The original code called V3Const::constifyParamsEdit() in
V3Width::visit(AstInsideRange*) to fold negative bin bounds
(e.g., NEGATE(100) -> -100) for covergroup bins. Since
constifyParamsEdit uses m_required=true, it errored on any
non-constant expression -- breaking the upstream test
t_constraint_array_index_simple which uses a foreach loop
variable 'i' in an inside constraint expression.
Fix: Remove the unconditional constifyParamsEdit calls from
V3Width::visit(AstInsideRange*). Instead:
- In V3Width: when m_vup is null (covergroup bin context,
not an expression context), use constifyEdit (m_required=false)
to fold constant arithmetic before iterating children.
- In V3Covergroup: call constifyEdit at the two sites that
cast InsideRange bounds to AstConst for bin expansion.
This moves the covergroup-specific concern to the right layer
(V3Covergroup) while keeping V3Width safe for all contexts.
Also update golden files for t_covergroup_trans_3value,
t_covergroup_trans_restart, and t_covergroup_with_sample_args_too_many_bad
whose line numbers shifted by 1 due to the SPDX copyright header
additions in the previous commit.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Replace Unicode em-dash and curly apostrophe with plain ASCII
equivalents to pass t_dist_whitespace CI check.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Coverage bin counters (__Vcov_* member variables) were being initialized
with VL_SCOPED_RAND_RESET_I because they are plain uint32 members with no
explicit initial value. This caused 0% coverage readings at runtime when
randReset != 0 (as used in CI).
Fix all 5 bin counter AstVar creation sites to set an explicit valuep of
AstConst(0), so V3EmitCFunc emits '= 0U' instead of VL_SCOPED_RAND_RESET_I.
Affects: regular bins, default bins, array bins, transition array bins,
and cross bins.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Commit b4244fc57 accidentally dropped the varp->lifetime(VLifetime::AUTOMATIC_EXPLICIT)
line when reverting V3LinkInc.cpp to a merge base. This caused __Vincrement* temp
variables to have VLifetime::NONE, triggering V3Fork's assertion that all variables
inside fork...join_none blocks have a known lifetime.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Two bugs in the covergroup -> AstClass transformation in V3LinkParse:
1. Infinite recursion: when a covergroup has a clocking event (e.g.
`@(posedge clk)`), visit(AstCovergroup*) embeds a sentinel
AstCovergroup node inside the new AstClass to carry the event for
V3Covergroup.cpp. The subsequent iterate(cgClassp) call then visits
the sentinel via visit(AstNodeModule*) -> iterateChildren -> which
hits visit(AstCovergroup*) again, creating another class with another
sentinel, infinitely.
Fix: skip transformation in visit(AstCovergroup*) when already inside
a covergroup class (m_modp->isCovergroup()), so sentinel nodes are
left alone.
2. Wrong fileline column: AstCovergroup was created with the fileline of
the identifier token ($<fl>2, the name position) rather than the
'covergroup' keyword token ($<fl>1). This caused warnings about the
covergroup to point to the name column instead of the keyword column.
Fix: use $<fl>1 (the 'covergroup' keyword fileline) when constructing
AstCovergroup in the parser.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
V3LinkParse's visit(AstCovergroup) creates std:: references and calls
setUsesStdPackage(). The previous removeStd() call happened before
V3LinkParse ran, so it deleted the std package before those references
were created, causing:
%Error: Package/class for ':: reference' not found: 'std'
Move removeStd() to immediately after V3LinkParse::linkParse() inside
process() so the std package is only pruned after all parse-time
transformations have had a chance to declare their std:: usage.
Fixes test failures:
- t_covergroup_option
- t_covergroup_with_sample_args_too_many_bad
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Three node deletions were accidentally dropped from the initial covergroup
commit as collateral damage:
- wait_order (no-stmt variant): restore DEL($3) for vrdList
- expect (no-stmt variant): restore DEL($3) for property_spec
- property_exprCaseIf yIF/yELSE: restore DEL($3) for condition expr
In all three cases $3 is an AstNode* that is not assigned to $$ and would
be leaked without the deletion.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Add v3Global.useCovergroup() flag (following the useRandSequence()
pattern) that is set to true when a covergroup_declaration is parsed.
Gate the V3Covergroup::covergroup() pass in Verilator.cpp on this flag
so the pass is skipped entirely for designs with no covergroups.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Two bugs fixed:
1. AstCReset: mark as ineligible for coverage expressions via
isExprCoverageEligible() override, preventing verilogForTree
from being called on CReset nodes (which has no V3EmitV handler).
2. generateCrossCode: when a cross references an unknown coverpoint,
don't delete the cross node early. The caller's cleanup loop
(in visit(AstClass*)) is responsible for deleting all coverpoints
and crosses. Early deletion left a dangling pointer in m_coverCrosses
causing a use-after-free segfault.
3. hasUnsupportedEvent path: added coverpoint/cross cleanup before
returning so AST nodes don't reach downstream passes (V3EmitCFunc,
V3MergeCond) which no longer have stub visitors for them.
All 60 covergroup tests now pass.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
The VLifetime::AUTOMATIC_EXPLICIT addition is not required for
functional coverage — all 60 covergroup tests pass without it.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
These changes are not required for functional coverage support and
should not be included in the upstream patch.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Introduce a new DfgAstRd vertex, which holds an AstNodeExpr that is a
reference to a variable. This enables tracking all read references in
Dfg, which both enables more optimization, and allows inlining of
expressions in place of the reference more intelligently (e.g, when the
expression is only used once, and is not in a loop). This can get rid of
20-30% of temporary variables introduced in Dfg in some designs. Note
V3Gate later got rid of a lot of those, this is a step towards making
V3Gate redundant. The more intelligent expression inlining is worth ~10%
runtime speed on some designs.
Matthew Ballance
github action
Wilson Snyder
Yilou Wang
em2machine
Geza Lore
Artur Bieniek