Skip to content

cleanup.md

Latest commit

 

History

History
149 lines (101 loc) · 4.77 KB

cleanup.md

File metadata and controls

149 lines (101 loc) · 4.77 KB
Raw

Small Cleanup Tasks

This document has small cleanup/refactoring tasks

Create a source_map out of the Token_list

Create a lh directory

-lh::tree -graph_core -str_tools

Basic data structures used all over but very LiveHD independent

Renaming

  • Change lgtuple to lbundle (Lgtupe -> Lbundle and lgtuple -> lbundle). The reason is that tuples are ordered, and the lbundle can be ordered and unordered. Also, the lg implies only a Lgraph structure but LNAST should use it too for emulation.

  • Rename lconst to lops (livehd cell ops). Other implementations would be sops (static C++), vops (verilog), and jops (javascript)

  • Rename TupAdd for Tup_add, TupGet for Tup_get, AttrSet for Attr_set, AttrGet for Attr_get

Pyrope parser

There are some bugs on the pyrope parser when dealing with ranges. This works:

foo = bar@(1:fcall(3))

But this fails (and it should work too):

foo = bar@(fcall(7):fcall(3))

Similarly, this works but it should fail (ambiguous):

foo = bar@(1+2:4) // bar@((1+2):4) is OK

Related, the bit selector should use [] not () and support commas. It should be:

foo = bar@[1:3]       // pick bits 1,2,3
foo = bar@[1+2:4, 10] // pick bits 3,4,10

The @[a,b] creates a one-hot bitmask to selects bits (feed to get_mask). The same is also used in the for loop.

for a in @[0xF] { // 0,1,2,3 iterations
for a in 0xF { // 15 iteration
for b in @[8,1:3,2] { // 1,2,3,8   iterations (bitmask order)
for b in (8,1:3,2)  { // 8,1,2,3,2 iterations (ordered tuple)

Allow tuples directly in brackets, no need for unnecessary parenthesis.

Currently, this is a compile error (parser)

xx = delay[1,2]

It requires this:

xx = delay[(1,2)]

Option 2 may be fine, but option 1 should be the same

Attr/Tup

  • Do not chain tuples without need (lnast.cpp). TupAdd/Get can handle position like foo.bar.xxx

  • Parser to directly generate Tupple_get, Tuple_set. This removes from LNAST the select field. Quite a bit of code from lnast.cpp like the set2attr_set_get could be removed

  • No attr_get/set from lnast_tolg. Also, no need to chain Attr like now. When cprop does the lgtuples, the attr fields gets generated as AttrSet/Get only when TupAdd/Get uses that attribute or the parent.

  • Once all the # is gone, the prp_lnast should create the __create_flop automatically out of #

  • Attr should be only for scalar

  • Ann_ssa does not need to be persistent

pass.compile

  • In FIRRTL no need to set firrtl:true. If input is a proto, it is firrtl
  • In FIRRTL no need to set the top always. If no top set, treat a proto as a top by itself

lgshell

  • autocompletion for LGDB names. E.g: lgraph.open name:a should match any module starting with a
  • fuzzy backward search

lnast-ssa

  • Clean up analysis based on dot, now the lnast should contains select [] only.
  • Modify fir_tolnast to generate select-only lnast tuple.

lconst

The lconst implementation of "has_unknowns" is not very efficient. It uses a string to represent things like 0b?0?. It may be more efficient to have 2 Numbers. The current num and unk_num. If the bit is set to 1, the bit is unknown. This binary representation allows for faster operations with uknowns (it should not be frequent, but it can happen).

In C++ generated simulation, the computation can be 2 or 4 bit logic. Inside LGraph/LNAST, the computation is always 4 bit logic.

node_type_area

The node_type_area is a "property" for each cell type (not per instance). It should be a per class attribute (not instance attribute). A small difference with normal attributes is that "base" cell types have also this property, not only the tech file. A tech file could also add area properties, so not just for cells.

  • Rename/rework node_type_area (sub_area.hpp).

  • Populate the default areas for each cell (this should add cells like __sum to the sub_node.

  • It is not a plain attribute because the value depends on the cell use. E.g: a sum with 2 bits is different than a sub with 100 bits.

  • There should be also be an option to have a per-instance area modification (factor?) but most cells should not have this, maybe some sub_node instances)

  • Summary: create a new sub_area class that tracks area per cell sub_node instance, and allows to adjust area per instance in some cases. The floorplanner could use this enhanced capability to track area.

pass/fplan

  • Many methods/variables use camelCase, it should be camel_case
  • "int" is used all over. Even for things that should be size_t (size of stuff)
  • node_tree is only used by pass/fplan. Move from core to pass/fplan
  • Lots of variable declarations c-style (early in the method, then assigned later. Use later on first assign)
  • Only classes can start with upper case (Name2Count -> name2count)
  • Classes should start uppercase. E.g: Ntype_area::dim