================
@@ -3167,3 +3167,30 @@ void tools::handleInterchangeLoopsArgs(const ArgList
&Args,
options::OPT_fno_loop_interchange, EnableInterchange))
CmdArgs.push_back("-floop-interchange");
}
+
+void tools::ParseMPreferVectorWidthOption(clang::DiagnosticsEngine &Diags,
+ const llvm::opt::ArgList &Args,
+ ArgStringList &CmdArgs,
+ bool isCompilerDriver) {
+ // If this was invoked by the Compiler Driver, we pass through the option
+ // as-is. Otherwise, if this is the Frontend Driver, we want just the value.
+ StringRef Out = (isCompilerDriver) ? "-mprefer-vector-width=" : "";
+
+ Arg *A =
Args.getLastArg(clang::driver::options::OPT_mprefer_vector_width_EQ);
+ if (!A)
+ return;
+
+ StringRef Value = A->getValue();
+ unsigned Width;
+
+ // Only "none" and Integer values are accepted by
+ // -mprefer-vector-width=<value>.
+ if (Value != "none" && Value.getAsInteger(10, Width)) {
+ Diags.Report(clang::diag::err_drv_invalid_value)
+ << A->getOption().getName() << Value;
+ return;
+ }
+
+ CmdArgs.push_back(Args.MakeArgString(Out + Value));
----------------
mcinally wrote:
> The function could take an ArgStringList*, only push Out + Value if it is
> provided, and always return Value as a StringRef.
That would work. It would disturb the Compiler Driver a bit to avoid warnings.
We'd need to explicitly discard the results when unwanted. E.g.:
`(void)ParseMPreferVectorWidthOption(...);`
I wonder if we could come up with a Frontend Driver mechanism to discard
everything up to and including the `=`. Then we'd just need to wrap the
parsing/handling call in that mechanism.
> Not sure if it would works with all options, I am not familiar with the
> drivers.
Same. I suggest that we delay this decision until there's a better view of the
entire problem.
https://github.com/llvm/llvm-project/pull/142800
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
