use li->roots instead of module constant table

[JeffBezanson]

Looking at #14556, I experimented with (1) disabling AST compression, and (2) returning constant tables to per-function instead of per-module. Here is the data:

	sysimg	tests	peak mem
master	29.49MB	1788 s	975MB
no compression	30.07MB	1787 s	1669MB
li->roots	29.83MB	1547 s	911MB

With the new GC, AST compression is not as crucial as it used to be. Performance is on par (some things were faster and some were slower) but memory use is way higher. However using function-local constant tables wins hands-down by removing most of the O(n^2) constant lookup while retaining the benefits of compression. The only benefit of per-module constant tables is system image size, but not by much. Also I suspect if the common symbols lists were recomputed for the no-compression case its system image would improve a lot.

The subarray test accounts for almost all of the improvement; it got ~100s faster. In that test the constant tables are huge, and consist almost entirely of types. In the no-compression case the subarray test is also 100s faster than master.

So for now I recommend ripping out the per-module constant table since it's an easy win. Long term, we should get rid of AST compression by addressing memory use in that case. Ideas there:

• Replace all remaining Exprs in our IR with more compact specialized objects.
• Delete more ASTs. There can be heuristics for ASTs not likely to be needed. For example specialized ASTs too big to be inlined (and not declared inline).
• Or... switch to a bytecode representation. Most efficient but difficult to do.

[simonster]

Does this fix #9375? cc @mlubin

[mlubin]

@simonster, that code isn't used anymore, so it would be a bit of work to dig it up and test

[tkelman]

to backport this to release-0.4 (for #14113), there are some remaining mentions of constant_table here

julia/src/gf.c

Lines 1598 to 1600 in 3b06c86

	if (m->constant_table != NULL) {
	for(i=0; i < jl_array_len(m->constant_table); i++) {
	jl_value_t *el = jl_cellref(m->constant_table,i);

- is that block safe to remove?

[JeffBezanson]

It won't affect normal operation, but will lead to --compile=all missing some functions.