[InferAddressSpaces] Support assumed addrspaces from addrspace predicates.

- CUDA cannot associate memory space with pointer types. Even though Clang could add extra attributes to specify the address space explicitly on a pointer type, it breaks the portability between Clang and NVCC.
- This change proposes to assume the address space from a pointer from the assumption built upon target-specific address space predicates, such as `__isGlobal` from CUDA. E.g.,

```
  foo(float *p) {
    __builtin_assume(__isGlobal(p));
    // From there, we could assume p is a global pointer instead of a
    // generic one.
  }
```

This makes the code portable without introducing the implementation-specific features.

Note that NVCC starts to support __builtin_assume from version 11.

Reviewed By: arsenm

Differential Revision: https://reviews.llvm.org/D112041

Author: darkbuck

clang/test/CodeGen/thinlto-distributed-newpm.ll

@@ -47,11 +47,11 @@
 ; CHECK-O: Running pass: PromotePass
 ; CHECK-O: Running analysis: DominatorTreeAnalysis on main
 ; CHECK-O: Running analysis: AssumptionAnalysis on main
+; CHECK-O: Running analysis: TargetIRAnalysis on main
 ; CHECK-O: Running pass: DeadArgumentEliminationPass
 ; CHECK-O: Running pass: InstCombinePass on main
 ; CHECK-O: Running analysis: TargetLibraryAnalysis on main
 ; CHECK-O: Running analysis: OptimizationRemarkEmitterAnalysis on main
-; CHECK-O: Running analysis: TargetIRAnalysis on main
 ; CHECK-O: Running analysis: AAManager on main
 ; CHECK-O: Running analysis: BasicAA on main
 ; CHECK-O: Running analysis: ScopedNoAliasAA on main

llvm/include/llvm/Analysis/AssumptionCache.h

@@ -29,6 +29,7 @@ namespace llvm {
 class AssumeInst;
 class Function;
 class raw_ostream;
+class TargetTransformInfo;
 class Value;
 
 /// A cache of \@llvm.assume calls within a function.
@@ -59,6 +60,8 @@ class AssumptionCache {
   /// We track this to lazily populate our assumptions.
   Function &F;
 
+  TargetTransformInfo *TTI;
+
   /// Vector of weak value handles to calls of the \@llvm.assume
   /// intrinsic.
   SmallVector<ResultElem, 4> AssumeHandles;
@@ -103,7 +106,8 @@ class AssumptionCache {
 public:
   /// Construct an AssumptionCache from a function by scanning all of
   /// its instructions.
-  AssumptionCache(Function &F) : F(F) {}
+  AssumptionCache(Function &F, TargetTransformInfo *TTI = nullptr)
+      : F(F), TTI(TTI) {}
 
   /// This cache is designed to be self-updating and so it should never be
   /// invalidated.
@@ -174,9 +178,7 @@ class AssumptionAnalysis : public AnalysisInfoMixin<AssumptionAnalysis> {
 public:
   using Result = AssumptionCache;
 
-  AssumptionCache run(Function &F, FunctionAnalysisManager &) {
-    return AssumptionCache(F);
-  }
+  AssumptionCache run(Function &F, FunctionAnalysisManager &);
 };
 
 /// Printer pass for the \c AssumptionAnalysis results.

llvm/include/llvm/Analysis/TargetTransformInfo.h

@@ -30,6 +30,7 @@
 #include "llvm/Support/DataTypes.h"
 #include "llvm/Support/InstructionCost.h"
 #include <functional>
+#include <utility>
 
 namespace llvm {
 
@@ -389,6 +390,9 @@ class TargetTransformInfo {
 
   unsigned getAssumedAddrSpace(const Value *V) const;
 
+  std::pair<const Value *, unsigned>
+  getPredicatedAddrSpace(const Value *V) const;
+
   /// Rewrite intrinsic call \p II such that \p OldV will be replaced with \p
   /// NewV, which has a different address space. This should happen for every
   /// operand index that collectFlatAddressOperands returned for the intrinsic.
@@ -1481,6 +1485,8 @@ class TargetTransformInfo::Concept {
   virtual bool
   canHaveNonUndefGlobalInitializerInAddressSpace(unsigned AS) const = 0;
   virtual unsigned getAssumedAddrSpace(const Value *V) const = 0;
+  virtual std::pair<const Value *, unsigned>
+  getPredicatedAddrSpace(const Value *V) const = 0;
   virtual Value *rewriteIntrinsicWithAddressSpace(IntrinsicInst *II,
                                                   Value *OldV,
                                                   Value *NewV) const = 0;
@@ -1824,6 +1830,11 @@ class TargetTransformInfo::Model final : public TargetTransformInfo::Concept {
     return Impl.getAssumedAddrSpace(V);
   }
 
+  std::pair<const Value *, unsigned>
+  getPredicatedAddrSpace(const Value *V) const override {
+    return Impl.getPredicatedAddrSpace(V);
+  }
+
   Value *rewriteIntrinsicWithAddressSpace(IntrinsicInst *II, Value *OldV,
                                           Value *NewV) const override {
     return Impl.rewriteIntrinsicWithAddressSpace(II, OldV, NewV);

llvm/include/llvm/Analysis/TargetTransformInfoImpl.h

@@ -24,6 +24,7 @@
 #include "llvm/IR/Operator.h"
 #include "llvm/IR/PatternMatch.h"
 #include "llvm/IR/Type.h"
+#include <utility>
 
 using namespace llvm::PatternMatch;
 
@@ -110,6 +111,11 @@ class TargetTransformInfoImplBase {
 
   unsigned getAssumedAddrSpace(const Value *V) const { return -1; }
 
+  std::pair<const Value *, unsigned>
+  getPredicatedAddrSpace(const Value *V) const {
+    return std::make_pair(nullptr, -1);
+  }
+
   Value *rewriteIntrinsicWithAddressSpace(IntrinsicInst *II, Value *OldV,
                                           Value *NewV) const {
     return nullptr;

llvm/include/llvm/CodeGen/BasicTTIImpl.h

@@ -283,6 +283,11 @@ class BasicTTIImplBase : public TargetTransformInfoImplCRTPBase<T> {
     return getTLI()->getTargetMachine().getAssumedAddrSpace(V);
   }
 
+  std::pair<const Value *, unsigned>
+  getPredicatedAddrSpace(const Value *V) const {
+    return getTLI()->getTargetMachine().getPredicatedAddrSpace(V);
+  }
+
   Value *rewriteIntrinsicWithAddressSpace(IntrinsicInst *II, Value *OldV,
                                           Value *NewV) const {
     return nullptr;

llvm/include/llvm/Target/TargetMachine.h

@@ -25,6 +25,7 @@
 #include "llvm/Target/CGPassBuilderOption.h"
 #include "llvm/Target/TargetOptions.h"
 #include <string>
+#include <utility>
 
 namespace llvm {
 
@@ -319,6 +320,18 @@ class TargetMachine {
   /// properties.
   virtual unsigned getAssumedAddrSpace(const Value *V) const { return -1; }
 
+  /// If the specified predicate checks whether a generic pointer falls within
+  /// a specified address space, return that generic pointer and the address
+  /// space being queried.
+  ///
+  /// Such predicates could be specified in @llvm.assume intrinsics for the
+  /// optimizer to assume that the given generic pointer always falls within
+  /// the address space based on that predicate.
+  virtual std::pair<const Value *, unsigned>
+  getPredicatedAddrSpace(const Value *V) const {
+    return std::make_pair(nullptr, -1);
+  }
+
   /// Get a \c TargetIRAnalysis appropriate for the target.
   ///
   /// This is used to construct the new pass manager's target IR analysis pass,

llvm/lib/Analysis/AssumptionCache.cpp

@@ -16,6 +16,7 @@
 #include "llvm/ADT/STLExtras.h"
 #include "llvm/ADT/SmallPtrSet.h"
 #include "llvm/ADT/SmallVector.h"
+#include "llvm/Analysis/TargetTransformInfo.h"
 #include "llvm/IR/BasicBlock.h"
 #include "llvm/IR/Function.h"
 #include "llvm/IR/InstrTypes.h"
@@ -56,7 +57,7 @@ AssumptionCache::getOrInsertAffectedValues(Value *V) {
 }
 
 static void
-findAffectedValues(CallBase *CI,
+findAffectedValues(CallBase *CI, TargetTransformInfo *TTI,
                    SmallVectorImpl<AssumptionCache::ResultElem> &Affected) {
   // Note: This code must be kept in-sync with the code in
   // computeKnownBitsFromAssume in ValueTracking.
@@ -124,11 +125,19 @@ findAffectedValues(CallBase *CI,
         match(B, m_ConstantInt()))
       AddAffected(X);
   }
+
+  if (TTI) {
+    const Value *Ptr;
+    unsigned AS;
+    std::tie(Ptr, AS) = TTI->getPredicatedAddrSpace(Cond);
+    if (Ptr)
+      AddAffected(const_cast<Value *>(Ptr->stripInBoundsOffsets()));
+  }
 }
 
 void AssumptionCache::updateAffectedValues(AssumeInst *CI) {
   SmallVector<AssumptionCache::ResultElem, 16> Affected;
-  findAffectedValues(CI, Affected);
+  findAffectedValues(CI, TTI, Affected);
 
   for (auto &AV : Affected) {
     auto &AVV = getOrInsertAffectedValues(AV.Assume);
@@ -141,7 +150,7 @@ void AssumptionCache::updateAffectedValues(AssumeInst *CI) {
 
 void AssumptionCache::unregisterAssumption(AssumeInst *CI) {
   SmallVector<AssumptionCache::ResultElem, 16> Affected;
-  findAffectedValues(CI, Affected);
+  findAffectedValues(CI, TTI, Affected);
 
   for (auto &AV : Affected) {
     auto AVI = AffectedValues.find_as(AV.Assume);
@@ -248,6 +257,12 @@ void AssumptionCache::registerAssumption(AssumeInst *CI) {
   updateAffectedValues(CI);
 }
 
+AssumptionCache AssumptionAnalysis::run(Function &F,
+                                        FunctionAnalysisManager &FAM) {
+  auto &TTI = FAM.getResult<TargetIRAnalysis>(F);
+  return AssumptionCache(F, &TTI);
+}
+
 AnalysisKey AssumptionAnalysis::Key;
 
 PreservedAnalyses AssumptionPrinterPass::run(Function &F,
@@ -278,10 +293,13 @@ AssumptionCache &AssumptionCacheTracker::getAssumptionCache(Function &F) {
   if (I != AssumptionCaches.end())
     return *I->second;
 
+  auto *TTIWP = getAnalysisIfAvailable<TargetTransformInfoWrapperPass>();
+  auto *TTI = TTIWP ? &TTIWP->getTTI(F) : nullptr;
+
   // Ok, build a new cache by scanning the function, insert it and the value
   // handle into our map, and return the newly populated cache.
   auto IP = AssumptionCaches.insert(std::make_pair(
-      FunctionCallbackVH(&F, this), std::make_unique<AssumptionCache>(F)));
+      FunctionCallbackVH(&F, this), std::make_unique<AssumptionCache>(F, TTI)));
   assert(IP.second && "Scanning function already in the map?");
   return *IP.first->second;
 }

llvm/lib/Analysis/TargetTransformInfo.cpp

@@ -268,6 +268,11 @@ unsigned TargetTransformInfo::getAssumedAddrSpace(const Value *V) const {
   return TTIImpl->getAssumedAddrSpace(V);
 }
 
+std::pair<const Value *, unsigned>
+TargetTransformInfo::getPredicatedAddrSpace(const Value *V) const {
+  return TTIImpl->getPredicatedAddrSpace(V);
+}
+
 Value *TargetTransformInfo::rewriteIntrinsicWithAddressSpace(
     IntrinsicInst *II, Value *OldV, Value *NewV) const {
   return TTIImpl->rewriteIntrinsicWithAddressSpace(II, OldV, NewV);

llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

@@ -36,8 +36,10 @@
 #include "llvm/CodeGen/Passes.h"
 #include "llvm/CodeGen/RegAllocRegistry.h"
 #include "llvm/CodeGen/TargetPassConfig.h"
+#include "llvm/IR/IntrinsicsAMDGPU.h"
 #include "llvm/IR/LegacyPassManager.h"
 #include "llvm/IR/PassManager.h"
+#include "llvm/IR/PatternMatch.h"
 #include "llvm/InitializePasses.h"
 #include "llvm/MC/TargetRegistry.h"
 #include "llvm/Passes/PassBuilder.h"
@@ -780,6 +782,33 @@ unsigned AMDGPUTargetMachine::getAssumedAddrSpace(const Value *V) const {
   return AMDGPUAS::GLOBAL_ADDRESS;
 }
 
+std::pair<const Value *, unsigned>
+AMDGPUTargetMachine::getPredicatedAddrSpace(const Value *V) const {
+  if (auto *II = dyn_cast<IntrinsicInst>(V)) {
+    switch (II->getIntrinsicID()) {
+    case Intrinsic::amdgcn_is_shared:
+      return std::make_pair(II->getArgOperand(0), AMDGPUAS::LOCAL_ADDRESS);
+    case Intrinsic::amdgcn_is_private:
+      return std::make_pair(II->getArgOperand(0), AMDGPUAS::PRIVATE_ADDRESS);
+    default:
+      break;
+    }
+    return std::make_pair(nullptr, -1);
+  }
+  // Check the global pointer predication based on
+  // (!is_share(p) && !is_private(p)). Note that logic 'and' is commutative and
+  // the order of 'is_shared' and 'is_private' is not significant.
+  Value *Ptr;
+  if (match(
+          const_cast<Value *>(V),
+          m_c_And(m_Not(m_Intrinsic<Intrinsic::amdgcn_is_shared>(m_Value(Ptr))),
+                  m_Not(m_Intrinsic<Intrinsic::amdgcn_is_private>(
+                      m_Deferred(Ptr))))))
+    return std::make_pair(Ptr, AMDGPUAS::GLOBAL_ADDRESS);
+
+  return std::make_pair(nullptr, -1);
+}
+
 //===----------------------------------------------------------------------===//
 // GCN Target Machine (SI+)
 //===----------------------------------------------------------------------===//

llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.h

@@ -17,6 +17,7 @@
 #include "GCNSubtarget.h"
 #include "llvm/CodeGen/TargetPassConfig.h"
 #include "llvm/Target/TargetMachine.h"
+#include <utility>
 
 namespace llvm {
 
@@ -63,6 +64,9 @@ class AMDGPUTargetMachine : public LLVMTargetMachine {
   bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override;
 
   unsigned getAssumedAddrSpace(const Value *V) const override;
+
+  std::pair<const Value *, unsigned>
+  getPredicatedAddrSpace(const Value *V) const override;
 };
 
 //===----------------------------------------------------------------------===//

llvm/lib/Target/NVPTX/NVPTXTargetMachine.cpp

@@ -23,6 +23,7 @@
 #include "llvm/Analysis/TargetTransformInfo.h"
 #include "llvm/CodeGen/Passes.h"
 #include "llvm/CodeGen/TargetPassConfig.h"
+#include "llvm/IR/IntrinsicsNVPTX.h"
 #include "llvm/IR/LegacyPassManager.h"
 #include "llvm/MC/TargetRegistry.h"
 #include "llvm/Pass.h"
@@ -240,6 +241,25 @@ NVPTXTargetMachine::getTargetTransformInfo(const Function &F) {
   return TargetTransformInfo(NVPTXTTIImpl(this, F));
 }
 
+std::pair<const Value *, unsigned>
+NVPTXTargetMachine::getPredicatedAddrSpace(const Value *V) const {
+  if (auto *II = dyn_cast<IntrinsicInst>(V)) {
+    switch (II->getIntrinsicID()) {
+    case Intrinsic::nvvm_isspacep_const:
+      return std::make_pair(II->getArgOperand(0), llvm::ADDRESS_SPACE_CONST);
+    case Intrinsic::nvvm_isspacep_global:
+      return std::make_pair(II->getArgOperand(0), llvm::ADDRESS_SPACE_GLOBAL);
+    case Intrinsic::nvvm_isspacep_local:
+      return std::make_pair(II->getArgOperand(0), llvm::ADDRESS_SPACE_LOCAL);
+    case Intrinsic::nvvm_isspacep_shared:
+      return std::make_pair(II->getArgOperand(0), llvm::ADDRESS_SPACE_SHARED);
+    default:
+      break;
+    }
+  }
+  return std::make_pair(nullptr, -1);
+}
+
 void NVPTXPassConfig::addEarlyCSEOrGVNPass() {
   if (getOptLevel() == CodeGenOpt::Aggressive)
     addPass(createGVNPass());

llvm/lib/Target/NVPTX/NVPTXTargetMachine.h

@@ -16,6 +16,7 @@
 #include "ManagedStringPool.h"
 #include "NVPTXSubtarget.h"
 #include "llvm/Target/TargetMachine.h"
+#include <utility>
 
 namespace llvm {
 
@@ -69,6 +70,9 @@ class NVPTXTargetMachine : public LLVMTargetMachine {
   bool isMachineVerifierClean() const override {
     return false;
   }
+
+  std::pair<const Value *, unsigned>
+  getPredicatedAddrSpace(const Value *V) const override;
 }; // NVPTXTargetMachine.
 
 class NVPTXTargetMachine32 : public NVPTXTargetMachine {