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

[via llvm-commits]

================
@@ -721,6 +721,17 @@ def TuningFastGather
     : SubtargetFeature<"fast-gather", "HasFastGather", "true",
                        "Indicates if gather is reasonably fast (this is true for Skylake client and all AVX-512 CPUs)">;
 
+// Use AMD Zen-tuned cost tables for masked gather/scatter intrinsics in the
+// X86 TargetTransformInfo cost model. Refines the flat overhead used by other
+// AVX-512 targets with per-element-type/per-VL costs measured on znver4 and
+// znver5. Inherited automatically by every znver4+ CPU via ZN4Tuning; not
+// applied to pre-AVX-512 Zen parts (znver1..3), which take the scalarise
+// path for masked gather anyway.
+def TuningPreferAMDZenGSCost
----------------
ganeshgit wrote:

This is a function of Vector Factor, Data Type, and the instruction. So, deducing a number from the latency of scheduler is not applicable directly. It seems more of a Feature Tuning. What I thought was something of below kind which can grow as and when required. (Can be even made as a table lookup later).

static const UnifiedVendorCostTable[] = {
  // Subtarget | ISD Opcode | Base Type | VF | Cost
  // --- Generation 1 Costs ---
  { ST_Gen1, <values>}, 
  { ST_Gen1, <values> },

  // --- Generation 2 Costs ---
  { ST_Gen2,    <values> },
  { ST_Gen2,    <values> },
  
  // --- Global Overrides (Optional Example) ---
  { ST_All,     ISD::SDIV,   MVT::i32,    1,   20 } 
}; 

Given that, I think the code changes looks okay now. If we don't want the Unified Vendor Cost Table, we can go back to Sumukh's first patch having TuningPreferAMDGSCostTable!

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