Investigate the need to cache

[abeldekat]

In cmp-luasnip, a cache is maintained by ft, both for snippets and docs. That might very well be done for performance reasons. Otherwise, perhaps because LuaSnips supports snippets to be dynamically added/changed?

On each completion, all snippets do need to be transformed into results for nvim-cmp.... I am still unsure if caching should be added.

See this answer from echasnovski

[echasnovski]

Following this comment, here a patch I used to profile default_prepare() with and without (comment out irrelevant parts) caching:

diff --git a/lua/mini/snippets.lua b/lua/mini/snippets.lua
index 8dafaef..59ba88d 100644
--- a/lua/mini/snippets.lua
+++ b/lua/mini/snippets.lua
@@ -993,11 +993,19 @@ end
 ---
 ---@return ... Array of snippets and supplied context (default if none was supplied).
 MiniSnippets.default_prepare = function(raw_snippets, opts)
+  local start_time = vim.loop.hrtime()
   if not H.islist(raw_snippets) then H.error('`raw_snippets` should be array') end
   opts = vim.tbl_extend('force', { context = nil }, opts or {})
   local context = opts.context
   if context == nil then context = H.get_default_context() end
 
+  local cache_id = context.buf_id .. context.lang
+  if H.prepare_cache[cache_id] then
+    table.insert(_G.durations, 0.000001 * (vim.loop.hrtime() - start_time))
+    return unpack(H.prepare_cache[cache_id])
+  end
+
   -- Traverse snippets to have unique non-empty prefixes
   local res = {}
   H.traverse_raw_snippets(raw_snippets, res, context)
@@ -1005,9 +1013,14 @@ MiniSnippets.default_prepare = function(raw_snippets, opts)
   -- Convert to array ordered by prefix
   res = vim.tbl_values(res)
   table.sort(res, function(a, b) return a.prefix < b.prefix end)
+  H.prepare_cache[cache_id] = { res, context }
+  table.insert(_G.durations, 0.000001 * (vim.loop.hrtime() - start_time))
   return res, context
 end
 
+H.prepare_cache = {}
+_G.durations = {}
+
 --- Default match
 ---
 --- Match snippets based on the line before cursor.

With benchmarking code present, open a separate nvim and do the expand <C-j> on blank line in Lua file 11 times (or more). Then :=durations will show how long default_prepare() took. For summary I use median from MiniMisc.stat_summary(): :=stat_summary(vim.list_slice(durations, 2)) (remove first entry as it reads data from the disk and thus visibly bigger).

Adding benchmarking code and then actually interactively performing the task to be benchmarked is usually a more accurate way of measuring things.

[abeldekat]

Use caching in the same way as implemented in this blink PR
Also, completion_item.documentation is reused.