build: use mimalloc

mimalloc is a compact general purpose allocator from Microsoft. It
consistently outperforms glibc's `malloc()` in allocation-heavy
workloads, such as Lix's evaluator

It's currently only linked in the main `nix` executable, as Boehm's GC
uses its own allocator. Other allocations that *do* go through glibc's
`malloc()` are still much faster, though

Benchmarked on x86_64-linux against Nixpkgs `bd07873`:

| attribute         | thunks  | lix@`1396012` | mimalloc | uplift |
|-------------------|---------|---------------|----------|--------|
| hello             | 206444  | 0.726s        | 0.620s   | 1.17x  |
| chromium          | 1177382 | 2.273s        | 2.136s   | 1.06x  |
| firefox-unwrapped | 1394797 | 2.521s        | 2.378s   | 1.06x  |
| texliveFull       | 3186440 | 4.992s        | 4.672s   | 1.07x  |
| nixosTests.gnome  | 7905808 | 6.115s        | 5.723s   | 1.07x  |

Based-on: https://github.com/NixOS/nix/pull/15596
Co-authored-by: Bernardo Meurer Costa <beme@anthropic.com>
Change-Id: I3ad92eacc075efeaf3d9f7730ada7b29835536a4

doc/manual/rl-next/mimalloc.md

@@ -0,0 +1,16 @@
+---
+synopsis: "Use mimalloc for faster evaluation"
+cls: [5645]
+category: Features
+credits: [getchoo, lovesegfault]
+---
+
+Lix now links with [mimalloc](https://github.com/microsoft/mimalloc),
+replacing the system's default `malloc()` for all non-GC allocations.
+
+This yields a **5–12% wall-clock improvement** on evaluation workloads,
+ranging from `nix-instantiate hello` to `nix-env -qa` and full NixOS
+configurations.
+
+The allocator can be disabled at build time with `-Dmimalloc=disabled`,
+or by passing the `useMimalloc = false` override to the `lix` package.

lix/nix/meson.build

@@ -172,6 +172,7 @@ nix = executable(
   dependencies : [
     liblix,
     boehm,
+    mimalloc,
     capnp_rpc,
     nlohmann_json,
     kj,

meson.build

@@ -352,6 +352,8 @@ gc_opt = get_option('gc').disable_if(
 boehm = dependency('bdw-gc', required : gc_opt, version : '>=8.2.6', include_type : 'system')
 configdata.set('HAVE_BOEHMGC', boehm.found().to_int())
 
+mimalloc = dependency('mimalloc', required: get_option('mimalloc'))
+
 boost = dependency('boost', required : true, include_type : 'system')
 kj = dependency('kj-async', required : true, include_type : 'system')
 capnp = dependency('capnp', required : true, include_type : 'system')

meson.options

@@ -8,6 +8,10 @@ option('gc', type : 'feature',
   description : 'enable garbage collection in the Nix expression evaluator (requires Boehm GC)',
 )
 
+option('mimalloc', type: 'feature', value: 'auto',
+  description: 'link against mimalloc to override the default memory allocator',
+)
+
 option('enable-embedded-sandbox-shell', type : 'boolean', value : false,
   description : 'include the sandbox shell in the Nix binary',
 )

package.nix

@@ -42,6 +42,7 @@
   cacert,
   mercurial,
   meson,
+  mimalloc,
   ninja,
   openssl,
   passt,
@@ -75,6 +76,9 @@
 
   # Support garbage collection in the evaluator.
   enableGC ? sanitize == null || !builtins.elem "address" sanitize,
+  # Whether to use mimalloc over the default `malloc`
+  # Significantly improves evaluation performance on allocation-heavy workloads
+  useMimalloc ? true,
   # List of Meson sanitize options. Accepts values of b_sanitize, e.g.
   # "address", "undefined", "thread".
   # Enabling the "address" sanitizer will disable garbage collection in the evaluator.
@@ -467,6 +471,7 @@ stdenv.mkDerivation (finalAttrs: {
       # so we must explicitly enable or disable features that we are not passing
       # dependencies for.
       (lib.mesonEnable "gc" enableGC)
+      (lib.mesonEnable "mimalloc" useMimalloc)
       (lib.mesonEnable "internal-api-docs" internalApiDocs)
       (lib.mesonEnable "dtrace-probes" withDtrace)
       (lib.mesonBool "enable-tests" (finalAttrs.finalPackage.doCheck || lintInsteadOfBuild))
@@ -574,7 +579,8 @@ stdenv.mkDerivation (finalAttrs: {
   ++ lib.optionals (finalAttrs.dontBuild) maybePropagatedInputs
   # I am so sorry. This is because checkInputs are required to pass
   # configure, but we don't actually want to *run* the checks here.
-  ++ lib.optionals lintInsteadOfBuild finalAttrs.checkInputs;
+  ++ lib.optionals lintInsteadOfBuild finalAttrs.checkInputs
+  ++ lib.optional useMimalloc mimalloc;
 
   checkInputs = [
     gtest