celox 0.10.1

Celox HDL Simulator
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251

//! Choose compiler concurrency from currently available host resources.

const MAX_WORKERS: usize = 4;

pub(super) fn workers(blocks: usize, instructions: usize, tasks: usize) -> usize {
    choose_workers(
        blocks,
        instructions,
        tasks,
        std::thread::available_parallelism().map_or(1, usize::from),
        available_memory(),
    )
}

fn choose_workers(
    blocks: usize,
    instructions: usize,
    tasks: usize,
    cpus: usize,
    available: Option<u64>,
) -> usize {
    let maximum = MAX_WORKERS.min(cpus.max(1)).min(tasks.max(1));
    let Some(available) = available else {
        // Retain the conservative large-design fallback when the OS cannot
        // report a usable budget; small designs keep their usual parallelism.
        return if blocks >= 65_536 { 1 } else { maximum };
    };
    // Budget the additional SIR/MIR copies and allocator analyses: the shared
    // input is already deducted from available memory. Three overlapping
    // eight-hart functions peaked at 6.54 GiB for the entire process tree.
    // Account for straight-line designs too, not only CFG size.
    // These are conservative estimates, not reservations or memory limits.
    let per_function = (blocks as u64)
        .saturating_mul(24 * 1024)
        .max((instructions as u64).saturating_mul(4096))
        .max(128 * 1024 * 1024);
    // Leave room for the running simulator and unrelated host activity.
    let budget = available - available / 5;
    (budget / per_function).min(maximum as u64).max(1) as usize
}

#[cfg(not(target_os = "linux"))]
fn available_memory() -> Option<u64> {
    None
}

#[cfg(target_os = "linux")]
fn available_memory() -> Option<u64> {
    host_available_memory()
        .into_iter()
        .chain(address_space_available())
        .min()
}

#[cfg(target_os = "linux")]
fn address_space_available() -> Option<u64> {
    let mut limit = libc::rlimit {
        rlim_cur: 0,
        rlim_max: 0,
    };
    // RLIMIT_AS covers allocator arenas, mapped code and thread stacks as well
    // as resident pages. Host/cgroup headroom can be much larger than it.
    if unsafe { libc::getrlimit(libc::RLIMIT_AS, &mut limit) } != 0
        || limit.rlim_cur == libc::RLIM_INFINITY
    {
        return None;
    }
    let status = std::fs::read_to_string("/proc/self/status").ok()?;
    let virtual_bytes = status.lines().find_map(|line| {
        line.strip_prefix("VmSize:")?
            .split_whitespace()
            .next()?
            .parse::<u64>()
            .ok()?
            .checked_mul(1024)
    })?;
    // rlim_t has different widths across Linux ABIs.
    #[allow(clippy::useless_conversion)]
    let limit = u64::try_from(limit.rlim_cur).ok()?;
    Some(limit.saturating_sub(virtual_bytes))
}

#[cfg(target_os = "linux")]
fn host_available_memory() -> Option<u64> {
    let meminfo = std::fs::read_to_string("/proc/meminfo").ok()?;
    let host = meminfo.lines().find_map(|line| {
        let value = line.strip_prefix("MemAvailable:")?;
        value
            .split_whitespace()
            .next()?
            .parse::<u64>()
            .ok()?
            .checked_mul(1024)
    })?;
    let membership = std::fs::read_to_string("/proc/self/cgroup").ok()?;
    let path = membership
        .lines()
        .find_map(|line| line.strip_prefix("0::"))?;
    let mounts = std::fs::read_to_string("/proc/self/mountinfo").ok()?;
    let (root, mount) = mounts.lines().find_map(|line| {
        let (fields, kind) = line.split_once(" - ")?;
        if kind.split_whitespace().next()? != "cgroup2" {
            return None;
        }
        let mut fields = fields.split_whitespace();
        Some((
            unescape_mount(fields.nth(3)?),
            unescape_mount(fields.next()?),
        ))
    })?;
    let relative = std::path::Path::new(path).strip_prefix(&root).ok()?;
    let mount = std::path::PathBuf::from(mount);
    let leaf = mount.join(relative);
    // A cgroup namespace may hide ancestors, but its mounted root still
    // contributes a limit. Walk every visible ancestor, not just the leaf.
    cgroup_available(host, &mount, &leaf)
}

#[cfg(target_os = "linux")]
fn unescape_mount(field: &str) -> String {
    field
        .replace("\\040", " ")
        .replace("\\011", "\t")
        .replace("\\012", "\n")
        .replace("\\134", "\\")
}

#[cfg(target_os = "linux")]
fn cgroup_available(host: u64, mount: &std::path::Path, leaf: &std::path::Path) -> Option<u64> {
    if !leaf.is_dir() || !leaf.starts_with(mount) {
        return None;
    }
    let mut available = host;
    for directory in leaf.ancestors().take_while(|path| path.starts_with(mount)) {
        match std::fs::read_to_string(directory.join("memory.max")) {
            Ok(limit) if limit.trim() == "max" => {}
            Ok(limit) => {
                let limit = limit.trim().parse::<u64>().ok()?;
                let current = std::fs::read_to_string(directory.join("memory.current")).ok()?;
                let current = current.trim().parse::<u64>().ok()?;
                available = available.min(limit.saturating_sub(current));
            }
            // The root of the host cgroup hierarchy has no memory.max.
            Err(error) if directory == mount && error.kind() == std::io::ErrorKind::NotFound => {}
            Err(_) => return None,
        }
    }
    Some(available)
}

#[cfg(test)]
mod tests {
    use super::*;

    const GIB: u64 = 1024 * 1024 * 1024;

    #[cfg(target_os = "linux")]
    #[test]
    fn finite_process_limit_caps_the_worker_memory_budget() {
        const CHILD: &str = "CELOX_TEST_ADDRESS_SPACE_LIMIT";
        // Test-process recursion marker, not a compiler diagnostic option.
        #[allow(clippy::disallowed_methods)]
        let is_child = std::env::var_os(CHILD).is_some();
        if is_child {
            let remaining = address_space_available().expect("finite process limit");
            assert!(remaining < GIB);
            let available = available_memory().expect("process budget is usable without cgroups");
            assert!(available <= GIB);
            assert_eq!(choose_workers(1, 200_000, 4, 16, Some(available)), 1);
            return;
        }
        use std::os::unix::process::CommandExt;
        let current = std::thread::current();
        let name = current.name().expect("named test thread");
        let mut child = std::process::Command::new(std::env::current_exe().unwrap());
        child
            .args(["--exact", name, "--test-threads=1", "--nocapture"])
            .env(CHILD, "1");
        unsafe {
            child.pre_exec(|| {
                let cap: libc::rlim_t = 1024 * 1024 * 1024;
                let limit = libc::rlimit {
                    rlim_cur: cap,
                    rlim_max: cap,
                };
                if libc::setrlimit(libc::RLIMIT_AS, &limit) != 0 {
                    return Err(std::io::Error::last_os_error());
                }
                Ok(())
            });
        }
        let result = child.output().unwrap();
        assert!(
            result.status.success(),
            "{}",
            String::from_utf8_lossy(&result.stderr)
        );
        assert!(String::from_utf8_lossy(&result.stdout).contains("1 passed"));
    }

    #[test]
    fn large_design_uses_host_memory_instead_of_always_serializing() {
        assert_eq!(choose_workers(120_000, 1_000_000, 5, 16, Some(38 * GIB)), 4);
        assert_eq!(choose_workers(120_000, 1_000_000, 5, 16, Some(5 * GIB)), 1);
        assert_eq!(choose_workers(120_000, 1_000_000, 5, 16, Some(10 * GIB)), 2);
        assert_eq!(choose_workers(120_000, 1_000_000, 5, 2, Some(11 * GIB)), 2);
    }

    #[test]
    fn respects_cpu_task_and_straight_line_memory_limits() {
        assert_eq!(choose_workers(1, 1, 5, 1, Some(38 * GIB)), 1);
        assert_eq!(choose_workers(1, 1, 2, 16, Some(38 * GIB)), 2);
        assert_eq!(choose_workers(1, 2_000_000, 5, 16, Some(5 * GIB)), 1);
        assert_eq!(choose_workers(1, 1, 5, 16, Some(0)), 1);
        assert_eq!(
            choose_workers(usize::MAX, usize::MAX, 5, 16, Some(u64::MAX)),
            1
        );
    }

    #[test]
    fn missing_memory_information_preserves_conservative_fallback() {
        assert_eq!(choose_workers(120_000, 1, 5, 16, None), 1);
        assert_eq!(choose_workers(100, 1000, 5, 16, None), 4);
    }

    #[cfg(target_os = "linux")]
    #[test]
    fn accounts_for_parent_usage_below_an_unlimited_child() {
        let temp = tempfile::tempdir().unwrap();
        let parent = temp.path().join("parent");
        let child = parent.join("child");
        std::fs::create_dir_all(&child).unwrap();
        std::fs::write(parent.join("memory.max"), (8 * GIB).to_string()).unwrap();
        std::fs::write(parent.join("memory.current"), (3 * GIB).to_string()).unwrap();
        std::fs::write(child.join("memory.max"), "max\n").unwrap();
        assert_eq!(
            cgroup_available(38 * GIB, temp.path(), &child),
            Some(5 * GIB)
        );
        assert_eq!(cgroup_available(GIB, temp.path(), &child), Some(GIB));
        std::fs::write(parent.join("memory.current"), (9 * GIB).to_string()).unwrap();
        assert_eq!(cgroup_available(38 * GIB, temp.path(), &child), Some(0));
        std::fs::remove_file(parent.join("memory.current")).unwrap();
        assert_eq!(cgroup_available(38 * GIB, temp.path(), &child), None);
        assert_eq!(
            cgroup_available(GIB, temp.path(), &child.join("missing")),
            None
        );
    }
}