Rollup merge of #119033 - Zalathar:unicode, r=davidtwco

coverage: `llvm-cov` expects column numbers to be bytes, not code points

Normally the compiler emits column numbers as a 1-based number of Unicode code points.

But when we embed coverage mappings for `-Cinstrument-coverage`, those mappings will ultimately be read by the `llvm-cov` tool. That tool assumes that column numbers are 1-based numbers of *bytes*, and relies on that assumption when slicing up source code to apply highlighting (in HTML reports, and in text-based reports with colour).

For the very common case of all-ASCII source code, bytes and code points are the same, so the difference isn't noticeable. But for code that contains non-ASCII characters, emitting column numbers as code points will result in `llvm-cov` slicing strings in the wrong places, producing mangled output or fatal errors.

(See taiki-e/cargo-llvm-cov#275 as an example of what can go wrong.)

compiler/rustc_mir_transform/src/coverage/mod.rs

@@ -23,7 +23,7 @@ use rustc_middle::mir::{
 use rustc_middle::ty::TyCtxt;
 use rustc_span::def_id::LocalDefId;
 use rustc_span::source_map::SourceMap;
-use rustc_span::{Span, Symbol};
+use rustc_span::{BytePos, Pos, RelativeBytePos, Span, Symbol};
 
 /// Inserts `StatementKind::Coverage` statements that either instrument the binary with injected
 /// counters, via intrinsic `llvm.instrprof.increment`, and/or inject metadata used during codegen
@@ -107,6 +107,12 @@ impl<'a, 'tcx> Instrumentor<'a, 'tcx> {
         );
 
         let mappings = self.create_mappings(&coverage_spans, &coverage_counters);
+        if mappings.is_empty() {
+            // No spans could be converted into valid mappings, so skip this function.
+            debug!("no spans could be converted into valid mappings; skipping");
+            return;
+        }
+
         self.inject_coverage_statements(bcb_has_coverage_spans, &coverage_counters);
 
         self.mir_body.function_coverage_info = Some(Box::new(FunctionCoverageInfo {
@@ -148,9 +154,9 @@ impl<'a, 'tcx> Instrumentor<'a, 'tcx> {
             // Flatten the spans into individual term/span pairs.
             .flat_map(|(term, spans)| spans.iter().map(move |&span| (term, span)))
             // Convert each span to a code region, and create the final mapping.
-            .map(|(term, span)| {
-                let code_region = make_code_region(source_map, file_name, span, body_span);
-                Mapping { term, code_region }
+            .filter_map(|(term, span)| {
+                let code_region = make_code_region(source_map, file_name, span, body_span)?;
+                Some(Mapping { term, code_region })
             })
             .collect::<Vec<_>>()
     }
@@ -252,41 +258,85 @@ fn inject_statement(mir_body: &mut mir::Body<'_>, counter_kind: CoverageKind, bb
     data.statements.insert(0, statement);
 }
 
-/// Convert the Span into its file name, start line and column, and end line and column
+/// Convert the Span into its file name, start line and column, and end line and column.
+///
+/// Line numbers and column numbers are 1-based. Unlike most column numbers emitted by
+/// the compiler, these column numbers are denoted in **bytes**, because that's what
+/// LLVM's `llvm-cov` tool expects to see in coverage maps.
+///
+/// Returns `None` if the conversion failed for some reason. This shouldn't happen,
+/// but it's hard to rule out entirely (especially in the presence of complex macros
+/// or other expansions), and if it does happen then skipping a span or function is
+/// better than an ICE or `llvm-cov` failure that the user might have no way to avoid.
 fn make_code_region(
     source_map: &SourceMap,
     file_name: Symbol,
     span: Span,
     body_span: Span,
-) -> CodeRegion {
+) -> Option<CodeRegion> {
     debug!(
         "Called make_code_region(file_name={}, span={}, body_span={})",
         file_name,
         source_map.span_to_diagnostic_string(span),
         source_map.span_to_diagnostic_string(body_span)
     );
 
-    let (file, mut start_line, mut start_col, mut end_line, mut end_col) =
-        source_map.span_to_location_info(span);
-    if span.hi() == span.lo() {
-        // Extend an empty span by one character so the region will be counted.
-        if span.hi() == body_span.hi() {
-            start_col = start_col.saturating_sub(1);
-        } else {
-            end_col = start_col + 1;
-        }
+    let lo = span.lo();
+    let hi = span.hi();
+
+    let file = source_map.lookup_source_file(lo);
+    if !file.contains(hi) {
+        debug!(?span, ?file, ?lo, ?hi, "span crosses multiple files; skipping");
+        return None;
+    }
+
+    // Column numbers need to be in bytes, so we can't use the more convenient
+    // `SourceMap` methods for looking up file coordinates.
+    let rpos_and_line_and_byte_column = |pos: BytePos| -> Option<(RelativeBytePos, usize, usize)> {
+        let rpos = file.relative_position(pos);
+        let line_index = file.lookup_line(rpos)?;
+        let line_start = file.lines()[line_index];
+        // Line numbers and column numbers are 1-based, so add 1 to each.
+        Some((rpos, line_index + 1, (rpos - line_start).to_usize() + 1))
     };
-    if let Some(file) = file {
-        start_line = source_map.doctest_offset_line(&file.name, start_line);
-        end_line = source_map.doctest_offset_line(&file.name, end_line);
+
+    let (lo_rpos, mut start_line, mut start_col) = rpos_and_line_and_byte_column(lo)?;
+    let (hi_rpos, mut end_line, mut end_col) = rpos_and_line_and_byte_column(hi)?;
+
+    // If the span is empty, try to expand it horizontally by one character's
+    // worth of bytes, so that it is more visible in `llvm-cov` reports.
+    // We do this after resolving line/column numbers, so that empty spans at the
+    // end of a line get an extra column instead of wrapping to the next line.
+    if span.is_empty()
+        && body_span.contains(span)
+        && let Some(src) = &file.src
+    {
+        // Prefer to expand the end position, if it won't go outside the body span.
+        if hi < body_span.hi() {
+            let hi_rpos = hi_rpos.to_usize();
+            let nudge_bytes = src.ceil_char_boundary(hi_rpos + 1) - hi_rpos;
+            end_col += nudge_bytes;
+        } else if lo > body_span.lo() {
+            let lo_rpos = lo_rpos.to_usize();
+            let nudge_bytes = lo_rpos - src.floor_char_boundary(lo_rpos - 1);
+            // Subtract the nudge, but don't go below column 1.
+            start_col = start_col.saturating_sub(nudge_bytes).max(1);
+        }
+        // If neither nudge could be applied, stick with the empty span coordinates.
     }
-    CodeRegion {
+
+    // Apply an offset so that code in doctests has correct line numbers.
+    // FIXME(#79417): Currently we have no way to offset doctest _columns_.
+    start_line = source_map.doctest_offset_line(&file.name, start_line);
+    end_line = source_map.doctest_offset_line(&file.name, end_line);
+
+    Some(CodeRegion {
         file_name,
         start_line: start_line as u32,
         start_col: start_col as u32,
         end_line: end_line as u32,
         end_col: end_col as u32,
-    }
+    })
 }
 
 fn is_eligible_for_coverage(tcx: TyCtxt<'_>, def_id: LocalDefId) -> bool {

compiler/rustc_mir_transform/src/lib.rs

@@ -9,6 +9,7 @@
 #![feature(min_specialization)]
 #![feature(never_type)]
 #![feature(option_get_or_insert_default)]
+#![feature(round_char_boundary)]
 #![feature(trusted_step)]
 #![feature(try_blocks)]
 #![feature(yeet_expr)]

tests/coverage/unicode.cov-map

@@ -0,0 +1,53 @@
+Function name: unicode::main
+Raw bytes (67): 0x[01, 01, 09, 01, 05, 03, 05, 1e, 0d, 22, 09, 03, 05, 11, 1b, 1e, 0d, 22, 09, 03, 05, 09, 01, 0e, 01, 00, 0b, 05, 01, 09, 00, 0c, 03, 00, 10, 00, 1b, 05, 00, 1c, 00, 28, 22, 02, 08, 00, 25, 09, 00, 29, 00, 46, 11, 00, 47, 02, 06, 1b, 02, 06, 00, 07, 17, 02, 05, 01, 02]
+Number of files: 1
+- file 0 => global file 1
+Number of expressions: 9
+- expression 0 operands: lhs = Counter(0), rhs = Counter(1)
+- expression 1 operands: lhs = Expression(0, Add), rhs = Counter(1)
+- expression 2 operands: lhs = Expression(7, Sub), rhs = Counter(3)
+- expression 3 operands: lhs = Expression(8, Sub), rhs = Counter(2)
+- expression 4 operands: lhs = Expression(0, Add), rhs = Counter(1)
+- expression 5 operands: lhs = Counter(4), rhs = Expression(6, Add)
+- expression 6 operands: lhs = Expression(7, Sub), rhs = Counter(3)
+- expression 7 operands: lhs = Expression(8, Sub), rhs = Counter(2)
+- expression 8 operands: lhs = Expression(0, Add), rhs = Counter(1)
+Number of file 0 mappings: 9
+- Code(Counter(0)) at (prev + 14, 1) to (start + 0, 11)
+- Code(Counter(1)) at (prev + 1, 9) to (start + 0, 12)
+- Code(Expression(0, Add)) at (prev + 0, 16) to (start + 0, 27)
+    = (c0 + c1)
+- Code(Counter(1)) at (prev + 0, 28) to (start + 0, 40)
+- Code(Expression(8, Sub)) at (prev + 2, 8) to (start + 0, 37)
+    = ((c0 + c1) - c1)
+- Code(Counter(2)) at (prev + 0, 41) to (start + 0, 70)
+- Code(Counter(4)) at (prev + 0, 71) to (start + 2, 6)
+- Code(Expression(6, Add)) at (prev + 2, 6) to (start + 0, 7)
+    = ((((c0 + c1) - c1) - c2) + c3)
+- Code(Expression(5, Add)) at (prev + 2, 5) to (start + 1, 2)
+    = (c4 + ((((c0 + c1) - c1) - c2) + c3))
+
+Function name: unicode::サビ
+Raw bytes (9): 0x[01, 01, 00, 01, 01, 1e, 14, 00, 18]
+Number of files: 1
+- file 0 => global file 1
+Number of expressions: 0
+Number of file 0 mappings: 1
+- Code(Counter(0)) at (prev + 30, 20) to (start + 0, 24)
+
+Function name: unicode::他 (unused)
+Raw bytes (9): 0x[01, 01, 00, 01, 00, 1e, 19, 00, 25]
+Number of files: 1
+- file 0 => global file 1
+Number of expressions: 0
+Number of file 0 mappings: 1
+- Code(Zero) at (prev + 30, 25) to (start + 0, 37)
+
+Function name: unicode::申し訳ございません
+Raw bytes (9): 0x[01, 01, 00, 01, 01, 18, 01, 02, 02]
+Number of files: 1
+- file 0 => global file 1
+Number of expressions: 0
+Number of file 0 mappings: 1
+- Code(Counter(0)) at (prev + 24, 1) to (start + 2, 2)
+

tests/coverage/unicode.coverage

@@ -0,0 +1,40 @@
+   LL|       |// edition: 2021
+   LL|       |// ignore-windows - we can't force `llvm-cov` to use ANSI escapes on Windows
+   LL|       |// llvm-cov-flags: --use-color
+   LL|       |
+   LL|       |// Check that column numbers are denoted in bytes, so that they don't cause
+   LL|       |// `llvm-cov` to fail or emit malformed output.
+   LL|       |//
+   LL|       |// Note that when `llvm-cov` prints ^ arrows on a subsequent line, it simply
+   LL|       |// inserts one space character for each "column", with no understanding of
+   LL|       |// Unicode or character widths. So those arrows will tend to be misaligned
+   LL|       |// for non-ASCII source code, regardless of whether column numbers are code
+   LL|       |// points or bytes.
+   LL|       |
+   LL|      1|fn main() {
+   LL|     33|    for _İ in 'А'..='Я' { /* Я */ }
+                      ^32                ^32
+   LL|       |
+   LL|      1|    if 申し訳ございません() && 申し訳ございません() {
+                                                      ^0
+   LL|      0|        println!("true");
+   LL|      1|    }
+   LL|       |
+   LL|      1|    サビ();
+   LL|      1|}
+   LL|       |
+   LL|      1|fn 申し訳ございません() -> bool {
+   LL|      1|    std::hint::black_box(false)
+   LL|      1|}
+   LL|       |
+   LL|       |macro_rules! macro_that_defines_a_function {
+   LL|       |    (fn $名:ident () $体:tt) => {
+   LL|      1|        fn $名 () $体 fn 他 () {}
+                                      ^0
+   LL|       |    }
+   LL|       |}
+   LL|       |
+   LL|       |macro_that_defines_a_function! {
+   LL|       |    fn サビ() {}
+   LL|       |}
+

tests/coverage/unicode.rs

@@ -0,0 +1,36 @@
+// edition: 2021
+// ignore-windows - we can't force `llvm-cov` to use ANSI escapes on Windows
+// llvm-cov-flags: --use-color
+
+// Check that column numbers are denoted in bytes, so that they don't cause
+// `llvm-cov` to fail or emit malformed output.
+//
+// Note that when `llvm-cov` prints ^ arrows on a subsequent line, it simply
+// inserts one space character for each "column", with no understanding of
+// Unicode or character widths. So those arrows will tend to be misaligned
+// for non-ASCII source code, regardless of whether column numbers are code
+// points or bytes.
+
+fn main() {
+    for _İ in 'А'..='Я' { /* Я */ }
+
+    if 申し訳ございません() && 申し訳ございません() {
+        println!("true");
+    }
+
+    サビ();
+}
+
+fn 申し訳ございません() -> bool {
+    std::hint::black_box(false)
+}
+
+macro_rules! macro_that_defines_a_function {
+    (fn $名:ident () $体:tt) => {
+        fn $名 () $体 fn 他 () {}
+    }
+}
+
+macro_that_defines_a_function! {
+    fn サビ() {}
+}