LinkDotFindVisitor queued an AstIface into m_ifaceModSyms once for every
symbol table entry it was reached through. When creating scopes,
LinkDotIfaceVisitor resolves the modport items against the AstVarScopes,
which LinkDotScopeVisitor only inserts under the entry of an AstScope. A
hierarchical block wrapper builds a second copy of the instance tree in
the symbol table, and an interface reached through that copy has no
AstVarScopes to resolve against.
Resolving is destructive (the AstModportVarRef is deleted once resolved),
so this only worked when a resolvable entry happened to be processed
first, and failed with "Modport item not found" otherwise. When creating
scopes, register from the AstScopes instead, as those are exactly the
entries that hold the AstVarScopes.
Reproduces on t_hier_block with -fno-inline, added as a test.
AstNodeCoverDecl::hier() is relative to the scope the declaration is
emitted from, as V3EmitCImp builds the reported path as 'vlNamep + hierp',
and that is how V3Coverage uses it. V3FsmDetect instead set it to the
absolute scope name, so the instance path was counted twice.
This was masked whenever the owning module was inlined, as the declaration
then ended up in the top scope. With inlining disabled, an FSM in an
instance reported 'top.t.forced_wide_u.t.forced_wide_u'. In the inlined
case it reported 'top.TOP', leaking the internal top wrapper name into
user visible coverage output, rather than plain 'top'.
Nodes were always cloned under the new scope, with the originals deleted
at the end of the pass. Count total module instantiations up front, then
move the nodes instead of cloning them when scoping the last instance.
Required to avoid a memory regression in a follow up, but also faster
and less peak memory overall.
Output is identical.
Each AstVarRef is visited exactly once (blocks are iterated under
their per-scope clone), and the fixups are independent, so no need
for the ordered set.
Store input edges of fixed arity vertices inline in the vertex class.
This reduces heap allocations, memory fragmentation, and pointer chasing
and speeds up Dfg passes. The extra branch introduced in inputEdgep() is
well predictable and profiling shows branchless alternatives are a loss.
Wilson Snyder
Kamil Danecki
Geza Lore
Kaleb Barrett
Shriyans Sahoo
Marco Bartoli
Kristof Marien
Jonathan Drolet
Nick Brereton
Jesús Alberto López Soto