[AMDGPU] Remove assert in canBreakPHINode

[Pierre-vh]

Summary

I thought this assert couldn't trip but I failed to
account for cases where some instruction gets
deleted and two PHIs suddenly become related.
It seems like its something that can happen
as this assert has been causing build failures in
some projects.

Fixes SWDEV-416336

[arsenm]

Testcase?

[Pierre-vh]

Actual fix is D157711

llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp

bool AMDGPUCodeGenPrepareImpl::canBreakPHINode(const PHINode &I) {
   // recursively consider all its users and incoming values that are also PHI
   // nodes. We then make a decision about all of those PHIs at once. Either they
   // all get broken up, or none of them do. That way, we avoid cases where a
   // single PHI is/is not broken and we end up reforming/exploding a vector
   // multiple times, or even worse, doing it in a loop.
   SmallPtrSet<const PHINode *, 8> WorkList;
   collectPHINodes(I, WorkList);
 
-#ifndef NDEBUG
-  // Check that none of the PHI nodes in the worklist are in the map. If some of
-  // them are, it means we're not good enough at collecting related PHIs.
-  for (const PHINode *WLP : WorkList) {
-    assert(BreakPhiNodesCache.count(WLP) == 0);
-  }
-#endif
-
   // To consider a PHI profitable to break, we need to see some interesting
   // incoming values. At least 2/3rd (rounded up) of all PHIs in the worklist
   // must have one to consider all PHIs breakable.
   //
   // This threshold has been determined through performance testing.
   //
   // Note that the computation below is equivalent to
   //