yaxunl marked 3 inline comments as done.
yaxunl added inline comments.
================
Comment at: clang/lib/Driver/OffloadBundler.cpp:1008
+ auto Output = Worklist.begin();
+ for (auto E = Worklist.end(); Output != E; Output++) {
+ if (isCodeObjectCompatible(
----------------
tra wrote:
> saiislam wrote:
> > tra wrote:
> > > The patch description implies that there are at least two classes of
> > > compatible objects -- the ones that match exactly and the ones that are
> > > not exact match, but are still compatible.
> > >
> > >
> > > Here we're iterating until we find the first compatible object. What if
> > > we also have the object that matches exactly, but it's further down the
> > > list. Is that a problem that we may pick one or the other, depending on
> > > the order they happen to appear in the worklist? It would be good to add
> > > a test case for this scenario.
> > Though it looks plausible, such a case is not possible.
> >
> > From [[
> > https://clang.llvm.org/docs/ClangOffloadBundler.html#bundle-entry-id |
> > Clang Offload Bundler's Documentation]]
> > > If there is an entry with a target feature specified as Any, then all
> > > entries must specify that target feature as Any for the same processor.
> >
> >
> Does it mean that the bundler is supposed to error out if I pass
> `-targets=hip-amdgcn-amd-amdhsa--gfx906,hip-amdgcn-amd-amdhsa--gfx906:xnack-`
> ?
>
> I've just tried it with a bundler built in recent LLVM tree and it accepts
> such a mis of targets without complaining.
>
>
Bundle entries in the same bundle should follow the constraints specified in
the clang-offload-bundler documentation.
Currently, clang enforces the rule when specifying `--offload-arch` options,
which is the most common approach to generate the clang-offload-bundler
bundles. However, if users use clang-offload-bundler directly to generate the
bundle, the rule is not enforced.
I will add the enforcement to clang-offload-bundler too.
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D134546/new/
https://reviews.llvm.org/D134546
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
