diff --git a/lld/utils/run_benchmark.py b/lld/utils/run_benchmark.py
new file mode 100755
index 000000000000..93702ba6e700
--- /dev/null
+++ b/lld/utils/run_benchmark.py
@@ -0,0 +1,151 @@
+#!/usr/bin/env python3
+#
+# Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+# See https://llvm.org/LICENSE.txt for license information.
+# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+#
+# ==------------------------------------------------------------------------==#
+
+import argparse
+import os
+import shutil
+import subprocess
+import tempfile
+
+# The purpose of this script is to measure the performance effect
+# of an lld change in a statistically sound way, automating all the
+# tedious parts of doing so. It copies the test case into /tmp as well as
+# running the test binaries from /tmp to reduce the influence on the test
+# machine's storage medium on the results. It accounts for measurement
+# bias caused by binary layout (using the --randomize-section-padding
+# flag to link the test binaries) and by environment variable size
+# (implemented by hyperfine [1]). Runs of the base and test case are
+# interleaved to account for environmental factors which may influence
+# the result due to the passage of time. The results of running hyperfine
+# are collected into a results.csv file in the output directory and may
+# be analyzed by the user with a tool such as ministat.
+#
+# Requirements: Linux host, hyperfine [2] in $PATH, run from a build directory
+# configured to use ninja and a recent version of lld that supports
+# --randomize-section-padding, /tmp is tmpfs.
+#
+# [1] https://github.com/sharkdp/hyperfine/blob/3cedcc38d0c430cbf38b4364b441c43a938d2bf3/src/util/randomized_environment_offset.rs#L1
+# [2] https://github.com/sharkdp/hyperfine
+#
+# Example invocation for comparing the performance of the current commit
+# against the previous commit which is treated as the baseline, without
+# linking debug info:
+#
+# lld/utils/run_benchmark.py \
+#   --base-commit HEAD^ \
+#   --test-commit HEAD \
+#   --test-case lld/utils/speed-test-reproducers/result/firefox-x64/response.txt \
+#   --num-iterations 512 \
+#   --num-binary-variants 16 \
+#   --output-dir outdir \
+#   --ldflags=-S
+#
+# Then this bash command will compare the real time of the base and test cases.
+#
+# ministat -A \
+#   <(grep lld-base outdir/results.csv | cut -d, -f2) \
+#   <(grep lld-test outdir/results.csv | cut -d, -f2)
+
+# We don't want to copy stat() information when we copy the reproducer
+# to the temporary directory. Files in the Nix store are read-only so this will
+# cause trouble when the linker writes the output file and when we want to clean
+# up the temporary directory. Python doesn't provide a way to disable copying
+# stat() information in shutil.copytree so we just monkeypatch shutil.copystat
+# to do nothing.
+shutil.copystat = lambda *args, **kwargs: 0
+
+parser = argparse.ArgumentParser(prog="benchmark_change.py")
+parser.add_argument("--base-commit", required=True)
+parser.add_argument("--test-commit", required=True)
+parser.add_argument("--test-case", required=True)
+parser.add_argument("--num-iterations", type=int, required=True)
+parser.add_argument("--num-binary-variants", type=int, required=True)
+parser.add_argument("--output-dir", required=True)
+parser.add_argument("--ldflags", required=False)
+args = parser.parse_args()
+
+test_dir = tempfile.mkdtemp()
+print(f"Using {test_dir} as temporary directory")
+
+os.makedirs(args.output_dir)
+print(f"Using {args.output_dir} as output directory")
+
+
+def extract_link_command(target):
+    # We assume that the last command printed by "ninja -t commands" containing a
+    # "-o" flag is the link command (we need to check for -o because subsequent
+    # commands create symlinks for ld.lld and so on). This is true for CMake and
+    # gn.
+    link_command = None
+    for line in subprocess.Popen(
+        ["ninja", "-t", "commands", target], stdout=subprocess.PIPE
+    ).stdout.readlines():
+        commands = line.decode("utf-8").split("&&")
+        for command in commands:
+            if " -o " in command:
+                link_command = command.strip()
+    return link_command
+
+
+def generate_binary_variants(case_name):
+    subprocess.run(["ninja", "lld"])
+    link_command = extract_link_command("lld")
+
+    for i in range(0, args.num_binary_variants):
+        print(f"Generating binary variant {i} for {case_name} case")
+        command = f"{link_command} -o {test_dir}/lld-{case_name}{i} -Wl,--randomize-section-padding={i}"
+        subprocess.run(command, check=True, shell=True)
+
+
+# Make sure that there are no local changes.
+subprocess.run(["git", "diff", "--exit-code", "HEAD"], check=True)
+
+# Resolve the base and test commit, since if they are relative to HEAD we will
+# check out the wrong commit below.
+resolved_base_commit = subprocess.check_output(
+    ["git", "rev-parse", args.base_commit]
+).strip()
+resolved_test_commit = subprocess.check_output(
+    ["git", "rev-parse", args.test_commit]
+).strip()
+
+test_case_dir = os.path.dirname(args.test_case)
+test_case_respfile = os.path.basename(args.test_case)
+
+test_dir_test_case_dir = f"{test_dir}/testcase"
+shutil.copytree(test_case_dir, test_dir_test_case_dir)
+
+subprocess.run(["git", "checkout", resolved_base_commit], check=True)
+generate_binary_variants("base")
+
+subprocess.run(["git", "checkout", resolved_test_commit], check=True)
+generate_binary_variants("test")
+
+
+def hyperfine_link_command(case_name):
+    return f'../lld-{case_name}$(({{iter}}%{args.num_binary_variants})) -flavor ld.lld @{test_case_respfile} {args.ldflags or ""}'
+
+
+results_csv = f"{args.output_dir}/results.csv"
+subprocess.run(
+    [
+        "hyperfine",
+        "--export-csv",
+        os.path.abspath(results_csv),
+        "-P",
+        "iter",
+        "0",
+        str(args.num_iterations - 1),
+        hyperfine_link_command("base"),
+        hyperfine_link_command("test"),
+    ],
+    check=True,
+    cwd=test_dir_test_case_dir,
+)
+
+shutil.rmtree(test_dir)