DevMeeting - 2021-06-29/30

[sequencer]

Pinned Items

• remove DefRegInit, change DefReg API with option definition. #1944: Unify Chisel internal IR of registers
  • Needs: Related PR to chisel-test to either stop using this API or fix the issue
• Module with multiple memories have no readmem statements firrtl#2168: Annoying issue where Dedup breaks some Verilog customization annotations
  • SiFive intern still making progress
• Espresso Decoder #1964:
  • [@sequencer]Status: Espresso now hosted by CHIPS Alliance (https://github.com/chipsalliance/espresso)
    • Version 2.4 released that compiles with modern C compilers
    • We could consider shipping binaries as resources (they’re ~250KB)
    • Alternatively just tell people how to install it themselves
    • Will be packaged by Arch and Debian soonish
• Fix CloneModuleAsRecord support for .toTarget #1974: Fix .toTarget and .toAbsoluteTarget on CloneModuleAsRecord Records
  • Highlights weirdness of internals and various important assumptions
  • Needs: Jack to fix 1 thing and would like review

Discussion 1

• Do another minor release? 3.4.4
  • [@jackkoenig] Would like firrtl#2168 and chisel3#1974
  • [@seldridge] There are regressions in FIRRT in 1.4.3 which nobody has had time to fix, yet, e.g., crash 2651 firrtl#2129
• Splitting rocket-chip stuff from dsptools Remove rocket sub project ucb-bar/dsptools#233
  • Pending PRs from jiuyang https://github.com/sequencer/chipsalliance-playground#pending-prs
• Feedback from China Users in CCC #1990
• Options for Verilog switch-case emission
  • Infer and emit from mux structure (easy, but perhaps brittle)
    • @seldridge got like 80% of the way to implementing this
    • [@seldridge] I have a branch: https://github.com/chipsalliance/firrtl/tree/case
  • Add FIRRTL IR nodes for this and Chisel emit them
  • FIRRTL IR is what it is, using CIRCT we could emit a different IR directly
  • [@sequencer] switch-case is basically a decoder,
  • [@jackkoenig] In my opinion, we should do (a), max benefit with minimal work

Discussion 2

• Discussions look great!
  • Pinned Announcement on top is really helpful: https://github.com/chipsalliance/chisel3/discussions
• [@jackkoenig] I broke Aspects in the CloneModuleAsRecord fix (Fix CloneModuleAsRecord support for .toTarget #1974)
  • 3 options
    • Does not show up in getDeep and collectDeep (old behavior) - Verdict
    • Return Nil for every API except ios when given CloneModule
    • Error for every API except ios when given CloneModule
  • Old behavior is most consistent with the behavior described below for InstanceAPI
• [@azidar] Should we deprecate .toAbsoluteTarget
  • In InstanceAPI prototype, Module definitions do not correspond to an instance, and thus .toAbsoluteTarget does not make sense for any of their components
    • .toTarget can still work
  • .toAbsoluteTarget (or some equivalent API) does make sense for Data from Instances
    • This API is just .toTarget, when doing instance access stuff (eg. a(_.b)(_.reg)), that accessing provides context such that .toTarget will return a non-local target (eg. A/b:B>reg)
  • How does this work in Aspects?
    • All methods in Aspects that take BaseModule will be overloaded to take Instance[T]
    • Select.instances will be deprecated in favor of .definitions(...): Iterable[BaseModule]
      • There will also be some new API (.insts?) that returns Iterable[Instance[BaseModule]]
      • Does Instance[A] need to be covariant?
    • Select will add instance equivalents for collectDeep and getDeep, those will provide absolute context such that .toTarget on - any Data from them will
  • Verdict
    • Deprecate .toAbsoluteTarget in 3.5 with introduction of Instance[] API
    • Deprecate CloneModuleAsRecord at same time
    • toTarget may (will?) return non-local targets for Instances[]
      What should the “template” or “definition” of a Module be called in the Instance API?
  • @jackkoenig keeps calling them definitions but what should we call them?
  • @azidar's original name is template, this makes sense from the context of metaprogramming
  • Arguably, the Scala Module class is the “definition” (generator definition)
  • The template/definition instantiation of that class could be called the module or IR definition, but perhaps “template” is better.

[seldridge]

As a total meta-comment, I really like this style of logging the meeting notes. Nice thinking @sequencer. 💯

[sequencer]

Finally we can truly log what happens for each dev-meetings!

[davidbiancolin]

For the uninitiated, is Discussion 1 = Monday vs Discussion 2 = Tuesday?

[seldridge]

Yes!