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

[Jay Foad 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")
----------------
jayfoad wrote:

At the hardware level it's a bit in the wave state register SCHED_MODE. In theory shader code could flip this bit whenever it likes, but the compiler only supports setting it once at the start of an entry function and leaving it set for the whole call graph. So yes disjoint call graphs with different settings could work.

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