[amdgpu] Skip non-instruction values in CF user tracing.

[hliao]

Summary

• CF users won't be non-instruction values. Skip them to save the compilation time. It's especially true when there are multiple functions in that module, where, says, a constant may be used in most functions. The current CF user tracing adds significant overhead.

[hliao]

one compilation is reduced from 460s to 121s after this trivial change.

[arsenm]

Where is this being called from that it's so expensive?

[hliao]

In D72174#1803485, @arsenm wrote:

Where is this being called from that it's so expensive?

the tracking through the D-U chain may step into other functions even though the lowering itself is a function pass. for modules with lots of functions, that implies the almost whole module IR is traversed each time a function is lowered. that's the unnecessary overhead.

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

if (!TRI->isSGPRClass(RC) && !isDivergent)
     return TRI->getEquivalentSGPRClass(RC);
   else if (TRI->isSGPRClass(RC) && isDivergent)
     return TRI->getEquivalentVGPRClass(RC);
 
   return RC;
 }
 
 static bool hasCFUser(const Value *V, SmallPtrSet<const Value *, 16> &Visited) {
+  if (!isa<Instruction>(V))
+    return false;
   if (!Visited.insert(V).second)
     return false;
   bool Result = false;
   for (auto U : V->users()) {
     if (const IntrinsicInst *Intrinsic = dyn_cast<IntrinsicInst>(U)) {
       if (V == U->getOperand(1)) {
         switch (Intrinsic->getIntrinsicID()) {
         default: