[CodeGen][NPM] Support CodeGenSCCOrder in pipeline

[optimisan]

Wrap passes into Post order CGSCC pass manager in codegen pass builder.

I am adding the pipeline test in this but it is not yet complete.

[optimisan]

• [CodeGen][NPM] Clear MachineFunctions without using PA #139517
• [AMDGPU][NPM] Fill in addPreSched2 passes #139516
• [CodeGen][NPM] Account inserted passes for -start/stop options #138830
• [CodeGen][NPM] Port ProcessImplicitDefs to NPM #138829
• [CodeGen][NPM] Register Function Passes #138828
• [CodeGen][NPM] Read TargetMachine's EnableIPRA option #138670
• [CodeGen][NPM] VirtRegRewriter: Set VirtReg flag #138660
• [CodeGen][NPM] Port PostRAMachineSinking to NPM #138497
• [AMDGPU][NPM] Register AMDGPUWaitSGPRHazards pass #138496
• [CodeGen][NPM] Port InitUndef to NPM #138495
• [AMDGPU][NPM] Complete optimized regalloc pipeline #138491
• [CodeGen][NPM] Support CodeGenSCCOrder in pipeline #136818 👈 (View in Graphite)
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[llvmbot]

@llvm/pr-subscribers-backend-amdgpu

Author: Akshat Oke (optimisan)

Changes

Wrap passes into Post order CGSCC pass manager in codegen pass builder.

---

Full diff: https://github.com/llvm/llvm-project/pull/136818.diff

3 Files Affected:

• (modified) llvm/include/llvm/Passes/CodeGenPassBuilder.h (+73-16)
• (modified) llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp (+2)
• (added) llvm/test/CodeGen/AMDGPU/llc-pipeline-npm.ll (+144)

diff --git a/llvm/include/llvm/Passes/CodeGenPassBuilder.h b/llvm/include/llvm/Passes/CodeGenPassBuilder.h
index 61e377de0c424..ce67cdb6eb8ff 100644
--- a/llvm/include/llvm/Passes/CodeGenPassBuilder.h
+++ b/llvm/include/llvm/Passes/CodeGenPassBuilder.h
@@ -18,6 +18,7 @@
 #include "llvm/ADT/StringRef.h"
 #include "llvm/Analysis/AliasAnalysis.h"
 #include "llvm/Analysis/BasicAliasAnalysis.h"
+#include "llvm/Analysis/CGSCCPassManager.h"
 #include "llvm/Analysis/ProfileSummaryInfo.h"
 #include "llvm/Analysis/ScopedNoAliasAA.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
@@ -207,10 +208,7 @@ template <typename DerivedT, typename TargetMachineT> class CodeGenPassBuilder {
   class AddIRPass {
   public:
     AddIRPass(ModulePassManager &MPM, const DerivedT &PB) : MPM(MPM), PB(PB) {}
-    ~AddIRPass() {
-      if (!FPM.isEmpty())
-        MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM)));
-    }
+    ~AddIRPass() { flushFPMToMPM(); }
 
     template <typename PassT>
     void operator()(PassT &&Pass, StringRef Name = PassT::name()) {
@@ -228,16 +226,40 @@ template <typename DerivedT, typename TargetMachineT> class CodeGenPassBuilder {
         FPM.addPass(std::forward<PassT>(Pass));
       } else {
         // Add Module Pass
-        if (!FPM.isEmpty()) {
-          MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM)));
-          FPM = FunctionPassManager();
-        }
-
+        flushFPMToMPM();
         MPM.addPass(std::forward<PassT>(Pass));
       }
     }
 
+    /// Setting this will add passes to the CGSCC pass manager.
+    void requireCGSCCOrder() {
+      if (PB.AddInCGSCCOrder)
+        return;
+      flushFPMToMPM();
+      PB.AddInCGSCCOrder = true;
+    }
+
+    /// Stop adding passes to the CGSCC pass manager.
+    /// Existing passes won't be removed.
+    void stopAddingInCGSCCOrder() {
+      if (!PB.AddInCGSCCOrder)
+        return;
+      flushFPMToMPM();
+      PB.AddInCGSCCOrder = false;
+    }
+
   private:
+    void flushFPMToMPM() {
+      if (!FPM.isEmpty()) {
+        if (PB.AddInCGSCCOrder) {
+          MPM.addPass(createModuleToPostOrderCGSCCPassAdaptor(
+              createCGSCCToFunctionPassAdaptor(std::move(FPM))));
+        } else {
+          MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM)));
+        }
+        FPM = FunctionPassManager();
+      }
+    }
     ModulePassManager &MPM;
     FunctionPassManager FPM;
     const DerivedT &PB;
@@ -254,7 +276,11 @@ template <typename DerivedT, typename TargetMachineT> class CodeGenPassBuilder {
         FPM.addPass(
             createFunctionToMachineFunctionPassAdaptor(std::move(MFPM)));
         FPM.addPass(InvalidateAnalysisPass<MachineFunctionAnalysis>());
-        MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM)));
+        if (this->PB.AddInCGSCCOrder) {
+          MPM.addPass(createModuleToPostOrderCGSCCPassAdaptor(
+              createCGSCCToFunctionPassAdaptor(std::move(FPM))));
+        } else
+          MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM)));
       }
     }
 
@@ -273,12 +299,7 @@ template <typename DerivedT, typename TargetMachineT> class CodeGenPassBuilder {
         MFPM.addPass(std::forward<PassT>(Pass));
       } else {
         // Add Module Pass
-        if (!MFPM.isEmpty()) {
-          MPM.addPass(createModuleToFunctionPassAdaptor(
-              createFunctionToMachineFunctionPassAdaptor(std::move(MFPM))));
-          MFPM = MachineFunctionPassManager();
-        }
-
+        flushMFPMToMPM();
         MPM.addPass(std::forward<PassT>(Pass));
       }
 
@@ -286,7 +307,39 @@ template <typename DerivedT, typename TargetMachineT> class CodeGenPassBuilder {
         C(Name, MFPM);
     }
 
+    /// Setting this will add passes to the CGSCC pass manager.
+    void requireCGSCCOrder() {
+      if (PB.AddInCGSCCOrder)
+        return;
+      flushMFPMToMPM();
+      PB.AddInCGSCCOrder = true;
+    }
+
+    /// Stop adding passes to the CGSCC pass manager.
+    /// Existing passes won't be removed.
+    void stopAddingInCGSCCOrder() {
+      if (!PB.AddInCGSCCOrder)
+        return;
+      flushMFPMToMPM();
+      PB.AddInCGSCCOrder = false;
+    }
+
   private:
+    void flushMFPMToMPM() {
+      if (!MFPM.isEmpty()) {
+        if (PB.AddInCGSCCOrder) {
+          MPM.addPass(createModuleToPostOrderCGSCCPassAdaptor(
+              createCGSCCToFunctionPassAdaptor(
+                  createFunctionToMachineFunctionPassAdaptor(
+                      std::move(MFPM)))));
+        } else {
+          MPM.addPass(createModuleToFunctionPassAdaptor(
+              createFunctionToMachineFunctionPassAdaptor(std::move(MFPM))));
+        }
+        MFPM = MachineFunctionPassManager();
+      }
+    }
+
     ModulePassManager &MPM;
     MachineFunctionPassManager MFPM;
     const DerivedT &PB;
@@ -552,6 +605,7 @@ template <typename DerivedT, typename TargetMachineT> class CodeGenPassBuilder {
   /// Helper variable for `-start-before/-start-after/-stop-before/-stop-after`
   mutable bool Started = true;
   mutable bool Stopped = true;
+  mutable bool AddInCGSCCOrder = false;
 };
 
 template <typename Derived, typename TargetMachineT>
@@ -810,6 +864,9 @@ void CodeGenPassBuilder<Derived, TargetMachineT>::addISelPrepare(
     AddIRPass &addPass) const {
   derived().addPreISel(addPass);
 
+  if (Opt.RequiresCodeGenSCCOrder)
+    addPass.requireCGSCCOrder();
+
   addPass(CallBrPreparePass());
   // Add both the safe stack and the stack protection passes: each of them will
   // only protect functions that have corresponding attributes.
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
index b6cc5137d711a..c2b89fd8188f8 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
@@ -2067,6 +2067,8 @@ void AMDGPUCodeGenPassBuilder::addCodeGenPrepare(AddIRPass &addPass) const {
   // being run on them, which causes crashes in the resource usage analysis).
   addPass(AMDGPULowerBufferFatPointersPass(TM));
 
+  addPass.requireCGSCCOrder();
+
   Base::addCodeGenPrepare(addPass);
 
   if (isPassEnabled(EnableLoadStoreVectorizer))
diff --git a/llvm/test/CodeGen/AMDGPU/llc-pipeline-npm.ll b/llvm/test/CodeGen/AMDGPU/llc-pipeline-npm.ll
new file mode 100644
index 0000000000000..96a533a19c88a
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/llc-pipeline-npm.ll
@@ -0,0 +1,144 @@
+; UNSUPPORTED: expensive_checks
+; RUN: llc -O3 -enable-new-pm -mtriple=amdgcn--amdhsa -disable-verify -print-pipeline-passes < %s 2>&1 \
+; RUN:   | tr ',' '\n' | FileCheck -check-prefix=GCN-O3 %s
+
+; REQUIRES: asserts
+
+; GCN-O3: require<MachineModuleAnalysis>
+; GCN-O3-NEXT: require<profile-summary>
+; GCN-O3-NEXT: require<collector-metadata>
+; GCN-O3-NEXT: pre-isel-intrinsic-lowering
+; GCN-O3-NEXT: function(expand-large-div-rem
+; GCN-O3-NEXT: expand-fp)
+; GCN-O3-NEXT: amdgpu-remove-incompatible-functions
+; GCN-O3-NEXT: amdgpu-printf-runtime-binding
+; GCN-O3-NEXT: amdgpu-lower-ctor-dtor
+; GCN-O3-NEXT: function(amdgpu-image-intrinsic-opt)
+; GCN-O3-NEXT: expand-variadics
+; GCN-O3-NEXT: amdgpu-always-inline
+; GCN-O3-NEXT: always-inline
+; GCN-O3-NEXT: amdgpu-export-kernel-runtime-handles
+; GCN-O3-NEXT: amdgpu-sw-lower-lds
+; GCN-O3-NEXT: amdgpu-lower-module-lds
+; GCN-O3-NEXT: function(infer-address-spaces
+; GCN-O3-NEXT: amdgpu-atomic-optimizer
+; GCN-O3-NEXT: atomic-expand
+; GCN-O3-NEXT: amdgpu-promote-alloca
+; GCN-O3-NEXT: separate-const-offset-from-gep<>
+; GCN-O3-NEXT: slsr
+; GCN-O3-NEXT: gvn<>
+; GCN-O3-NEXT: nary-reassociate
+; GCN-O3-NEXT: early-cse<>
+; GCN-O3-NEXT: amdgpu-codegenprepare
+; GCN-O3-NEXT: loop-mssa(loop-reduce)
+; GCN-O3-NEXT: mergeicmps
+; GCN-O3-NEXT: expand-memcmp
+; GCN-O3-NEXT: gc-lowering
+; GCN-O3-NEXT: lower-constant-intrinsics
+; GCN-O3-NEXT: UnreachableBlockElimPass
+; GCN-O3-NEXT: consthoist
+; GCN-O3-NEXT: ReplaceWithVeclib
+; GCN-O3-NEXT: partially-inline-libcalls
+; GCN-O3-NEXT: ee-instrument<post-inline>
+; GCN-O3-NEXT: scalarize-masked-mem-intrin
+; GCN-O3-NEXT: ExpandReductionsPass
+; GCN-O3-NEXT: gvn<>
+; GCN-O3-NEXT: amdgpu-lower-kernel-arguments)
+; GCN-O3-NEXT: amdgpu-lower-buffer-fat-pointers
+; GCN-O3-NEXT: cgscc(function(codegenprepare
+; GCN-O3-NEXT: load-store-vectorizer
+; GCN-O3-NEXT: lower-switch
+; GCN-O3-NEXT: lower-invoke
+; GCN-O3-NEXT: UnreachableBlockElimPass
+; GCN-O3-NEXT: flatten-cfg
+; GCN-O3-NEXT: sink
+; GCN-O3-NEXT: amdgpu-late-codegenprepare
+; GCN-O3-NEXT: amdgpu-unify-divergent-exit-nodes
+; GCN-O3-NEXT: fix-irreducible
+; GCN-O3-NEXT: unify-loop-exits
+; GCN-O3-NEXT: StructurizeCFGPass
+; GCN-O3-NEXT: amdgpu-annotate-uniform
+; GCN-O3-NEXT: si-annotate-control-flow
+; GCN-O3-NEXT: amdgpu-rewrite-undef-for-phi
+; GCN-O3-NEXT: lcssa))
+; GCN-O3-NEXT: amdgpu-perf-hint
+; GCN-O3-NEXT: cgscc(function(require<uniformity>
+; GCN-O3-NEXT: callbr-prepare
+; GCN-O3-NEXT: safe-stack
+; GCN-O3-NEXT: stack-protector))
+; GCN-O3-NEXT: cgscc(function(machine-function(amdgpu-isel
+; GCN-O3-NEXT: si-fix-sgpr-copies
+; GCN-O3-NEXT: si-i1-copies
+; GCN-O3-NEXT: finalize-isel
+; GCN-O3-NEXT: early-tailduplication
+; GCN-O3-NEXT: opt-phis
+; GCN-O3-NEXT: stack-coloring
+; GCN-O3-NEXT: localstackalloc
+; GCN-O3-NEXT: dead-mi-elimination
+; GCN-O3-NEXT: early-machinelicm
+; GCN-O3-NEXT: machine-cse
+; GCN-O3-NEXT: machine-sink
+; GCN-O3-NEXT: peephole-opt
+; GCN-O3-NEXT: dead-mi-elimination
+; GCN-O3-NEXT: si-fold-operands
+; GCN-O3-NEXT: gcn-dpp-combine
+; GCN-O3-NEXT: si-load-store-opt
+; GCN-O3-NEXT: si-peephole-sdwa
+; GCN-O3-NEXT: early-machinelicm
+; GCN-O3-NEXT: machine-cse
+; GCN-O3-NEXT: si-fold-operands
+; GCN-O3-NEXT: dead-mi-elimination
+; GCN-O3-NEXT: si-shrink-instructions
+; GCN-O3-NEXT: detect-dead-lanes
+; GCN-O3-NEXT: InitUndefPass
+; GCN-O3-NEXT: ProcessImplicitDefsPass
+; GCN-O3-NEXT: unreachable-mbb-elimination
+; GCN-O3-NEXT: require<live-vars>
+; GCN-O3-NEXT: require<machine-loops>
+; GCN-O3-NEXT: phi-node-elimination
+; GCN-O3-NEXT: two-address-instruction
+; GCN-O3-NEXT: register-coalescer
+; GCN-O3-NEXT: rename-independent-subregs
+; GCN-O3-NEXT: machine-scheduler
+; GCN-O3-NEXT: greedy<all>
+; GCN-O3-NEXT: amdgpu-nsa-reassign
+; GCN-O3-NEXT: VirtRegRewriterPass
+; GCN-O3-NEXT: stack-slot-coloring
+; GCN-O3-NEXT: machine-cp
+; GCN-O3-NEXT: machinelicm
+; GCN-O3-NEXT: si-fix-vgpr-copies
+; GCN-O3-NEXT: si-optimize-exec-masking
+; GCN-O3-NEXT: remove-redundant-debug-values
+; GCN-O3-NEXT: fixup-statepoint-caller-saved
+; GCN-O3-NEXT: PostRAMachineSinkingPass
+; GCN-O3-NEXT: ShrinkWrapPass
+; GCN-O3-NEXT: PrologEpilogInserterPass
+; GCN-O3-NEXT: branch-folder
+; GCN-O3-NEXT: tailduplication
+; GCN-O3-NEXT: machine-latecleanup
+; GCN-O3-NEXT: machine-cp
+; GCN-O3-NEXT: post-ra-pseudos
+; GCN-O3-NEXT: postmisched
+; GCN-O3-NEXT: block-placement
+; GCN-O3-NEXT: fentry-insert
+; GCN-O3-NEXT: xray-instrumentation
+; GCN-O3-NEXT: patchable-function
+; GCN-O3-NEXT: gcn-create-vopd
+; GCN-O3-NEXT: si-memory-legalizer
+; GCN-O3-NEXT: si-insert-waitcnts
+; GCN-O3-NEXT: si-late-branch-lowering
+; GCN-O3-NEXT: si-pre-emit-peephole
+; GCN-O3-NEXT: post-RA-hazard-rec
+; GCN-O3-NEXT: AMDGPUWaitSGPRHazardsPass
+; GCN-O3-NEXT: amdgpu-insert-delay-alu
+; GCN-O3-NEXT: branch-relaxation
+; GCN-O3-NEXT: remove-loads-into-fake-uses
+; GCN-O3-NEXT: live-debug-values
+; GCN-O3-NEXT: machine-sanmd
+; GCN-O3-NEXT: stack-frame-layout)
+; GCN-O3-NEXT: invalidate<machine-function-info>))
+
+
+define void @empty() {
+  ret void
+}

[cdevadas]

LGTM.