[llvm-branch-commits] [llvm] AMDGPU/GlobalISel: AMDGPURegBankSelect (PR #112863)

[Matt Arsenault via llvm-branch-commits]

================
@@ -66,9 +81,208 @@ FunctionPass *llvm::createAMDGPURegBankSelectPass() {
   return new AMDGPURegBankSelect();
 }
 
+class RegBankSelectHelper {
+  MachineIRBuilder &B;
+  MachineRegisterInfo &MRI;
+  AMDGPU::IntrinsicLaneMaskAnalyzer &ILMA;
+  const MachineUniformityInfo &MUI;
+  const SIRegisterInfo &TRI;
+  const RegisterBank *SgprRB;
+  const RegisterBank *VgprRB;
+  const RegisterBank *VccRB;
+
+public:
+  RegBankSelectHelper(MachineIRBuilder &B,
+                      AMDGPU::IntrinsicLaneMaskAnalyzer &ILMA,
+                      const MachineUniformityInfo &MUI,
+                      const SIRegisterInfo &TRI, const RegisterBankInfo &RBI)
+      : B(B), MRI(*B.getMRI()), ILMA(ILMA), MUI(MUI), TRI(TRI),
+        SgprRB(&RBI.getRegBank(AMDGPU::SGPRRegBankID)),
+        VgprRB(&RBI.getRegBank(AMDGPU::VGPRRegBankID)),
+        VccRB(&RBI.getRegBank(AMDGPU::VCCRegBankID)) {}
+
+  // Temporal divergence copy: COPY to vgpr with implicit use of $exec inside of
+  // the cycle
+  // Note: uniformity analysis does not consider that registers with vgpr def
+  // are divergent (you can have uniform value in vgpr).
+  // - TODO: implicit use of $exec could be implemented as indicator that
+  //   instruction is divergent
+  bool isTemporalDivergenceCopy(Register Reg) {
----------------
arsenm wrote:

Special casing a copy with an implicit exec operand doesn't feel hermetic. We already have ugly hacks to insert these copies at one point, and it's not a wholistic strategy. There are ample opportunities to lose the operand, such as introducing new derived operations which do not have the use 

Where are these exec operands getting inserted? Should we have a different pseudo instead? 

https://github.com/llvm/llvm-project/pull/112863