simon_tatham added inline comments.
================ Comment at: clang/include/clang/Basic/arm_mve.td:289 + def vmaxaq: Intrinsic<UVector, (args UVector:$a, Vector:$b), + (IRInt<"vmaxa", [Vector,UVector]> $a, $b)>; } ---------------- I wonder if we should implement at least the simple case (integer and unpredicated) using standard IR nodes instead of an IR intrinsic? We already implement `vmaxq` using an icmp and a select. We haven't implemented `vabsq` yet, but when we do, it will surely be done in a similar way, to take advantage of the existing pattern matching showcased in `llvm/test/CodeGen/Thumb2/mve-abs.ll`. So possibly we should code-generate `vmaxaq(a,b)` as if it was `vmaxq(a, vabsq(b))`, and write a more complicated isel pattern that will match that whole tree? The advantage would be that if a user had //literally// written a combination of `vmaxq` and `vabsq`, codegen would be able to fold them together into a single instruction at compile time. The FP versions //might// make sense to do the same way, using the standard `@llvm.fabs` IR intrinsic for the abs part. Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION https://reviews.llvm.org/D72761/new/ https://reviews.llvm.org/D72761 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
