Implement vector and node conversion in Lua

[TurkeyMcMac]

As @paradust7 found and reported in #12438, constructing Lua strings and tables from C++ is slow. Looking at the code of PUC Lua and LuaJIT, this slowness affects calls such as lua_getfield, which have to construct/hash strings every time they are called. On the other hand, LuaJIT can optimize regular field accesses by storing strings and their hashes inline etc. This PR reimplements the functions for pushing and reading nodes and vectors using Lua. Unlike #12438, it does not touch specific API functions. Therefore, it will have wider reaching effects, but will probably not give as much improvement for get_node in particular. I have found that it indeed provides significant improvement.

To do

This PR is Ready for Review.

How to test

Run the unit tests, and maybe some other mods.

Benchmarks

First apply this patch for some extra benchmarking commands:

diff --git a/games/devtest/mods/experimental/commands.lua b/games/devtest/mods/experimental/commands.lua
index e42ae954d..1b67add1c 100644
--- a/games/devtest/mods/experimental/commands.lua
+++ b/games/devtest/mods/experimental/commands.lua
@@ -86,6 +86,60 @@ minetest.register_chatcommand("bench_bulk_set_node", {
 	end,
 })
 
+minetest.register_chatcommand("bench_get_node", {
+	params = "",
+	description = "Benchmark: get node",
+	func = function(name, param)
+		local player = minetest.get_player_by_name(name)
+		if not player then
+			return false, "No player."
+		end
+		local pos = player:get_pos()
+
+		minetest.chat_send_player(name, "Benchmarking minetest.get_node. Warming up ...");
+
+		for i = 1, 1000000 do
+			minetest.get_node(pos)
+			pos = pos:add(1)
+		end
+
+		minetest.chat_send_player(name, "Warming up finished, now benchmarking ...");
+
+		local middle_time = minetest.get_us_time()
+		pos = player:get_pos()
+		for i = 1, 1000000 do
+			minetest.get_node(pos)
+			pos = pos:add(1)
+		end
+		local end_time = minetest.get_us_time()
+		return true, ("Finished. Time: %.2fms"):format((end_time - middle_time) / 1000)
+	end,
+})
+
+minetest.register_chatcommand("bench_pos", {
+	params = "",
+	description = "Benchmark: get node",
+	func = function(name, param)
+		local player = minetest.get_player_by_name(name)
+		if not player then
+			return false, "No player."
+		end
+		local pos1 = player:get_pos():round()
+		local pos2 = pos1:add(110)
+
+		minetest.chat_send_player(name, "Benchmarking position creation. Warming up ...");
+
+		minetest.find_nodes_in_area(pos1, pos2, "air")
+
+		minetest.chat_send_player(name, "Warming up finished, now benchmarking ...");
+
+		local middle_time = minetest.get_us_time()
+		minetest.find_nodes_in_area(pos1, pos2, "air")
+		local end_time = minetest.get_us_time()
+		return true, ("Finished. Time: %.2fms"):format((end_time - middle_time) / 1000)
+	end,
+})
+
 local function advance_pos(pos, start_pos, advance_z)
 	if advance_z then
 		pos.z = pos.z + 2

Then run these benchmarks:

1. /bench_bulk_set_node: This benchmarks the reading of nodes and vectors. The set_node loop tests both in equal proportion, while bulk_set_node mainly just benchmarks reading vectors. I have noticed a 10-15% decrease in runtime of the set_node loop, and a 5-10% decrease in runtime of bulk_set_node.
2. /bench_get_node: This benchmarks the reading of vectors and the pushing of nodes. I have noticed a 10-40% decrease in the runtime of this benchmark.
3. /bench_pos: This benchmarks the pushing of vectors. I have noticed a 5-10% decrease in its runtime.

[Desour]

Some suggestions:

• The push and read functions could be called in lua to reduce the number of C++ to Lua calls (which are expensive, AFAIK).
  For example:

  objref_methods.set_pos = function(v) original_set_pos(read_vector(v)) end

  This would, however, require to wrap all those functions in builtin. (Or just do this for commonly used functions.)
• Instead of content_id, param1, param2, you could pack the node into one 32 bit number: content_id + 65536 * (param1 + 256 * param2)).
• Same for node positions: One could use hash_node_position after round.

[TurkeyMcMac]

@Desour

• I'll leave changes to specific functions to be done in Implement core.get_node() in Lua so that it can be optimized by LuaJIT #12438 and similar PRs.
• Perhaps I can try this, but it may not have a lot of impact relative to other changes.
• Same here.

[sfan5]

• Instead of content_id, param1, param2, you could pack the node into one 32 bit number: content_id + 65536 * (param1 + 256 * param2)).

• Same for node positions: One could use hash_node_position after round.

I don't see the point here. You're trading two lua_pushinteger calls for having to do several bit operations that LuaJIT probably doesn't optimize as well and RAM usage is the same if not temporarily higher.

[Desour]

I thought using the Lua stack less could maybe result in a performance benefit. But those suggestions weren't based on any real knowledge. They were meant as something that one could try out. I should've made that more clear, sorry.

[Zughy]

@TurkeyMcMac please rebase

[Desour]

Code looks fine otherwise.

Haven't tested.

[sfan5]

The only thing I'm wondering about is whether we should have a short snippet in C++ that checks that READ_NODE, PUSH_NODE, READ_VECTOR, PUSH_VECTOR are actually set on startup.
Plus I wanted to test this.
And lastly put a comment in a suitable place detailing why we're doing the extra work of having table creation done by lua code.

[sfan5]

Works, measured a 30% improvement to bench_get_node

[sfan5]

This makes error messages quite less friendly:

minetest/bin/../builtin/common/item_s.lua:229: bad argument #1 to '__index' (string expected, got nil)

happens when you pass a string, empty table or other stuff to a function expecting a node