weiweichen wrote: > The only thing `AArch64TargetInfo::initFeatureMap` adds is features from the > selected CPU. IMHO this was not an appropriate place to be doing that. Since > #94279 the CPU features are added either by > `AArch64TargetInfo::parseTargetAttr` when dealing with > `__attribute(target(...))`, and by `getAArch64ArchFeaturesFromMcpu` if you > are in the clang driver dealing with `-mcpu`. Basically anywhere you see > `AArch64::ExtensionSet::addCPUDefaults` used. This approach properly handles > feature dependencies, which the old `AArch64TargetInfo::initFeatureMap` > approach did not. > > What inputs are you giving when you create the `TargetInfo`? What features do > you expect to see that are missing? A concrete example would help to > understand.
We are using `llvm::sys::getDefaultTargetTriple()` and `llvm::sys::getHostCPUName()` to get TargetInfo, and was getting 👇 features before the aarch64 PR landed. (We still get features correctly for other backends). ``` +aes +bf16 +complxnum +crc +dotprod +fp-armv8 +fp16fml +fullfp16 +i8mm +jsconv +lse +neon +pauth +rand +ras +rcpc +rdm +sha2 +sha3 +sm4 +spe +ssbs +sve ``` https://github.com/llvm/llvm-project/pull/96832 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
