[PowerPC] Scalar IBM MASS library conversion pass

This patch introduces the conversions from math function calls
to MASS library calls. To resolves calls generated with these conversions, one
need to link libxlopt.a library. This patch is tested on PowerPC Linux and AIX.

Differential: https://reviews.llvm.org/D101759

Reviewer: bmahjour

Author: msdataei

llvm/include/llvm/Analysis/ScalarFuncs.def

@@ -0,0 +1,117 @@
+//===-- ScalarFuncs.def - Library information ----------*- C++ -*----------===//
+//
+// 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
+//
+//===----------------------------------------------------------------------===//
+
+// This .def file creates mapping from standard IEEE math functions
+// their corresponding entries in the IBM MASS (scalar) library.
+// LLVM intrinsic math functions will be handled in PPCISelLowing to
+// allow existing optimizations like pow(x,0.5) --> sqrt(x).
+
+#if defined(TLI_DEFINE_SCALAR_MASS_FUNCS)
+#define TLI_DEFINE_SCALAR_MASS_FUNC(SCAL, MASSENTRY) {SCAL, MASSENTRY},
+#endif
+
+TLI_DEFINE_SCALAR_MASS_FUNC("acosf", "__xl_acosf")
+TLI_DEFINE_SCALAR_MASS_FUNC("__acosf_finite", "__xl_acosf")
+TLI_DEFINE_SCALAR_MASS_FUNC("acos", "__xl_acos")
+TLI_DEFINE_SCALAR_MASS_FUNC("__acos_finite", "__xl_acos")
+
+TLI_DEFINE_SCALAR_MASS_FUNC("acoshf", "__xl_acoshf")
+TLI_DEFINE_SCALAR_MASS_FUNC("__acoshf_finite", "__xl_acoshf")
+TLI_DEFINE_SCALAR_MASS_FUNC("acosh", "__xl_acosh")
+TLI_DEFINE_SCALAR_MASS_FUNC("__acosh_finite", "__xl_acosh")
+
+TLI_DEFINE_SCALAR_MASS_FUNC("asinf", "__xl_asinf")
+TLI_DEFINE_SCALAR_MASS_FUNC("__asinf_finite", "__xl_asinf")
+TLI_DEFINE_SCALAR_MASS_FUNC("asin", "__xl_asin")
+TLI_DEFINE_SCALAR_MASS_FUNC("__asin_finite", "__xl_asin")
+
+TLI_DEFINE_SCALAR_MASS_FUNC("asinhf", "__xl_asinhf")
+TLI_DEFINE_SCALAR_MASS_FUNC("asinh", "__xl_asinh")
+
+TLI_DEFINE_SCALAR_MASS_FUNC("atanf", "__xl_atanf")
+TLI_DEFINE_SCALAR_MASS_FUNC("atan", "__xl_atan")
+
+TLI_DEFINE_SCALAR_MASS_FUNC("atan2f", "__xl_atan2f")
+TLI_DEFINE_SCALAR_MASS_FUNC("__atan2f_finite", "__xl_atan2f")
+TLI_DEFINE_SCALAR_MASS_FUNC("atan2", "__xl_atan2")
+TLI_DEFINE_SCALAR_MASS_FUNC("__atan2_finite", "__xl_atan2")
+
+TLI_DEFINE_SCALAR_MASS_FUNC("atanhf", "__xl_atanhf")
+TLI_DEFINE_SCALAR_MASS_FUNC("__atanhf_finite", "__xl_atanhf")
+TLI_DEFINE_SCALAR_MASS_FUNC("atanh", "__xl_atanh")
+TLI_DEFINE_SCALAR_MASS_FUNC("__atanh_finite", "__xl_atanh")
+
+TLI_DEFINE_SCALAR_MASS_FUNC("cbrtf", "__xl_cbrtf")
+TLI_DEFINE_SCALAR_MASS_FUNC("cbrt", "__xl_cbrt")
+
+TLI_DEFINE_SCALAR_MASS_FUNC("cosf", "__xl_cosf")
+TLI_DEFINE_SCALAR_MASS_FUNC("cos", "__xl_cos")
+
+TLI_DEFINE_SCALAR_MASS_FUNC("coshf", "__xl_coshf")
+TLI_DEFINE_SCALAR_MASS_FUNC("__coshf_finite", "__xl_coshf")
+TLI_DEFINE_SCALAR_MASS_FUNC("cosh", "__xl_cosh")
+TLI_DEFINE_SCALAR_MASS_FUNC("__cosh_finite", "__xl_cosh")
+
+TLI_DEFINE_SCALAR_MASS_FUNC("erff", "__xl_erff")
+TLI_DEFINE_SCALAR_MASS_FUNC("erf", "__xl_erf")
+
+TLI_DEFINE_SCALAR_MASS_FUNC("erfcf", "__xl_erfcf")
+TLI_DEFINE_SCALAR_MASS_FUNC("erfc", "__xl_erfc")
+
+TLI_DEFINE_SCALAR_MASS_FUNC("expf", "__xl_expf")
+TLI_DEFINE_SCALAR_MASS_FUNC("__expf_finite", "__xl_expf")
+TLI_DEFINE_SCALAR_MASS_FUNC("exp", "__xl_exp")
+TLI_DEFINE_SCALAR_MASS_FUNC("__exp_finite", "__xl_exp")
+
+TLI_DEFINE_SCALAR_MASS_FUNC("expm1f", "__xl_expm1f")
+TLI_DEFINE_SCALAR_MASS_FUNC("expm1", "__xl_expm1")
+
+TLI_DEFINE_SCALAR_MASS_FUNC("hypotf", "__xl_hypotf")
+TLI_DEFINE_SCALAR_MASS_FUNC("hypot", "__xl_hypot")
+
+TLI_DEFINE_SCALAR_MASS_FUNC("lgammaf", "__xl_lgammaf")
+TLI_DEFINE_SCALAR_MASS_FUNC("lgamma", "__xl_lgamma")
+
+TLI_DEFINE_SCALAR_MASS_FUNC("logf", "__xl_logf")
+TLI_DEFINE_SCALAR_MASS_FUNC("__logf_finite", "__xl_logf")
+TLI_DEFINE_SCALAR_MASS_FUNC("log", "__xl_log")
+TLI_DEFINE_SCALAR_MASS_FUNC("__log_finite", "__xl_log")
+
+TLI_DEFINE_SCALAR_MASS_FUNC("log10f", "__xl_log10f")
+TLI_DEFINE_SCALAR_MASS_FUNC("__log10f_finite", "__xl_log10f")
+TLI_DEFINE_SCALAR_MASS_FUNC("log10", "__xl_log10")
+TLI_DEFINE_SCALAR_MASS_FUNC("__log10_finite", "__xl_log10")
+
+TLI_DEFINE_SCALAR_MASS_FUNC("log1pf", "__xl_log1pf")
+TLI_DEFINE_SCALAR_MASS_FUNC("log1p", "__xl_log1p")
+
+TLI_DEFINE_SCALAR_MASS_FUNC("powf", "__xl_powf")
+TLI_DEFINE_SCALAR_MASS_FUNC("__powf_finite", "__xl_powf")
+TLI_DEFINE_SCALAR_MASS_FUNC("pow", "__xl_pow")
+TLI_DEFINE_SCALAR_MASS_FUNC("__pow_finite", "__xl_pow")
+
+TLI_DEFINE_SCALAR_MASS_FUNC("rsqrt", "__xl_rsqrt")
+
+TLI_DEFINE_SCALAR_MASS_FUNC("sinf", "__xl_sinf")
+TLI_DEFINE_SCALAR_MASS_FUNC("sin", "__xl_sin")
+
+TLI_DEFINE_SCALAR_MASS_FUNC("sinhf", "__xl_sinhf")
+TLI_DEFINE_SCALAR_MASS_FUNC("__sinhf_finite", "__xl_sinhf")
+TLI_DEFINE_SCALAR_MASS_FUNC("sinh", "__xl_sinh")
+TLI_DEFINE_SCALAR_MASS_FUNC("__sinh_finite", "__xl_sinh")
+
+TLI_DEFINE_SCALAR_MASS_FUNC("sqrt", "__xl_sqrt")
+
+TLI_DEFINE_SCALAR_MASS_FUNC("tanf", "__xl_tanf")
+TLI_DEFINE_SCALAR_MASS_FUNC("tan", "__xl_tan")
+
+TLI_DEFINE_SCALAR_MASS_FUNC("tanhf", "__xl_tanhf")
+TLI_DEFINE_SCALAR_MASS_FUNC("tanh", "__xl_tanh")
+
+#undef TLI_DEFINE_SCALAR_MASS_FUNCS
+#undef TLI_DEFINE_SCALAR_MASS_FUNC

llvm/include/llvm/CodeGen/CommandFlags.h

@@ -62,6 +62,8 @@ bool getEnableNoNaNsFPMath();
 
 bool getEnableNoSignedZerosFPMath();
 
+bool getEnableApproxFuncFPMath();
+
 bool getEnableNoTrappingFPMath();
 
 DenormalMode::DenormalModeKind getDenormalFPMath();

llvm/include/llvm/IR/Attributes.td

@@ -294,6 +294,7 @@ def MustProgress : EnumAttr<"mustprogress", [FnAttr]>;
 def LessPreciseFPMAD : StrBoolAttr<"less-precise-fpmad">;
 def NoInfsFPMath : StrBoolAttr<"no-infs-fp-math">;
 def NoNansFPMath : StrBoolAttr<"no-nans-fp-math">;
+def ApproxFuncFPMath : StrBoolAttr<"approx-func-fp-math">;
 def NoSignedZerosFPMath : StrBoolAttr<"no-signed-zeros-fp-math">;
 def UnsafeFPMath : StrBoolAttr<"unsafe-fp-math">;
 def NoJumpTables : StrBoolAttr<"no-jump-tables">;
@@ -333,6 +334,7 @@ class MergeRule<string F> {
 def : MergeRule<"setAND<LessPreciseFPMADAttr>">;
 def : MergeRule<"setAND<NoInfsFPMathAttr>">;
 def : MergeRule<"setAND<NoNansFPMathAttr>">;
+def : MergeRule<"setAND<ApproxFuncFPMathAttr>">;
 def : MergeRule<"setAND<NoSignedZerosFPMathAttr>">;
 def : MergeRule<"setAND<UnsafeFPMathAttr>">;
 def : MergeRule<"setOR<NoImplicitFloatAttr>">;

llvm/lib/CodeGen/CommandFlags.cpp

@@ -58,6 +58,7 @@ CGOPT(bool, EnableUnsafeFPMath)
 CGOPT(bool, EnableNoInfsFPMath)
 CGOPT(bool, EnableNoNaNsFPMath)
 CGOPT(bool, EnableNoSignedZerosFPMath)
+CGOPT(bool, EnableApproxFuncFPMath)
 CGOPT(bool, EnableNoTrappingFPMath)
 CGOPT(bool, EnableAIXExtendedAltivecABI)
 CGOPT(DenormalMode::DenormalModeKind, DenormalFPMath)
@@ -218,6 +219,12 @@ codegen::RegisterCodeGenFlags::RegisterCodeGenFlags() {
       cl::init(false));
   CGBINDOPT(EnableNoSignedZerosFPMath);
 
+  static cl::opt<bool> EnableApproxFuncFPMath(
+      "enable-approx-func-fp-math",
+      cl::desc("Enable FP math optimizations that assume approx func"),
+      cl::init(false));
+  CGBINDOPT(EnableApproxFuncFPMath);
+
   static cl::opt<bool> EnableNoTrappingFPMath(
       "enable-no-trapping-fp-math",
       cl::desc("Enable setting the FP exceptions build "
@@ -493,6 +500,7 @@ codegen::InitTargetOptionsFromCodeGenFlags(const Triple &TheTriple) {
   Options.NoInfsFPMath = getEnableNoInfsFPMath();
   Options.NoNaNsFPMath = getEnableNoNaNsFPMath();
   Options.NoSignedZerosFPMath = getEnableNoSignedZerosFPMath();
+  Options.ApproxFuncFPMath = getEnableApproxFuncFPMath();
   Options.NoTrappingFPMath = getEnableNoTrappingFPMath();
 
   DenormalMode::DenormalModeKind DenormKind = getDenormalFPMath();
@@ -643,6 +651,7 @@ void codegen::setFunctionAttributes(StringRef CPU, StringRef Features,
   HANDLE_BOOL_ATTR(EnableNoInfsFPMathView, "no-infs-fp-math");
   HANDLE_BOOL_ATTR(EnableNoNaNsFPMathView, "no-nans-fp-math");
   HANDLE_BOOL_ATTR(EnableNoSignedZerosFPMathView, "no-signed-zeros-fp-math");
+  HANDLE_BOOL_ATTR(EnableApproxFuncFPMathView, "approx-func-fp-math");
 
   if (DenormalFPMathView->getNumOccurrences() > 0 &&
       !F.hasFnAttribute("denormal-fp-math")) {

llvm/lib/Target/PowerPC/CMakeLists.txt

@@ -55,6 +55,7 @@ add_llvm_target(PowerPCCodeGen
   PPCExpandISEL.cpp
   PPCPreEmitPeephole.cpp
   PPCLowerMASSVEntries.cpp
+  PPCGenScalarMASSEntries.cpp
   GISel/PPCCallLowering.cpp
   GISel/PPCRegisterBankInfo.cpp
   GISel/PPCLegalizerInfo.cpp

llvm/lib/Target/PowerPC/PPC.h

@@ -84,6 +84,10 @@ FunctionPass *createPPCCTRLoops();
   void initializePPCLowerMASSVEntriesPass(PassRegistry &);
   extern char &PPCLowerMASSVEntriesID;
 
+  ModulePass *createPPCGenScalarMASSEntriesPass();
+  void initializePPCGenScalarMASSEntriesPass(PassRegistry &);
+  extern char &PPCGenScalarMASSEntriesID;
+
   InstructionSelector *
   createPPCInstructionSelector(const PPCTargetMachine &, const PPCSubtarget &,
                                const PPCRegisterBankInfo &);

llvm/lib/Target/PowerPC/PPCGenScalarMASSEntries.cpp

@@ -0,0 +1,141 @@
+//===-- PPCGenScalarMASSEntries.cpp ---------------------------------------===//
+//
+// 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
+//
+//===----------------------------------------------------------------------===//
+//
+// This transformation converts standard math functions into their
+// corresponding MASS (scalar) entries for PowerPC targets.
+// Following are examples of such conversion:
+//     tanh ---> __xl_tanh_finite
+// Such lowering is legal under the fast-math option.
+//
+//===----------------------------------------------------------------------===//
+
+#include "PPC.h"
+#include "PPCSubtarget.h"
+#include "PPCTargetMachine.h"
+#include "llvm/Analysis/TargetTransformInfo.h"
+#include "llvm/CodeGen/TargetPassConfig.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/IR/Module.h"
+
+#define DEBUG_TYPE "ppc-gen-scalar-mass"
+
+using namespace llvm;
+
+namespace {
+
+class PPCGenScalarMASSEntries : public ModulePass {
+public:
+  static char ID;
+
+  PPCGenScalarMASSEntries() : ModulePass(ID) {
+    ScalarMASSFuncs = {
+#define TLI_DEFINE_SCALAR_MASS_FUNCS
+#include "llvm/Analysis/ScalarFuncs.def"
+    };
+  }
+
+  bool runOnModule(Module &M) override;
+
+  StringRef getPassName() const override {
+    return "PPC Generate Scalar MASS Entries";
+  }
+
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    AU.addRequired<TargetTransformInfoWrapperPass>();
+  }
+
+private:
+  std::map<StringRef, StringRef> ScalarMASSFuncs;
+  bool isCandidateSafeToLower(const CallInst &CI) const;
+  bool isFiniteCallSafe(const CallInst &CI) const;
+  bool createScalarMASSCall(StringRef MASSEntry, CallInst &CI,
+                            Function &Func) const;
+};
+
+} // namespace
+
+// Returns true if 'afn' flag exists on the call instruction with the math
+// function
+bool PPCGenScalarMASSEntries::isCandidateSafeToLower(const CallInst &CI) const {
+  return CI.hasApproxFunc();
+}
+
+// Returns true if 'nnan', 'ninf' and 'nsz' flags exist on the call instruction
+// with the math function
+bool PPCGenScalarMASSEntries::isFiniteCallSafe(const CallInst &CI) const {
+  // FIXME: no-errno and trapping-math need to be set for MASS converstion
+  // but they don't have IR representation.
+  return CI.hasNoNaNs() && CI.hasNoInfs() && CI.hasNoSignedZeros();
+}
+
+/// Lowers scalar math functions to scalar MASS functions.
+///     e.g.: tanh         --> __xl_tanh_finite or __xl_tanh
+/// Both function prototype and its callsite is updated during lowering.
+bool PPCGenScalarMASSEntries::createScalarMASSCall(StringRef MASSEntry,
+                                                   CallInst &CI,
+                                                   Function &Func) const {
+  if (CI.use_empty())
+    return false;
+
+  Module *M = Func.getParent();
+  assert(M && "Expecting a valid Module");
+
+  std::string MASSEntryStr = MASSEntry.str();
+  if (isFiniteCallSafe(CI))
+    MASSEntryStr += "_finite";
+
+  FunctionCallee FCache = M->getOrInsertFunction(
+      MASSEntryStr, Func.getFunctionType(), Func.getAttributes());
+
+  CI.setCalledFunction(FCache);
+
+  return true;
+}
+
+bool PPCGenScalarMASSEntries::runOnModule(Module &M) {
+  bool Changed = false;
+
+  auto *TPC = getAnalysisIfAvailable<TargetPassConfig>();
+  if (!TPC || skipModule(M))
+    return false;
+
+  for (Function &Func : M) {
+    if (!Func.isDeclaration())
+      continue;
+
+    auto Iter = ScalarMASSFuncs.find(Func.getName());
+    if (Iter == ScalarMASSFuncs.end())
+      continue;
+
+    // The call to createScalarMASSCall() invalidates the iterator over users
+    // upon replacing the users. Precomputing the current list of users allows
+    // us to replace all the call sites.
+    SmallVector<User *, 4> TheUsers;
+    for (auto *User : Func.users())
+      TheUsers.push_back(User);
+
+    for (auto *User : TheUsers)
+      if (auto *CI = dyn_cast_or_null<CallInst>(User)) {
+        if (isCandidateSafeToLower(*CI))
+          Changed |= createScalarMASSCall(Iter->second, *CI, Func);
+      }
+  }
+
+  return Changed;
+}
+
+char PPCGenScalarMASSEntries::ID = 0;
+
+char &llvm::PPCGenScalarMASSEntriesID = PPCGenScalarMASSEntries::ID;
+
+INITIALIZE_PASS(PPCGenScalarMASSEntries, DEBUG_TYPE,
+                "Generate Scalar MASS entries", false, false)
+
+ModulePass *llvm::createPPCGenScalarMASSEntriesPass() {
+  return new PPCGenScalarMASSEntries();
+}