[PATCH] D124435: [X86] Always extend the integer parameters in callee
LiuChen via Phabricator via cfe-commits
cfe-commits at lists.llvm.org
Wed Apr 27 00:28:49 PDT 2022
LiuChen3 added inline comments.
Comment at: clang/docs/ClangCommandLineReference.rst:2988-2992
+.. option:: -mconservative-extend
+Always extend the integer parameter both in the callee and caller.
+.. option:: -mno-conservative-extend
+Keep the original integer parameter passing behavior.
> rjmccall wrote:
> > pengfei wrote:
> > > Combine like others?
> > How about:
> > ```
> > In the past, Clang passed small integer arguments on certain targets using a
> > parameter convention in which the caller was assumed to have sign-extended
> > or zero-extended the argument to a certain width. This convention was not
> > conformant with the documented ABI on these platforms, which does not
> > require the caller to perform this extension. Clang no longer assumes that
> > callers perform this extension, but for compatibility with code compiled by
> > previous releases of Clang, Clang defaults to still extending the argument in the
> > caller. `-mno-conservative-extend` disables this, which may improve
> > performance and code size if compatibility with old versions of Clang is not
> > required.
> > This affects most 32-bit and 64-bit x86 targets, except:
> > - Windows, which Clang has never assumed extension on
> > - Apple platforms, which use a non-standard ABI that unconditionally assumes extension
> > ```
> > Note that I need to check that that's what Apple actually wants to do. You should also reach out to the Sony folks to see what they want to do, but I expect that it's to assume extension unconditionally.
> Thanks a lot! It's much clearer.
> Just a small correction for windows: only windows64 is not affected.
Hi, @RKSimon , @probinson. This patch will extend the integer parameters in the caller. Is there any concern about this for Sony?
rG LLVM Github Monorepo
CHANGES SINCE LAST ACTION
More information about the cfe-commits