[llvm] [VPlan] Speed up VPSlotTracker by using ModuleSlotTracker (PR #139881)

[Igor Kirillov via llvm-commits]

================
@@ -1441,7 +1441,23 @@ void VPSlotTracker::assignName(const VPValue *V) {
   std::string Name;
   if (UV) {
     raw_string_ostream S(Name);
-    UV->printAsOperand(S, false);
+    if (MST) {
+      UV->printAsOperand(S, false, *MST);
+    } else if (isa<Instruction>(UV) && !UV->hasName()) {
+      // Lazily create the ModuleSlotTracker when we first hit an unnamed
+      // instruction
+      auto *IUV = cast<Instruction>(UV);
+      // This check is required to support unit tests with incomplete IR.
+      if (IUV->getParent()) {
+        MST = std::make_unique<ModuleSlotTracker>(IUV->getModule());
----------------
igogo-x86 wrote:

The main slowdown comes from `VPRecipeBase::cost`, which prints debug output like:

```
dbgs() << "Cost of " << RecipeCost << " for VF " << VF << ": ";
dump();
```

or:

```
Cost of 1 for VF 2: WIDEN ir<%1547> = add nsw ir<%1546>, ir<%10>
```

This is where we create and destroy a new `VPSlotTracker` for each line of output. `VPSlotTracker` assigns names to all VPValues in the VPlan every time it’s instantiated. That does not look okay, but I guess we can live with that.

The real issue is that for each underlying Value, `VPSlotTracker `internally creates a `SlotTracker`, which in turn traverses all instructions in the function. That makes dumping a value quadratic in complexity. The time spent initializing `SlotTracker` itself is minor compared to the much larger cost of running `VPSlotTracker::assignNames`.

https://github.com/llvm/llvm-project/pull/139881