Relax the live variable analysis performed by Dfg to bail on fewer
cases. This analysis was already conservative (meaning it might think
variables are live when they are not), which is good enough for Dfg use.
This change in particular enables synthesizing more complex logic
involving arrays, e.g. those introduce by V3Table creating lookup
tables.
Teach DFG about CReset. This is not so much to optimize CReset itself, but to enable synthesizing logic involving CReset, which does appear with automatic variables used only in certain branches
Combine consecutive assertOn() checks into one, and hoist past enclosing
'if' statements if possible. This enables combining a lot of them, which
can be worth 10% performance on some assertion heavy designs depending
on how the assertions are written.
Turn `x = $sformatf(...)` into `$sformat(x, ...)`. The former requires
checking and running a destructor for `x` at the call site, the later
does it in the callee VL_SFORMAT. This reduces the size of the call
site, which can be significant e.g. in the presence of many assertions.
Also added a rewrite of `$sformat(x, "const-string")` back into `x =
"const-string"` for the cases where the `$sformatf` would have been
folded into a constant string.
Treat AstSFormat as a special form of assignment in V3Life. This allows
eliminating earlier redundant assignments to strings when an $sformat
later sets the string. UVM has lot of these.
V3Premit extracts wide sub-expressions via temporaries, which is needed
for emitting wide operations to C++ (calls to `VL_*_W`). The previous
version used to extract both branches of an AstCond unconditionally,
meaning both branches were fully evaluated. Rewriting the AstCond into
an AstIf instead enables evaluating only the required branch. While
this does limit V3Subst, overall the resulting code is ~3% faster,
and contains ~25% fewer branches on a large design.
If the same statements appears in both branches of an 'if', put a single
copy after the 'if', apply recursively. This also has the effect of
getting rid of conditionals with identical branches, but is more widely
applicable.
Geza Lore
pawelktk
Yilin Li
Artur Bieniek
Marco Bartoli
em2machine
github action
Nikolai Kumar
Ryszard Rozak
Matthew Ballance
Mateusz Gancarz
Yogish Sekhar
Nick Brereton
Todd Strader
Igor Zaworski
Krzysztof Bieganski
Kamil Danecki
Yilou Wang
Wilson Snyder
Muzaffer Kal
Zubin Jain