================
@@ -4303,21 +4288,15 @@ void CodeGenModule::emitMultiVersionFunctions() {
continue;
llvm::Function *Func = createFunction(CurFD, I);
- StringRef Architecture;
Feats.clear();
- if (getTarget().getTriple().isAArch64())
- TC->getFeatures(Feats, I);
- else if (getTarget().getTriple().isRISCV()) {
- StringRef Version = TC->getFeatureStr(I);
- Feats.push_back(Version);
+ if (getTarget().getTriple().isX86()) {
+ TC->getAddedFeatures(Feats, I);
----------------
labrinea wrote:
No, `I` goes from 0 to 2 in your example. The proplem with these parsing
functions in the .td file is that they are shared among all targets but they
are not virtual. Each target has slightly different semantics though. If you
look at line
https://github.com/llvm/llvm-project/pull/116257/files#diff-e724febedab9c1a2832bf2056d208ff02ddcb2e6f90b5a653afc9b19ac78a5d7L4283
I am basically adding the same functions to the `target_clones` attribute that
are used for the `target` attribute, so that the code is cleaner across
targets, hiding target-specific parsing details from
CodeGenModule::emitMultiVersionFunctions.
https://github.com/llvm/llvm-project/pull/116257
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
