[llvm] [AMDGPU] Do not propagate vgpr count in dVGPR mode (PR #187078)

[Mirko Brkušanin via llvm-commits]

================
@@ -298,9 +299,37 @@ void MCResourceInfo::gatherResourceInfo(
     }
   };
 
+  auto SetToLocal = [&](int64_t LocalValue, ResourceInfoKind RIK) {
+    MCSymbol *Sym = getSymbol(FnSym->getName(), RIK, OutContext, IsLocal);
+    LLVM_DEBUG(
+        dbgs() << "MCResUse:   " << Sym->getName() << ": Adding " << LocalValue
+               << ", no further propagation as indirect callee found within\n");
+    Sym->setVariableValue(MCConstantExpr::create(LocalValue, OutContext));
+  };
+
   LLVM_DEBUG(dbgs() << "MCResUse: " << FnSym->getName() << '\n');
-  SetMaxReg(MaxVGPRSym, FRI.NumVGPR, RIK_NumVGPR);
-  SetMaxReg(MaxAGPRSym, FRI.NumAGPR, RIK_NumAGPR);
+
+  CallingConv::ID CC = MF.getFunction().getCallingConv();
+
+  // When DynamicVGPR is enabled, chain functions should not propagate VGPR
+  // counts from other chain callees since each chain function can have its own
+  // VGPR allocation, but should still propagate from non-chain callees.
+  if (MF.getInfo<SIMachineFunctionInfo>()->isDynamicVGPREnabled() &&
+      (CC == CallingConv::AMDGPU_CS_Chain || CC == CallingConv::AMDGPU_CS)) {
+    SmallVector<const Function *, 16> NonChainCallees;
+    for (const Function *Callee : FRI.Callees) {
+      if (!AMDGPU::isChainCC(Callee->getCallingConv()) &&
+          !Callee->isDeclaration())
+        NonChainCallees.push_back(Callee);
+    }
----------------
mbrkusanin wrote:

Issue is that the way chain functions are set up they will always have `has_indirect_call`.
Conservatively we should be okay with setting the `num_vgpr` for current function as max of current and of all other non-chain functions, if there is a non-chain indirect call. Otherwise use max of all callees as it currently is.

https://github.com/llvm/llvm-project/pull/187078