print/multiversion: align adjacent columns' lines, to match up their …

…anchors.

Cargo.toml

@@ -27,6 +27,7 @@ indexmap = "1.7.0"
 internal-iterator = "0.2.0"
 itertools = "0.10.3"
 lazy_static = "1.4.0"
+longest-increasing-subsequence = "0.1.0"
 rustc-hash = "1.1.0"
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1.0"

src/print/multiversion.rs

@@ -1,9 +1,14 @@
 //! Multi-version pretty-printing support (e.g. for comparing the IR between passes).
 
-use crate::print::pretty;
+use crate::print::pretty::{self, TextOp};
+use crate::FxIndexMap;
+use internal_iterator::{
+    FromInternalIterator, InternalIterator, IntoInternalIterator, IteratorExt,
+};
+use itertools::Itertools;
 use smallvec::SmallVec;
-use std::fmt;
 use std::fmt::Write;
+use std::{fmt, mem};
 
 #[allow(rustdoc::private_intra_doc_links)]
 /// Wrapper for handling the difference between single-version and multi-version
@@ -153,14 +158,25 @@ impl Versions<pretty::FragmentPostLayout> {
                     }
                     last_was_uniform = is_uniform;
 
+                    // Attempt to align as many anchors as possible between the
+                    // columns, to improve legibility (see also `AnchorAligner`).
+                    let mut anchor_aligner = AnchorAligner::default();
+                    for (fragment, _) in versions_with_repeat_count {
+                        anchor_aligner
+                            .add_column_and_align_anchors(fragment.render_to_text_ops().collect());
+                    }
+
                     html.body += "<tr>\n";
-                    for (fragment, repeat_count) in versions_with_repeat_count {
+                    for ((_, repeat_count), column) in versions_with_repeat_count
+                        .iter()
+                        .zip(anchor_aligner.merged_columns())
+                    {
                         writeln!(html.body, "<td colspan=\"{repeat_count}\">").unwrap();
 
                         let pretty::HtmlSnippet {
                             head_deduplicatable_elements: fragment_head,
                             body: fragment_body,
-                        } = fragment.render_to_html();
+                        } = column.into_internal().collect();
                         html.head_deduplicatable_elements.extend(fragment_head);
                         html.body += &fragment_body;
 
@@ -198,3 +214,260 @@ impl<PF> Versions<PF> {
         }
     }
 }
+
+/// Tool for adjusting pretty-printed columns, so that their anchors line up
+/// (by adding empty lines to whichever side "is behind").
+#[derive(Default)]
+struct AnchorAligner<'a> {
+    merged_lines: Vec<AAMergedLine>,
+
+    /// Current ("rightmost") column's anchor definitions (with indices pointing
+    /// into `merged_lines`), which the next column will align to.
+    //
+    // FIXME(eddyb) does this need additional interning?
+    anchor_def_to_merged_line_idx: FxIndexMap<&'a String, usize>,
+
+    // FIXME(eddyb) fine-tune this inline size.
+    // FIXME(eddyb) maybe don't keep most of this data around anyway?
+    original_columns: SmallVec<[AAColumn<'a>; 4]>,
+}
+
+/// Abstraction for one "physical" line spanning all columns, after alignment.
+struct AAMergedLine {
+    // FIXME(eddyb) fine-tune this inline size.
+    // FIXME(eddyb) consider using `u32` here?
+    per_column_line_lengths: SmallVec<[usize; 4]>,
+}
+
+struct AAColumn<'a> {
+    /// All `TextOp`s in all lines from this column, concatenated together.
+    text_ops: Vec<TextOp<'a>>,
+
+    /// The length, in `TextOp`s (from `text_ops`), of each line.
+    //
+    // FIXME(eddyb) consider using `u32` here?
+    line_lengths: Vec<usize>,
+}
+
+impl<'a> AAColumn<'a> {
+    /// Reconstruct lines (made of `TextOp`s) from line lengths.
+    fn lines(
+        &self,
+        line_lengths: impl Iterator<Item = usize>,
+    ) -> impl Iterator<Item = &[TextOp<'a>]> {
+        let mut next_start = 0;
+        line_lengths.map(move |len| {
+            let start = next_start;
+            let end = start + len;
+            next_start = end;
+            &self.text_ops[start..end]
+        })
+    }
+}
+
+// FIXME(eddyb) is this impl the best way? (maybe it should be a inherent method)
+impl<'a> FromInternalIterator<TextOp<'a>> for AAColumn<'a> {
+    fn from_iter<T>(text_ops: T) -> Self
+    where
+        T: IntoInternalIterator<Item = TextOp<'a>>,
+    {
+        let mut column = AAColumn {
+            text_ops: vec![],
+            line_lengths: vec![0],
+        };
+        text_ops.into_internal_iter().for_each(|op| {
+            if let TextOp::Text("\n") = op {
+                column.line_lengths.push(0);
+            } else {
+                // FIXME(eddyb) this *happens* to be true,
+                // but the `LineOp`/`TextOp` split could be
+                // improved to avoid such sanity checks.
+                if let TextOp::Text(text) = op {
+                    assert!(!text.contains('\n'));
+                }
+                column.text_ops.push(op);
+                *column.line_lengths.last_mut().unwrap() += 1;
+            }
+        });
+        column
+    }
+}
+
+impl<'a> AnchorAligner<'a> {
+    /// Flatten all columns to `TextOp`s (including line separators).
+    fn merged_columns(&self) -> impl Iterator<Item = impl Iterator<Item = TextOp<'a>> + '_> {
+        self.original_columns
+            .iter()
+            .enumerate()
+            .map(|(column_idx, column)| {
+                let line_lengths = self
+                    .merged_lines
+                    .iter()
+                    .map(move |line| line.per_column_line_lengths[column_idx]);
+
+                // HACK(eddyb) trim all trailing empty lines (which is done on
+                // a `peekable` of the reverse, followed by reversing *again*,
+                // equivalent to a hypothetical `peekable_back` and no reversing).
+                let mut rev_line_lengths = line_lengths.rev().peekable();
+                while rev_line_lengths.peek() == Some(&0) {
+                    rev_line_lengths.next().unwrap();
+                }
+                let line_lengths = rev_line_lengths.rev();
+
+                column
+                    .lines(line_lengths)
+                    .intersperse(&[TextOp::Text("\n")])
+                    .flatten()
+                    .copied()
+            })
+    }
+
+    /// Merge `new_column` into the current set of columns, aligning as many
+    /// anchors as possible, between it, and the most recent column.
+    fn add_column_and_align_anchors(&mut self, new_column: AAColumn<'a>) {
+        // NOTE(eddyb) "old" and "new" are used to refer to the two columns being
+        // aligned, but "old" maps to the *merged* lines, not its original ones.
+
+        let old_lines = mem::take(&mut self.merged_lines);
+        let old_anchor_def_to_line_idx = mem::take(&mut self.anchor_def_to_merged_line_idx);
+
+        // Index all the anchor definitions in the new column.
+        let mut new_anchor_def_to_line_idx = FxIndexMap::default();
+        for (new_line_idx, new_line_text_ops) in new_column
+            .lines(new_column.line_lengths.iter().copied())
+            .enumerate()
+        {
+            for op in new_line_text_ops {
+                if let TextOp::PushStyles(styles) = op {
+                    if let Some(anchor) = &styles.anchor {
+                        if styles.anchor_is_def {
+                            new_anchor_def_to_line_idx
+                                .entry(anchor)
+                                .or_insert(new_line_idx);
+                        }
+                    }
+                }
+            }
+        }
+
+        // Find all the possible anchor alignments (i.e. anchors defined in both
+        // "old" and "new") as pairs of line indices in "old" and "new".
+        //
+        // HACK(eddyb) the order is given by the "new" line index, implicitly.
+        // FIXME(eddyb) fine-tune this inline size.
+        let common_anchors: SmallVec<[_; 8]> = new_anchor_def_to_line_idx
+            .iter()
+            .filter_map(|(anchor, &new_line_idx)| {
+                Some((*old_anchor_def_to_line_idx.get(anchor)?, new_line_idx))
+            })
+            .collect();
+
+        // Fast-path: if all the "old" line indices are already in (increasing)
+        // order (i.e. "monotonic"), they can all be used directly for alignment.
+        let is_already_monotonic = {
+            // FIXME(eddyb) should be `.is_sorted_by_key(|&(old_line_idx, _)| old_line_idx)`
+            // but that slice method is still unstable.
+            common_anchors.windows(2).all(|w| w[0].0 <= w[1].0)
+        };
+        let monotonic_common_anchors = if is_already_monotonic {
+            common_anchors
+        } else {
+            // FIXME(eddyb) this could maybe avoid all the unnecessary allocations.
+            longest_increasing_subsequence::lis(&common_anchors)
+                .into_iter()
+                .map(|i| common_anchors[i])
+                .collect()
+        };
+
+        // Allocate space for the merge of "old" and "new".
+        let mut merged_lines = Vec::with_capacity({
+            // Cheap conservative estimate, based on the last anchor (i.e. the
+            // final position of the last anchor is *at least* `min_before_last`).
+            let &(old_last, new_last) = monotonic_common_anchors.last().unwrap_or(&(0, 0));
+            let min_before_last = old_last.max(new_last);
+            let after_last =
+                (old_lines.len() - old_last).max(new_column.line_lengths.len() - new_last);
+            (min_before_last + after_last).next_power_of_two()
+        });
+
+        // Build the merged lines using (partially) lockstep iteration to pull
+        // the relevant data out of either side, and update "new" line indices.
+        let mut old_lines = old_lines.into_iter().enumerate().peekable();
+        let mut new_lines = new_column
+            .line_lengths
+            .iter()
+            .copied()
+            .enumerate()
+            .peekable();
+        let mut monotonic_common_anchors = monotonic_common_anchors.into_iter().peekable();
+        let mut fixup_new_to_merged = new_anchor_def_to_line_idx.values_mut().peekable();
+        while old_lines.len() > 0 || new_lines.len() > 0 {
+            let old_line_idx = old_lines.peek().map(|&(i, _)| i);
+            let new_line_idx = new_lines.peek().map(|&(i, _)| i);
+            let mut next_anchor = monotonic_common_anchors.peek().copied();
+
+            // Discard anchor alignments that have been used already, and also
+            // any others that cannot be relevant anymore - this can occur when
+            // multiple anchors coincide on the same line.
+            while let Some((anchor_old, anchor_new)) = next_anchor {
+                let obsolete = old_line_idx.map_or(false, |old| old > anchor_old)
+                    || new_line_idx.map_or(false, |new| new > anchor_new);
+                if !obsolete {
+                    break;
+                }
+                monotonic_common_anchors.next().unwrap();
+                next_anchor = monotonic_common_anchors.peek().copied();
+            }
+
+            // Figure out which side has to wait, to align an upcoming anchor.
+            let (old_at_anchor, new_at_anchor) =
+                next_anchor.map_or((false, false), |(anchor_old, anchor_new)| {
+                    (
+                        old_line_idx.map_or(false, |old| old == anchor_old),
+                        new_line_idx.map_or(false, |new| new == anchor_new),
+                    )
+                });
+            let old_line = if old_at_anchor && !new_at_anchor {
+                // Pausing "old", waiting for "new".
+                None
+            } else {
+                old_lines.next().map(|(_, old_line)| old_line)
+            };
+            let new_line_len = if !old_at_anchor && new_at_anchor {
+                // Pausing "new", waiting for "old".
+                None
+            } else {
+                new_lines.next().map(|(_, new_line_len)| new_line_len)
+            };
+
+            // When the "new" side is advanced, that "sets" the merged line index
+            // of the consumed line, which can then be used for fixing up indices.
+            if new_line_len.is_some() {
+                let new_line_idx = new_line_idx.unwrap();
+                let merged_line_idx = merged_lines.len();
+                while fixup_new_to_merged.peek().map(|i| **i) == Some(new_line_idx) {
+                    *fixup_new_to_merged.next().unwrap() = merged_line_idx;
+                }
+            }
+
+            let new_line_len = new_line_len.unwrap_or(0);
+            let merged_line = match old_line {
+                Some(mut line) => {
+                    line.per_column_line_lengths.push(new_line_len);
+                    line
+                }
+                None => AAMergedLine {
+                    per_column_line_lengths: (0..self.original_columns.len())
+                        .map(|_| 0)
+                        .chain([new_line_len])
+                        .collect(),
+                },
+            };
+            merged_lines.push(merged_line);
+        }
+
+        self.merged_lines = merged_lines;
+        self.anchor_def_to_merged_line_idx = new_anchor_def_to_line_idx;
+        self.original_columns.push(new_column);
+    }
+}

src/print/pretty.rs

@@ -160,7 +160,6 @@ pub struct FragmentPostLayout(Fragment);
 impl fmt::Display for FragmentPostLayout {
     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
         let result = self
-            .0
             .render_to_text_ops()
             .filter_map(|op| match op {
                 TextOp::Text(text) => Some(text),
@@ -178,11 +177,16 @@ impl fmt::Display for FragmentPostLayout {
 }
 
 impl FragmentPostLayout {
+    /// Flatten the [`Fragment`] to [`TextOp`]s.
+    pub(super) fn render_to_text_ops(&self) -> impl InternalIterator<Item = TextOp<'_>> {
+        self.0.render_to_text_ops()
+    }
+
     /// Flatten the [`Fragment`] to HTML, producing a [`HtmlSnippet`].
     //
     // FIXME(eddyb) provide a non-allocating version.
     pub fn render_to_html(&self) -> HtmlSnippet {
-        self.0.render_to_text_ops().collect()
+        self.render_to_text_ops().collect()
     }
 }
 
@@ -270,6 +274,7 @@ impl HtmlSnippet {
     }
 }
 
+// FIXME(eddyb) is this impl the best way? (maybe it should be a inherent method)
 impl<'a> FromInternalIterator<TextOp<'a>> for HtmlSnippet {
     fn from_iter<T>(text_ops: T) -> Self
     where
@@ -821,7 +826,8 @@ impl Fragment {
 }
 
 /// Text-oriented operation (plain text snippets interleaved with style push/pop).
-enum TextOp<'a> {
+#[derive(Copy, Clone)]
+pub(super) enum TextOp<'a> {
     PushStyles(&'a Styles),
     PopStyles(&'a Styles),