serialize bits values and QuoteNodes in IR encoding

[JeffBezanson]

Using this instead of the roots array avoids scaling problems in functions with many constants. This might also be better for precompiling. Grows the system image by 1.6%, but is probably worth it. Copying small bits values in code just seems like the right thing.

[JeffBezanson]

This seems to be a bit crashy? I can reproduce it sometimes but not always. E.g.

LinearAlgebra/addmul  (1) |        started at 2022-05-04T14:47:32.609

signal (11): Segmentation fault
in expression starting at /home/jeff/src/julia/usr/share/julia/stdlib/v1.9/LinearAlgebra/test/addmul.jl:133
symtab_lookup at /home/jeff/src/julia/src/symbol.c:59
_jl_symbol at /home/jeff/src/julia/src/symbol.c:89
ijl_uncompress_argnames at /home/jeff/src/julia/src/ircode.c:980
ijl_uncompress_ir at /home/jeff/src/julia/src/ircode.c:849
InliningTodo at ./compiler/ssair/inlining.jl:921 [inlined]
resolve_todo at ./compiler/ssair/inlining.jl:855
analyze_method! at ./compiler/ssair/inlining.jl:912

or

LinearAlgebra/addmul    (1) |        started at 2022-05-04T14:32:09.368

signal (11): Segmentation fault
in expression starting at /home/jeff/src/julia/usr/share/julia/stdlib/v1.9/LinearAlgebra/test/addmul.jl:133
jl_gc_wb at /home/jeff/src/julia/src/julia.h:911
Allocations: 1114935379 (Pool: 1112683407; Big: 2251972); GC: 1619
Segmentation fault (core dumped)

[JeffBezanson]

Ok, what happened was this: if a value is not in a method roots list, but is later needed by codegen, codegen will add it to the roots list, but de-duplicating values only by address instead of using ===. That can lead to codegen inserting huge numbers of copies of the same value as method roots. There were also cases where method roots contained QuoteNode(x), but codegen of course wants to emit x, so it would add a root for that as well leading to more duplication.

This PR now fully fixes that situation. Codegen doesn't need integer indices for the values it roots; it just needs them rooted somewhere. So I added a global IdDict of values that might be needed by generated code, so they are de-duplicated by === and there can be lots of sharing. That also avoids adding more entries to the method roots list, which is good because they use O(n) lookup. Plus the logic for getting a globally-rooted version of a value is now centralized. As a result the PR more than pays for itself as the system image is slightly smaller 🎉 .

Some possible future improvements to constant handling:

• Codegen adds a global root as soon as it sees a value, but it should do it more lazily, only if it really needs a boxed version of the value.
• CodeInfo could have its own list of constants, since if a constant is generated by compile-time evaluation it is probably specific to a certain specialization.
• Add special representations for Int8, Int16, UInt16, UInt32, UInt64, and Char to staticdata.c so we can take advantage of those caches as global roots (currently Int32, Int64, and UInt8 are handled).

[vtjnash]

sgtm

[lassepe]

Is there any chance that this gets backported to 1.8.x given that it also fixes some bugs we observed in Symbolics.jl/RuntimeGeneratedFunctions or is this considered a latency improvement only and thus waits until 1.9?

[JeffBezanson]

I think this can be backported.