[llvm] [AMDGPU] Add support for GFX12 expert scheduling mode 2 (PR #170319)

[Matt Arsenault via llvm-commits]

================
@@ -2741,7 +2979,14 @@ bool SIInsertWaitcnts::run(MachineFunction &MF) {
   AMDGPU::IsaVersion IV = AMDGPU::getIsaVersion(ST->getCPU());
 
   if (ST->hasExtendedWaitCounts()) {
-    MaxCounter = NUM_EXTENDED_INST_CNTS;
+    if (ST->hasSoftwareHazardMode() &&
+        (MF.getFunction()
+             .getFnAttribute("amdgpu-software-hazard-mode")
----------------
arsenm wrote:

I don't know what "expert scheduling mode 2" means. I take it it's another mode bit in the kernel, which every function it calls needs to know about? 

What about supporting multiple entry points in the module? If disjoint call graphs can work, it would have to be an attribute




https://github.com/llvm/llvm-project/pull/170319