The DFG peephole pass converts all associative trees into right leaning,
which is good for simplifying pattern recognition, but can lead to an
excessive amount of wide intermediate results being constructed for
right leaning concatenations.
Add a new pass to balance concatenation trees by trying to:
- Create VL_EDATASIZE (32-bit) sub-terms, so words can then be packed
easily afterwards
- Try to ensure the operands of a concat are roughly the same width
within a concatenation tree. This does not yield the shortest tree,
but it ensures it has many sub-nodes that are small enough to fit into
machine registers.
This can eliminate a lot of wide intermediate results, which would need
temporaries, and also increases ILP within sub-expressions (assuming the
C compiler can't figure that out itself).
This is over 2x run-time speedup on the high_perf configuration of
VeeR EH2 (which you could arguably also get with -fno-dfg, but oh well).
This functionality used to be distributed in the removeVars pass and the
final dfgToAst conversion. Instead added a new 'regularize' pass to
convert DFGs into forms that can be trivially converted back to Ast, and
a new 'eliminateVars' pass to remove/repalce redundant variables. This
simplifies dfgToAst significantly and makes the code a bit easier to
follow.
The new 'regularize' pass will ensure that every sub-expression with
multiple uses is assigned to a temporary (unless it's a trivial memory
reference or constant), and will also eliminate or replace redundant
variables. Overall it is a performance neutral change but it does
enable some later improvements which required the graph to be in this
form, and this also happens to be the form required for the dfgToAst
conversion.
Replace the 'run to fixed point' algorithm with a work list driven
approach. Instead of marking the graph as changed, we explicitly add
vertices to the work list, to be visited, when a vertex is changed. This
improves both memory locality (as the work list is processed in last in
first out order), and removed unnecessary visitations when only a few
nodes changes.
Apart from the representational changes below, this patch renames
AstNodeMath to AstNodeExpr, and AstCMath to AstCExpr.
Now every expression (i.e.: those AstNodes that represent a [possibly
void] value, with value being interpreted in a very general sense) has
AstNodeExpr as a super class. This necessitates the introduction of an
AstStmtExpr, which represents an expression in statement position, e.g :
'foo();' would be represented as AstStmtExpr(AstCCall(foo)). In exchange
we can get rid of isStatement() in AstNodeStmt, which now really always
represent a statement
Peak memory consumption and verilation speed are not measurably changed.
Partial step towards #3420
Vertices representing variables (DfgVertexVar) and constants (DfgConst)
are very common (40-50% of all vertices created in some large designs),
and we also need to, or can treat them specially in algorithms. Keep
these as separate lists in DfgGraph for direct access to them. This
improve verilation speed.
Cyclic components are now extracted separately, so there is no
functional reason to have to do a topological sort (previously we used it
to detect cyclic graphs). Removing it to gain some speed.
Added a DfgVertex::user() mechanism for storing data in vertices.
Similar in spirit to AstNode user data, but the generation counter is
stored in the DfgGraph the vertex is held under. Use this to cache
DfgVertex::hash results, and also speed up DfgVertex hashing in general.
Use these and additional improvements to speed up CSE.
Allow constant folding through adjacent nodes of all associative
operations, for example '((a & 2) & 3)' or '(3 & (2 & a))' can now be
folded into '(a & 2)' and '(2 & a)' respectively. Also improve speed of
making associative expression trees right leaning by using rotation of
the existing vertices whenever instead of allocation of new nodes.
Use the same style, and reuse the bulk of astgen to generate DfgVertex
related code. In particular allow for easier definition of custom
DfgVertex sub-types that do not directly correspond to an AstNode
sub-type. Also introduces specific names for the fixed arity vertices.
No functional change intended.
A lot of optimizations in DFG assume a DAG, but the more things are
representable, the more likely it is that a small cyclic sub-graph is
present in an otherwise very large graph that is mostly acyclic. In
order to avoid loosing optimization opportunities, we explicitly extract
the cyclic sub-graphs (which are the strongly connected components +
anything feeing them, up to variable boundaries) and treat them
separately. This enables optimization of the remaining input.
Added DfgVertexVariadic to represent DFG vetices with a varying number
of source operands. Converted DfgVar to be a variadic vertex, with each
driver corresponding to a fixed range of bits in the packed variable.
This allows us to handle AstSel on the LHS of assignments. Also added
support for AstConcat on the LHS by selecting into the RHS as
appropriate.
This improves OpenTitan ST speed by ~13%
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.
Wilson Snyder
Geza Lore
Bartłomiej Chmiel
Mariusz Glebocki
Anthony Donlon
Kamil Rakoczy