4 files changed, 19 insertions, 2 deletions

diff --git a/crates/profile/Cargo.toml b/crates/profile/Cargo.toml
index 9c7ce26be..cc7da27f7 100644
--- a/crates/profile/Cargo.toml
+++ b/crates/profile/Cargo.toml
@@ -14,12 +14,15 @@ once_cell = "1.3.1"
 cfg-if = "1"
 libc = "0.2.73"
 la-arena = { version = "0.2.0", path = "../../lib/arena" }
+countme = { version = "2.0.0-pre.2", features = ["enable"] }
+jemalloc-ctl = { version = "0.3.3", optional = true }
 
 [target.'cfg(target_os = "linux")'.dependencies]
 perf-event = "0.4"
 
 [features]
 cpu_profiler = []
+jemalloc = ["jemalloc-ctl"]
 
 # Uncomment to enable for the whole crate graph
 # default = [ "cpu_profiler" ]
diff --git a/crates/profile/src/hprof.rs b/crates/profile/src/hprof.rs
index 8957ea016..29d2ed518 100644
--- a/crates/profile/src/hprof.rs
+++ b/crates/profile/src/hprof.rs
@@ -3,6 +3,7 @@ use once_cell::sync::Lazy;
 use std::{
     cell::RefCell,
     collections::{BTreeMap, HashSet},
+    env,
     io::{stderr, Write},
     sync::{
         atomic::{AtomicBool, Ordering},
@@ -18,7 +19,8 @@ use crate::tree::{Idx, Tree};
 /// env RA_PROFILE=foo|bar|baz   // enabled only selected entries
 /// env RA_PROFILE=*@3>10        // dump everything, up to depth 3, if it takes more than 10 ms
 pub fn init() {
-    let spec = std::env::var("RA_PROFILE").unwrap_or_default();
+    countme::enable(env::var("RA_COUNT").is_ok());
+    let spec = env::var("RA_PROFILE").unwrap_or_default();
     init_from(&spec);
 }
 
diff --git a/crates/profile/src/lib.rs b/crates/profile/src/lib.rs
index aa6ccc36c..79dba47d5 100644
--- a/crates/profile/src/lib.rs
+++ b/crates/profile/src/lib.rs
@@ -15,6 +15,13 @@ pub use crate::{
     stop_watch::{StopWatch, StopWatchSpan},
 };
 
+pub use countme;
+/// Include `_c: Count<Self>` field in important structs to count them.
+///
+/// To view the counts, run with `RA_COUNT=1`. The overhead of disabled count is
+/// almost zero.
+pub use countme::Count;
+
 thread_local!(static IN_SCOPE: RefCell<bool> = RefCell::new(false));
 
 /// Allows to check if the current code is withing some dynamic scope, can be
diff --git a/crates/profile/src/memory_usage.rs b/crates/profile/src/memory_usage.rs
index 83390212a..cb4e54447 100644
--- a/crates/profile/src/memory_usage.rs
+++ b/crates/profile/src/memory_usage.rs
@@ -24,7 +24,12 @@ impl std::ops::Sub for MemoryUsage {
 impl MemoryUsage {
     pub fn current() -> MemoryUsage {
         cfg_if! {
-            if #[cfg(all(target_os = "linux", target_env = "gnu"))] {
+            if #[cfg(all(feature = "jemalloc", not(target_env = "msvc")))] {
+                jemalloc_ctl::epoch::advance().unwrap();
+                MemoryUsage {
+                    allocated: Bytes(jemalloc_ctl::stats::allocated::read().unwrap() as isize),
+                }
+            } else if #[cfg(all(target_os = "linux", target_env = "gnu"))] {
                 // Note: This is incredibly slow.
                 let alloc = unsafe { libc::mallinfo() }.uordblks as isize;
                 MemoryUsage { allocated: Bytes(alloc) }