fallback");
Check(SpirvInstructionOpcodeCount(result.spirv, 247) == 3u &&
SpirvInstructionOpcodeCount(result.spirv, 246) == 1u &&
SpirvInstructionOpcodeCount(result.spirv, 251) == 0u,
"nested early-exit SPIR-V has the wrong structured control flow");
CheckSpirvBinaryValidates(result.spirv);
}
void TestNewShaderRecompilerCfgExecSccSharedArm() {
const uint32_t shader[] = {
EncodeSop2(0x15, 126, 4, 126), // s_andn2_b64 exec, s4, exec
@@ -9333,6 +9407,7 @@ int main() {
TestNewShaderRecompilerCfgConditionalLoopHeaderSelection();
TestNewShaderRecompilerCfgMultipleLoopLatches();
TestNewShaderRecompilerCfgDuplicateMergeStructuredSplit();
TestNewShaderRecompilerCfgNestedEarlyExitLoopForwarders();
TestNewShaderRecompilerCfgExecSccSharedArm();
TestNewShaderRecompilerCfgOverlappingEarlyExitLadder();
TestNewShaderRecompilerCfgExternallyEnteredSelectionDispatcher();