[llvm-branch-commits] [llvm] [AMDGPU] Limit allocation of lo128 registers for occupancy (PR #173100)

[Matt Arsenault via llvm-branch-commits]

================
@@ -570,6 +570,15 @@ unsigned GCNSubtarget::getMaxNumVGPRs(const MachineFunction &MF) const {
   return getMaxNumVGPRs(MF.getFunction());
 }
 
+unsigned GCNSubtarget::getSoftMaxNumVGPRs(const MachineFunction &MF) const {
+  unsigned N = getRegisterInfo()
+                   ->getAllocatableSet(MF, &AMDGPU::VGPR_32RegClass)
+                   .count();
----------------
arsenm wrote:

This needs to be spread everywhere (i.e., we're desperate for https://github.com/llvm/llvm-project/pull/164877 to finally be completed)

https://github.com/llvm/llvm-project/pull/173100