Merge #30085 #trim! all whitespace

This commit is contained in:
Justin M. Keyes
2024-12-06 10:08:20 -08:00
committed by GitHub
5 changed files with 168 additions and 39 deletions

View File

@@ -572,13 +572,17 @@ local directive_handlers = {
metadata[id].text = text:gsub(pattern, replacement)
end,
-- Trim blank lines from end of the node
-- Example: (#trim! @fold)
-- TODO(clason): generalize to arbitrary whitespace removal
-- Trim whitespace from both sides of the node
-- Example: (#trim! @fold 1 1 1 1)
['trim!'] = function(match, _, bufnr, pred, metadata)
local capture_id = pred[2]
assert(type(capture_id) == 'number')
local trim_start_lines = pred[3] == '1'
local trim_start_cols = pred[4] == '1'
local trim_end_lines = pred[5] == '1' or not pred[3] -- default true for backwards compatibility
local trim_end_cols = pred[6] == '1'
local nodes = match[capture_id]
if not nodes or #nodes == 0 then
return
@@ -588,20 +592,36 @@ local directive_handlers = {
local start_row, start_col, end_row, end_col = node:range()
-- Don't trim if region ends in middle of a line
if end_col ~= 0 then
return
local node_text = vim.split(vim.treesitter.get_node_text(node, bufnr), '\n')
local end_idx = #node_text
local start_idx = 1
if trim_end_lines then
while end_idx > 0 and node_text[end_idx]:find('^%s*$') do
end_idx = end_idx - 1
end_row = end_row - 1
end
end
if trim_end_cols then
if end_idx == 0 then
end_row = start_row
end_col = start_col
else
local whitespace_start = node_text[end_idx]:find('(%s*)$')
end_col = (whitespace_start - 1) + (end_idx == 1 and start_col or 0)
end
end
while end_row >= start_row do
-- As we only care when end_col == 0, always inspect one line above end_row.
local end_line = api.nvim_buf_get_lines(bufnr, end_row - 1, end_row, true)[1]
if end_line ~= '' then
break
if trim_start_lines then
while start_idx <= end_idx and node_text[start_idx]:find('^%s*$') do
start_idx = start_idx + 1
start_row = start_row + 1
end
end_row = end_row - 1
end
if trim_start_cols and node_text[start_idx] then
local _, whitespace_end = node_text[start_idx]:find('^(%s*)')
whitespace_end = whitespace_end or 0
start_col = (start_idx == 1 and start_col or 0) + whitespace_end
end
-- If this produces an invalid range, we just skip it.