verilator/src/V3DfgPasses.cpp

214 lines
8.6 KiB
C++
Raw Normal View History

Introduce DFG based combinational logic optimizer (#3527) Added a new data-flow graph (DFG) based combinational logic optimizer. The capabilities of this covers a combination of V3Const and V3Gate, but is also more capable of transforming combinational logic into simplified forms and more. This entail adding a new internal representation, `DfgGraph`, and appropriate `astToDfg` and `dfgToAst` conversion functions. The graph represents some of the combinational equations (~continuous assignments) in a module, and for the duration of the DFG passes, it takes over the role of AstModule. A bulk of the Dfg vertices represent expressions. These vertex classes, and the corresponding conversions to/from AST are mostly auto-generated by astgen, together with a DfgVVisitor that can be used for dynamic dispatch based on vertex (operation) types. The resulting combinational logic graph (a `DfgGraph`) is then optimized in various ways. Currently we perform common sub-expression elimination, variable inlining, and some specific peephole optimizations, but there is scope for more optimizations in the future using the same representation. The optimizer is run directly before and after inlining. The pre inline pass can operate on smaller graphs and hence converges faster, but still has a chance of substantially reducing the size of the logic on some designs, making inlining both faster and less memory intensive. The post inline pass can then optimize across the inlined module boundaries. No optimization is performed across a module boundary. For debugging purposes, each peephole optimization can be disabled individually via the -fno-dfg-peepnole-<OPT> option, where <OPT> is one of the optimizations listed in V3DfgPeephole.h, for example -fno-dfg-peephole-remove-not-not. The peephole patterns currently implemented were mostly picked based on the design that inspired this work, and on that design the optimizations yields ~30% single threaded speedup, and ~50% speedup on 4 threads. As you can imagine not having to haul around redundant combinational networks in the rest of the compilation pipeline also helps with memory consumption, and up to 30% peak memory usage of Verilator was observed on the same design. Gains on other arbitrary designs are smaller (and can be improved by analyzing those designs). For example OpenTitan gains between 1-15% speedup depending on build type.
2022-09-23 17:46:22 +02:00
// -*- mode: C++; c-file-style: "cc-mode" -*-
//*************************************************************************
// DESCRIPTION: Verilator: Implementations of simple passes over DfgGraph
//
// Code available from: https://verilator.org
//
//*************************************************************************
//
// Copyright 2003-2022 by Wilson Snyder. This program is free software; you
// can redistribute it and/or modify it under the terms of either the GNU
// Lesser General Public License Version 3 or the Perl Artistic License
// Version 2.0.
// SPDX-License-Identifier: LGPL-3.0-only OR Artistic-2.0
//
//*************************************************************************
#include "config_build.h"
#include "V3DfgPasses.h"
#include "V3Dfg.h"
#include "V3Global.h"
#include "V3String.h"
#include <algorithm>
VL_DEFINE_DEBUG_FUNCTIONS;
V3DfgCseContext::~V3DfgCseContext() {
V3Stats::addStat("Optimizations, DFG " + m_label + " CSE, expressions eliminated",
m_eliminated);
}
DfgRemoveVarsContext::~DfgRemoveVarsContext() {
V3Stats::addStat("Optimizations, DFG " + m_label + " Remove vars, variables removed",
m_removed);
}
static std::string getPrefix(const std::string& label) {
if (label.empty()) return "";
std::string str = VString::removeWhitespace(label);
std::transform(str.begin(), str.end(), str.begin(), [](unsigned char c) { //
return c == ' ' ? '-' : std::tolower(c);
});
str += "-";
return str;
}
V3DfgOptimizationContext::V3DfgOptimizationContext(const std::string& label)
: m_label{label}
, m_prefix{getPrefix(label)} {}
V3DfgOptimizationContext::~V3DfgOptimizationContext() {
const string prefix = "Optimizations, DFG " + m_label + " ";
V3Stats::addStat(prefix + "General, modules", m_modules);
V3Stats::addStat(prefix + "Ast2Dfg, representable", m_representable);
V3Stats::addStat(prefix + "Ast2Dfg, non-representable (dtype)", m_nonRepDType);
V3Stats::addStat(prefix + "Ast2Dfg, non-representable (impure)", m_nonRepImpure);
V3Stats::addStat(prefix + "Ast2Dfg, non-representable (timing)", m_nonRepTiming);
V3Stats::addStat(prefix + "Ast2Dfg, non-representable (lhs)", m_nonRepLhs);
V3Stats::addStat(prefix + "Ast2Dfg, non-representable (node)", m_nonRepNode);
V3Stats::addStat(prefix + "Ast2Dfg, non-representable (unknown)", m_nonRepUnknown);
V3Stats::addStat(prefix + "Ast2Dfg, non-representable (var ref)", m_nonRepVarRef);
V3Stats::addStat(prefix + "Ast2Dfg, non-representable (width)", m_nonRepWidth);
V3Stats::addStat(prefix + "Dfg2Ast, intermediate variables", m_intermediateVars);
V3Stats::addStat(prefix + "Dfg2Ast, replaced variables", m_replacedVars);
V3Stats::addStat(prefix + "Dfg2Ast, result equations", m_resultEquations);
}
// 'Inline' DfgVar nodes with known drivers
void V3DfgPasses::inlineVars(DfgGraph& dfg) {
dfg.forEachVertex([](DfgVertex& vtx) {
// For each DfgVar that has a known driver
if (DfgVar* const varVtxp = vtx.cast<DfgVar>()) {
if (DfgVertex* const driverp = varVtxp->driverp()) {
// Make consumers of the DfgVar consume the driver directly
varVtxp->forEachSinkEdge([=](DfgEdge& edge) { edge.relinkSource(driverp); });
}
}
});
}
// Common subexpression elimination
void V3DfgPasses::cse(DfgGraph& dfg, V3DfgCseContext& ctx) {
DfgVertex::HashCache hashCache;
DfgVertex::EqualsCache equalsCache;
std::unordered_multimap<V3Hash, DfgVertex*> verticesWithEqualHashes;
// In reverse, as the graph is sometimes in reverse topological order already
dfg.forEachVertexInReverse([&](DfgVertex& vtx) {
// Don't merge constants
if (vtx.is<DfgConst>()) return;
// For everything else...
const V3Hash hash = vtx.hash(hashCache);
auto pair = verticesWithEqualHashes.equal_range(hash);
for (auto it = pair.first, end = pair.second; it != end; ++it) {
DfgVertex* const candidatep = it->second;
if (candidatep->equals(vtx, equalsCache)) {
++ctx.m_eliminated;
vtx.replaceWith(candidatep);
vtx.unlinkDelete(dfg);
return;
}
}
verticesWithEqualHashes.emplace(hash, &vtx);
});
}
void V3DfgPasses::removeVars(DfgGraph& dfg, DfgRemoveVarsContext& ctx) {
dfg.forEachVertex([&](DfgVertex& vtx) {
// We can eliminate certain redundant DfgVar vertices
DfgVar* const varp = vtx.cast<DfgVar>();
if (!varp) return;
// Can't remove if it has consumers
if (varp->hasSinks()) return;
// Can't remove if read in the module and driven here (i.e.: it's an output of the DFG)
if (varp->hasModRefs() && varp->driverp()) return;
// Can't remove if referenced externally, or other special reasons
if (varp->keep()) return;
// If the driver of this variable has multiple non-variable sinks, then we would need
// a temporary when rendering the graph. Instead of introducing a temporary, keep the
// first variable that is driven by that driver
if (DfgVertex* const driverp = varp->driverp()) {
unsigned nonVarSinks = 0;
const DfgVar* firstSinkVarp = nullptr;
const bool keepFirst = driverp->findSink<DfgVertex>([&](const DfgVertex& sink) {
if (const DfgVar* const sinkVarp = sink.cast<DfgVar>()) {
if (!firstSinkVarp) firstSinkVarp = sinkVarp;
} else {
++nonVarSinks;
}
// We can stop as soon as we found the first var, and 2 non-var sinks
return firstSinkVarp && nonVarSinks >= 2;
});
// Keep this DfgVar if needed
if (keepFirst && firstSinkVarp == varp) return;
}
// OK, we can delete this DfgVar!
++ctx.m_removed;
// If not referenced outside the DFG, then also delete the referenced AstVar,
// as it is now unused.
if (!varp->hasRefs()) varp->varp()->unlinkFrBack()->deleteTree();
// Unlink and delete vertex
vtx.unlinkDelete(dfg);
});
}
void V3DfgPasses::removeUnused(DfgGraph& dfg) {
const auto processVertex = [&](DfgVertex& vtx) {
// Keep variables
if (vtx.is<DfgVar>()) return false;
// Keep if it has sinks
if (vtx.hasSinks()) return false;
// Unlink and delete vertex
vtx.unlinkDelete(dfg);
return true;
};
dfg.runToFixedPoint(processVertex);
}
void V3DfgPasses::optimize(DfgGraph& dfg, V3DfgOptimizationContext& ctx) {
// There is absolutely nothing useful we can do with a graph of size 2 or less
if (dfg.size() <= 2) return;
// We consider a DFG trivial if it contains no more than 1 non-variable, non-constant vertex
unsigned excitingVertices = 0;
const bool isTrivial = !dfg.findVertex<DfgVertex>([&](const DfgVertex& vtx) { //
if (vtx.is<DfgVar>()) return false;
if (vtx.is<DfgConst>()) return false;
return ++excitingVertices >= 2;
});
int passNumber = 0;
const auto apply = [&](int dumpLevel, const string name, std::function<void()> pass) {
pass();
if (dumpDfg() >= dumpLevel) {
const string strippedName = VString::removeWhitespace(name);
const string label
= ctx.prefix() + "pass-" + cvtToStr(passNumber) + "-" + strippedName;
dfg.dumpDotFilePrefixed(label);
}
++passNumber;
};
if (!isTrivial) {
// Optimize non-trivial graph
if (dumpDfg() >= 8) { dfg.dumpDotAllVarConesPrefixed(ctx.prefix() + "input"); }
apply(3, "input ", [&]() {});
apply(4, "inlineVars ", [&]() { inlineVars(dfg); });
apply(4, "cse ", [&]() { cse(dfg, ctx.m_cseContext0); });
if (v3Global.opt.fDfgPeephole()) {
apply(4, "peephole ", [&]() { peephole(dfg, ctx.m_peepholeContext); });
}
apply(4, "removeVars ", [&]() { removeVars(dfg, ctx.m_removeVarsContext); });
apply(4, "cse ", [&]() { cse(dfg, ctx.m_cseContext1); });
apply(3, "optimized ", [&]() { removeUnused(dfg); });
if (dumpDfg() >= 8) { dfg.dumpDotAllVarConesPrefixed(ctx.prefix() + "optimized"); }
} else {
// We can still eliminate redundancies from trivial graphs
apply(5, "trivial-input ", [&]() {});
apply(6, "trivial-inlineVars ", [&]() { inlineVars(dfg); });
apply(5, "trivial-optimized ", [&]() { removeVars(dfg, ctx.m_removeVarsContext); });
}
}