fix: switch ld with lld on linux, fixes OOM

We've recently taken to limit parallelism during builds due to Out Of
Memory errors we got while running `cargo test` on 16-core machine,
or 8-core machines running with `-j16`.

We hit OOM due to two reasons:
1. Commit fae5aed introduced `llvm`
   dependencies to the blockifier, which notably increased compilation
   and linking time.
2. Before fae5aed, we already had a quite
memory-heavy linking stage, specifically ~7-8 `ld` instances running in
parallel, each with ~2-3GB of memory, which was just enough not to hit
OOM on most 16core machines.
This was primarily due to a combination of factors:
   2.1 the blockifier being a large lib that needs a 2-3GB `ld` process
   to link in isolation.
   2.2 many crates have the blockifier as a dependency (transitively).
   2.3 `ld` naively parallelizes linking tasks.
The combination of the above three issues caused a heavily parallelized
environment to create many heavy linking jobs that shared no
memory (blockifier linking was duplicated).

This PR switches to the rust linker from `ld` to `lld`, thus solving the
OOM issue without artificially bounding the amount of
parallelization.

`lld` handles parallelizing better and reduces memory stress
from ~10 2-3GB `ld` linker instances post fae5aed,
to ~2 1-2GB `lld` linker instances for the same tasks, which has a
minimal memory impact.

Moreover, `lld` is already the default linker on `rust` nightly, and
will eventually become the default on stable (see code comment for link
to tracking issue).

.cargo/config.toml

@@ -3,10 +3,10 @@ LLVM_SYS_181_PREFIX = "/usr/lib/llvm-18/"
 MLIR_SYS_180_PREFIX = "/usr/lib/llvm-18/"
 TABLEGEN_180_PREFIX = "/usr/lib/llvm-18/"
 
-# Limit concurrency to prevent crash during "cargo build" due to excessive memory usage. Should be removed in the future. Slows down performance.
-[build]
-jobs = 8
-
-# Limit concurrency to prevent crash during "cargo test" due to excessive memory usage. Should be removed in the future. Slows down performance.
-[test]
-jobs = 8
+# Use `lld` for linking instead of `ld`, since we run out of memory while linking with `ld` on
+# 16-cores linux machines, see:
+# https://nnethercote.github.io/perf-book/build-configuration.html#linking.
+# TODO: remove this once `rust` stabilizes `lld` as the default linker, currently only on nightly:
+# https://github.com/rust-lang/rust/issues/39915#issuecomment-618726211
+[target.x86_64-unknown-linux-gnu]
+rustflags = ["-Clink-arg=-fuse-ld=lld"]

.github/workflows/committer_ci.yml

@@ -48,6 +48,7 @@ jobs:
       - uses: actions/checkout@v4
         with:
           ref: ${{ github.base_ref }}
+      - uses: ./.github/actions/bootstrap
 
       # Download the old benchmark inputs.
       - id: auth

scripts/dependencies.sh

@@ -36,7 +36,8 @@ function setup_llvm_deps() {
             libmlir-18-dev \
             libpolly-18-dev \
             libzstd-dev \
-            mlir-18-tools
+            mlir-18-tools \
+            lld
         '
         ;;
     *)