[llvm] [AMDGPU][Scheduler] Support for rematerializing SGPRs and AGPRs (PR #140036)

[Jeffrey Byrnes via llvm-commits]

================
@@ -1707,26 +1709,34 @@ struct ExcessRP {
   bool UnifiedRF;
 
   /// Constructs the excess RP model; determines the excess pressure w.r.t. a
-  /// maximum number of allowed VGPRs.
-  ExcessRP(const GCNSubtarget &ST, const GCNRegPressure &RP, unsigned MaxVGPRs);
+  /// maximum number of allowed SGPRs/VGPRs.
+  ExcessRP(const GCNSubtarget &ST, const GCNRegPressure &RP, unsigned MaxSGPRs,
+           unsigned MaxVGPRs);
+
+  /// Accounts for \p NumRegs saved SGPRs in the model. Returns whether saving
+  /// these SGPRs helped reduce excess pressure.
+  bool saveSGPRs(unsigned NumRegs) { return saveRegs(SGPRs, NumRegs); }
 
   /// Accounts for \p NumRegs saved ArchVGPRs in the model. If \p
   /// UseArchVGPRForAGPRSpill is true, saved ArchVGPRs are used to save excess
   /// AGPRs once excess ArchVGPR pressure has been eliminated. Returns whether
   /// saving these ArchVGPRs helped reduce excess pressure.
   bool saveArchVGPRs(unsigned NumRegs, bool UseArchVGPRForAGPRSpill);
 
-  /// Accounts for \p NumRegs saved AGPRS in the model. Returns whether saving
-  /// these ArchVGPRs helped reduce excess pressure.
-  bool saveAGPRs(unsigned NumRegs);
+  /// Accounts for \p NumRegs saved AGPRs in the model. Returns whether saving
+  /// these AGPRs helped reduce excess pressure.
+  bool saveAGPRs(unsigned NumRegs) {
+    return saveRegs(AGPRs, NumRegs) || saveRegs(VGPRs, NumRegs);
----------------
jrbyrnes wrote:

Should UnifiedRF be a precondition to saving VGPRs?

https://github.com/llvm/llvm-project/pull/140036