Snapshot reference Markdown output

Add elapsed time to progress bar

doc/hyperfine.1

@@ -116,13 +116,17 @@ not every time as would happen with the \fB\-\-prepare\fR option.
 .HP
 \fB\-\-reference\fR \fICMD\fP
 .IP
-Benchmark \fICMD\fP as the reference for the relative comparison of results.
-If this option is not specified, the fastest command is used as the reference.
+Use \fICMD\fP as the reference for the relative comparison of results. Without
+parameters, benchmark \fICMD\fP as a separate reference command. If this option
+is not specified, the fastest command is used as the reference.
+With \fB\-\-parameter\-scan\fR or \fB\-\-parameter\-list\fR, select an existing
+benchmark by its exact printed name. The name must match exactly one benchmark.
 .HP
 \fB\-\-reference\-name\fR \fINAME\fP
 .IP
 Give a meaningful \fINAME\fP to the reference command.
 Requires \fB\-\-reference\fR.
+Cannot be used with parameterized benchmarks; use \fB\-\-command\-name\fR instead.
 .HP
 \fB\-p\fR, \fB\-\-prepare\fR \fICMD...\fP
 .IP

src/benchmark/relative_speed.rs

@@ -16,6 +16,20 @@ pub struct BenchmarkResultWithRelativeSpeed<'a> {
     pub relative_ordering: Ordering,
 }
 
+impl BenchmarkResultWithRelativeSpeed<'_> {
+    pub fn reference_label(&self) -> &'static str {
+        if self.is_reference {
+            " (reference)"
+        } else {
+            match self.relative_ordering {
+                Ordering::Less => " (faster)",
+                Ordering::Equal => " (same speed)",
+                Ordering::Greater => " (slower)",
+            }
+        }
+    }
+}
+
 pub fn compare_mean_time(l: &BenchmarkResult, r: &BenchmarkResult) -> Ordering {
     l.mean_wall_clock_time()
         .partial_cmp(&r.mean_wall_clock_time())
@@ -37,10 +51,13 @@ fn compute_relative_speeds<'a>(
     let mut results: Vec<_> = results
         .iter()
         .map(|result| {
-            let is_reference = result == reference;
+            // Separate benchmarks can have identical names and measurements.
+            let is_reference = std::ptr::eq(result, reference);
             let relative_ordering = compare_mean_time(result, reference);
 
-            if result.mean_wall_clock_time() == Time::zero() {
+            if result.mean_wall_clock_time() == Time::zero()
+                || reference.mean_wall_clock_time() == Time::zero()
+            {
                 return BenchmarkResultWithRelativeSpeed {
                     result,
                     relative_speed: if is_reference { 1.0 } else { f64::INFINITY },
@@ -64,11 +81,11 @@ fn compute_relative_speeds<'a>(
                 result.measurements.stddev(),
                 reference.measurements.stddev(),
             ) {
-                (Some(result_stddev), Some(fastest_stddev)) => Some(
+                (Some(result_stddev), Some(reference_stddev)) => Some(
                     ratio
                         * ((result_stddev / result.mean_wall_clock_time())
                             .powi(uom::typenum::P2::new())
-                            + (fastest_stddev / reference.mean_wall_clock_time())
+                            + (reference_stddev / reference.mean_wall_clock_time())
                                 .powi(uom::typenum::P2::new()))
                         .sqrt(),
                 ),
@@ -122,14 +139,14 @@ pub fn compute_with_check(
     Some(compute_relative_speeds(results, fastest, sort_order))
 }
 
-/// Same as compute_with_check, potentially resulting in relative speeds of infinity
-pub fn compute(
-    results: &[BenchmarkResult],
+/// Compute relative speeds against the given reference, or the fastest result.
+pub fn compute<'a>(
+    results: &'a [BenchmarkResult],
     sort_order: SortOrder,
-) -> Vec<BenchmarkResultWithRelativeSpeed<'_>> {
-    let fastest = fastest_of(results);
-
-    compute_relative_speeds(results, fastest, sort_order)
+    reference: Option<&'a BenchmarkResult>,
+) -> Vec<BenchmarkResultWithRelativeSpeed<'a>> {
+    let reference = reference.unwrap_or_else(|| fastest_of(results));
+    compute_relative_speeds(results, reference, sort_order)
 }
 
 #[cfg(test)]
@@ -192,3 +209,28 @@ fn test_compute_relative_speed_for_zero_times() {
 
     assert!(annotated_results.is_none());
 }
+
+#[test]
+fn reference_identity_distinguishes_equal_results() {
+    let results = vec![create_result("same", 1.0), create_result("same", 1.0)];
+    let entries = compute_relative_speeds(&results, &results[0], SortOrder::Command);
+
+    assert!(entries[0].is_reference);
+    assert!(!entries[1].is_reference);
+}
+
+#[test]
+fn reference_ratios_handle_zero_times() {
+    let mut results = vec![create_result("reference", 2.0), create_result("zero", 0.0)];
+    for result in &mut results {
+        let measurement = result.measurements.measurements[0];
+        result.measurements.measurements.push(measurement);
+    }
+    let entries = compute_relative_speeds(&results, &results[0], SortOrder::Command);
+    assert_eq!(entries[1].relative_speed, f64::INFINITY);
+
+    let entries = compute_relative_speeds(&results, &results[1], SortOrder::Command);
+    assert_eq!(entries[0].relative_speed, f64::INFINITY);
+    assert_eq!(entries[0].relative_speed_stddev, None);
+    assert_eq!(entries[1].relative_speed, 1.0);
+}

src/benchmark/scheduler.rs

@@ -54,7 +54,8 @@ impl<'a> Scheduler<'a> {
 
             // We export results after each individual benchmark, because
             // we would risk losing them if a later benchmark fails.
-            self.export_manager.write_results(&self.results, true)?;
+            self.export_manager
+                .write_results(&self.results, true, self.options.reference_index)?;
         }
 
         Ok(())
@@ -71,9 +72,8 @@ impl<'a> Scheduler<'a> {
 
         let reference = self
             .options
-            .reference_command
-            .as_ref()
-            .map(|_| &self.results[0])
+            .reference_index
+            .map(|index| &self.results[index])
             .unwrap_or_else(|| relative_speed::fastest_of(&self.results));
 
         if let Some(annotated_results) = relative_speed::compute_with_check_from_reference(
@@ -130,9 +130,14 @@ impl<'a> Scheduler<'a> {
                     console_writeln!(stdout, "{}", "Relative speed comparison".bold())?;
 
                     for item in annotated_results {
+                        let relationship = if self.options.reference_index.is_some() {
+                            item.reference_label()
+                        } else {
+                            ""
+                        };
                         console_writeln!(
                             stdout,
-                            "  {}{}  {}",
+                            "  {}{}  {}{}",
                             format!("{:10.2}", item.relative_speed).bold().green(),
                             if item.is_reference {
                                 "        ".into()
@@ -142,6 +147,7 @@ impl<'a> Scheduler<'a> {
                                 "        ".into()
                             },
                             item.result.command_with_unused_parameters,
+                            relationship,
                         )?;
                     }
                 }
@@ -163,7 +169,8 @@ impl<'a> Scheduler<'a> {
     }
 
     pub fn final_export(&self) -> Result<()> {
-        self.export_manager.write_results(&self.results, false)
+        self.export_manager
+            .write_results(&self.results, false, self.options.reference_index)
     }
 }
 

src/cli.rs

@@ -105,8 +105,11 @@ fn build_command() -> Command {
                 .action(ArgAction::Set)
                 .value_name("CMD")
                 .help(
-                    "The reference command for the relative comparison of results. \
-                    If this is unset, results are compared with the fastest command as reference."
+                    "The reference for the relative comparison of results. Without parameters, \
+                    CMD is run as a separate reference command. With --parameter-scan or \
+                    --parameter-list, CMD must exactly match one unique benchmark name as \
+                    shown in the output. If this is unset, results are compared with the \
+                    fastest command as reference."
                 )
         )
         .arg(
@@ -114,7 +117,8 @@ fn build_command() -> Command {
                 .long("reference-name")
                 .action(ArgAction::Set)
                 .value_name("CMD")
-                .help("Give a meaningful name to the reference command.")
+                .help("Give a meaningful name to the reference command. This cannot be used \
+                       with parameterized benchmarks; use --command-name instead.")
                 .requires("reference")
         )
         .arg(

src/export/csv.rs

@@ -18,6 +18,7 @@ impl Exporter for CsvExporter {
         results: &[BenchmarkResult],
         _unit: Option<TimeUnit>,
         _sort_order: SortOrder,
+        _reference_index: Option<usize>,
     ) -> Result<Vec<u8>> {
         const CSV_UNIT: TimeUnit = TimeUnit::Second;
         const CSV_PRECISION: usize = 6;
@@ -172,7 +173,7 @@ fn test_csv() {
 
     let actual = String::from_utf8(
         exporter
-            .serialize(&results, Some(TimeUnit::Second), SortOrder::Command)
+            .serialize(&results, Some(TimeUnit::Second), SortOrder::Command, None)
             .unwrap(),
     )
     .unwrap();

src/export/json.rs

@@ -22,6 +22,7 @@ impl Exporter for JsonExporter {
         results: &[BenchmarkResult],
         _unit: Option<TimeUnit>,
         _sort_order: SortOrder,
+        _reference_index: Option<usize>,
     ) -> Result<Vec<u8>> {
         let mut output = to_vec_pretty(&HyperfineSummary { results });
         if let Ok(ref mut content) = output {

src/export/markup.rs

@@ -1,10 +1,11 @@
 use crate::benchmark::relative_speed::BenchmarkResultWithRelativeSpeed;
 use crate::benchmark::{benchmark_result::BenchmarkResult, relative_speed};
 use crate::options::SortOrder;
-use crate::quantity::{FormatQuantity, IsUnit, TimeUnit};
+use crate::quantity::{FormatQuantity, IsUnit, Time, TimeUnit, Zero};
 
 use super::Exporter;
 use anyhow::Result;
+use std::cmp::Ordering;
 
 pub enum Alignment {
     Left,
@@ -16,6 +17,8 @@ pub trait MarkupExporter {
         &self,
         entries: &[BenchmarkResultWithRelativeSpeed],
         unit: TimeUnit,
+        reference_pending: bool,
+        explicit_reference: bool,
     ) -> String {
         // prepare table header strings
         let notation = format!("[{}]", unit.short_name());
@@ -58,13 +61,25 @@ pub trait MarkupExporter {
             };
             let min_str = measurement.measurements.min().format_value(unit);
             let max_str = measurement.measurements.max().format_value(unit);
-            let rel_str = format!("{:.2}", entry.relative_speed);
-            let rel_stddev_str = if entry.is_reference {
-                "".into()
-            } else if let Some(stddev) = entry.relative_speed_stddev {
-                format!(" ± {stddev:.2}")
+            // The ratio of two zero times is undefined, even if they compare equal.
+            let relative_unavailable = reference_pending
+                || (explicit_reference
+                    && !entry.is_reference
+                    && entry.relative_ordering == Ordering::Equal
+                    && measurement.mean_wall_clock_time() == Time::zero());
+            let relative = if relative_unavailable {
+                "N/A".to_string()
             } else {
-                "".into()
+                let stddev = match (entry.is_reference, entry.relative_speed_stddev) {
+                    (false, Some(stddev)) => format!(" ± {stddev:.2}"),
+                    _ => String::new(),
+                };
+                let label = if explicit_reference {
+                    entry.reference_label()
+                } else {
+                    ""
+                };
+                format!("{:.2}{stddev}{label}", entry.relative_speed)
             };
 
             // prepare table row entries
@@ -73,7 +88,7 @@ pub trait MarkupExporter {
                 &format!("{mean_str}{stddev_str}"),
                 &min_str,
                 &max_str,
-                &format!("{rel_str}{rel_stddev_str}"),
+                &relative,
             ]))
         }
 
@@ -114,11 +129,20 @@ impl<T: MarkupExporter> Exporter for T {
         results: &[BenchmarkResult],
         unit: Option<TimeUnit>,
         sort_order: SortOrder,
+        reference_index: Option<usize>,
     ) -> Result<Vec<u8>> {
         let unit = unit.unwrap_or_else(|| determine_unit_from_results(results));
-        let entries = relative_speed::compute(results, sort_order);
+        // Do not report ratios against another benchmark while the selected
+        // reference is still pending in an intermediate export.
+        let reference_pending = reference_index.is_some_and(|i| i >= results.len());
+        let entries = relative_speed::compute(
+            results,
+            sort_order,
+            reference_index.and_then(|i| results.get(i)),
+        );
 
-        let table = self.table_results(&entries, unit);
+        let table =
+            self.table_results(&entries, unit, reference_pending, reference_index.is_some());
         Ok(table.as_bytes().to_vec())
     }
 }

src/export/mod.rs

@@ -51,6 +51,7 @@ trait Exporter {
         results: &[BenchmarkResult],
         unit: Option<TimeUnit>,
         sort_order: SortOrder,
+        reference_index: Option<usize>,
     ) -> Result<Vec<u8>>;
 }
 
@@ -130,11 +131,18 @@ impl ExportManager {
     /// results are written to all file targets (to always have them up to date, even
     /// if a benchmark fails). In the latter case, we only print to stdout targets (in
     /// order not to clutter the output of hyperfine with intermediate results).
-    pub fn write_results(&self, results: &[BenchmarkResult], intermediate: bool) -> Result<()> {
+    /// `reference_index` refers to the complete benchmark sequence, so it can
+    /// point past `results` while that benchmark is still pending.
+    pub fn write_results(
+        &self,
+        results: &[BenchmarkResult],
+        intermediate: bool,
+        reference_index: Option<usize>,
+    ) -> Result<()> {
         for e in &self.exporters {
             let content = || {
                 e.exporter
-                    .serialize(results, self.time_unit, self.sort_order)
+                    .serialize(results, self.time_unit, self.sort_order, reference_index)
             };
 
             match e.target {

src/export/tests.rs

@@ -14,7 +14,7 @@ fn get_output<E: Exporter + Default>(
     sort_order: SortOrder,
 ) -> String {
     let exporter = E::default();
-    String::from_utf8(exporter.serialize(results, unit, sort_order).unwrap()).unwrap()
+    String::from_utf8(exporter.serialize(results, unit, sort_order, None).unwrap()).unwrap()
 }
 
 /// Ensure the makrup output includes the table header and the multiple

src/options.rs

@@ -4,7 +4,7 @@ use std::path::PathBuf;
 use std::process::{Command, Stdio};
 use std::{cmp, env, fmt, io};
 
-use anyhow::ensure;
+use anyhow::{bail, ensure};
 use clap::ArgMatches;
 
 use crate::command::Commands;
@@ -208,12 +208,15 @@ pub struct Options {
     /// Whether or not to ignore non-zero exit codes
     pub command_failure_action: CmdFailureAction,
 
-    // Command to use as a reference for relative speed comparison
+    // Standalone reference command, cleared when an existing benchmark is selected.
     pub reference_command: Option<String>,
 
     // Name of the reference command
     pub reference_name: Option<String>,
 
+    // Index of the reference in the benchmark sequence, if any.
+    pub reference_index: Option<usize>,
+
     /// Command(s) to run before each timing run
     pub preparation_command: Option<Vec<String>>,
 
@@ -257,6 +260,7 @@ impl Default for Options {
             command_failure_action: CmdFailureAction::RaiseError,
             reference_command: None,
             reference_name: None,
+            reference_index: None,
             preparation_command: None,
             conclusion_command: None,
             setup_command: None,
@@ -464,6 +468,35 @@ impl Options {
     }
 
     pub fn validate_against_command_list(&mut self, commands: &Commands) -> Result<()> {
+        if let Some(reference) = &self.reference_command {
+            if commands
+                .iter()
+                .next()
+                .is_some_and(|command| !command.get_parameters().is_empty())
+            {
+                ensure!(
+                    self.reference_name.is_none(),
+                    "--reference-name cannot be used with a parameterized reference; use --command-name to name the benchmark"
+                );
+                let mut matches = commands
+                    .iter()
+                    .enumerate()
+                    .filter(|(_, command)| command.get_name_with_unused_parameters() == *reference);
+                let Some((index, _)) = matches.next() else {
+                    bail!(
+                        "Reference '{reference}' does not match any parameterized benchmark. Use its full displayed benchmark name."
+                    );
+                };
+                ensure!(
+                    matches.next().is_none(),
+                    "Reference '{reference}' matches multiple parameterized benchmarks. Use --command-name to give them unique names."
+                );
+                self.reference_index = Some(index);
+                self.reference_command = None;
+            } else {
+                self.reference_index = Some(0);
+            }
+        }
         let has_reference_command = self.reference_command.is_some();
         let num_commands = commands.num_commands(has_reference_command);
 

src/timer/windows_timer.rs

@@ -34,11 +34,11 @@ use crate::quantity::{nanosecond, Information, Time, Zero};
 
 #[cfg(not(feature = "windows_process_extensions_main_thread_handle"))]
 #[allow(non_upper_case_globals)]
-static NtResumeProcess: Lazy<unsafe extern "system" fn(ProcessHandle: HANDLE) -> NTSTATUS> =
+static NtResumeProcess: Lazy<unsafe extern "system" fn(process_handle: HANDLE) -> NTSTATUS> =
     Lazy::new(|| {
         // SAFETY: Getting the module handle for ntdll.dll is safe
         let ntdll = unsafe { GetModuleHandleW(w!("ntdll.dll")) };
-        assert!(ntdll != std::ptr::null_mut(), "GetModuleHandleW failed");
+        assert!(!ntdll.is_null(), "GetModuleHandleW failed");
 
         // SAFETY: The ntdll handle is valid
         let nt_resume_process = unsafe { GetProcAddress(ntdll, s!("NtResumeProcess")) };
@@ -57,10 +57,7 @@ impl CPUTimer {
 
         // SAFETY: Creating a new job object is safe
         let job_object = unsafe { CreateJobObjectW(ptr::null_mut(), ptr::null_mut()) };
-        assert!(
-            job_object != std::ptr::null_mut(),
-            "CreateJobObjectW failed"
-        );
+        assert!(!job_object.is_null(), "CreateJobObjectW failed");
 
         // SAFETY: The job object handle is valid
         let ret = unsafe { AssignProcessToJobObject(job_object, child_handle) };

tests/execution_order_tests.rs

@@ -383,26 +383,6 @@ fn reference_is_executed_first() {
         .run();
 }
 
-#[test]
-fn reference_is_executed_first_parameter_value() {
-    ExecutionOrderTest::new()
-        .arg("--runs=2")
-        .reference("reference")
-        .arg("--parameter-list")
-        .arg("number")
-        .arg("1,2,3")
-        .command("command {number}")
-        .expect_output("reference")
-        .expect_output("reference")
-        .expect_output("command 1")
-        .expect_output("command 1")
-        .expect_output("command 2")
-        .expect_output("command 2")
-        .expect_output("command 3")
-        .expect_output("command 3")
-        .run();
-}
-
 #[test]
 fn reference_is_executed_separately_from_commands() {
     ExecutionOrderTest::new()

tests/integration_tests.rs

@@ -705,6 +705,68 @@ fn shows_benchmark_comparison_relative_to_reference() {
     ");
 }
 
+#[test]
+fn command_sorted_comparison_and_markup_identify_reference() {
+    let _settings = snapshot_settings().bind_to_scope();
+    let directory = tempfile::tempdir().unwrap();
+    let export_path = directory.path().join("results.md");
+    let output = hyperfine_debug()
+        .args([
+            "--style=basic",
+            "--runs=1",
+            "--sort=command",
+            "--reference",
+            "sleep 2",
+            "--reference-name",
+            "baseline",
+            "--export-markdown",
+        ])
+        .arg(&export_path)
+        .args(["sleep 1", "sleep 2", "sleep 3"])
+        .output()
+        .unwrap();
+    assert!(output.status.success(), "{:?}", output);
+    let stdout = String::from_utf8(output.stdout).unwrap();
+    let comparison = stdout.split_once("Relative speed comparison\n").unwrap().1;
+    insta::assert_snapshot!(comparison, @r"
+            1.00          baseline (reference)
+            2.00          sleep 1 (faster)
+            1.00          sleep 2 (same speed)
+            1.50          sleep 3 (slower)
+    ");
+    let markdown = std::fs::read_to_string(export_path).unwrap();
+    insta::assert_snapshot!(markdown, @r"
+    | Command | Mean [s] | Min [s] | Max [s] | Relative |
+    |:---|---:|---:|---:|---:|
+    | `baseline` | 2.000 | 2.000 | 2.000 | 1.00 (reference) |
+    | `sleep 1` | 1.000 | 1.000 | 1.000 | 2.00 (faster) |
+    | `sleep 2` | 2.000 | 2.000 | 2.000 | 1.00 (same speed) |
+    | `sleep 3` | 3.000 | 3.000 | 3.000 | 1.50 (slower) |
+    ");
+
+    // Equal zero times have no meaningful relative factor.
+    let output = hyperfine_debug()
+        .args([
+            "--style=none",
+            "--runs=1",
+            "--export-markdown=-",
+            "--reference",
+            "sleep 0",
+            "--reference-name",
+            "baseline",
+            "sleep 0",
+        ])
+        .output()
+        .unwrap();
+    assert!(output.status.success(), "{:?}", output);
+    let markdown = String::from_utf8(output.stdout).unwrap();
+    let row = markdown
+        .lines()
+        .find(|line| line.contains("sleep 0"))
+        .unwrap();
+    assert!(row.ends_with("| N/A |"), "{}", row);
+}
+
 #[test]
 fn shows_reference_name() {
     let _settings = snapshot_settings().bind_to_scope();
@@ -795,48 +857,180 @@ fn rejects_negative_parameter_steps() {
     }
 }
 
+#[cfg(unix)]
 #[test]
-fn performs_reference_and_all_benchmarks_in_parameter_scan() {
+fn selects_reference_with_parameterized_prepare() {
+    let directory = tempfile::tempdir().unwrap();
+    let csv_path = directory.path().join("results.csv");
+    let json_path = directory.path().join("results.json");
+    let output = hyperfine_raw_command()
+        .args([
+            "--style=basic",
+            "--shell=none",
+            "--runs=1",
+            "-L",
+            "delay",
+            "0.2,0.4,0.6",
+            "--prepare",
+            "sleep {delay}",
+            "--reference",
+            "echo a (delay = 0.4)",
+            "--export-csv",
+        ])
+        .arg(&csv_path)
+        .arg("--export-json")
+        .arg(&json_path)
+        .arg("echo a")
+        .output()
+        .unwrap();
+    assert!(output.status.success(), "{:?}", output);
+    let stdout = String::from_utf8(output.stdout).unwrap();
+    assert_eq!(stdout.matches("Benchmark ").count(), 3);
+    for (number, delay) in ["0.2", "0.4", "0.6"].iter().enumerate() {
+        assert!(
+            stdout.contains(&format!(
+                "Benchmark {}: echo a (delay = {})",
+                number + 1,
+                delay
+            )),
+            "{}",
+            stdout
+        );
+    }
+    assert!(stdout.contains("echo a (delay = 0.4) ran"), "{}", stdout);
+
+    let mut csv = csv::Reader::from_path(csv_path).unwrap();
+    assert!(csv
+        .headers()
+        .unwrap()
+        .iter()
+        .any(|header| header == "parameter_delay"));
+    assert_eq!(
+        csv.records().collect::<Result<Vec<_>, _>>().unwrap().len(),
+        3
+    );
+    let json: serde_json::Value =
+        serde_json::from_slice(&std::fs::read(json_path).unwrap()).unwrap();
+    let results = json["results"].as_array().unwrap();
+    assert_eq!(results.len(), 3);
+    assert_eq!(
+        results
+            .iter()
+            .map(|result| result["parameters"]["delay"].as_str().unwrap())
+            .collect::<Vec<_>>(),
+        ["0.2", "0.4", "0.6"]
+    );
+}
+
+#[test]
+fn rejects_unmatched_reference_before_export_or_setup() {
+    let directory = tempfile::tempdir().unwrap();
+    let export_path = directory.path().join("results.csv");
+    let marker_path = directory.path().join("setup-ran");
+    std::fs::write(&export_path, "previous contents").unwrap();
+
+    hyperfine()
+        .args(["--reference", "sleep 1", "-P", "secs", "2", "3", "--setup"])
+        .arg(format!("echo touched > \"{}\"", marker_path.display()))
+        .arg("--export-csv")
+        .arg(&export_path)
+        .arg("sleep {secs}")
+        .assert()
+        .failure()
+        .stdout(predicate::str::is_empty())
+        .stderr(predicate::str::contains(
+            "does not match any parameterized benchmark",
+        ));
+    assert_eq!(
+        std::fs::read_to_string(export_path).unwrap(),
+        "previous contents"
+    );
+    assert!(!marker_path.exists());
+}
+
+#[test]
+fn rejects_ambiguous_parameterized_reference() {
+    hyperfine_debug()
+        .args([
+            "-L",
+            "x",
+            "1,2",
+            "--command-name",
+            "case",
+            "--reference",
+            "case",
+            "sleep {x}",
+        ])
+        .assert()
+        .failure()
+        .stderr(predicate::str::contains(
+            "matches multiple parameterized benchmarks",
+        ));
+}
+
+#[cfg(unix)]
+#[test]
+fn intermediate_markdown_waits_for_selected_reference() {
+    let directory = tempfile::tempdir().unwrap();
+    let export_path = directory.path().join("results.md");
+    hyperfine()
+        .args([
+            "--style=none",
+            "--shell=none",
+            "--runs=1",
+            "-L",
+            "delay",
+            "0.01,0.02,0.03",
+            "--reference",
+            "sleep 0.02",
+            "--prepare",
+            "true",
+            "--prepare",
+            "false",
+            "--prepare",
+            "true",
+            "--export-markdown",
+        ])
+        .arg(&export_path)
+        .arg("sleep {delay}")
+        .assert()
+        .failure();
+    let markdown = std::fs::read_to_string(export_path).unwrap();
+    let row = markdown
+        .lines()
+        .find(|line| line.contains("sleep 0.01"))
+        .unwrap();
+    assert!(row.ends_with("| N/A |"), "{}", row);
+}
+
+#[test]
+fn markdown_export_uses_selected_parameterized_reference() {
     let _settings = snapshot_settings().bind_to_scope();
     assert_cmd_snapshot!(hyperfine_debug()
-        .arg("--style=basic")
-        .arg("--reference=sleep 25")
-        .arg("--parameter-scan")
-        .arg("time")
-        .arg("30")
-        .arg("45")
-        .arg("--parameter-step-size")
-        .arg("5")
-        .arg("sleep {time}"), @r"
+        .args([
+            "--style=none",
+            "--runs=1",
+            "--export-markdown=-",
+            "-P",
+            "x",
+            "1",
+            "3",
+            "--command-name",
+            "case-{x}",
+            "--reference",
+            "case-2",
+            "sleep {x}",
+        ]), @r"
     success: true
     exit_code: 0
     ----- stdout -----
-    Benchmark 1: sleep 25
-      Time (mean ± σ):     25.000 s ±  0.000 s    [User: 0.000 s, System: 0.000 s]
-      Range (min … max):   25.000 s … 25.000 s    10 runs
 
-    Benchmark 2: sleep 30
-      Time (mean ± σ):     30.000 s ±  0.000 s    [User: 0.000 s, System: 0.000 s]
-      Range (min … max):   30.000 s … 30.000 s    10 runs
+    | Command | Mean [s] | Min [s] | Max [s] | Relative |
+    |:---|---:|---:|---:|---:|
+    | `case-1` | 1.000 | 1.000 | 1.000 | 2.00 (faster) |
+    | `case-2` | 2.000 | 2.000 | 2.000 | 1.00 (reference) |
+    | `case-3` | 3.000 | 3.000 | 3.000 | 1.50 (slower) |
 
-    Benchmark 3: sleep 35
-      Time (mean ± σ):     35.000 s ±  0.000 s    [User: 0.000 s, System: 0.000 s]
-      Range (min … max):   35.000 s … 35.000 s    10 runs
-
-    Benchmark 4: sleep 40
-      Time (mean ± σ):     40.000 s ±  0.000 s    [User: 0.000 s, System: 0.000 s]
-      Range (min … max):   40.000 s … 40.000 s    10 runs
-
-    Benchmark 5: sleep 45
-      Time (mean ± σ):     45.000 s ±  0.000 s    [User: 0.000 s, System: 0.000 s]
-      Range (min … max):   45.000 s … 45.000 s    10 runs
-
-    Summary
-      sleep 25 ran
-        1.20 ± 0.00 times faster than sleep 30
-        1.40 ± 0.00 times faster than sleep 35
-        1.60 ± 0.00 times faster than sleep 40
-        1.80 ± 0.00 times faster than sleep 45
 
     ----- stderr -----
     ");