[AMDGPU] Insert inliner anchor earlier

llvm/include/llvm/CodeGen/TargetPassConfig.h

@@ -364,6 +364,15 @@ class LLVM_ABI TargetPassConfig : public ImmutablePass {
     return true;
   }
 
+  /// Add passes at the start of the function pass manager created after
+  /// enforcing CGSCC ordering. This hook is only called when
+  /// requiresCodeGenSCCOrder() returns true.
+  ///
+  /// Targets can use this to insert passes that need to run at the beginning
+  /// of the codegen function pass manager, after the CGSCC boundary has been
+  /// established.
+  virtual void addInitialCGSCCCodeGenPasses() {}
+
   /// addMachineSSAOptimization - Add standard passes that optimize machine
   /// instructions in SSA form.
   virtual void addMachineSSAOptimization();

llvm/lib/CodeGen/TargetPassConfig.cpp

@@ -966,8 +966,10 @@ void TargetPassConfig::addISelPrepare() {
   addPreISel();
 
   // Force codegen to run according to the callgraph.
-  if (requiresCodeGenSCCOrder())
+  if (requiresCodeGenSCCOrder()) {
     addPass(new DummyCGSCCPass);
+    addInitialCGSCCCodeGenPasses();
+  }
 
   if (getOptLevel() != CodeGenOptLevel::None)
     addPass(createObjCARCContractPass());

llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

@@ -1245,6 +1245,7 @@ class GCNPassConfig final : public AMDGPUPassConfig {
   }
 
   bool addPreISel() override;
+  void addInitialCGSCCCodeGenPasses() override;
   void addMachineSSAOptimization() override;
   bool addILPOpts() override;
   bool addInstSelector() override;
@@ -1504,6 +1505,11 @@ bool GCNPassConfig::addPreISel() {
   return false;
 }
 
+void GCNPassConfig::addInitialCGSCCCodeGenPasses() {
+  if (EnableAMDGPUMachineLevelInliner)
+    addPass(createAMDGPUInliningAnchorPass());
+}
+
 void GCNPassConfig::addMachineSSAOptimization() {
   TargetPassConfig::addMachineSSAOptimization();
 

llvm/test/CodeGen/AMDGPU/llc-pipeline.ll

@@ -1614,7 +1614,8 @@
 ; INLINER-NEXT:        Loop-Closed SSA Form Pass
 ; INLINER-NEXT:      Analysis if a function is memory bound
 ; INLINER-NEXT:      DummyCGSCCPass
-; INLINER-NEXT:      FunctionPass Manager
+; INLINER-NEXT:      AMDGPU Inlining Pass Manager
+; INLINER-NEXT:        AMDGPU Inlining Anchor
 ; INLINER-NEXT:        Dominator Tree Construction
 ; INLINER-NEXT:        Basic Alias Analysis (stateless AA impl)
 ; INLINER-NEXT:        Function Alias Analysis Results
@@ -1745,15 +1746,11 @@
 ; INLINER-NEXT:        Machine Copy Propagation Pass
 ; INLINER-NEXT:        Post-RA pseudo instruction expansion pass
 ; INLINER-NEXT:        SI Shrink Instructions
-; INLINER-NEXT:      AMDGPU Inlining Pass Manager
 ; INLINER-NEXT:        AMDGPU Inlining Anchor
 ; INLINER-NEXT:        AMDGPU Machine Level Inliner
 ; INLINER-NEXT:        SI post-RA bundler
 ; INLINER-NEXT:        MachineDominator Tree Construction
 ; INLINER-NEXT:        Machine Natural Loop Construction
-; INLINER-NEXT:        Dominator Tree Construction
-; INLINER-NEXT:        Basic Alias Analysis (stateless AA impl)
-; INLINER-NEXT:        Function Alias Analysis Results
 ; INLINER-NEXT:        PostRA Machine Instruction Scheduler
 ; INLINER-NEXT:        Machine Block Frequency Analysis
 ; INLINER-NEXT:        MachinePostDominator Tree Construction