Open Quuxplusone opened 2 years ago
https://godbolt.org/z/er1Wsb3da
Efficiently calculating the interleaved load/store costs has been greatly hindered as the llvm-mca region markers aren't great barriers.
https://reviews.llvm.org/D111945
Even if it affects codegen, it'd be very useful if we can get llvm-mca to give us at least an estimate of costs so we can automate this a little.
+1, the UX has been rather pathetic so far.
Adding {rsp} to the asm capture does help a little, but is probably adding more memory traffic to the region block than necessary.
https://godbolt.org/z/4Pf5dncv3
https://godbolt.org/z/er1Wsb3da
Efficiently calculating the interleaved load/store costs has been greatly hindered as the llvm-mca region markers aren't great barriers.
https://reviews.llvm.org/D111945
Even if it affects codegen, it'd be very useful if we can get llvm-mca to give us at least an estimate of costs so we can automate this a little.