[llvm] [X86][CostModel] Add per-shape gather/scatter cost tables for AMD znver4+ (PR #199488)

[Simon Pilgrim via llvm-commits]

https://github.com/RKSimon commented:

I love seeing realistic numbers but I'd really like to avoid the tuning flag.

We're tried very hard to keep to the worse case costs for the middle-end cost tables - target specific costs have proven nothing but trouble in terms of test coverage - especially in terms of full phase ordering etc. without many notable perf gains. A more manageable analogue has proven to be based off ISA levels (SSE2 vs SSSE3 vs AVX1 vs AVX2 vs AVX512 etc.) which keep in sync with latest similar hardware performance across vendors.

I'd much prefer we just use these new costs as the default for AVX512 targets, AFAICT a lot of Intel numbers aren't that dissimilar for many of these, once all the ucode changes have been applied for speculative exploits etc.


https://github.com/llvm/llvm-project/pull/199488