[SVFS] Inject TLI Mappings in VFABI attribute.

This patch introduces a function pass to inject the scalar-to-vector
mappings stored in the TargetLIbraryInfo (TLI) into the Vector
Function ABI (VFABI) variants attribute.

The test is testing the injection for three vector libraries supported
by the TLI (Accelerate, SVML, MASSV).

The pass does not change any of the analysis associated to the
function.

Differential Revision: https://reviews.llvm.org/D70107

include/llvm/Analysis/TargetLibraryInfo.h

@@ -197,6 +197,10 @@ public:
   /// Returns the size of the wchar_t type in bytes or 0 if the size is unknown.
   /// This queries the 'wchar_size' metadata.
   unsigned getWCharSize(const Module &M) const;
+
+  /// Returns the largest vectorization factor used in the list of
+  /// vector functions.
+  unsigned getWidestVF(StringRef ScalarF) const;
 };
 
 /// Provides information about what library functions are available for
@@ -337,6 +341,12 @@ public:
                   FunctionAnalysisManager::Invalidator &) {
     return false;
   }
+
+  /// Returns the largest vectorization factor used in the list of
+  /// vector functions.
+  unsigned getWidestVF(StringRef ScalarF) const {
+    return Impl->getWidestVF(ScalarF);
+  }
 };
 
 /// Analysis pass providing the \c TargetLibraryInfo.

include/llvm/InitializePasses.h

@@ -180,6 +180,7 @@ void initializeIndVarSimplifyLegacyPassPass(PassRegistry&);
 void initializeIndirectBrExpandPassPass(PassRegistry&);
 void initializeInferAddressSpacesPass(PassRegistry&);
 void initializeInferFunctionAttrsLegacyPassPass(PassRegistry&);
+void initializeInjectTLIMappingsLegacyPass(PassRegistry &);
 void initializeInlineCostAnalysisPass(PassRegistry&);
 void initializeInstCountPass(PassRegistry&);
 void initializeInstNamerPass(PassRegistry&);

include/llvm/LinkAllPasses.h

@@ -225,6 +225,7 @@ namespace {
       (void) llvm::createScalarizeMaskedMemIntrinPass();
       (void) llvm::createWarnMissedTransformationsPass();
       (void) llvm::createHardwareLoopsPass();
+      (void)llvm::createInjectTLIMappingsLegacyPass();
 
       (void)new llvm::IntervalPartition();
       (void)new llvm::ScalarEvolutionWrapperPass();

include/llvm/Transforms/Utils.h

@@ -119,6 +119,13 @@ ModulePass *createStripNonLineTableDebugInfoPass();
 // number of conditional branches in the hot paths based on profiles.
 //
 FunctionPass *createControlHeightReductionLegacyPass();
+
+//===----------------------------------------------------------------------===//
+//
+// InjectTLIMappingsLegacy - populates the VFABI attribute with the
+// scalar-to-vector mappings from the TargetLibraryInfo.
+//
+FunctionPass *createInjectTLIMappingsLegacyPass();
 }
 
 #endif

include/llvm/Transforms/Utils/InjectTLIMappings.h

@@ -0,0 +1,37 @@
+//===- InjectTLIMAppings.h - TLI to VFABI attribute injection  ------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// Populates the VFABI attribute with the scalar-to-vector mappings
+// from the TargetLibraryInfo.
+//
+//===----------------------------------------------------------------------===//
+#ifndef LLVM_TRANSFORMS_UTILS_INJECTTLIMAPPINGS_H
+#define LLVM_TRANSFORMS_UTILS_INJECTTLIMAPPINGS_H
+
+#include "llvm/IR/PassManager.h"
+#include "llvm/InitializePasses.h"
+
+namespace llvm {
+class InjectTLIMappings : public PassInfoMixin<InjectTLIMappings> {
+public:
+  PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM);
+};
+
+// Legacy pass
+class InjectTLIMappingsLegacy : public FunctionPass {
+public:
+  static char ID;
+  InjectTLIMappingsLegacy() : FunctionPass(ID) {
+    initializeInjectTLIMappingsLegacyPass(*PassRegistry::getPassRegistry());
+  }
+  void getAnalysisUsage(AnalysisUsage &AU) const override;
+  bool runOnFunction(Function &F) override;
+};
+
+} // End namespace llvm
+#endif // LLVM_TRANSFORMS_UTILS_INJECTTLIMAPPINGS_H

lib/Analysis/TargetLibraryInfo.cpp

@@ -1637,3 +1637,19 @@ INITIALIZE_PASS(TargetLibraryInfoWrapperPass, "targetlibinfo",
 char TargetLibraryInfoWrapperPass::ID = 0;
 
 void TargetLibraryInfoWrapperPass::anchor() {}
+
+unsigned TargetLibraryInfoImpl::getWidestVF(StringRef ScalarF) const {
+  ScalarF = sanitizeFunctionName(ScalarF);
+  if (ScalarF.empty())
+    return 1;
+
+  unsigned VF = 1;
+  std::vector<VecDesc>::const_iterator I =
+      llvm::lower_bound(VectorDescs, ScalarF, compareWithScalarFnName);
+  while (I != VectorDescs.end() && StringRef(I->ScalarFnName) == ScalarF) {
+    if (I->VectorizationFactor > VF)
+      VF = I->VectorizationFactor;
+    ++I;
+  }
+  return VF;
+}

lib/Analysis/VectorUtils.cpp

@@ -1166,6 +1166,9 @@ void VFABI::getVectorVariantNames(
   const StringRef S =
       CI.getAttribute(AttributeList::FunctionIndex, VFABI::MappingsAttrName)
           .getValueAsString();
+  if (S.empty())
+    return;
+
   SmallVector<StringRef, 8> ListAttr;
   S.split(ListAttr, ",");
 

lib/Passes/PassBuilder.cpp

@@ -170,6 +170,7 @@
 #include "llvm/Transforms/Utils/BreakCriticalEdges.h"
 #include "llvm/Transforms/Utils/CanonicalizeAliases.h"
 #include "llvm/Transforms/Utils/EntryExitInstrumenter.h"
+#include "llvm/Transforms/Utils/InjectTLIMappings.h"
 #include "llvm/Transforms/Utils/LCSSA.h"
 #include "llvm/Transforms/Utils/LibCallsShrinkWrap.h"
 #include "llvm/Transforms/Utils/LoopSimplify.h"

lib/Passes/PassRegistry.def

@@ -184,6 +184,7 @@ FUNCTION_PASS("invalidate<all>", InvalidateAllAnalysesPass())
 FUNCTION_PASS("float2int", Float2IntPass())
 FUNCTION_PASS("no-op-function", NoOpFunctionPass())
 FUNCTION_PASS("libcalls-shrinkwrap", LibCallsShrinkWrapPass())
+FUNCTION_PASS("inject-tli-mappings", InjectTLIMappings())
 FUNCTION_PASS("loweratomic", LowerAtomicPass())
 FUNCTION_PASS("lower-expect", LowerExpectIntrinsicPass())
 FUNCTION_PASS("lower-guard-intrinsic", LowerGuardIntrinsicPass())

lib/Transforms/Utils/CMakeLists.txt

@@ -23,6 +23,7 @@ add_llvm_library(LLVMTransformUtils
   GuardUtils.cpp
   InlineFunction.cpp
   ImportedFunctionsInliningStatistics.cpp
+  InjectTLIMappings.cpp
   InstructionNamer.cpp
   IntegerDivision.cpp
   LCSSA.cpp

lib/Transforms/Utils/InjectTLIMappings.cpp

@@ -0,0 +1,186 @@
+//===- InjectTLIMAppings.cpp - TLI to VFABI attribute injection  ----------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// Populates the VFABI attribute with the scalar-to-vector mappings
+// from the TargetLibraryInfo.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Transforms/Utils/InjectTLIMappings.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/Analysis/VectorUtils.h"
+#include "llvm/IR/InstIterator.h"
+#include "llvm/Transforms/Utils.h"
+#include "llvm/Transforms/Utils/ModuleUtils.h"
+
+using namespace llvm;
+
+#define DEBUG_TYPE "inject-tli-mappings"
+
+STATISTIC(NumCallInjected,
+          "Number of calls in which the mappings have been injected.");
+
+STATISTIC(NumVFDeclAdded,
+          "Number of function declarations that have been added.");
+STATISTIC(NumCompUsedAdded,
+          "Number of `@llvm.compiler.used` operands that have been added.");
+
+/// Helper function to map the TLI name to a strings that holds
+/// scalar-to-vector mapping.
+///
+///    _ZGV<isa><mask><vlen><vparams>_<scalarname>(<vectorname>)
+///
+/// where:
+///
+/// <isa> = "_LLVM_"
+/// <mask> = "N". Note: TLI does not support masked interfaces.
+/// <vlen> = Number of concurrent lanes, stored in the `VectorizationFactor`
+///          field of the `VecDesc` struct.
+/// <vparams> = "v", as many as are the number of parameters of CI.
+/// <scalarname> = the name of the scalar function called by CI.
+/// <vectorname> = the name of the vector function mapped by the TLI.
+static std::string mangleTLIName(StringRef VectorName, const CallInst &CI,
+                                 unsigned VF) {
+  SmallString<256> Buffer;
+  llvm::raw_svector_ostream Out(Buffer);
+  Out << "_ZGV" << VFABI::_LLVM_ << "N" << VF;
+  for (unsigned I = 0; I < CI.getNumArgOperands(); ++I)
+    Out << "v";
+  Out << "_" << CI.getCalledFunction()->getName() << "(" << VectorName << ")";
+  return Out.str();
+}
+
+/// A helper function for converting Scalar types to vector types.
+/// If the incoming type is void, we return void. If the VF is 1, we return
+/// the scalar type.
+static Type *ToVectorTy(Type *Scalar, unsigned VF, bool isScalable = false) {
+  if (Scalar->isVoidTy() || VF == 1)
+    return Scalar;
+  return VectorType::get(Scalar, {VF, isScalable});
+}
+
+/// A helper function that adds the vector function declaration that
+/// vectorizes the CallInst CI with a vectorization factor of VF
+/// lanes. The TLI assumes that all parameters and the return type of
+/// CI (other than void) need to be widened to a VectorType of VF
+/// lanes.
+static void addVariantDeclaration(CallInst &CI, const unsigned VF,
+                                  const StringRef VFName) {
+  Module *M = CI.getModule();
+
+  // Add function declaration.
+  Type *RetTy = ToVectorTy(CI.getType(), VF);
+  SmallVector<Type *, 4> Tys;
+  for (Value *ArgOperand : CI.arg_operands())
+    Tys.push_back(ToVectorTy(ArgOperand->getType(), VF));
+  assert(!CI.getFunctionType()->isVarArg() &&
+         "VarArg functions are not supported.");
+  FunctionType *FTy = FunctionType::get(RetTy, Tys, /*isVarArg=*/false);
+  Function *VectorF =
+      Function::Create(FTy, Function::ExternalLinkage, VFName, M);
+  VectorF->copyAttributesFrom(CI.getCalledFunction());
+  ++NumVFDeclAdded;
+  LLVM_DEBUG(dbgs() << DEBUG_TYPE << ": Added to the module: `" << VFName
+                    << "` of type " << *(VectorF->getType()) << "\n");
+
+  // Make function declaration (without a body) "sticky" in the IR by
+  // listing it in the @llvm.compiler.used intrinsic.
+  assert(!VectorF->size() && "VFABI attribute requires `@llvm.compiler.used` "
+                             "only on declarations.");
+  appendToCompilerUsed(*M, {VectorF});
+  LLVM_DEBUG(dbgs() << DEBUG_TYPE << ": Adding `" << VFName
+                    << "` to `@llvm.compiler.used`.\n");
+  ++NumCompUsedAdded;
+}
+
+static void addMappingsFromTLI(const TargetLibraryInfo &TLI, CallInst &CI) {
+  // This is needed to make sure we don't query the TLI for calls to
+  // bitcast of function pointers, like `%call = call i32 (i32*, ...)
+  // bitcast (i32 (...)* @goo to i32 (i32*, ...)*)(i32* nonnull %i)`,
+  // as such calls make the `isFunctionVectorizable` raise an
+  // exception.
+  if (CI.isNoBuiltin() || !CI.getCalledFunction())
+    return;
+
+  const std::string ScalarName = CI.getCalledFunction()->getName();
+  // Nothing to be done if the TLI thinks the function is not
+  // vectorizable.
+  if (!TLI.isFunctionVectorizable(ScalarName))
+    return;
+  SmallVector<std::string, 8> Mappings;
+  VFABI::getVectorVariantNames(CI, Mappings);
+  Module *M = CI.getModule();
+  const SetVector<StringRef> OriginalSetOfMappings(Mappings.begin(),
+                                                   Mappings.end());
+  //  All VFs in the TLI are powers of 2.
+  for (unsigned VF = 2, WidestVF = TLI.getWidestVF(ScalarName); VF <= WidestVF;
+       VF *= 2) {
+    const std::string TLIName = TLI.getVectorizedFunction(ScalarName, VF);
+    if (!TLIName.empty()) {
+      std::string MangledName = mangleTLIName(TLIName, CI, VF);
+      if (!OriginalSetOfMappings.count(MangledName)) {
+        Mappings.push_back(MangledName);
+        ++NumCallInjected;
+      }
+      Function *VariantF = M->getFunction(TLIName);
+      if (!VariantF)
+        addVariantDeclaration(CI, VF, TLIName);
+    }
+  }
+
+  VFABI::setVectorVariantNames(&CI, Mappings);
+}
+
+static bool runImpl(const TargetLibraryInfo &TLI, Function &F) {
+  for (auto &I : instructions(F))
+    if (auto CI = dyn_cast<CallInst>(&I))
+      addMappingsFromTLI(TLI, *CI);
+  // Even if the pass adds IR attributes, the analyses are preserved.
+  return false;
+}
+
+////////////////////////////////////////////////////////////////////////////////
+// New pass manager implementation.
+////////////////////////////////////////////////////////////////////////////////
+PreservedAnalyses InjectTLIMappings::run(Function &F,
+                                         FunctionAnalysisManager &AM) {
+  const TargetLibraryInfo &TLI = AM.getResult<TargetLibraryAnalysis>(F);
+  runImpl(TLI, F);
+  // Even if the pass adds IR attributes, the analyses are preserved.
+  return PreservedAnalyses::all();
+}
+
+////////////////////////////////////////////////////////////////////////////////
+// Legacy PM Implementation.
+////////////////////////////////////////////////////////////////////////////////
+bool InjectTLIMappingsLegacy::runOnFunction(Function &F) {
+  const TargetLibraryInfo &TLI =
+      getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
+  return runImpl(TLI, F);
+}
+
+void InjectTLIMappingsLegacy::getAnalysisUsage(AnalysisUsage &AU) const {
+  AU.setPreservesCFG();
+  AU.addRequired<TargetLibraryInfoWrapperPass>();
+  AU.addPreserved<TargetLibraryInfoWrapperPass>();
+}
+
+////////////////////////////////////////////////////////////////////////////////
+// Legacy Pass manager initialization
+////////////////////////////////////////////////////////////////////////////////
+char InjectTLIMappingsLegacy::ID = 0;
+
+INITIALIZE_PASS_BEGIN(InjectTLIMappingsLegacy, DEBUG_TYPE,
+                      "Inject TLI Mappings", false, false)
+INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
+INITIALIZE_PASS_END(InjectTLIMappingsLegacy, DEBUG_TYPE, "Inject TLI Mappings",
+                    false, false)
+
+FunctionPass *llvm::createInjectTLIMappingsLegacyPass() {
+  return new InjectTLIMappingsLegacy();
+}

lib/Transforms/Utils/Utils.cpp

@@ -39,6 +39,7 @@ void llvm::initializeTransformUtils(PassRegistry &Registry) {
   initializeMetaRenamerPass(Registry);
   initializeStripGCRelocatesPass(Registry);
   initializePredicateInfoPrinterLegacyPassPass(Registry);
+  initializeInjectTLIMappingsLegacyPass(Registry);
 }
 
 /// LLVMInitializeTransformUtils - C binding for initializeTransformUtilsPasses.

test/Transforms/Util/add-TLI-mappings.ll

@@ -0,0 +1,61 @@
+; RUN: opt -vector-library=SVML       -inject-tli-mappings        -S < %s | FileCheck %s  --check-prefixes=COMMON,SVML
+; RUN: opt -vector-library=SVML       -passes=inject-tli-mappings -S < %s | FileCheck %s  --check-prefixes=COMMON,SVML
+; RUN: opt -vector-library=MASSV      -inject-tli-mappings        -S < %s | FileCheck %s  --check-prefixes=COMMON,MASSV
+; RUN: opt -vector-library=MASSV      -passes=inject-tli-mappings -S < %s | FileCheck %s  --check-prefixes=COMMON,MASSV
+; RUN: opt -vector-library=Accelerate -inject-tli-mappings        -S < %s | FileCheck %s  --check-prefixes=COMMON,ACCELERATE
+; RUN: opt -vector-library=Accelerate -passes=inject-tli-mappings -S < %s | FileCheck %s  --check-prefixes=COMMON,ACCELERATE
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+; COMMON-LABEL: @llvm.compiler.used = appending global
+; SVML-SAME:        [3 x i8*] [
+; SVML-SAME:          i8* bitcast (<2 x double> (<2 x double>)* @__svml_sin2 to i8*),
+; SVML-SAME:          i8* bitcast (<4 x double> (<4 x double>)* @__svml_sin4 to i8*),
+; SVML-SAME:          i8* bitcast (<8 x double> (<8 x double>)* @__svml_sin8 to i8*)
+; MASSV-SAME:       [2 x i8*] [
+; MASSV-SAME:         i8* bitcast (<2 x double> (<2 x double>)* @__sind2_massv to i8*),
+; MASSV-SAME:         i8* bitcast (<4 x float> (<4 x float>)* @__log10f4_massv to i8*)
+; ACCELERATE-SAME:  [1 x i8*] [
+; ACCELERATE-SAME:    i8* bitcast (<4 x float> (<4 x float>)* @vlog10f to i8*)
+; COMMON-SAME:      ], section "llvm.metadata"
+
+define double @sin_f64(double %in) {
+; COMMON-LABEL: @sin_f64(
+; SVML:         call double @sin(double %{{.*}}) #[[SIN:[0-9]+]]
+; MASSV:        call double @sin(double %{{.*}}) #[[SIN:[0-9]+]]
+; ACCELERATE:   call double @sin(double %{{.*}})
+; No mapping of "sin" to a vector function for Accelerate.
+; ACCELERATE-NOT: _ZGV_LLVM_{{.*}}_sin({{.*}})
+  %call = tail call double @sin(double %in)
+  ret double %call
+}
+
+declare double @sin(double) #0
+
+define float @call_llvm.log10.f32(float %in) {
+; COMMON-LABEL: @call_llvm.log10.f32(
+; SVML:         call float @llvm.log10.f32(float %{{.*}})
+; MASSV:        call float @llvm.log10.f32(float %{{.*}}) #[[LOG10:[0-9]+]]
+; ACCELERATE:   call float @llvm.log10.f32(float %{{.*}}) #[[LOG10:[0-9]+]]
+; No mapping of "llvm.log10.f32" to a vector function for SVML.
+; SVML-NOT:     _ZGV_LLVM_{{.*}}_llvm.log10.f32({{.*}})
+  %call = tail call float @llvm.log10.f32(float %in)
+  ret float %call
+}
+
+declare float @llvm.log10.f32(float) #0
+attributes #0 = { nounwind readnone }
+
+; SVML:      attributes #[[SIN]] = { "vector-function-abi-variant"=
+; SVML-SAME:   "_ZGV_LLVM_N2v_sin(__svml_sin2),
+; SVML-SAME:   _ZGV_LLVM_N4v_sin(__svml_sin4),
+; SVML-SAME:   _ZGV_LLVM_N8v_sin(__svml_sin8)" }
+
+; MASSV:      attributes #[[SIN]] = { "vector-function-abi-variant"=
+; MASSV-SAME:   "_ZGV_LLVM_N2v_sin(__sind2_massv)" }
+; MASSV:      attributes #[[LOG10]] = { "vector-function-abi-variant"=
+; MASSV-SAME:   "_ZGV_LLVM_N4v_llvm.log10.f32(__log10f4_massv)" }
+
+; ACCELERATE:      attributes #[[LOG10]] = { "vector-function-abi-variant"=
+; ACCELERATE-SAME:   "_ZGV_LLVM_N4v_llvm.log10.f32(vlog10f)" }