[llvm] Split vgpr regalloc pipeline (PR #93526)

[Christudasan Devadasan via llvm-commits]

================
@@ -32,12 +33,18 @@ using MBBVector = SmallVector<MachineBasicBlock *, 4>;
 
 namespace {
 
+static cl::opt<unsigned> MaxNumVGPRsForWwmAllocation(
+    "amdgpu-num-vgprs-for-wwm-alloc",
+    cl::desc("Max num VGPRs for whole-wave register allocation."),
+    cl::ReallyHidden, cl::init(5));
----------------
cdevadas wrote:

I see that. I initially wanted this value between 5 - 7 based on the experience with SGPR spilling on heavy workloads earlier. In the internal testing, I used 5 and 10 for the upper cap and they both worked without any issues. I should have kept the higher value here. Will change it to 10.

https://github.com/llvm/llvm-project/pull/93526