Add tool src/tools/coverage-dump for use by some new coverage tests

Cargo.lock

@@ -722,6 +722,18 @@ version = "0.8.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "e496a50fda8aacccc86d7529e2c1e0892dbd0f898a6b5645b5561b89c3210efa"
 
+[[package]]
+name = "coverage-dump"
+version = "0.1.0"
+dependencies = [
+ "anyhow",
+ "leb128",
+ "md-5",
+ "miniz_oxide",
+ "regex",
+ "rustc-demangle",
+]
+
 [[package]]
 name = "coverage_test_macros"
 version = "0.0.0"
@@ -2041,6 +2053,12 @@ version = "1.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "830d08ce1d1d941e6b30645f1a0eb5643013d835ce3779a5fc208261dbe10f55"
 
+[[package]]
+name = "leb128"
+version = "0.2.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "884e2677b40cc8c339eaefcb701c32ef1fd2493d71118dc0ca4b6a736c93bd67"
+
 [[package]]
 name = "levenshtein"
 version = "1.0.5"

Cargo.toml

@@ -43,6 +43,7 @@ members = [
   "src/tools/generate-windows-sys",
   "src/tools/rustdoc-gui-test",
   "src/tools/opt-dist",
+  "src/tools/coverage-dump",
 ]
 
 exclude = [

src/bootstrap/builder.rs

@@ -703,7 +703,8 @@ macro_rules! describe {
                 llvm::Lld,
                 llvm::CrtBeginEnd,
                 tool::RustdocGUITest,
-                tool::OptimizedDist
+                tool::OptimizedDist,
+                tool::CoverageDump,
             ),
             Kind::Check | Kind::Clippy | Kind::Fix => describe!(
                 check::Std,

src/bootstrap/tool.rs

@@ -306,6 +306,7 @@ fn run(self, builder: &Builder<'_>) -> PathBuf {
     GenerateWindowsSys, "src/tools/generate-windows-sys", "generate-windows-sys";
     RustdocGUITest, "src/tools/rustdoc-gui-test", "rustdoc-gui-test", is_unstable_tool = true, allow_features = "test";
     OptimizedDist, "src/tools/opt-dist", "opt-dist";
+    CoverageDump, "src/tools/coverage-dump", "coverage-dump";
 );
 
 #[derive(Debug, Copy, Clone, Hash, PartialEq, Eq, Ord, PartialOrd)]

src/tools/coverage-dump/Cargo.toml

@@ -0,0 +1,14 @@
+[package]
+name = "coverage-dump"
+version = "0.1.0"
+edition = "2021"
+
+# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
+
+[dependencies]
+anyhow = "1.0.71"
+leb128 = "0.2.5"
+md5 = { package = "md-5" , version = "0.10.5" }
+miniz_oxide = "0.7.1"
+regex = "1.8.4"
+rustc-demangle = "0.1.23"

src/tools/coverage-dump/README.md

@@ -0,0 +1,8 @@
+This tool extracts coverage mapping information from an LLVM IR assembly file
+(`.ll`), and prints it in a more human-readable form that can be used for
+snapshot tests.
+
+The output format is mostly arbitrary, so it's OK to change the output as long
+as any affected tests are also re-blessed. However, the output should be
+consistent across different executions on different platforms, so avoid
+printing any information that is platform-specific or non-deterministic.

src/tools/coverage-dump/src/covfun.rs

@@ -0,0 +1,296 @@
+use crate::parser::{unescape_llvm_string_contents, Parser};
+use anyhow::{anyhow, Context};
+use regex::Regex;
+use std::collections::HashMap;
+use std::fmt::{self, Debug, Write as _};
+use std::sync::OnceLock;
+
+pub(crate) fn dump_covfun_mappings(
+    llvm_ir: &str,
+    function_names: &HashMap<u64, String>,
+) -> anyhow::Result<()> {
+    // Extract function coverage entries from the LLVM IR assembly, and associate
+    // each entry with its (demangled) name.
+    let mut covfun_entries = llvm_ir
+        .lines()
+        .filter_map(covfun_line_data)
+        .map(|line_data| (function_names.get(&line_data.name_hash).map(String::as_str), line_data))
+        .collect::<Vec<_>>();
+    covfun_entries.sort_by(|a, b| {
+        // Sort entries primarily by name, to help make the order consistent
+        // across platforms and relatively insensitive to changes.
+        // (Sadly we can't use `sort_by_key` because we would need to return references.)
+        Ord::cmp(&a.0, &b.0)
+            .then_with(|| Ord::cmp(&a.1.is_used, &b.1.is_used))
+            .then_with(|| Ord::cmp(a.1.payload.as_slice(), b.1.payload.as_slice()))
+    });
+
+    for (name, line_data) in &covfun_entries {
+        let name = name.unwrap_or("(unknown)");
+        let unused = if line_data.is_used { "" } else { " (unused)" };
+        println!("Function name: {name}{unused}");
+
+        let payload: &[u8] = &line_data.payload;
+        println!("Raw bytes ({len}): 0x{payload:02x?}", len = payload.len());
+
+        let mut parser = Parser::new(payload);
+
+        let num_files = parser.read_uleb128_u32()?;
+        println!("Number of files: {num_files}");
+
+        for i in 0..num_files {
+            let global_file_id = parser.read_uleb128_u32()?;
+            println!("- file {i} => global file {global_file_id}");
+        }
+
+        let num_expressions = parser.read_uleb128_u32()?;
+        println!("Number of expressions: {num_expressions}");
+
+        let mut expression_resolver = ExpressionResolver::new();
+        for i in 0..num_expressions {
+            let lhs = parser.read_simple_term()?;
+            let rhs = parser.read_simple_term()?;
+            println!("- expression {i} operands: lhs = {lhs:?}, rhs = {rhs:?}");
+            expression_resolver.push_operands(lhs, rhs);
+        }
+
+        for i in 0..num_files {
+            let num_mappings = parser.read_uleb128_u32()?;
+            println!("Number of file {i} mappings: {num_mappings}");
+
+            for _ in 0..num_mappings {
+                let (kind, region) = parser.read_mapping_kind_and_region()?;
+                println!("- {kind:?} at {region:?}");
+
+                match kind {
+                    // Also print expression mappings in resolved form.
+                    MappingKind::Code(term @ CovTerm::Expression { .. })
+                    | MappingKind::Gap(term @ CovTerm::Expression { .. }) => {
+                        println!("    = {}", expression_resolver.format_term(term));
+                    }
+                    // If the mapping is a branch region, print both of its arms
+                    // in resolved form (even if they aren't expressions).
+                    MappingKind::Branch { r#true, r#false } => {
+                        println!("    true  = {}", expression_resolver.format_term(r#true));
+                        println!("    false = {}", expression_resolver.format_term(r#false));
+                    }
+                    _ => (),
+                }
+            }
+        }
+
+        parser.ensure_empty()?;
+        println!();
+    }
+    Ok(())
+}
+
+struct CovfunLineData {
+    name_hash: u64,
+    is_used: bool,
+    payload: Vec<u8>,
+}
+
+/// Checks a line of LLVM IR assembly to see if it contains an `__llvm_covfun`
+/// entry, and if so extracts relevant data in a `CovfunLineData`.
+fn covfun_line_data(line: &str) -> Option<CovfunLineData> {
+    let re = {
+        // We cheat a little bit and match variable names `@__covrec_[HASH]u`
+        // rather than the section name, because the section name is harder to
+        // extract and differs across Linux/Windows/macOS. We also extract the
+        // symbol name hash from the variable name rather than the data, since
+        // it's easier and both should match.
+        static RE: OnceLock<Regex> = OnceLock::new();
+        RE.get_or_init(|| {
+            Regex::new(
+                r#"^@__covrec_(?<name_hash>[0-9A-Z]+)(?<is_used>u)? = .*\[[0-9]+ x i8\] c"(?<payload>[^"]*)".*$"#,
+            )
+            .unwrap()
+        })
+    };
+
+    let captures = re.captures(line)?;
+    let name_hash = u64::from_str_radix(&captures["name_hash"], 16).unwrap();
+    let is_used = captures.name("is_used").is_some();
+    let payload = unescape_llvm_string_contents(&captures["payload"]);
+
+    Some(CovfunLineData { name_hash, is_used, payload })
+}
+
+// Extra parser methods only needed when parsing `covfun` payloads.
+impl<'a> Parser<'a> {
+    fn read_simple_term(&mut self) -> anyhow::Result<CovTerm> {
+        let raw_term = self.read_uleb128_u32()?;
+        CovTerm::decode(raw_term).context("decoding term")
+    }
+
+    fn read_mapping_kind_and_region(&mut self) -> anyhow::Result<(MappingKind, MappingRegion)> {
+        let mut kind = self.read_raw_mapping_kind()?;
+        let mut region = self.read_raw_mapping_region()?;
+
+        const HIGH_BIT: u32 = 1u32 << 31;
+        if region.end_column & HIGH_BIT != 0 {
+            region.end_column &= !HIGH_BIT;
+            kind = match kind {
+                MappingKind::Code(term) => MappingKind::Gap(term),
+                // LLVM's coverage mapping reader will actually handle this
+                // case without complaint, but the result is almost certainly
+                // a meaningless implementation artifact.
+                _ => return Err(anyhow!("unexpected base kind for gap region: {kind:?}")),
+            }
+        }
+
+        Ok((kind, region))
+    }
+
+    fn read_raw_mapping_kind(&mut self) -> anyhow::Result<MappingKind> {
+        let raw_mapping_kind = self.read_uleb128_u32()?;
+        if let Some(term) = CovTerm::decode(raw_mapping_kind) {
+            return Ok(MappingKind::Code(term));
+        }
+
+        assert_eq!(raw_mapping_kind & 0b11, 0);
+        assert_ne!(raw_mapping_kind, 0);
+
+        let (high, is_expansion) = (raw_mapping_kind >> 3, raw_mapping_kind & 0b100 != 0);
+        if is_expansion {
+            Ok(MappingKind::Expansion(high))
+        } else {
+            match high {
+                0 => unreachable!("zero kind should have already been handled as a code mapping"),
+                2 => Ok(MappingKind::Skip),
+                4 => {
+                    let r#true = self.read_simple_term()?;
+                    let r#false = self.read_simple_term()?;
+                    Ok(MappingKind::Branch { r#true, r#false })
+                }
+                _ => Err(anyhow!("unknown mapping kind: {raw_mapping_kind:#x}")),
+            }
+        }
+    }
+
+    fn read_raw_mapping_region(&mut self) -> anyhow::Result<MappingRegion> {
+        let start_line_offset = self.read_uleb128_u32()?;
+        let start_column = self.read_uleb128_u32()?;
+        let end_line_offset = self.read_uleb128_u32()?;
+        let end_column = self.read_uleb128_u32()?;
+        Ok(MappingRegion { start_line_offset, start_column, end_line_offset, end_column })
+    }
+}
+
+/// Enum that can hold a constant zero value, the ID of an physical coverage
+/// counter, or the ID (and operation) of a coverage-counter expression.
+///
+/// Terms are used as the operands of coverage-counter expressions, as the arms
+/// of branch mappings, and as the value of code/gap mappings.
+#[derive(Clone, Copy, Debug)]
+pub(crate) enum CovTerm {
+    Zero,
+    Counter(u32),
+    Expression(u32, Op),
+}
+
+/// Operator (addition or subtraction) used by an expression.
+#[derive(Clone, Copy, Debug)]
+pub(crate) enum Op {
+    Sub,
+    Add,
+}
+
+impl CovTerm {
+    pub(crate) fn decode(input: u32) -> Option<Self> {
+        let (high, tag) = (input >> 2, input & 0b11);
+        match tag {
+            0b00 if high == 0 => Some(Self::Zero),
+            0b01 => Some(Self::Counter(high)),
+            0b10 => Some(Self::Expression(high, Op::Sub)),
+            0b11 => Some(Self::Expression(high, Op::Add)),
+            // When reading expression operands or branch arms, the LLVM coverage
+            // mapping reader will always interpret a `0b00` tag as a zero
+            // term, even when the high bits are non-zero.
+            // We treat that case as failure instead, so that this code can be
+            // shared by the full mapping-kind reader as well.
+            _ => None,
+        }
+    }
+}
+
+#[derive(Debug)]
+enum MappingKind {
+    Code(CovTerm),
+    Gap(CovTerm),
+    Expansion(u32),
+    Skip,
+    // Using raw identifiers here makes the dump output a little bit nicer
+    // (via the derived Debug), at the expense of making this tool's source
+    // code a little bit uglier.
+    Branch { r#true: CovTerm, r#false: CovTerm },
+}
+
+struct MappingRegion {
+    /// Offset of this region's start line, relative to the *start line* of
+    /// the *previous mapping* (or 0). Line numbers are 1-based.
+    start_line_offset: u32,
+    /// This region's start column, absolute and 1-based.
+    start_column: u32,
+    /// Offset of this region's end line, relative to the *this mapping's*
+    /// start line. Line numbers are 1-based.
+    end_line_offset: u32,
+    /// This region's end column, absolute, 1-based, and exclusive.
+    ///
+    /// If the highest bit is set, that bit is cleared and the associated
+    /// mapping becomes a gap region mapping.
+    end_column: u32,
+}
+
+impl Debug for MappingRegion {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        write!(
+            f,
+            "(prev + {}, {}) to (start + {}, {})",
+            self.start_line_offset, self.start_column, self.end_line_offset, self.end_column
+        )
+    }
+}
+
+/// Helper type that prints expressions in a "resolved" form, so that
+/// developers reading the dump don't need to resolve expressions by hand.
+struct ExpressionResolver {
+    operands: Vec<(CovTerm, CovTerm)>,
+}
+
+impl ExpressionResolver {
+    fn new() -> Self {
+        Self { operands: Vec::new() }
+    }
+
+    fn push_operands(&mut self, lhs: CovTerm, rhs: CovTerm) {
+        self.operands.push((lhs, rhs));
+    }
+
+    fn format_term(&self, term: CovTerm) -> String {
+        let mut output = String::new();
+        self.write_term(&mut output, term);
+        output
+    }
+
+    fn write_term(&self, output: &mut String, term: CovTerm) {
+        match term {
+            CovTerm::Zero => output.push_str("Zero"),
+            CovTerm::Counter(id) => write!(output, "c{id}").unwrap(),
+            CovTerm::Expression(id, op) => {
+                let (lhs, rhs) = self.operands[id as usize];
+                let op = match op {
+                    Op::Sub => "-",
+                    Op::Add => "+",
+                };
+
+                output.push('(');
+                self.write_term(output, lhs);
+                write!(output, " {op} ").unwrap();
+                self.write_term(output, rhs);
+                output.push(')');
+            }
+        }
+    }
+}

src/tools/coverage-dump/src/main.rs

@@ -0,0 +1,17 @@
+mod covfun;
+mod parser;
+mod prf_names;
+
+fn main() -> anyhow::Result<()> {
+    use anyhow::Context as _;
+
+    let args = std::env::args().collect::<Vec<_>>();
+
+    let llvm_ir_path = args.get(1).context("LLVM IR file not specified")?;
+    let llvm_ir = std::fs::read_to_string(llvm_ir_path).context("couldn't read LLVM IR file")?;
+
+    let function_names = crate::prf_names::make_function_names_table(&llvm_ir)?;
+    crate::covfun::dump_covfun_mappings(&llvm_ir, &function_names)?;
+
+    Ok(())
+}

src/tools/coverage-dump/src/parser.rs

@@ -0,0 +1,80 @@
+#[cfg(test)]
+mod tests;
+
+use anyhow::ensure;
+use regex::bytes;
+use std::sync::OnceLock;
+
+/// Given the raw contents of a string literal in LLVM IR assembly, decodes any
+/// backslash escapes and returns a vector containing the resulting byte string.
+pub(crate) fn unescape_llvm_string_contents(contents: &str) -> Vec<u8> {
+    let escape_re = {
+        static RE: OnceLock<bytes::Regex> = OnceLock::new();
+        // LLVM IR supports two string escapes: `\\` and `\xx`.
+        RE.get_or_init(|| bytes::Regex::new(r"\\\\|\\([0-9A-Za-z]{2})").unwrap())
+    };
+
+    fn u8_from_hex_digits(digits: &[u8]) -> u8 {
+        // We know that the input contains exactly 2 hex digits, so these calls
+        // should never fail.
+        assert_eq!(digits.len(), 2);
+        let digits = std::str::from_utf8(digits).unwrap();
+        u8::from_str_radix(digits, 16).unwrap()
+    }
+
+    escape_re
+        .replace_all(contents.as_bytes(), |captures: &bytes::Captures<'_>| {
+            let byte = match captures.get(1) {
+                None => b'\\',
+                Some(hex_digits) => u8_from_hex_digits(hex_digits.as_bytes()),
+            };
+            [byte]
+        })
+        .into_owned()
+}
+
+pub(crate) struct Parser<'a> {
+    rest: &'a [u8],
+}
+
+impl<'a> Parser<'a> {
+    pub(crate) fn new(input: &'a [u8]) -> Self {
+        Self { rest: input }
+    }
+
+    pub(crate) fn ensure_empty(self) -> anyhow::Result<()> {
+        ensure!(self.rest.is_empty(), "unparsed bytes: 0x{:02x?}", self.rest);
+        Ok(())
+    }
+
+    pub(crate) fn read_n_bytes(&mut self, n: usize) -> anyhow::Result<&'a [u8]> {
+        ensure!(n <= self.rest.len());
+
+        let (bytes, rest) = self.rest.split_at(n);
+        self.rest = rest;
+        Ok(bytes)
+    }
+
+    pub(crate) fn read_uleb128_u32(&mut self) -> anyhow::Result<u32> {
+        self.read_uleb128_u64_and_convert()
+    }
+
+    pub(crate) fn read_uleb128_usize(&mut self) -> anyhow::Result<usize> {
+        self.read_uleb128_u64_and_convert()
+    }
+
+    fn read_uleb128_u64_and_convert<T>(&mut self) -> anyhow::Result<T>
+    where
+        T: TryFrom<u64> + 'static,
+        T::Error: std::error::Error + Send + Sync,
+    {
+        let mut temp_rest = self.rest;
+        let raw_value: u64 = leb128::read::unsigned(&mut temp_rest)?;
+        let converted_value = T::try_from(raw_value)?;
+
+        // Only update `self.rest` if the above steps succeeded, so that the
+        // parser position can be used for error reporting if desired.
+        self.rest = temp_rest;
+        Ok(converted_value)
+    }
+}

src/tools/coverage-dump/src/parser/tests.rs

@@ -0,0 +1,38 @@
+use super::unescape_llvm_string_contents;
+
+// WARNING: These tests don't necessarily run in CI, and were mainly used to
+// help track down problems when originally developing this tool.
+// (The tool is still tested indirectly by snapshot tests that rely on it.)
+
+// Tests for `unescape_llvm_string_contents`:
+
+#[test]
+fn unescape_empty() {
+    assert_eq!(unescape_llvm_string_contents(""), &[]);
+}
+
+#[test]
+fn unescape_noop() {
+    let input = "The quick brown fox jumps over the lazy dog.";
+    assert_eq!(unescape_llvm_string_contents(input), input.as_bytes());
+}
+
+#[test]
+fn unescape_backslash() {
+    let input = r"\\Hello\\world\\";
+    assert_eq!(unescape_llvm_string_contents(input), r"\Hello\world\".as_bytes());
+}
+
+#[test]
+fn unescape_hex() {
+    let input = r"\01\02\03\04\0a\0b\0C\0D\fd\fE\FF";
+    let expected: &[u8] = &[0x01, 0x02, 0x03, 0x04, 0x0a, 0x0b, 0x0c, 0x0d, 0xfd, 0xfe, 0xff];
+    assert_eq!(unescape_llvm_string_contents(input), expected);
+}
+
+#[test]
+fn unescape_mixed() {
+    let input = r"\\01.\5c\5c";
+    let expected: &[u8] = br"\01.\\";
+    assert_eq!(unescape_llvm_string_contents(input), expected);
+}

src/tools/coverage-dump/src/prf_names.rs

@@ -0,0 +1,87 @@
+use crate::parser::{unescape_llvm_string_contents, Parser};
+use anyhow::{anyhow, ensure};
+use regex::Regex;
+use std::collections::HashMap;
+use std::sync::OnceLock;
+
+/// Scans through the contents of an LLVM IR assembly file to find `__llvm_prf_names`
+/// entries, decodes them, and creates a table that maps name hash values to
+/// (demangled) function names.
+pub(crate) fn make_function_names_table(llvm_ir: &str) -> anyhow::Result<HashMap<u64, String>> {
+    fn prf_names_payload(line: &str) -> Option<&str> {
+        let re = {
+            // We cheat a little bit and match the variable name `@__llvm_prf_nm`
+            // rather than the section name, because the section name is harder
+            // to extract and differs across Linux/Windows/macOS.
+            static RE: OnceLock<Regex> = OnceLock::new();
+            RE.get_or_init(|| {
+                Regex::new(r#"^@__llvm_prf_nm =.*\[[0-9]+ x i8\] c"([^"]*)".*$"#).unwrap()
+            })
+        };
+
+        let payload = re.captures(line)?.get(1).unwrap().as_str();
+        Some(payload)
+    }
+
+    /// LLVM's profiler/coverage metadata often uses an MD5 hash truncated to
+    /// 64 bits as a way to associate data stored in different tables/sections.
+    fn truncated_md5(bytes: &[u8]) -> u64 {
+        use md5::{Digest, Md5};
+        let mut hasher = Md5::new();
+        hasher.update(bytes);
+        let hash: [u8; 8] = hasher.finalize().as_slice()[..8].try_into().unwrap();
+        // The truncated hash is explicitly little-endian, regardless of host
+        // or target platform. (See `MD5Result::low` in LLVM's `MD5.h`.)
+        u64::from_le_bytes(hash)
+    }
+
+    fn demangle_if_able(symbol_name_bytes: &[u8]) -> anyhow::Result<String> {
+        // In practice, raw symbol names should always be ASCII.
+        let symbol_name_str = std::str::from_utf8(symbol_name_bytes)?;
+        match rustc_demangle::try_demangle(symbol_name_str) {
+            Ok(d) => Ok(format!("{d:#}")),
+            // If demangling failed, don't treat it as an error. This lets us
+            // run the dump tool against non-Rust coverage maps produced by
+            // `clang`, for testing purposes.
+            Err(_) => Ok(format!("(couldn't demangle) {symbol_name_str}")),
+        }
+    }
+
+    let mut map = HashMap::new();
+
+    for payload in llvm_ir.lines().filter_map(prf_names_payload).map(unescape_llvm_string_contents)
+    {
+        let mut parser = Parser::new(&payload);
+        let uncompressed_len = parser.read_uleb128_usize()?;
+        let compressed_len = parser.read_uleb128_usize()?;
+
+        let uncompressed_bytes_vec;
+        let uncompressed_bytes: &[u8] = if compressed_len == 0 {
+            // The symbol name bytes are uncompressed, so read them directly.
+            parser.read_n_bytes(uncompressed_len)?
+        } else {
+            // The symbol name bytes are compressed, so read and decompress them.
+            let compressed_bytes = parser.read_n_bytes(compressed_len)?;
+
+            uncompressed_bytes_vec = miniz_oxide::inflate::decompress_to_vec_zlib_with_limit(
+                compressed_bytes,
+                uncompressed_len,
+            )
+            .map_err(|e| anyhow!("{e:?}"))?;
+            ensure!(uncompressed_bytes_vec.len() == uncompressed_len);
+
+            &uncompressed_bytes_vec
+        };
+
+        // Symbol names in the payload are separated by `0x01` bytes.
+        for raw_name in uncompressed_bytes.split(|&b| b == 0x01) {
+            let hash = truncated_md5(raw_name);
+            let demangled = demangle_if_able(raw_name)?;
+            map.insert(hash, demangled);
+        }
+
+        parser.ensure_empty()?;
+    }
+
+    Ok(map)
+}