[azchohfi]

Summary

Reduce Windows proxy startup time by delay-loading system DLLs that the proxy path doesn't use.

This adds user32, shell32, ole32, oleaut32, crypt32, secur32, and ws2_32 to the existing delay-load list in build.rs. The change is limited to linker configuration; no runtime code changes are needed.

Motivation

Rustup proxies currently load several Windows system DLLs at process startup, even though the proxy path never uses them. Loading these DLLs and their transitive dependencies adds unnecessary startup overhead to every invocation of cargo, rustc, and other proxied tools.

Delay-loading these DLLs defers that cost until a code path actually needs them, such as rustup check or a download.

Performance

Benchmarked on Windows 11 with an i9-14900K, using rustup 1.29.1 built from the release tag with +crt-static (matching CI). Measurements used fresh binary copies, 40 interleaved passes, paired medians, and 95% confidence intervals.

cargo +1.96.0 --version	Time
Before	43.5 ms
After	39.2 ms
Paired median improvement	4.1 ms (95% CI: 3.1–5.1 ms)

The change was faster in 35 of 40 passes.

The number of loaded images dropped from 25 to 14, eliminating startup loads of:

user32, gdi32, gdi32full, win32u, imm32, shell32, ole32, oleaut32, crypt32, iphlpapi, and msvcp_win.

For context, the proxy itself adds approximately 24.0 ms on Windows versus 3.1 ms on Linux. This change removes roughly one-sixth of the Windows proxy overhead.

Safety

• rustup-init.rs calls SetDefaultDllDirectories(LOAD_LIBRARY_SEARCH_SYSTEM32) before the delay-loaded DLLs are needed, preserving the existing System32-only loading behavior.
• delayimp.lib is already linked.
• The build uses /WX, so an unused /DELAYLOAD entry triggers LNK4199 and fails the build rather than silently becoming stale.
• DLLs needed by other rustup commands remain available and are loaded on demand.

Validation

Verified that the following produce the same output as before:

• cargo --version
• rustc -vV
• rustup show
• rustup which cargo

Also verified that rustup check continues to work over TLS.

On current main (5c8ff3f) the change also links with /WX. There the proxy goes from 27 to 14 loaded images, and cargo +1.96.0 --version takes 36.9 ms instead of 40.8 ms (paired median improvement 4.5 ms, 95% CI 3.7–6.0 ms, faster in 36 of 40 passes).

Related work

• improve proxy_mode (e.g. rustc wrapper) startup time? #2626 — Proxy startup performance
• Windows library loading #3493 — Existing Windows DLL-loading safeguards

[ChrisDenton]

Delay loading the DLLs is great but the fix in 958542d does not look right to me. It would be better to do something like gate it on if env::var_os("CARGO_FEATURE_TEST").is_none() so that we only use it in release builds.