[MarcroSoft]

GCC-built NSIS stubs are ~116 KB versus ~38 KB for the official MSVC builds, and every generated installer carries one. Most of the gap is mingw-w64's monolithic libuuid.a; -ffunction-sections + -Wl,--gc-sections plus -Oz, -fno-ident and -fno-asynchronous-unwind-tables bring stubs to ~50 KB on x86. Plugins get the same flags except the unwind one (NSISdl uses C++ exceptions). makensis switches from -O2 to -Os. -fdata-sections is deliberately left out; see the patch header.

Verified on x86: zlib/lzma/bzip2 installers built with the patched stubs install correctly.

[MarcroSoft]

Since this PR has no checks here: the same branch (d9f00c1, upstream master plus only this patch) built successfully for both x64 and x86 on my fork's CI: https://github.com/MarcroSoft/w64devkit/actions/runs/39411671387

[Peter0x44]

I approved your workflow so it should now run, and also any future PRs you make

[skeeto]

Thanks, @MarcroSoft! I pushed a fixup that backs off some of your changes, being a little more conservative. -fno-asynchronous-unwind-tables is definitely unsafe for programs linking the CRT, including this one, and it's caused problems with busybox-w32. -Oz makes the stubs 1-3% slower on my systems, and makesis 6%-38% slower on my systems, for very small size improvements. -Wl,--gc-sections is the real heavy hitter, and I kept -ffunction-sections for free, modest gains on x86.

More important for makensis is that it wasn't stripped! So I just pushed that: 9ad0368. While investigating Claude also found an NSIS+Mingw-w64 bug (f915d8e).

I'm ready to merge this if you're satisfied. Any thoughts, @Peter0x44, since you're here?

[MarcroSoft]

Thanks, that all makes sense — happy with the more conservative version, and good catch on stripping makensis. Fine to merge from my side.