[llvm] [AArch64] Give a higher cost for more expensive SVE FCMP instructions (PR #153816)

[David Green via llvm-commits]

================
@@ -4365,6 +4365,34 @@ AArch64TTIImpl::getAddressComputationCost(Type *PtrTy, ScalarEvolution *SE,
   return 1;
 }
 
+/// Check whether Opcode1 has less throughput according to the scheduling
+/// model than Opcode2.
+bool AArch64TTIImpl::hasLessThroughputFromSchedulingModel(
+    unsigned Opcode1, unsigned Opcode2) const {
+  const MCSchedModel &Sched = ST->getSchedModel();
+  const TargetInstrInfo *TII = ST->getInstrInfo();
+  if (!Sched.hasInstrSchedModel())
+    return false;
+
+  auto ResolveVariant = [&](unsigned Opcode) {
+    unsigned SCIdx = TII->get(Opcode).getSchedClass();
+    while (SCIdx && Sched.getSchedClassDesc(SCIdx)->isVariant())
+      SCIdx = ST->resolveVariantSchedClass(SCIdx, nullptr, TII,
+                                           Sched.getProcessorID());
+    assert(SCIdx);
+    const MCSchedClassDesc &SCDesc = Sched.SchedClassTable[SCIdx];
+    const MCWriteProcResEntry *B = ST->getWriteProcResBegin(&SCDesc);
+    const MCWriteProcResEntry *E = ST->getWriteProcResEnd(&SCDesc);
+    unsigned Min = Sched.IssueWidth;
+    for (const MCWriteProcResEntry *I = B; I != E; I++)
+      Min = std::min(Min, Sched.getProcResource(I->ProcResourceIdx)->NumUnits /
+                              (I->ReleaseAtCycle - I->AcquireAtCycle));
+    return Min;
+  };
+
+  return ResolveVariant(Opcode1) < ResolveVariant(Opcode2);
----------------
davemgreen wrote:

It gives incorrect results for models with invalid sched desc's (for schedules without SVE info), and I don't love that it uses 1/*MinThroughput for the same reasons as above (and I didn't think that ReleaseAtCycle == AcquireAtCycle would happen, maybe I'm wrong about that). I've had a go at making it work. I'm not sure if using the sched info is a really great idea as it locks us in to not using variants, but we can try it and change it if it causes problems.

https://github.com/llvm/llvm-project/pull/153816