Port -lower-matrix-intrinsics-minimal to NPM

[aeubanks]

Summary

This reuses the existing lower-matrix-intrinsics pass rather than going
the legacy pass route of creating a new pass.

Use this new variant in the NPM -O0 pipeline.

[asbirlea]

llvm/lib/Transforms/Scalar/LowerMatrixIntrinsics.cpp

1954

I'm missing context here...
Looking at the code for LowerMatrixIntrinsics I don't see checks for the analyses being nullptr. I don't see where visiting is limited when analyses are not retrieved.
Even if checks existed, the analyses cannot be preserved below if IR changes occur.

[aeubanks]

fixed preserved analyses

llvm/lib/Transforms/Scalar/LowerMatrixIntrinsics.cpp

1954

I'm following LowerMatrixIntrinsicsMinimalLegacyPass below, the code was changed in https://reviews.llvm.org/D76867.
Modified the PreservedAnalyses.

[asbirlea]

llvm/lib/Transforms/Scalar/LowerMatrixIntrinsics.cpp

1347–1349

Could you add an assert on AA and LI here? The DT check seems to imply AA is guaranteed to exist, and we also mark LI as preserved on the same premise.

llvm/include/llvm/Transforms/Scalar/LowerMatrixIntrinsics.h

 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_TRANSFORMS_SCALAR_LOWERMATRIXINTRINSICSPASS_H
 #define LLVM_TRANSFORMS_SCALAR_LOWERMATRIXINTRINSICSPASS_H
 
 #include "llvm/IR/PassManager.h"
 
 namespace llvm {
-struct LowerMatrixIntrinsicsPass : PassInfoMixin<LowerMatrixIntrinsicsPass> {
+class LowerMatrixIntrinsicsPass
+    : public PassInfoMixin<LowerMatrixIntrinsicsPass> {
+  bool Minimal;
+
+public:
+  LowerMatrixIntrinsicsPass(bool Minimal = false) : Minimal(Minimal) {}
   PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM);
   static bool isRequired() { return true; }
 };
 } // namespace llvm
 
 #endif

llvm/lib/Passes/PassBuilder.cpp

ModulePassManager PassBuilder::buildO0DefaultPipeline(OptimizationLevel Level,
   // lifetime intrinsics to avoid enabling further optimizations during
   // code generation.
   // However, we need to insert lifetime intrinsics to avoid invalid access
   // caused by multithreaded coroutines.
   MPM.addPass(AlwaysInlinerPass(
       /*InsertLifetimeIntrinsics=*/PTO.Coroutines));
 
   if (EnableMatrix)
-    MPM.addPass(createModuleToFunctionPassAdaptor(LowerMatrixIntrinsicsPass()));
+    MPM.addPass(
+        createModuleToFunctionPassAdaptor(LowerMatrixIntrinsicsPass(true)));
 
   if (!LateLoopOptimizationsEPCallbacks.empty()) {
     LoopPassManager LPM(DebugLogging);
     for (auto &C : LateLoopOptimizationsEPCallbacks)
       C(LPM, Level);
     if (!LPM.isEmpty()) {
       MPM.addPass(createModuleToFunctionPassAdaptor(
           createFunctionToLoopPassAdaptor(std::move(LPM))));

llvm/lib/Passes/PassRegistry.def

 FUNCTION_PASS("lint", LintPass())
 FUNCTION_PASS("inject-tli-mappings", InjectTLIMappings())
 FUNCTION_PASS("instnamer", InstructionNamerPass())
 FUNCTION_PASS("loweratomic", LowerAtomicPass())
 FUNCTION_PASS("lower-expect", LowerExpectIntrinsicPass())
 FUNCTION_PASS("lower-guard-intrinsic", LowerGuardIntrinsicPass())
 FUNCTION_PASS("lower-constant-intrinsics", LowerConstantIntrinsicsPass())
 FUNCTION_PASS("lower-matrix-intrinsics", LowerMatrixIntrinsicsPass())
+FUNCTION_PASS("lower-matrix-intrinsics-minimal", LowerMatrixIntrinsicsPass(true))
 FUNCTION_PASS("lower-widenable-condition", LowerWidenableConditionPass())
 FUNCTION_PASS("guard-widening", GuardWideningPass())
 FUNCTION_PASS("load-store-vectorizer", LoadStoreVectorizerPass())
 FUNCTION_PASS("loop-simplify", LoopSimplifyPass())
 FUNCTION_PASS("loop-sink", LoopSinkPass())
 FUNCTION_PASS("loop-unroll-and-jam", LoopUnrollAndJamPass())
 FUNCTION_PASS("loop-flatten", LoopFlattenPass())
 FUNCTION_PASS("lowerinvoke", LowerInvokePass())

llvm/lib/Transforms/Scalar/LowerMatrixIntrinsics.cpp

public:
   /// No need to return a MatrixTy object for the result of the operation, since
   /// the single store user will be lowered as part of this. Instructions that
   /// are completely eliminated by fusion are added to \p FusedInsts.
   void LowerMatrixMultiplyFused(CallInst *MatMul,
                                 SmallPtrSetImpl<Instruction *> &FusedInsts) {
     if (!FuseMatrix || !MatMul->hasOneUse() ||
         MatrixLayout != MatrixLayoutTy::ColumnMajor || !DT)
       return;
 
+    assert(AA && LI && "Analyses should be available");
+
     auto *LoadOp0 = dyn_cast<LoadInst>(MatMul->getOperand(0));
     auto *LoadOp1 = dyn_cast<LoadInst>(MatMul->getOperand(1));
     auto *Store = dyn_cast<StoreInst>(*MatMul->user_begin());
     if (LoadOp0 && LoadOp1 && Store) {
       // The store address must dominate the MatMul instruction, otherwise
       // we create invalid IR.
       // FIXME: See if we can hoist the store address computation.
       auto *AddrI = dyn_cast<Instruction>(Store->getOperand(1));
struct RemarkGenerator {
     }
   };
 };
 } // namespace
 
 PreservedAnalyses LowerMatrixIntrinsicsPass::run(Function &F,
                                                  FunctionAnalysisManager &AM) {
   auto &TTI = AM.getResult<TargetIRAnalysis>(F);
-  auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
-  auto &AA = AM.getResult<AAManager>(F);
-  auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
-  auto &LI = AM.getResult<LoopAnalysis>(F);
+  OptimizationRemarkEmitter *ORE = nullptr;
+  AAResults *AA = nullptr;
+  DominatorTree *DT = nullptr;
+  LoopInfo *LI = nullptr;
+
+  if (!Minimal) {
+    ORE = &AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
+    AA = &AM.getResult<AAManager>(F);
+    DT = &AM.getResult<DominatorTreeAnalysis>(F);
+    LI = &AM.getResult<LoopAnalysis>(F);
+  }
 
-  LowerMatrixIntrinsics LMT(F, TTI, &AA, &DT, &LI, &ORE);
+  LowerMatrixIntrinsics LMT(F, TTI, AA, DT, LI, ORE);
   if (LMT.Visit()) {
     PreservedAnalyses PA;
+    if (!Minimal) {
       PA.preserve<LoopAnalysis>();
       PA.preserve<DominatorTreeAnalysis>();
+    }
     return PA;
   }
   return PreservedAnalyses::all();
 }
 
 namespace {
 
 class LowerMatrixIntrinsicsLegacyPass : public FunctionPass {

llvm/test/Other/new-pm-O0-defaults.ll

 ; The IR below was crafted so as:
 ; 1) To have a loop, so we create a loop pass manager
 ; 2) To be "immutable" in the sense that no pass in the standard
 ;    pipeline will modify it.
 ; Since no transformations take place, we don't expect any analyses
 ; to be invalidated.
 ; Any invalidation that shows up here is a bug, unless we started modifying
 ; the IR, in which case we need to make it immutable harder.
 
 ; RUN: opt -disable-verify -debug-pass-manager \
 ; RUN:     -passes='default<O0>' -S %s 2>&1 \
 ; RUN:     | FileCheck %s --check-prefixes=CHECK,CHECK-DEFAULT
+; RUN: opt -disable-verify -debug-pass-manager -enable-matrix \
+; RUN:     -passes='default<O0>' -S %s 2>&1 \
+; RUN:     | FileCheck %s --check-prefixes=CHECK,CHECK-DEFAULT,CHECK-MATRIX
 ; RUN: opt -disable-verify -debug-pass-manager \
 ; RUN:     -passes='thinlto-pre-link<O0>' -S %s 2>&1 \
 ; RUN:     | FileCheck %s --check-prefixes=CHECK,CHECK-DEFAULT,CHECK-PRE-LINK
 ; RUN: opt -disable-verify -debug-pass-manager \
 ; RUN:     -passes='lto-pre-link<O0>' -S %s 2>&1 \
 ; RUN:     | FileCheck %s --check-prefixes=CHECK,CHECK-DEFAULT,CHECK-PRE-LINK
 ; RUN: opt -disable-verify -debug-pass-manager \
 ; RUN:     -passes='thinlto<O0>' -S %s 2>&1 \
 ; RUN:     | FileCheck %s --check-prefixes=CHECK,CHECK-THINLTO
 ; RUN: opt -disable-verify -debug-pass-manager \
 ; RUN:     -passes='lto<O0>' -S %s 2>&1 \
 ; RUN:     | FileCheck %s --check-prefixes=CHECK,CHECK-LTO
 
 ; CHECK: Starting llvm::Module pass manager run.
 ; CHECK-DEFAULT-NEXT: Running pass: AlwaysInlinerPass
 ; CHECK-DEFAULT-NEXT: Running analysis: InnerAnalysisManagerProxy
 ; CHECK-DEFAULT-NEXT: Running analysis: ProfileSummaryAnalysis
+; CHECK-MATRIX-NEXT: Running pass: LowerMatrixIntrinsicsPass
+; CHECK-MATRIX-NEXT: Running analysis: TargetIRAnalysis
 ; CHECK-PRE-LINK-NEXT: Running pass: CanonicalizeAliasesPass
 ; CHECK-PRE-LINK-NEXT: Running pass: NameAnonGlobalPass
 ; CHECK-THINLTO-NEXT: Running pass: Annotation2MetadataPass
 ; CHECK-LTO-NEXT: Running pass: Annotation2MetadataPass
 ; CHECK-LTO-NEXT: Running pass: WholeProgramDevirtPass
 ; CHECK-LTO-NEXT: Running analysis: InnerAnalysisManagerProxy
 ; CHECK-LTO-NEXT: Running pass: LowerTypeTestsPass
 ; CHECK-LTO-NEXT: Running pass: LowerTypeTestsPass

llvm/test/Transforms/LowerMatrixIntrinsics/multiply-minimal.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt -lower-matrix-intrinsics-minimal -fuse-matrix-tile-size=2 -matrix-allow-contract -force-fuse-matrix -instcombine -verify-dom-info %s -S | FileCheck %s
+; RUN: opt -passes='lower-matrix-intrinsics-minimal,instcombine,verify<dom-tree>' -fuse-matrix-tile-size=2 -matrix-allow-contract -force-fuse-matrix %s -S | FileCheck %s
 
 ; Test for the minimal version of the matrix lowering pass, which does not
 ; require DT or AA. Make sure no tiling is happening, even though it was
 ; requested.
 
 ; REQUIRES: aarch64-registered-target
 
 target datalayout = "e-m:o-i64:64-f80:128-n8:8:32:64-S128"