So many of the decisions within this PR are really careless compared to the rest of the TCC zeitgeist of carefully thinking about what needs to be done and finding compact approaches. It matters far more that TCC is stable, small, and concise than it having features IMO. Even though I have no real sway TCC (as I'm just a power user), I would strongly prefer if this kind of content was not up-streamed, since it eats into the precious bytes I am trying to leverage when I've chosen TCC over GCC.
If this kind of content were upstreamed, I may have to just start working and iterating from a fork. It's very difficult to sift through and pair things back down from something this voluminous. I maintain an interpreter (cilbox) that I've recently had to do a lot of back-and-forth with someone who uses AI coding tools on. We normally do come to solutions which get to the point of being able to be upstreamed, but it's not without significant effort, and pairing things down to PRs of about 10-150 lines of code at most. I just haven't found a way for maintainers (like me) to thoughtfully provide the needed input to make the code quality high, unless it's on very small sections of code at a time. I feel like a more appropriate use-case for something like this would be to make a TCC-armwin64 port that does what you need. On Sun, Apr 5, 2026 at 7:08 PM Benjamin Oldenburg via Tinycc-devel < [email protected]> wrote: > Hi gr, > > thanks for the feedback! > > I took a closer look at the size question. > > From fada98b1 to win_arm64_reviewable the branch is currently about: > > 73 files changed, 9232 insertions(+), 358 deletions(-) > So net roughly +8870 lines. > > For comparison, your arm64-win32 reference branch is about: > > 48 files changed, 5202 insertions(+), 578 deletions(-) > So net roughly +4620 lines. > > However, I think the important point is that your arm64-win32 branch is > not simply “the same result with less code”, but closer to a compacted > snapshot of an earlier stage of the same Windows ARM64 work. In other > words, a large part of the current size difference is not just history > noise, but functionality and validation that was added later. > > The reason is that a large part of the extra size is not just noise, but > actual functionality and coverage that is beyond the smaller subset. > > The “more code” currently covers in particular: > > - fuller Windows ARM64 hosted `-run` support > - `-run` argv/wargv/env/wenv handling > - stdin and exit-path coverage for `-run` > - nested `tcc -run tcc.c ... -run ...` coverage, including > `tests/test2` and `tests/test3` > - Windows ARM64 backtrace coverage > - ARM64 libtcc context save/restore coverage > - PE/alignment and Win32 runtime smoke tests > - broader GNU-style ARM64 inline asm support, including more > operand/constraint/clobber handling and additional inline asm validation > tests > - the Windows-specific reference outputs used by the test suite > > > One especially large part is the two reference files: > > - tests/test.ref.win32-arm64 > - tests/test.ref.win32-x86_64-ucrt > > > Those two are mostly the same. > > My estimate is with the current behavior preserved, I can probably reduce > the branch somewhat further, but the realistic floor is still much closer > to ~+7000 than to ~+4000 to get down near your target, I would likely have > to cut scope, not just rewrite smaller > > So I think the choice is roughly this: > > 1. Keep the current behavior and test coverage. Then I can still try > to trim the branch further, but not to half its current net size. > 2. Aim for the smaller upstreamable subset closer to your arm64-win32 > branch. Then I would deliberately drop or postpone part of the current > runtime/test/validation surface and keep only the minimal core needed for > the port. > > > I am fine with either direction, but I think it is better to say this > explicitly rather than pretend that “same features, same tests, half the > code” is easy here. > > If your preference is the smaller subset, I can work toward that and split > the omitted parts into later follow-up work. > > Cheers, > > Ben > > On 6 Apr 2026, at 01:45, grischka <[email protected]> wrote: > > On April 5, 2026 1:07:33 PM UTC, Benjamin Oldenburg < > [email protected]> wrote: > > Hi gr, > > I rewrote the history on a new review branch: > > https://github.com/bold84/tinycc/tree/win_arm64_reviewable > [...] > How to proceed? I seem to not be able to push to the mob branch… 🥺 > > > Well, in this specific case I think this piece of work could still benefit > from r9eview before being pushed on mob anyway. > > For example: > git diff --shortstat mob win_arm64_reviewable > currently gives: > 73 files changed, 9182 insertions(+), 499 deletions(-) > > Which is about +8700 lines net. > > Now I believe this could be safely reduced to about the half of it, say > +4000 lines. > > (After all, tinycc is all about supporting features with the least > possible amount of code) > > I will however not have time to do this within the next 2 months. If in > between you feel to add or change something on your branch, no problem at > all. > > Maybe AI could do that too. Like: "Now give me same results with only half > of the code plus still readable by humans, plus running the test suite on > GHActions, green, please..." ;) > > Thanks, > > -- grischka > > > > Some personal context on why I ended up doing the work here: > > My original goal was to get OpenCode running on Windows ARM64, which first > led me to work on the Bun side. I submitted that work there. Instead of > discussing it normally, the maintainer took that work, opened a separate > PR, and merged it under his own name. Several of those commits were direct > one-to-one copies of patches I had already authored and submitted. If > useful, I can point to the corresponding commit pairs. > > Once that happened, I had no interest in continuing to contribute the TCC > side through Bun’s fork, because I did not want to spend more time > preparing another PR there only to watch it get consumed in the same way > again. By then I had already realized that Bun was only half of the problem > anyway, because TCC also needed proper Windows ARM64 support. > > So I made a deliberate decision not to continue downstream at Bun, but to > contribute the compiler work at the actual upstream source instead. That is > why I came to TinyCC directly and why I cleaned up the history here for > review. > > I enjoy saying openly that I used OpenAI Codex, specifically the GPT-5.4 > model, heavily while developing and validating this contribution. Claude > Opus just couldn’t do it. So there is a certain irony in watching code > developed that way later getting merged downstream by Bun, which is now > part of Anthropic. 🤓 > > In any case, the important part is that the work exists upstream in the > right place. > > Greetings, > > Ben > > On 5 Apr 2026, at 01:07, Benjamin Oldenburg < > [email protected]> wrote: > > Ah.. “success” means built without warnings and all tests pass. > > On 5 Apr 2026, at 01:07, Benjamin Oldenburg < > [email protected]> wrote: > > Hi gr, > > alright. I will rewrite the history accordingly. > > The current state for > https://github.com/bold84/tinycc/tree/win_arm64_clean : > > - arm64-win32 > - Clang: success > - MinGW GCC: not run > - MSVC cl: success > - x86_64-win32 > - Clang: not run > - MinGW GCC: success > - MSVC cl: success > Linux x86_64 > - GCC: success > - Clang: not installed > Linux aarch64 > - GCC: success > - Clang: not installed > macOS ARM64 (Darwin) > - Clang: success > > Any other targets that need to be green? > > Kind regards, > > Ben > > > On 5 Apr 2026, at 00:52, grischka <[email protected]> wrote: > > On April 4, 2026 11:10:11 AM UTC, Benjamin Oldenburg < > [email protected]> wrote: > > I can complete the implementation in > https://github.com/bold84/tinycc/tree/win_arm64_clean if there’s a chance > it will actually be merged. > > > Hi Ben, > sure it will, great work. > > It would make sense though if we could somewhat compress the history to > say less than 10 commits each addressing a theme such as for example > generator, asm, runtime, symbolic replacement, etc. (while omitting > unrelated whitespace changes and some stuff that we maybe dont really need > also). Just a suggestion though ;) > > Some first approach of mine here: > https://github.com/gr8001/tinycc/tree/arm64-win32 > > Greetings, > -- gr > > > Kind regards, > > Ben > > > > > > > _______________________________________________ > Tinycc-devel mailing list > [email protected] > https://lists.nongnu.org/mailman/listinfo/tinycc-devel >
_______________________________________________ Tinycc-devel mailing list [email protected] https://lists.nongnu.org/mailman/listinfo/tinycc-devel
