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
|
|
|
|
|
//
|
|
|
|
|
//*************************************************************************
|
|
|
|
|
//
|
2025-01-01 14:30:25 +01:00
|
|
|
// Copyright 2003-2025 by Wilson Snyder. This program is free software; you
|
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
|
|
|
// 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
|
|
|
|
|
//
|
|
|
|
|
//*************************************************************************
|
|
|
|
|
|
2023-10-18 12:37:46 +02:00
|
|
|
#include "V3PchAstNoMT.h" // VL_MT_DISABLED_CODE_UNIT
|
|
|
|
|
|
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
|
|
|
#include "V3DfgPasses.h"
|
|
|
|
|
|
|
|
|
|
#include "V3Dfg.h"
|
2024-03-02 20:49:29 +01:00
|
|
|
#include "V3File.h"
|
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
|
|
|
#include "V3Global.h"
|
|
|
|
|
#include "V3String.h"
|
|
|
|
|
|
|
|
|
|
VL_DEFINE_DEBUG_FUNCTIONS;
|
|
|
|
|
|
2022-11-06 16:42:01 +01:00
|
|
|
// Common sub-expression elimination
|
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
|
|
|
void V3DfgPasses::cse(DfgGraph& dfg, V3DfgCseContext& ctx) {
|
2022-11-06 16:42:01 +01:00
|
|
|
// Remove common sub-expressions
|
|
|
|
|
{
|
|
|
|
|
// Used by DfgVertex::hash
|
|
|
|
|
const auto userDataInUse = dfg.userDataInUse();
|
|
|
|
|
|
|
|
|
|
DfgVertex::EqualsCache equalsCache;
|
|
|
|
|
std::unordered_map<V3Hash, std::vector<DfgVertex*>> verticesWithEqualHashes;
|
|
|
|
|
verticesWithEqualHashes.reserve(dfg.size());
|
|
|
|
|
|
|
|
|
|
// Pre-hash variables, these are all unique, so just set their hash to a unique value
|
|
|
|
|
uint32_t varHash = 0;
|
2024-03-26 00:06:25 +01:00
|
|
|
for (DfgVertexVar& vtx : dfg.varVertices()) vtx.user<V3Hash>() = V3Hash{++varHash};
|
2022-10-06 12:26:11 +02:00
|
|
|
|
2022-11-06 16:42:01 +01:00
|
|
|
// Similarly pre-hash constants for speed. While we don't combine constants, we do want
|
|
|
|
|
// expressions using the same constants to be combined, so we do need to hash equal
|
|
|
|
|
// constants to equal values.
|
2024-03-26 00:06:25 +01:00
|
|
|
for (DfgConst* const vtxp : dfg.constVertices().unlinkable()) {
|
2022-11-06 16:42:01 +01:00
|
|
|
// Delete unused constants while we are at it.
|
|
|
|
|
if (!vtxp->hasSinks()) {
|
2024-03-26 00:06:25 +01:00
|
|
|
VL_DO_DANGLING(vtxp->unlinkDelete(dfg), vtxp);
|
2022-11-06 16:42:01 +01:00
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
vtxp->user<V3Hash>() = vtxp->num().toHash() + varHash;
|
2022-10-07 17:13:01 +02:00
|
|
|
}
|
2022-11-06 16:42:01 +01:00
|
|
|
|
|
|
|
|
// Combine operation vertices
|
2024-03-26 00:06:25 +01:00
|
|
|
for (DfgVertex* const vtxp : dfg.opVertices().unlinkable()) {
|
2022-11-06 16:42:01 +01:00
|
|
|
// Delete unused nodes while we are at it.
|
|
|
|
|
if (!vtxp->hasSinks()) {
|
2022-10-06 12:26:11 +02:00
|
|
|
vtxp->unlinkDelete(dfg);
|
2022-11-06 16:42:01 +01:00
|
|
|
continue;
|
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
|
|
|
}
|
2022-11-06 16:42:01 +01:00
|
|
|
const V3Hash hash = vtxp->hash();
|
|
|
|
|
std::vector<DfgVertex*>& vec = verticesWithEqualHashes[hash];
|
|
|
|
|
bool replaced = false;
|
|
|
|
|
for (DfgVertex* const candidatep : vec) {
|
|
|
|
|
if (candidatep->equals(*vtxp, equalsCache)) {
|
|
|
|
|
++ctx.m_eliminated;
|
|
|
|
|
vtxp->replaceWith(candidatep);
|
2024-03-26 00:06:25 +01:00
|
|
|
VL_DO_DANGLING(vtxp->unlinkDelete(dfg), vtxp);
|
2022-11-06 16:42:01 +01:00
|
|
|
replaced = true;
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
if (replaced) continue;
|
|
|
|
|
vec.push_back(vtxp);
|
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
|
|
|
}
|
2022-10-06 12:26:11 +02:00
|
|
|
}
|
2022-11-06 16:42:01 +01:00
|
|
|
|
|
|
|
|
// Prune unused nodes
|
|
|
|
|
removeUnused(dfg);
|
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
|
|
|
}
|
|
|
|
|
|
2024-03-26 00:06:25 +01:00
|
|
|
void V3DfgPasses::inlineVars(DfgGraph& dfg) {
|
|
|
|
|
for (DfgVertexVar& vtx : dfg.varVertices()) {
|
|
|
|
|
if (DfgVarPacked* const varp = vtx.cast<DfgVarPacked>()) {
|
2025-07-21 19:32:08 +02:00
|
|
|
if (varp->hasSinks() && varp->isDrivenFullyByDfg()) {
|
2025-07-14 23:09:34 +02:00
|
|
|
DfgVertex* const driverp = varp->srcp();
|
2024-04-27 14:41:10 +02:00
|
|
|
varp->forEachSinkEdge([=](DfgEdge& edge) { edge.relinkSource(driverp); });
|
2022-10-07 17:13:01 +02:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
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
|
|
|
void V3DfgPasses::removeUnused(DfgGraph& dfg) {
|
2022-11-06 15:13:42 +01:00
|
|
|
// DfgVertex::user is the next pointer of the work list elements
|
|
|
|
|
const auto userDataInUse = dfg.userDataInUse();
|
|
|
|
|
|
|
|
|
|
// Head of work list. Note that we want all next pointers in the list to be non-zero (including
|
|
|
|
|
// that of the last element). This allows as to do two important things: detect if an element
|
2022-12-03 00:46:38 +01:00
|
|
|
// is in the list by checking for a non-zero next pointer, and easy prefetching without
|
2022-11-06 15:13:42 +01:00
|
|
|
// conditionals. The address of the graph is a good sentinel as it is a valid memory address,
|
|
|
|
|
// and we can easily check for the end of the list.
|
|
|
|
|
DfgVertex* const sentinelp = reinterpret_cast<DfgVertex*>(&dfg);
|
|
|
|
|
DfgVertex* workListp = sentinelp;
|
|
|
|
|
|
|
|
|
|
// Add all unused vertices to the work list. This also allocates all DfgVertex::user.
|
2024-03-26 00:06:25 +01:00
|
|
|
for (DfgVertex& vtx : dfg.opVertices()) {
|
|
|
|
|
if (vtx.hasSinks()) {
|
2022-11-06 15:13:42 +01:00
|
|
|
// This vertex is used. Allocate user, but don't add to work list.
|
2024-03-26 00:06:25 +01:00
|
|
|
vtx.setUser<DfgVertex*>(nullptr);
|
2022-11-06 15:13:42 +01:00
|
|
|
} else {
|
|
|
|
|
// This vertex is unused. Add to work list.
|
2024-03-26 00:06:25 +01:00
|
|
|
vtx.setUser<DfgVertex*>(workListp);
|
|
|
|
|
workListp = &vtx;
|
2022-10-07 17:13:01 +02:00
|
|
|
}
|
2022-11-06 15:13:42 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Process the work list
|
|
|
|
|
while (workListp != sentinelp) {
|
|
|
|
|
// Pick up the head
|
|
|
|
|
DfgVertex* const vtxp = workListp;
|
|
|
|
|
// Detach the head
|
|
|
|
|
workListp = vtxp->getUser<DfgVertex*>();
|
|
|
|
|
// Prefetch next item
|
|
|
|
|
VL_PREFETCH_RW(workListp);
|
|
|
|
|
// If used, then nothing to do, so move on
|
|
|
|
|
if (vtxp->hasSinks()) continue;
|
|
|
|
|
// Add sources of unused vertex to work list
|
|
|
|
|
vtxp->forEachSource([&](DfgVertex& src) {
|
|
|
|
|
// We only remove actual operation vertices in this loop
|
|
|
|
|
if (src.is<DfgConst>() || src.is<DfgVertexVar>()) return;
|
|
|
|
|
// If already in work list then nothing to do
|
|
|
|
|
if (src.getUser<DfgVertex*>()) return;
|
|
|
|
|
// Actually add to work list.
|
|
|
|
|
src.setUser<DfgVertex*>(workListp);
|
|
|
|
|
workListp = &src;
|
|
|
|
|
});
|
|
|
|
|
// Remove the unused vertex
|
|
|
|
|
vtxp->unlinkDelete(dfg);
|
2022-10-07 17:13:01 +02:00
|
|
|
}
|
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
|
|
|
|
2022-10-07 17:13:01 +02:00
|
|
|
// Finally remove unused constants
|
2024-03-26 00:06:25 +01:00
|
|
|
for (DfgConst* const vtxp : dfg.constVertices().unlinkable()) {
|
|
|
|
|
if (!vtxp->hasSinks()) VL_DO_DANGLING(vtxp->unlinkDelete(dfg), vtxp);
|
2022-10-07 17:13:01 +02:00
|
|
|
}
|
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
|
|
|
}
|
|
|
|
|
|
2025-06-17 00:14:24 +02:00
|
|
|
void V3DfgPasses::binToOneHot(DfgGraph& dfg, V3DfgBinToOneHotContext& ctx) {
|
2025-07-01 23:55:08 +02:00
|
|
|
UASSERT(dfg.modulep(), "binToOneHot only works with unscoped DfgGraphs for now");
|
|
|
|
|
|
2025-06-17 00:14:24 +02:00
|
|
|
const auto userDataInUse = dfg.userDataInUse();
|
|
|
|
|
|
|
|
|
|
// Structure to keep track of comparison details
|
|
|
|
|
struct Term final {
|
|
|
|
|
DfgVertex* m_vtxp; // Vertex to replace
|
|
|
|
|
bool m_inv; // '!=', instead of '=='
|
|
|
|
|
Term() = default;
|
|
|
|
|
Term(DfgVertex* vtxp, bool inv)
|
|
|
|
|
: m_vtxp{vtxp}
|
|
|
|
|
, m_inv{inv} {}
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
// Map from 'value beign compared' -> 'terms', stored in DfgVertex::user()
|
|
|
|
|
using Val2Terms = std::map<uint32_t, std::vector<Term>>;
|
|
|
|
|
// Allocator for Val2Terms, so it's cleaned up on return
|
|
|
|
|
std::deque<Val2Terms> val2TermsAllocator;
|
|
|
|
|
// List of vertices that are used as sources
|
|
|
|
|
std::vector<DfgVertex*> srcps;
|
|
|
|
|
|
|
|
|
|
// Only consider input variables from a reasonable range:
|
|
|
|
|
// - not too big to avoid huge tables, you are doomed anyway at that point..
|
|
|
|
|
// - not too small, as it's probably not worth it
|
|
|
|
|
constexpr uint32_t WIDTH_MIN = 7;
|
|
|
|
|
constexpr uint32_t WIDTH_MAX = 20;
|
|
|
|
|
const auto widthOk = [](const DfgVertex* vtxp) {
|
|
|
|
|
const uint32_t width = vtxp->width();
|
|
|
|
|
return WIDTH_MIN <= width && width <= WIDTH_MAX;
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
// Do not convert terms that look like they are in a Cond tree
|
|
|
|
|
// the C++ compiler can generate jump tables for these
|
|
|
|
|
const std::function<bool(const DfgVertex*, bool)> useOk
|
|
|
|
|
= [&](const DfgVertex* vtxp, bool inv) -> bool {
|
|
|
|
|
// Go past a single 'Not' sink, which is common
|
|
|
|
|
if (DfgVertex* const sinkp = vtxp->singleSink()) {
|
|
|
|
|
if (sinkp->is<DfgNot>()) return useOk(sinkp, !inv);
|
|
|
|
|
}
|
|
|
|
|
return !vtxp->findSink<DfgCond>([vtxp, inv](const DfgCond& sink) {
|
|
|
|
|
if (sink.condp() != vtxp) return false;
|
|
|
|
|
return inv ? sink.thenp()->is<DfgCond>() : sink.elsep()->is<DfgCond>();
|
|
|
|
|
});
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
// Look at all comparison nodes and build the 'Val2Terms' map for each source vertex
|
|
|
|
|
uint32_t nTerms = 0;
|
|
|
|
|
for (DfgVertex& vtx : dfg.opVertices()) {
|
|
|
|
|
DfgVertex* srcp = nullptr;
|
|
|
|
|
uint32_t val = 0;
|
|
|
|
|
bool inv = false;
|
|
|
|
|
if (DfgEq* const eqp = vtx.cast<DfgEq>()) {
|
|
|
|
|
DfgConst* const constp = eqp->lhsp()->cast<DfgConst>();
|
|
|
|
|
if (!constp || !widthOk(constp) || !useOk(eqp, false)) continue;
|
|
|
|
|
srcp = eqp->rhsp();
|
|
|
|
|
val = constp->toU32();
|
|
|
|
|
inv = false;
|
|
|
|
|
} else if (DfgNeq* const neqp = vtx.cast<DfgNeq>()) {
|
|
|
|
|
DfgConst* const constp = neqp->lhsp()->cast<DfgConst>();
|
|
|
|
|
if (!constp || !widthOk(constp) || !useOk(neqp, true)) continue;
|
|
|
|
|
srcp = neqp->rhsp();
|
|
|
|
|
val = constp->toU32();
|
|
|
|
|
inv = true;
|
|
|
|
|
} else if (DfgRedAnd* const redAndp = vtx.cast<DfgRedAnd>()) {
|
|
|
|
|
srcp = redAndp->srcp();
|
|
|
|
|
if (!widthOk(srcp) || !useOk(redAndp, false)) continue;
|
|
|
|
|
val = (1U << srcp->width()) - 1;
|
|
|
|
|
inv = false;
|
|
|
|
|
} else if (DfgRedOr* const redOrp = vtx.cast<DfgRedOr>()) {
|
|
|
|
|
srcp = redOrp->srcp();
|
|
|
|
|
if (!widthOk(srcp) || !useOk(redOrp, true)) continue;
|
|
|
|
|
val = 0;
|
|
|
|
|
inv = true;
|
|
|
|
|
} else {
|
|
|
|
|
// Not a comparison-like vertex
|
|
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
// Grab the Val2Terms entry
|
|
|
|
|
Val2Terms*& val2Termspr = srcp->user<Val2Terms*>();
|
|
|
|
|
if (!val2Termspr) {
|
|
|
|
|
// Remeber and allocate on first encounter
|
|
|
|
|
srcps.emplace_back(srcp);
|
|
|
|
|
val2TermsAllocator.emplace_back();
|
|
|
|
|
val2Termspr = &val2TermsAllocator.back();
|
|
|
|
|
}
|
|
|
|
|
// Record term
|
|
|
|
|
(*val2Termspr)[val].emplace_back(&vtx, inv);
|
|
|
|
|
++nTerms;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Somewhat arbitrarily, only apply if more than 64 unique comparisons are required
|
|
|
|
|
constexpr uint32_t TERM_LIMIT = 65;
|
|
|
|
|
// This should hold, otherwise we do redundant work gathering terms that will never be used
|
|
|
|
|
static_assert((1U << WIDTH_MIN) >= TERM_LIMIT, "TERM_LIMIT too big relative to 2**WIDTH_MIN");
|
|
|
|
|
|
|
|
|
|
// Fast path exit if we surely don't need to convet anything
|
|
|
|
|
if (nTerms < TERM_LIMIT) return;
|
|
|
|
|
|
|
|
|
|
// Sequence numbers for name generation
|
|
|
|
|
size_t nTables = 0;
|
|
|
|
|
|
|
|
|
|
// Create decoders for each srcp
|
|
|
|
|
for (DfgVertex* const srcp : srcps) {
|
|
|
|
|
const Val2Terms& val2Terms = *srcp->getUser<Val2Terms*>();
|
|
|
|
|
|
|
|
|
|
// If not enough terms in this vertex, ignore
|
|
|
|
|
if (val2Terms.size() < TERM_LIMIT) continue;
|
|
|
|
|
|
|
|
|
|
// Width of the decoded binary value
|
|
|
|
|
const uint32_t width = srcp->width();
|
|
|
|
|
// Number of bits in the input operand
|
|
|
|
|
const uint32_t nBits = 1U << width;
|
|
|
|
|
|
|
|
|
|
// Construct the decoder by converting many "const == vtx" by:
|
|
|
|
|
// - Adding a single decoder block, where 'tab' is zero initialized:
|
|
|
|
|
// always_comb begin
|
|
|
|
|
// tab[pre] = 0;
|
|
|
|
|
// tab[vtx] = 1;
|
|
|
|
|
// pre = vtx;
|
|
|
|
|
// end
|
|
|
|
|
// We mark 'pre' so the write is ignored during scheduling, so this
|
|
|
|
|
// won't cause a combinational cycle.
|
|
|
|
|
// Note that albeit this looks like partial udpates to 'tab', the
|
|
|
|
|
// actual result is that only one value in 'tab' is ever one, while
|
|
|
|
|
// all the others are always zero.
|
|
|
|
|
// - and replace the comparisons with 'tab[const]'
|
|
|
|
|
|
|
|
|
|
FileLine* const flp = srcp->fileline();
|
|
|
|
|
|
|
|
|
|
// Required data types
|
|
|
|
|
AstNodeDType* const idxDTypep = srcp->dtypep();
|
|
|
|
|
AstNodeDType* const bitDTypep = DfgVertex::dtypeForWidth(1);
|
|
|
|
|
AstUnpackArrayDType* const tabDTypep = new AstUnpackArrayDType{
|
|
|
|
|
flp, bitDTypep, new AstRange{flp, static_cast<int>(nBits - 1), 0}};
|
|
|
|
|
v3Global.rootp()->typeTablep()->addTypesp(tabDTypep);
|
|
|
|
|
|
|
|
|
|
// The index variable
|
|
|
|
|
DfgVarPacked* const idxVtxp = [&]() {
|
|
|
|
|
// If there is an existing result variable, use that, otherwise create a new variable
|
2025-07-14 23:09:34 +02:00
|
|
|
DfgVarPacked* varp = nullptr;
|
|
|
|
|
if (DfgVertexVar* const vp = srcp->getResultVar()) {
|
|
|
|
|
varp = vp->as<DfgVarPacked>();
|
|
|
|
|
} else {
|
2025-06-17 00:14:24 +02:00
|
|
|
const std::string name = dfg.makeUniqueName("BinToOneHot_Idx", nTables);
|
2025-07-01 23:55:08 +02:00
|
|
|
varp = dfg.makeNewVar(flp, name, idxDTypep, nullptr)->as<DfgVarPacked>();
|
2025-06-17 00:14:24 +02:00
|
|
|
varp->varp()->isInternal(true);
|
2025-07-14 23:09:34 +02:00
|
|
|
varp->srcp(srcp);
|
2025-06-17 00:14:24 +02:00
|
|
|
}
|
|
|
|
|
varp->setHasModRefs();
|
|
|
|
|
return varp;
|
|
|
|
|
}();
|
|
|
|
|
// The previous index variable - we don't need a vertex for this
|
|
|
|
|
AstVar* const preVarp = [&]() {
|
|
|
|
|
const std::string name = dfg.makeUniqueName("BinToOneHot_Pre", nTables);
|
|
|
|
|
AstVar* const varp = new AstVar{flp, VVarType::MODULETEMP, name, idxDTypep};
|
|
|
|
|
dfg.modulep()->addStmtsp(varp);
|
|
|
|
|
varp->isInternal(true);
|
|
|
|
|
varp->noReset(true);
|
|
|
|
|
varp->setIgnoreSchedWrite();
|
|
|
|
|
return varp;
|
|
|
|
|
}();
|
|
|
|
|
// The table variable
|
|
|
|
|
DfgVarArray* const tabVtxp = [&]() {
|
|
|
|
|
const std::string name = dfg.makeUniqueName("BinToOneHot_Tab", nTables);
|
2025-07-01 23:55:08 +02:00
|
|
|
DfgVarArray* const varp
|
|
|
|
|
= dfg.makeNewVar(flp, name, tabDTypep, nullptr)->as<DfgVarArray>();
|
2025-06-17 00:14:24 +02:00
|
|
|
varp->varp()->isInternal(true);
|
|
|
|
|
varp->varp()->noReset(true);
|
|
|
|
|
varp->setHasModRefs();
|
|
|
|
|
return varp;
|
|
|
|
|
}();
|
|
|
|
|
|
|
|
|
|
++nTables;
|
|
|
|
|
++ctx.m_decodersCreated;
|
|
|
|
|
|
|
|
|
|
// Initialize 'tab' and 'pre' variables statically
|
|
|
|
|
AstInitialStatic* const initp = new AstInitialStatic{flp, nullptr};
|
|
|
|
|
dfg.modulep()->addStmtsp(initp);
|
|
|
|
|
{ // pre = 0
|
|
|
|
|
initp->addStmtsp(new AstAssign{
|
|
|
|
|
flp, //
|
|
|
|
|
new AstVarRef{flp, preVarp, VAccess::WRITE}, //
|
|
|
|
|
new AstConst{flp, AstConst::WidthedValue{}, static_cast<int>(width), 0}});
|
|
|
|
|
}
|
|
|
|
|
{ // tab.fill(0)
|
|
|
|
|
AstCMethodHard* const callp = new AstCMethodHard{
|
|
|
|
|
flp, new AstVarRef{flp, tabVtxp->varp(), VAccess::WRITE}, "fill"};
|
|
|
|
|
callp->addPinsp(new AstConst{flp, AstConst::BitFalse{}});
|
|
|
|
|
callp->dtypeSetVoid();
|
|
|
|
|
initp->addStmtsp(callp->makeStmt());
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Build the decoder logic
|
|
|
|
|
AstAlways* const logicp = new AstAlways{flp, VAlwaysKwd::ALWAYS_COMB, nullptr, nullptr};
|
|
|
|
|
dfg.modulep()->addStmtsp(logicp);
|
|
|
|
|
{ // tab[pre] = 0;
|
|
|
|
|
logicp->addStmtsp(new AstAssign{
|
|
|
|
|
flp, //
|
|
|
|
|
new AstArraySel{flp, new AstVarRef{flp, tabVtxp->varp(), VAccess::WRITE},
|
|
|
|
|
new AstVarRef{flp, preVarp, VAccess::READ}}, //
|
|
|
|
|
new AstConst{flp, AstConst::BitFalse{}}});
|
|
|
|
|
}
|
|
|
|
|
{ // tab[idx] = 1
|
|
|
|
|
logicp->addStmtsp(new AstAssign{
|
|
|
|
|
flp, //
|
|
|
|
|
new AstArraySel{flp, new AstVarRef{flp, tabVtxp->varp(), VAccess::WRITE},
|
|
|
|
|
new AstVarRef{flp, idxVtxp->varp(), VAccess::READ}}, //
|
|
|
|
|
new AstConst{flp, AstConst::BitTrue{}}});
|
|
|
|
|
}
|
|
|
|
|
{ // pre = idx
|
|
|
|
|
logicp->addStmtsp(new AstAssign{flp, //
|
|
|
|
|
new AstVarRef{flp, preVarp, VAccess::WRITE}, //
|
|
|
|
|
new AstVarRef{flp, idxVtxp->varp(), VAccess::READ}});
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Replace terms with ArraySels
|
|
|
|
|
for (const auto& pair : val2Terms) {
|
|
|
|
|
const uint32_t val = pair.first;
|
|
|
|
|
const std::vector<Term>& terms = pair.second;
|
|
|
|
|
// Create the ArraySel
|
2025-06-28 18:29:41 +02:00
|
|
|
FileLine* const aflp = terms.front().m_vtxp->fileline();
|
|
|
|
|
DfgArraySel* const aselp = new DfgArraySel{dfg, aflp, bitDTypep};
|
2025-06-17 00:14:24 +02:00
|
|
|
aselp->fromp(tabVtxp);
|
2025-06-28 18:29:41 +02:00
|
|
|
aselp->bitp(new DfgConst{dfg, aflp, width, val});
|
2025-06-17 00:14:24 +02:00
|
|
|
// The inverted value, if needed
|
|
|
|
|
DfgNot* notp = nullptr;
|
|
|
|
|
// Repalce the terms
|
|
|
|
|
for (const Term& term : terms) {
|
|
|
|
|
if (term.m_inv) {
|
|
|
|
|
if (!notp) {
|
2025-06-28 18:29:41 +02:00
|
|
|
notp = new DfgNot{dfg, aflp, bitDTypep};
|
2025-06-17 00:14:24 +02:00
|
|
|
notp->srcp(aselp);
|
|
|
|
|
}
|
|
|
|
|
term.m_vtxp->replaceWith(notp);
|
|
|
|
|
} else {
|
|
|
|
|
term.m_vtxp->replaceWith(aselp);
|
|
|
|
|
}
|
|
|
|
|
VL_DO_DANGLING(term.m_vtxp->unlinkDelete(dfg), term.m_vtxp);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2024-03-02 20:49:29 +01:00
|
|
|
void V3DfgPasses::eliminateVars(DfgGraph& dfg, V3DfgEliminateVarsContext& ctx) {
|
|
|
|
|
const auto userDataInUse = dfg.userDataInUse();
|
|
|
|
|
|
|
|
|
|
// Head of work list. Note that we want all next pointers in the list to be non-zero
|
|
|
|
|
// (including that of the last element). This allows us to do two important things: detect
|
|
|
|
|
// if an element is in the list by checking for a non-zero next pointer, and easy
|
|
|
|
|
// prefetching without conditionals. The address of the graph is a good sentinel as it is a
|
|
|
|
|
// valid memory address, and we can easily check for the end of the list.
|
|
|
|
|
DfgVertex* const sentinelp = reinterpret_cast<DfgVertex*>(&dfg);
|
|
|
|
|
DfgVertex* workListp = sentinelp;
|
|
|
|
|
|
|
|
|
|
// Add all variables to the initial work list
|
2024-03-26 00:06:25 +01:00
|
|
|
for (DfgVertexVar& vtx : dfg.varVertices()) {
|
|
|
|
|
vtx.setUser<DfgVertex*>(workListp);
|
|
|
|
|
workListp = &vtx;
|
2024-03-02 20:49:29 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
const auto addToWorkList = [&](DfgVertex& vtx) {
|
|
|
|
|
// If already in work list then nothing to do
|
|
|
|
|
DfgVertex*& nextInWorklistp = vtx.user<DfgVertex*>();
|
|
|
|
|
if (nextInWorklistp) return;
|
|
|
|
|
// Actually add to work list.
|
|
|
|
|
nextInWorklistp = workListp;
|
|
|
|
|
workListp = &vtx;
|
|
|
|
|
};
|
|
|
|
|
|
2025-07-01 23:55:08 +02:00
|
|
|
// List of variables (AstVar or AstVarScope) we are replacing
|
|
|
|
|
std::vector<AstNode*> replacedVariables;
|
2024-03-05 19:48:31 +01:00
|
|
|
// AstVar::user1p() : AstVar* -> The replacement variables
|
2025-07-01 23:55:08 +02:00
|
|
|
// AstVarScope::user1p() : AstVarScope* -> The replacement variables
|
2024-03-05 19:48:31 +01:00
|
|
|
const VNUser1InUse user1InUse;
|
2024-03-02 20:49:29 +01:00
|
|
|
|
|
|
|
|
// Process the work list
|
|
|
|
|
while (workListp != sentinelp) {
|
|
|
|
|
// Pick up the head of the work list
|
|
|
|
|
DfgVertex* const vtxp = workListp;
|
|
|
|
|
// Detach the head
|
|
|
|
|
workListp = vtxp->getUser<DfgVertex*>();
|
2025-06-15 19:12:37 +02:00
|
|
|
// Reset user pointer so it can be added back to the work list later
|
|
|
|
|
vtxp->setUser<DfgVertex*>(nullptr);
|
2024-03-02 20:49:29 +01:00
|
|
|
// Prefetch next item
|
|
|
|
|
VL_PREFETCH_RW(workListp);
|
|
|
|
|
|
|
|
|
|
// Remove unused non-variable vertices
|
|
|
|
|
if (!vtxp->is<DfgVertexVar>() && !vtxp->hasSinks()) {
|
|
|
|
|
// Add sources of removed vertex to work list
|
|
|
|
|
vtxp->forEachSource(addToWorkList);
|
|
|
|
|
// Remove the unused vertex
|
|
|
|
|
vtxp->unlinkDelete(dfg);
|
2024-03-23 23:12:43 +01:00
|
|
|
continue;
|
2024-03-02 20:49:29 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// We can only eliminate DfgVarPacked vertices at the moment
|
|
|
|
|
DfgVarPacked* const varp = vtxp->cast<DfgVarPacked>();
|
|
|
|
|
if (!varp) continue;
|
|
|
|
|
|
|
|
|
|
// Can't remove if it has external drivers
|
|
|
|
|
if (!varp->isDrivenFullyByDfg()) continue;
|
|
|
|
|
|
|
|
|
|
// Can't remove if must be kept (including external, non module references)
|
|
|
|
|
if (varp->keep()) continue;
|
|
|
|
|
|
|
|
|
|
// Can't remove if referenced in other DFGs of the same module (otherwise might rm twice)
|
|
|
|
|
if (varp->hasDfgRefs()) continue;
|
|
|
|
|
|
|
|
|
|
// If it has multiple sinks, it can't be eliminated
|
|
|
|
|
if (varp->hasMultipleSinks()) continue;
|
|
|
|
|
|
|
|
|
|
if (!varp->hasModRefs()) {
|
|
|
|
|
// If it is only referenced in this DFG, it can be removed
|
|
|
|
|
++ctx.m_varsRemoved;
|
2025-07-14 23:09:34 +02:00
|
|
|
varp->replaceWith(varp->srcp());
|
2025-07-01 23:55:08 +02:00
|
|
|
varp->nodep()->unlinkFrBack()->deleteTree();
|
2025-07-14 23:09:34 +02:00
|
|
|
} else if (DfgVarPacked* const driverp = varp->srcp()->cast<DfgVarPacked>()) {
|
2025-07-21 19:32:08 +02:00
|
|
|
// If it's driven from another variable, it can be replaced by that.
|
2024-03-05 19:48:31 +01:00
|
|
|
// Mark it for replacement
|
2024-03-02 20:49:29 +01:00
|
|
|
++ctx.m_varsReplaced;
|
|
|
|
|
UASSERT_OBJ(!varp->hasSinks(), varp, "Variable inlining should make this impossible");
|
2025-07-01 23:55:08 +02:00
|
|
|
// Grab the AstVar/AstVarScope
|
|
|
|
|
AstNode* const nodep = varp->nodep();
|
|
|
|
|
UASSERT_OBJ(!nodep->user1p(), nodep, "Replacement already exists");
|
|
|
|
|
replacedVariables.emplace_back(nodep);
|
|
|
|
|
nodep->user1p(driverp->nodep());
|
2024-03-02 20:49:29 +01:00
|
|
|
} else {
|
|
|
|
|
// Otherwise this *is* the canonical var
|
|
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Add sources of redundant variable to the work list
|
|
|
|
|
vtxp->forEachSource(addToWorkList);
|
|
|
|
|
// Remove the redundant variable
|
|
|
|
|
vtxp->unlinkDelete(dfg);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Job done if no replacements possible
|
2024-03-05 19:48:31 +01:00
|
|
|
if (replacedVariables.empty()) return;
|
2024-03-02 20:49:29 +01:00
|
|
|
|
2025-07-01 23:55:08 +02:00
|
|
|
// Apply variable replacements
|
|
|
|
|
if (AstModule* const modp = dfg.modulep()) {
|
|
|
|
|
modp->foreach([&](AstVarRef* refp) {
|
|
|
|
|
AstVar* varp = refp->varp();
|
|
|
|
|
while (AstVar* const replacep = VN_AS(varp->user1p(), Var)) varp = replacep;
|
|
|
|
|
refp->varp(varp);
|
|
|
|
|
});
|
|
|
|
|
} else {
|
|
|
|
|
v3Global.rootp()->foreach([&](AstVarRef* refp) {
|
|
|
|
|
AstVarScope* vscp = refp->varScopep();
|
|
|
|
|
while (AstVarScope* const replacep = VN_AS(vscp->user1p(), VarScope)) vscp = replacep;
|
|
|
|
|
refp->varScopep(vscp);
|
|
|
|
|
refp->varp(vscp->varp());
|
|
|
|
|
});
|
|
|
|
|
}
|
2024-03-02 20:49:29 +01:00
|
|
|
|
|
|
|
|
// Remove the replaced variables
|
2025-07-01 23:55:08 +02:00
|
|
|
for (AstNode* const nodep : replacedVariables) nodep->unlinkFrBack()->deleteTree();
|
2024-03-02 20:49:29 +01:00
|
|
|
}
|
|
|
|
|
|
2025-07-26 21:37:01 +02:00
|
|
|
void V3DfgPasses::optimize(DfgGraph& dfg, V3DfgContext& ctx) {
|
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
|
|
|
// There is absolutely nothing useful we can do with a graph of size 2 or less
|
|
|
|
|
if (dfg.size() <= 2) return;
|
|
|
|
|
|
|
|
|
|
int passNumber = 0;
|
|
|
|
|
|
2022-11-27 11:52:40 +01:00
|
|
|
const auto apply = [&](int dumpLevel, const string& name, std::function<void()> pass) {
|
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
|
|
|
pass();
|
2023-05-04 00:04:10 +02:00
|
|
|
if (dumpDfgLevel() >= dumpLevel) {
|
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
|
|
|
const string strippedName = VString::removeWhitespace(name);
|
|
|
|
|
const string label
|
|
|
|
|
= ctx.prefix() + "pass-" + cvtToStr(passNumber) + "-" + strippedName;
|
|
|
|
|
dfg.dumpDotFilePrefixed(label);
|
|
|
|
|
}
|
|
|
|
|
++passNumber;
|
|
|
|
|
};
|
|
|
|
|
|
2023-05-04 00:04:10 +02:00
|
|
|
if (dumpDfgLevel() >= 8) dfg.dumpDotAllVarConesPrefixed(ctx.prefix() + "input");
|
2022-11-06 16:42:01 +01:00
|
|
|
apply(3, "input ", [&]() {});
|
|
|
|
|
apply(4, "inlineVars ", [&]() { inlineVars(dfg); });
|
|
|
|
|
apply(4, "cse0 ", [&]() { cse(dfg, ctx.m_cseContext0); });
|
2025-07-01 23:55:08 +02:00
|
|
|
if (dfg.modulep()) {
|
|
|
|
|
apply(4, "binToOneHot ", [&]() { binToOneHot(dfg, ctx.m_binToOneHotContext); });
|
|
|
|
|
}
|
2022-09-27 14:53:07 +02:00
|
|
|
if (v3Global.opt.fDfgPeephole()) {
|
2022-11-06 16:42:01 +01:00
|
|
|
apply(4, "peephole ", [&]() { peephole(dfg, ctx.m_peepholeContext); });
|
|
|
|
|
// We just did CSE above, so without peephole there is no need to run it again these
|
|
|
|
|
apply(4, "cse1 ", [&]() { cse(dfg, ctx.m_cseContext1); });
|
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
|
|
|
}
|
2024-03-02 20:49:29 +01:00
|
|
|
// Accumulate patterns for reporting
|
|
|
|
|
if (v3Global.opt.stats()) ctx.m_patternStats.accumulate(dfg);
|
|
|
|
|
apply(4, "regularize", [&]() { regularize(dfg, ctx.m_regularizeContext); });
|
2023-05-04 00:04:10 +02:00
|
|
|
if (dumpDfgLevel() >= 8) dfg.dumpDotAllVarConesPrefixed(ctx.prefix() + "optimized");
|
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
|
|
|
}
|