[PATCH] D64095: SVFS implementation according to RFC: Interface user provided vector functions with the vectorizer.

Sumedh Arani via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Fri Aug 9 12:20:33 PDT 2019


aranisumedh marked 8 inline comments as done.
aranisumedh added inline comments.


================
Comment at: llvm/lib/Analysis/SearchVectorFunctionSystem.cpp:266
+  assert(Records.size() <= 2 && "Invalid Table Entries");
+  if (Records.size() == 2) {
+    StringRef VecName0 = VFABI::getVectorName(Records[0].VectorName);
----------------
sdesmalen wrote:
> I think you should either support 1 mapping, or support multiple mappings with a clear mechanism to prioritise which mapping takes priority. This function seems to prioritise based on whether or not a function-name is mangled which makes little sense to me.
The priority is given on the basis that there exists a redirection to a custom user-defined function. This scenario arises from the different uses of  `declare simd` and `declare variant`. 

And the implementation follows that we save the VectorName as the mangledName if there is no redirection. Hence we prioritize `declare variant` over `declare simd`


Repository:
  rL LLVM

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D64095/new/

https://reviews.llvm.org/D64095





More information about the llvm-commits mailing list