Convert calls to __sinpi and __cospi into __sincospi_stret

This adds an SimplifyLibCalls case which converts the special __sinpi and
__cospi (float & double variants) into a __sincospi_stret where appropriate to
remove duplicated work.

Patch by Tim Northover

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@193943 91177308-0d34-0410-b5e6-96231b3b80d8

include/llvm/Target/TargetLibraryInfo.h

@@ -46,6 +46,10 @@ namespace llvm {
       Znwm,
       /// void *new(unsigned long, nothrow);
       ZnwmRKSt9nothrow_t,
+      /// double __cospi(double x);
+      cospi,
+      /// float __cospif(float x);
+      cospif,
       /// int __cxa_atexit(void (*f)(void *), void *p, void *d);
       cxa_atexit,
       /// void __cxa_guard_abort(guard_t *guard);
@@ -61,6 +65,14 @@ namespace llvm {
       dunder_isoc99_sscanf,
       /// void *__memcpy_chk(void *s1, const void *s2, size_t n, size_t s1size);
       memcpy_chk,
+      /// double __sincospi_stret(double x);
+      sincospi_stret,
+      /// float __sincospi_stretf(float x);
+      sincospi_stretf,
+      /// double __sinpi(double x);
+      sinpi,
+      /// float __sinpif(float x);
+      sinpif,
       /// double __sqrt_finite(double x);
       sqrt_finite,
       /// float __sqrt_finite(float x);

lib/Target/TargetLibraryInfo.cpp

@@ -38,6 +38,8 @@ const char* TargetLibraryInfo::StandardNames[LibFunc::NumLibFuncs] =
     "_ZnwjRKSt9nothrow_t",
     "_Znwm",
     "_ZnwmRKSt9nothrow_t",
+    "__cospi",
+    "__cospif",
     "__cxa_atexit",
     "__cxa_guard_abort",
     "__cxa_guard_acquire",
@@ -45,6 +47,10 @@ const char* TargetLibraryInfo::StandardNames[LibFunc::NumLibFuncs] =
     "__isoc99_scanf",
     "__isoc99_sscanf",
     "__memcpy_chk",
+    "__sincospi_stret",
+    "__sincospi_stretf",
+    "__sinpi",
+    "__sinpif",
     "__sqrt_finite",
     "__sqrtf_finite",
     "__sqrtl_finite",
@@ -331,6 +337,24 @@ const char* TargetLibraryInfo::StandardNames[LibFunc::NumLibFuncs] =
     "write"
   };
 
+static bool hasSinCosPiStret(const Triple &T) {
+  // Only Darwin variants have _stret versions of combined trig functions.
+  if (!T.isMacOSX() && T.getOS() != Triple::IOS)
+    return false;
+
+  // The ABI is rather complicated on x86, so don't do anything special there.
+  if (T.getArch() == Triple::x86)
+    return false;
+
+  if (T.isMacOSX() && T.isMacOSXVersionLT(10, 9))
+    return false;
+
+  if (T.getOS() == Triple::IOS && T.isOSVersionLT(7, 0))
+    return false;
+
+  return true;
+}
+
 /// initialize - Initialize the set of available library functions based on the
 /// specified target triple.  This should be carefully written so that a missing
 /// target triple gets a sane set of defaults.
@@ -357,6 +381,15 @@ static void initialize(TargetLibraryInfo &TLI, const Triple &T,
     TLI.setUnavailable(LibFunc::memset_pattern16);
   }
 
+  if (!hasSinCosPiStret(T)) {
+    TLI.setUnavailable(LibFunc::sinpi);
+    TLI.setUnavailable(LibFunc::sinpif);
+    TLI.setUnavailable(LibFunc::cospi);
+    TLI.setUnavailable(LibFunc::cospif);
+    TLI.setUnavailable(LibFunc::sincospi_stret);
+    TLI.setUnavailable(LibFunc::sincospi_stretf);
+  }
+
   if (T.isMacOSX() && T.getArch() == Triple::x86 &&
       !T.isMacOSXVersionLT(10, 7)) {
     // x86-32 OSX has a scheme where fwrite and fputs (and some other functions

lib/Transforms/Utils/SimplifyLibCalls.cpp

@@ -17,6 +17,7 @@
 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
 #include "llvm/ADT/SmallString.h"
 #include "llvm/ADT/StringMap.h"
+#include "llvm/ADT/Triple.h"
 #include "llvm/Analysis/ValueTracking.h"
 #include "llvm/IR/DataLayout.h"
 #include "llvm/IR/Function.h"
@@ -1252,6 +1253,155 @@ struct Exp2Opt : public UnsafeFPLibCallOptimization {
   }
 };
 
+struct SinCosPiOpt : public LibCallOptimization {
+  SinCosPiOpt() {}
+
+  virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) {
+    // Make sure the prototype is as expected, otherwise the rest of the
+    // function is probably invalid and likely to abort.
+    if (!isTrigLibCall(CI))
+      return 0;
+
+    Value *Arg = CI->getArgOperand(0);
+    SmallVector<CallInst *, 1> SinCalls;
+    SmallVector<CallInst *, 1> CosCalls;
+    SmallVector<CallInst *, 1> SinCosCalls;
+
+    bool IsFloat = Arg->getType()->isFloatTy();
+
+    // Look for all compatible sinpi, cospi and sincospi calls with the same
+    // argument. If there are enough (in some sense) we can make the
+    // substitution.
+    for (Value::use_iterator UI = Arg->use_begin(), UE = Arg->use_end();
+         UI != UE; ++UI)
+      classifyArgUse(*UI, CI->getParent(), IsFloat, SinCalls, CosCalls,
+                     SinCosCalls);
+
+    // It's only worthwhile if both sinpi and cospi are actually used.
+    if (SinCosCalls.empty() && (SinCalls.empty() || CosCalls.empty()))
+      return 0;
+
+    Value *Sin, *Cos, *SinCos;
+    insertSinCosCall(B, CI->getCalledFunction(), Arg, IsFloat, Sin, Cos,
+                     SinCos);
+
+    replaceTrigInsts(SinCalls, Sin);
+    replaceTrigInsts(CosCalls, Cos);
+    replaceTrigInsts(SinCosCalls, SinCos);
+
+    return 0;
+  }
+
+  bool isTrigLibCall(CallInst *CI) {
+    Function *Callee = CI->getCalledFunction();
+    FunctionType *FT = Callee->getFunctionType();
+
+    // We can only hope to do anything useful if we can ignore things like errno
+    // and floating-point exceptions.
+    bool AttributesSafe = CI->hasFnAttr(Attribute::NoUnwind) &&
+                          CI->hasFnAttr(Attribute::ReadNone);
+
+    // Other than that we need float(float) or double(double)
+    return AttributesSafe && FT->getNumParams() == 1 &&
+           FT->getReturnType() == FT->getParamType(0) &&
+           (FT->getParamType(0)->isFloatTy() ||
+            FT->getParamType(0)->isDoubleTy());
+  }
+
+  void classifyArgUse(Value *Val, BasicBlock *BB, bool IsFloat,
+                      SmallVectorImpl<CallInst *> &SinCalls,
+                      SmallVectorImpl<CallInst *> &CosCalls,
+                      SmallVectorImpl<CallInst *> &SinCosCalls) {
+    CallInst *CI = dyn_cast<CallInst>(Val);
+
+    if (!CI)
+      return;
+
+    Function *Callee = CI->getCalledFunction();
+    StringRef FuncName = Callee->getName();
+    LibFunc::Func Func;
+    if (!TLI->getLibFunc(FuncName, Func) || !TLI->has(Func) ||
+        !isTrigLibCall(CI))
+      return;
+
+    if (IsFloat) {
+      if (Func == LibFunc::sinpif)
+        SinCalls.push_back(CI);
+      else if (Func == LibFunc::cospif)
+        CosCalls.push_back(CI);
+      else if (Func == LibFunc::sincospi_stretf)
+        SinCosCalls.push_back(CI);
+    } else {
+      if (Func == LibFunc::sinpi)
+        SinCalls.push_back(CI);
+      else if (Func == LibFunc::cospi)
+        CosCalls.push_back(CI);
+      else if (Func == LibFunc::sincospi_stret)
+        SinCosCalls.push_back(CI);
+    }
+  }
+
+  void replaceTrigInsts(SmallVectorImpl<CallInst*> &Calls, Value *Res) {
+    for (SmallVectorImpl<CallInst*>::iterator I = Calls.begin(),
+           E = Calls.end();
+         I != E; ++I) {
+      LCS->replaceAllUsesWith(*I, Res);
+    }
+  }
+
+  void insertSinCosCall(IRBuilder<> &B, Function *OrigCallee, Value *Arg,
+                        bool UseFloat, Value *&Sin, Value *&Cos,
+                        Value *&SinCos) {
+    Type *ArgTy = Arg->getType();
+    Type *ResTy;
+    StringRef Name;
+
+    Triple T(OrigCallee->getParent()->getTargetTriple());
+    if (UseFloat) {
+      Name = "__sincospi_stretf";
+
+      assert(T.getArch() != Triple::x86 && "x86 messy and unsupported for now");
+      // x86_64 can't use {float, float} since that would be returned in both
+      // xmm0 and xmm1, which isn't what a real struct would do.
+      ResTy = T.getArch() == Triple::x86_64
+                  ? static_cast<Type *>(VectorType::get(ArgTy, 2))
+                  : static_cast<Type *>(StructType::get(ArgTy, ArgTy, NULL));
+    } else {
+      Name = "__sincospi_stret";
+      ResTy = StructType::get(ArgTy, ArgTy, NULL);
+    }
+
+    Module *M = OrigCallee->getParent();
+    Value *Callee = M->getOrInsertFunction(Name, OrigCallee->getAttributes(),
+                                           ResTy, ArgTy, NULL);
+
+    if (Instruction *ArgInst = dyn_cast<Instruction>(Arg)) {
+      // If the argument is an instruction, it must dominate all uses so put our
+      // sincos call there.
+      BasicBlock::iterator Loc = ArgInst;
+      B.SetInsertPoint(ArgInst->getParent(), ++Loc);
+    } else {
+      // Otherwise (e.g. for a constant) the beginning of the function is as
+      // good a place as any.
+      BasicBlock &EntryBB = B.GetInsertBlock()->getParent()->getEntryBlock();
+      B.SetInsertPoint(&EntryBB, EntryBB.begin());
+    }
+
+    SinCos = B.CreateCall(Callee, Arg, "sincospi");
+
+    if (SinCos->getType()->isStructTy()) {
+      Sin = B.CreateExtractValue(SinCos, 0, "sinpi");
+      Cos = B.CreateExtractValue(SinCos, 1, "cospi");
+    } else {
+      Sin = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 0),
+                                   "sinpi");
+      Cos = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 1),
+                                   "cospi");
+    }
+  }
+
+};
+
 //===----------------------------------------------------------------------===//
 // Integer Library Call Optimizations
 //===----------------------------------------------------------------------===//
@@ -1764,6 +1914,7 @@ static MemSetOpt MemSet;
 // Math library call optimizations.
 static UnaryDoubleFPOpt UnaryDoubleFP(false);
 static UnaryDoubleFPOpt UnsafeUnaryDoubleFP(true);
+static SinCosPiOpt SinCosPi;
 
   // Integer library call optimizations.
 static FFSOpt FFS;
@@ -1848,6 +1999,11 @@ LibCallOptimization *LibCallSimplifierImpl::lookupOptimization(CallInst *CI) {
       case LibFunc::cos:
       case LibFunc::cosl:
         return &Cos;
+      case LibFunc::sinpif:
+      case LibFunc::sinpi:
+      case LibFunc::cospif:
+      case LibFunc::cospi:
+        return &SinCosPi;
       case LibFunc::powf:
       case LibFunc::pow:
       case LibFunc::powl:

test/Transforms/InstCombine/sincospi.ll

@@ -0,0 +1,91 @@
+; RUN: opt -instcombine -S < %s -mtriple=x86_64-apple-macosx10.9 | FileCheck %s --check-prefix=CHECK-FLOAT-IN-VEC
+; RUN: opt -instcombine -S < %s -mtriple=arm-apple-ios7.0 | FileCheck %s
+; RUN: opt -instcombine -S < %s -mtriple=x86_64-apple-macosx10.8 | FileCheck %s --check-prefix=CHECK-NO-SINCOS
+; RUN: opt -instcombine -S < %s -mtriple=arm-apple-ios6.0 | FileCheck %s --check-prefix=CHECK-NO-SINCOS
+; RUN: opt -instcombine -S < %s -mtriple=x86_64-none-linux-gnu | FileCheck %s --check-prefix=CHECK-NO-SINCOS
+
+
+attributes #0 = { readnone nounwind }
+
+declare float @__sinpif(float %x) #0
+declare float @__cospif(float %x) #0 
+
+declare double @__sinpi(double %x) #0
+declare double @__cospi(double %x) #0 
+
+@var32 = global float 0.0
+@var64 = global double 0.0
+
+define float @test_instbased_f32() {
+       %val = load float* @var32
+       %sin = call float @__sinpif(float %val) #0
+       %cos = call float @__cospif(float %val) #0
+       %res = fadd float %sin, %cos
+       ret float %res
+; CHECK-FLOAT-IN-VEC: [[VAL:%[a-z0-9]+]] = load float* @var32
+; CHECK-FLOAT-IN-VEC: [[SINCOS:%[a-z0-9]+]] = call <2 x float> @__sincospi_stretf(float [[VAL]])
+; CHECK-FLOAT-IN-VEC: extractelement <2 x float> [[SINCOS]], i32 0
+; CHECK-FLOAT-IN-VEC: extractelement <2 x float> [[SINCOS]], i32 1
+
+; CHECK: [[VAL:%[a-z0-9]+]] = load float* @var32
+; CHECK: [[SINCOS:%[a-z0-9]+]] = call { float, float } @__sincospi_stretf(float [[VAL]])
+; CHECK: extractvalue { float, float } [[SINCOS]], 0
+; CHECK: extractvalue { float, float } [[SINCOS]], 1
+
+; CHECK-NO-SINCOS: call float @__sinpif
+; CHECK-NO-SINCOS: call float @__cospif
+}
+
+define float @test_constant_f32() {
+       %sin = call float @__sinpif(float 1.0) #0
+       %cos = call float @__cospif(float 1.0) #0
+       %res = fadd float %sin, %cos
+       ret float %res
+; CHECK-FLOAT-IN-VEC: [[SINCOS:%[a-z0-9]+]] = call <2 x float> @__sincospi_stretf(float 1.000000e+00)
+; CHECK-FLOAT-IN-VEC: extractelement <2 x float> [[SINCOS]], i32 0
+; CHECK-FLOAT-IN-VEC: extractelement <2 x float> [[SINCOS]], i32 1
+
+; CHECK: [[SINCOS:%[a-z0-9]+]] = call { float, float } @__sincospi_stretf(float 1.000000e+00)
+; CHECK: extractvalue { float, float } [[SINCOS]], 0
+; CHECK: extractvalue { float, float } [[SINCOS]], 1
+
+; CHECK-NO-SINCOS: call float @__sinpif
+; CHECK-NO-SINCOS: call float @__cospif
+}
+
+define double @test_instbased_f64() {
+       %val = load double* @var64
+       %sin = call double @__sinpi(double %val) #0
+       %cos = call double @__cospi(double %val) #0
+       %res = fadd double %sin, %cos
+       ret double %res
+; CHECK-FLOAT-IN-VEC: [[VAL:%[a-z0-9]+]] = load double* @var64
+; CHECK-FLOAT-IN-VEC: [[SINCOS:%[a-z0-9]+]] = call { double, double } @__sincospi_stret(double [[VAL]])
+; CHECK-FLOAT-IN-VEC: extractvalue { double, double } [[SINCOS]], 0
+; CHECK-FLOAT-IN-VEC: extractvalue { double, double } [[SINCOS]], 1
+
+; CHECK: [[VAL:%[a-z0-9]+]] = load double* @var64
+; CHECK: [[SINCOS:%[a-z0-9]+]] = call { double, double } @__sincospi_stret(double [[VAL]])
+; CHECK: extractvalue { double, double } [[SINCOS]], 0
+; CHECK: extractvalue { double, double } [[SINCOS]], 1
+
+; CHECK-NO-SINCOS: call double @__sinpi
+; CHECK-NO-SINCOS: call double @__cospi
+}
+
+define double @test_constant_f64() {
+       %sin = call double @__sinpi(double 1.0) #0
+       %cos = call double @__cospi(double 1.0) #0
+       %res = fadd double %sin, %cos
+       ret double %res
+; CHECK-FLOAT-IN-VEC: [[SINCOS:%[a-z0-9]+]] = call { double, double } @__sincospi_stret(double 1.000000e+00)
+; CHECK-FLOAT-IN-VEC: extractvalue { double, double } [[SINCOS]], 0
+; CHECK-FLOAT-IN-VEC: extractvalue { double, double } [[SINCOS]], 1
+
+; CHECK: [[SINCOS:%[a-z0-9]+]] = call { double, double } @__sincospi_stret(double 1.000000e+00)
+; CHECK: extractvalue { double, double } [[SINCOS]], 0
+; CHECK: extractvalue { double, double } [[SINCOS]], 1
+
+; CHECK-NO-SINCOS: call double @__sinpi
+; CHECK-NO-SINCOS: call double @__cospi
+}