Add strictfp attribute to prevent unwanted optimizations of libm calls

Differential Revision: https://reviews.llvm.org/D34163

llvm-svn: 310885

docs/LangRef.rst

@@ -1644,6 +1644,12 @@ example:
     If a function that has an ``sspstrong`` attribute is inlined into a
     function that doesn't have an ``sspstrong`` attribute, then the
     resulting function will have an ``sspstrong`` attribute.
+``strictfp``
+    This attribute indicates that the function was called from a scope that
+    requires strict floating point semantics.  LLVM will not attempt any
+    optimizations that require assumptions about the floating point rounding
+    mode or that might alter the state of floating point status flags that
+    might otherwise be set or cleared by calling this function.
 ``"thunk"``
     This attribute indicates that the function will delegate to some other
     function with a tail call. The prototype of a thunk should not be used for

include/llvm/Bitcode/LLVMBitCodes.h

@@ -558,7 +558,8 @@ enum AttributeKindCodes {
   ATTR_KIND_INACCESSIBLEMEM_OR_ARGMEMONLY = 50,
   ATTR_KIND_ALLOC_SIZE = 51,
   ATTR_KIND_WRITEONLY = 52,
-  ATTR_KIND_SPECULATABLE = 53
+  ATTR_KIND_SPECULATABLE = 53,
+  ATTR_KIND_STRICT_FP = 54,
 };
 
 enum ComdatSelectionKindCodes {

include/llvm/IR/Attributes.td

@@ -149,6 +149,9 @@ def StackProtectReq : EnumAttr<"sspreq">;
 /// Strong Stack protection.
 def StackProtectStrong : EnumAttr<"sspstrong">;
 
+/// Function was called in a scope requiring strict floating point semantics.
+def StrictFP : EnumAttr<"strictfp">;
+
 /// Hidden pointer to structure to return.
 def StructRet : EnumAttr<"sret">;
 

include/llvm/IR/CallSite.h

@@ -426,6 +426,11 @@ public:
     CALLSITE_DELEGATE_GETTER(isNoBuiltin());
   }
 
+  /// Return true if the call requires strict floating point semantics.
+  bool isStrictFP() const {
+    CALLSITE_DELEGATE_GETTER(isStrictFP());
+  }
+
   /// Return true if the call should not be inlined.
   bool isNoInline() const {
     CALLSITE_DELEGATE_GETTER(isNoInline());

include/llvm/IR/Instructions.h

@@ -1757,6 +1757,9 @@ public:
       !hasFnAttrImpl(Attribute::Builtin);
   }
 
+  /// Determine if the call requires strict floating point semantics.
+  bool isStrictFP() const { return hasFnAttr(Attribute::StrictFP); }
+
   /// Return true if the call should not be inlined.
   bool isNoInline() const { return hasFnAttr(Attribute::NoInline); }
   void setIsNoInline() {
@@ -3844,6 +3847,9 @@ public:
       !hasFnAttrImpl(Attribute::Builtin);
   }
 
+  /// Determine if the call requires strict floating point semantics.
+  bool isStrictFP() const { return hasFnAttr(Attribute::StrictFP); }
+
   /// Return true if the call should not be inlined.
   bool isNoInline() const { return hasFnAttr(Attribute::NoInline); }
   void setIsNoInline() {

include/llvm/Transforms/Utils/SimplifyLibCalls.h

@@ -137,6 +137,9 @@ private:
   Value *optimizeSqrt(CallInst *CI, IRBuilder<> &B);
   Value *optimizeSinCosPi(CallInst *CI, IRBuilder<> &B);
   Value *optimizeTan(CallInst *CI, IRBuilder<> &B);
+  // Wrapper for all floating point library call optimizations
+  Value *optimizeFloatingPointLibCall(CallInst *CI, LibFunc Func,
+                                      IRBuilder<> &B);
 
   // Integer Library Call Optimizations
   Value *optimizeFFS(CallInst *CI, IRBuilder<> &B);

lib/Analysis/ConstantFolding.cpp

@@ -1359,7 +1359,7 @@ llvm::ConstantFoldLoadThroughGEPIndices(Constant *C,
 //
 
 bool llvm::canConstantFoldCallTo(ImmutableCallSite CS, const Function *F) {
-  if (CS.isNoBuiltin())
+  if (CS.isNoBuiltin() || CS.isStrictFP())
     return false;
   switch (F->getIntrinsicID()) {
   case Intrinsic::fabs:
@@ -2066,7 +2066,7 @@ Constant *
 llvm::ConstantFoldCall(ImmutableCallSite CS, Function *F,
                        ArrayRef<Constant *> Operands,
                        const TargetLibraryInfo *TLI) {
-  if (CS.isNoBuiltin())
+  if (CS.isNoBuiltin() || CS.isStrictFP())
     return nullptr;
   if (!F->hasName())
     return nullptr;
@@ -2084,7 +2084,7 @@ llvm::ConstantFoldCall(ImmutableCallSite CS, Function *F,
 bool llvm::isMathLibCallNoop(CallSite CS, const TargetLibraryInfo *TLI) {
   // FIXME: Refactor this code; this duplicates logic in LibCallsShrinkWrap
   // (and to some extent ConstantFoldScalarCall).
-  if (CS.isNoBuiltin())
+  if (CS.isNoBuiltin() || CS.isStrictFP())
     return false;
   Function *F = CS.getCalledFunction();
   if (!F)

lib/AsmParser/LLLexer.cpp

@@ -654,6 +654,7 @@ lltok::Kind LLLexer::LexIdentifier() {
   KEYWORD(ssp);
   KEYWORD(sspreq);
   KEYWORD(sspstrong);
+  KEYWORD(strictfp);
   KEYWORD(safestack);
   KEYWORD(sanitize_address);
   KEYWORD(sanitize_thread);

lib/AsmParser/LLParser.cpp

@@ -1121,6 +1121,7 @@ bool LLParser::ParseFnAttributeValuePairs(AttrBuilder &B,
       B.addAttribute(Attribute::SanitizeThread); break;
     case lltok::kw_sanitize_memory:
       B.addAttribute(Attribute::SanitizeMemory); break;
+    case lltok::kw_strictfp: B.addAttribute(Attribute::StrictFP); break;
     case lltok::kw_uwtable: B.addAttribute(Attribute::UWTable); break;
     case lltok::kw_writeonly: B.addAttribute(Attribute::WriteOnly); break;
 
@@ -1446,6 +1447,7 @@ bool LLParser::ParseOptionalParamAttrs(AttrBuilder &B) {
     case lltok::kw_sspreq:
     case lltok::kw_sspstrong:
     case lltok::kw_safestack:
+    case lltok::kw_strictfp:
     case lltok::kw_uwtable:
       HaveError |= Error(Lex.getLoc(), "invalid use of function-only attribute");
       break;
@@ -1537,6 +1539,7 @@ bool LLParser::ParseOptionalReturnAttrs(AttrBuilder &B) {
     case lltok::kw_sspreq:
     case lltok::kw_sspstrong:
     case lltok::kw_safestack:
+    case lltok::kw_strictfp:
     case lltok::kw_uwtable:
       HaveError |= Error(Lex.getLoc(), "invalid use of function-only attribute");
       break;

lib/AsmParser/LLToken.h

@@ -207,6 +207,7 @@ enum Kind {
   kw_sret,
   kw_sanitize_thread,
   kw_sanitize_memory,
+  kw_strictfp,
   kw_swifterror,
   kw_swiftself,
   kw_uwtable,

lib/Bitcode/Reader/BitcodeReader.cpp

@@ -1137,6 +1137,7 @@ static uint64_t getRawAttributeMask(Attribute::AttrKind Val) {
   case Attribute::SwiftError:      return 1ULL << 52;
   case Attribute::WriteOnly:       return 1ULL << 53;
   case Attribute::Speculatable:    return 1ULL << 54;
+  case Attribute::StrictFP:        return 1ULL << 55;
   case Attribute::Dereferenceable:
     llvm_unreachable("dereferenceable attribute not supported in raw format");
     break;
@@ -1345,6 +1346,8 @@ static Attribute::AttrKind getAttrFromCode(uint64_t Code) {
     return Attribute::StackProtectStrong;
   case bitc::ATTR_KIND_SAFESTACK:
     return Attribute::SafeStack;
+  case bitc::ATTR_KIND_STRICT_FP:
+    return Attribute::StrictFP;
   case bitc::ATTR_KIND_STRUCT_RET:
     return Attribute::StructRet;
   case bitc::ATTR_KIND_SANITIZE_ADDRESS:

lib/Bitcode/Writer/BitcodeWriter.cpp

@@ -610,6 +610,8 @@ static uint64_t getAttrKindEncoding(Attribute::AttrKind Kind) {
     return bitc::ATTR_KIND_STACK_PROTECT_STRONG;
   case Attribute::SafeStack:
     return bitc::ATTR_KIND_SAFESTACK;
+  case Attribute::StrictFP:
+    return bitc::ATTR_KIND_STRICT_FP;
   case Attribute::StructRet:
     return bitc::ATTR_KIND_STRUCT_RET;
   case Attribute::SanitizeAddress:

lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

@@ -6555,10 +6555,10 @@ void SelectionDAGBuilder::visitCall(const CallInst &I) {
 
     // Check for well-known libc/libm calls.  If the function is internal, it
     // can't be a library call.  Don't do the check if marked as nobuiltin for
-    // some reason.
+    // some reason or the call site requires strict floating point semantics.
     LibFunc Func;
-    if (!I.isNoBuiltin() && !F->hasLocalLinkage() && F->hasName() &&
-        LibInfo->getLibFunc(*F, Func) &&
+    if (!I.isNoBuiltin() && !I.isStrictFP() && !F->hasLocalLinkage() &&
+        F->hasName() && LibInfo->getLibFunc(*F, Func) &&
         LibInfo->hasOptimizedCodeGen(Func)) {
       switch (Func) {
       default: break;

lib/IR/Attributes.cpp

@@ -327,6 +327,8 @@ std::string Attribute::getAsString(bool InAttrGrp) const {
     return "sspstrong";
   if (hasAttribute(Attribute::SafeStack))
     return "safestack";
+  if (hasAttribute(Attribute::StrictFP))
+    return "strictfp";
   if (hasAttribute(Attribute::StructRet))
     return "sret";
   if (hasAttribute(Attribute::SanitizeThread))

lib/IR/Verifier.cpp

@@ -1377,6 +1377,7 @@ static bool isFuncOnlyAttr(Attribute::AttrKind Kind) {
   case Attribute::InaccessibleMemOrArgMemOnly:
   case Attribute::AllocSize:
   case Attribute::Speculatable:
+  case Attribute::StrictFP:
     return true;
   default:
     break;

lib/Transforms/IPO/ForceFunctionAttrs.cpp

@@ -57,6 +57,7 @@ static Attribute::AttrKind parseAttrKind(StringRef Kind) {
       .Case("ssp", Attribute::StackProtect)
       .Case("sspreq", Attribute::StackProtectReq)
       .Case("sspstrong", Attribute::StackProtectStrong)
+      .Case("strictfp", Attribute::StrictFP)
       .Case("uwtable", Attribute::UWTable)
       .Default(Attribute::None);
 }

lib/Transforms/Utils/SimplifyLibCalls.cpp

@@ -2041,13 +2041,103 @@ Value *LibCallSimplifier::optimizeStringMemoryLibCall(CallInst *CI,
   return nullptr;
 }
 
+Value *LibCallSimplifier::optimizeFloatingPointLibCall(CallInst *CI,
+                                                       LibFunc Func,
+                                                       IRBuilder<> &Builder) {
+  // Don't optimize calls that require strict floating point semantics.
+  if (CI->isStrictFP())
+    return nullptr;
+
+  switch (Func) {
+  case LibFunc_cosf:
+  case LibFunc_cos:
+  case LibFunc_cosl:
+    return optimizeCos(CI, Builder);
+  case LibFunc_sinpif:
+  case LibFunc_sinpi:
+  case LibFunc_cospif:
+  case LibFunc_cospi:
+    return optimizeSinCosPi(CI, Builder);
+  case LibFunc_powf:
+  case LibFunc_pow:
+  case LibFunc_powl:
+    return optimizePow(CI, Builder);
+  case LibFunc_exp2l:
+  case LibFunc_exp2:
+  case LibFunc_exp2f:
+    return optimizeExp2(CI, Builder);
+  case LibFunc_fabsf:
+  case LibFunc_fabs:
+  case LibFunc_fabsl:
+    return replaceUnaryCall(CI, Builder, Intrinsic::fabs);
+  case LibFunc_sqrtf:
+  case LibFunc_sqrt:
+  case LibFunc_sqrtl:
+    return optimizeSqrt(CI, Builder);
+  case LibFunc_log:
+  case LibFunc_log10:
+  case LibFunc_log1p:
+  case LibFunc_log2:
+  case LibFunc_logb:
+    return optimizeLog(CI, Builder);
+  case LibFunc_tan:
+  case LibFunc_tanf:
+  case LibFunc_tanl:
+    return optimizeTan(CI, Builder);
+  case LibFunc_ceil:
+    return replaceUnaryCall(CI, Builder, Intrinsic::ceil);
+  case LibFunc_floor:
+    return replaceUnaryCall(CI, Builder, Intrinsic::floor);
+  case LibFunc_round:
+    return replaceUnaryCall(CI, Builder, Intrinsic::round);
+  case LibFunc_nearbyint:
+    return replaceUnaryCall(CI, Builder, Intrinsic::nearbyint);
+  case LibFunc_rint:
+    return replaceUnaryCall(CI, Builder, Intrinsic::rint);
+  case LibFunc_trunc:
+    return replaceUnaryCall(CI, Builder, Intrinsic::trunc);
+  case LibFunc_acos:
+  case LibFunc_acosh:
+  case LibFunc_asin:
+  case LibFunc_asinh:
+  case LibFunc_atan:
+  case LibFunc_atanh:
+  case LibFunc_cbrt:
+  case LibFunc_cosh:
+  case LibFunc_exp:
+  case LibFunc_exp10:
+  case LibFunc_expm1:
+  case LibFunc_sin:
+  case LibFunc_sinh:
+  case LibFunc_tanh:
+    if (UnsafeFPShrink && hasFloatVersion(CI->getCalledFunction()->getName()))
+      return optimizeUnaryDoubleFP(CI, Builder, true);
+    return nullptr;
+  case LibFunc_copysign:
+    if (hasFloatVersion(CI->getCalledFunction()->getName()))
+      return optimizeBinaryDoubleFP(CI, Builder);
+    return nullptr;
+  case LibFunc_fminf:
+  case LibFunc_fmin:
+  case LibFunc_fminl:
+  case LibFunc_fmaxf:
+  case LibFunc_fmax:
+  case LibFunc_fmaxl:
+    return optimizeFMinFMax(CI, Builder);
+  default:
+    return nullptr;
+  }
+}
+
 Value *LibCallSimplifier::optimizeCall(CallInst *CI) {
+  // TODO: Split out the code below that operates on FP calls so that
+  //       we can all non-FP calls with the StrictFP attribute to be
+  //       optimized.
   if (CI->isNoBuiltin())
     return nullptr;
 
   LibFunc Func;
   Function *Callee = CI->getCalledFunction();
-  StringRef FuncName = Callee->getName();
 
   SmallVector<OperandBundleDef, 2> OpBundles;
   CI->getOperandBundlesAsDefs(OpBundles);
@@ -2055,6 +2145,8 @@ Value *LibCallSimplifier::optimizeCall(CallInst *CI) {
   bool isCallingConvC = isCallingConvCCompatible(CI);
 
   // Command-line parameter overrides instruction attribute.
+  // This can't be moved to optimizeFloatingPointLibCall() because it may be
+  // used by the intrinsic optimizations. 
   if (EnableUnsafeFPShrink.getNumOccurrences() > 0)
     UnsafeFPShrink = EnableUnsafeFPShrink;
   else if (isa<FPMathOperator>(CI) && CI->hasUnsafeAlgebra())
@@ -2064,6 +2156,8 @@ Value *LibCallSimplifier::optimizeCall(CallInst *CI) {
   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
     if (!isCallingConvC)
       return nullptr;
+    // The FP intrinsics have corresponding constrained versions so we don't
+    // need to check for the StrictFP attribute here.
     switch (II->getIntrinsicID()) {
     case Intrinsic::pow:
       return optimizePow(CI, Builder);
@@ -2104,32 +2198,9 @@ Value *LibCallSimplifier::optimizeCall(CallInst *CI) {
       return nullptr;
     if (Value *V = optimizeStringMemoryLibCall(CI, Builder))
       return V;
+    if (Value *V = optimizeFloatingPointLibCall(CI, Func, Builder))
+      return V;
     switch (Func) {
-    case LibFunc_cosf:
-    case LibFunc_cos:
-    case LibFunc_cosl:
-      return optimizeCos(CI, Builder);
-    case LibFunc_sinpif:
-    case LibFunc_sinpi:
-    case LibFunc_cospif:
-    case LibFunc_cospi:
-      return optimizeSinCosPi(CI, Builder);
-    case LibFunc_powf:
-    case LibFunc_pow:
-    case LibFunc_powl:
-      return optimizePow(CI, Builder);
-    case LibFunc_exp2l:
-    case LibFunc_exp2:
-    case LibFunc_exp2f:
-      return optimizeExp2(CI, Builder);
-    case LibFunc_fabsf:
-    case LibFunc_fabs:
-    case LibFunc_fabsl:
-      return replaceUnaryCall(CI, Builder, Intrinsic::fabs);
-    case LibFunc_sqrtf:
-    case LibFunc_sqrt:
-    case LibFunc_sqrtl:
-      return optimizeSqrt(CI, Builder);
     case LibFunc_ffs:
     case LibFunc_ffsl:
     case LibFunc_ffsll:
@@ -2158,18 +2229,8 @@ Value *LibCallSimplifier::optimizeCall(CallInst *CI) {
       return optimizeFWrite(CI, Builder);
     case LibFunc_fputs:
       return optimizeFPuts(CI, Builder);
-    case LibFunc_log:
-    case LibFunc_log10:
-    case LibFunc_log1p:
-    case LibFunc_log2:
-    case LibFunc_logb:
-      return optimizeLog(CI, Builder);
     case LibFunc_puts:
       return optimizePuts(CI, Builder);
-    case LibFunc_tan:
-    case LibFunc_tanf:
-    case LibFunc_tanl:
-      return optimizeTan(CI, Builder);
     case LibFunc_perror:
       return optimizeErrorReporting(CI, Builder);
     case LibFunc_vfprintf:
@@ -2177,46 +2238,6 @@ Value *LibCallSimplifier::optimizeCall(CallInst *CI) {
       return optimizeErrorReporting(CI, Builder, 0);
     case LibFunc_fputc:
       return optimizeErrorReporting(CI, Builder, 1);
-    case LibFunc_ceil:
-      return replaceUnaryCall(CI, Builder, Intrinsic::ceil);
-    case LibFunc_floor:
-      return replaceUnaryCall(CI, Builder, Intrinsic::floor);
-    case LibFunc_round:
-      return replaceUnaryCall(CI, Builder, Intrinsic::round);
-    case LibFunc_nearbyint:
-      return replaceUnaryCall(CI, Builder, Intrinsic::nearbyint);
-    case LibFunc_rint:
-      return replaceUnaryCall(CI, Builder, Intrinsic::rint);
-    case LibFunc_trunc:
-      return replaceUnaryCall(CI, Builder, Intrinsic::trunc);
-    case LibFunc_acos:
-    case LibFunc_acosh:
-    case LibFunc_asin:
-    case LibFunc_asinh:
-    case LibFunc_atan:
-    case LibFunc_atanh:
-    case LibFunc_cbrt:
-    case LibFunc_cosh:
-    case LibFunc_exp:
-    case LibFunc_exp10:
-    case LibFunc_expm1:
-    case LibFunc_sin:
-    case LibFunc_sinh:
-    case LibFunc_tanh:
-      if (UnsafeFPShrink && hasFloatVersion(FuncName))
-        return optimizeUnaryDoubleFP(CI, Builder, true);
-      return nullptr;
-    case LibFunc_copysign:
-      if (hasFloatVersion(FuncName))
-        return optimizeBinaryDoubleFP(CI, Builder);
-      return nullptr;
-    case LibFunc_fminf:
-    case LibFunc_fmin:
-    case LibFunc_fminl:
-    case LibFunc_fmaxf:
-    case LibFunc_fmax:
-    case LibFunc_fmaxl:
-      return optimizeFMinFMax(CI, Builder);
     default:
       return nullptr;
     }

test/Bitcode/compatibility.ll

@@ -608,6 +608,7 @@ declare void @f.inaccessiblememonly() inaccessiblememonly
 ; CHECK: declare void @f.inaccessiblememonly() #33
 declare void @f.inaccessiblemem_or_argmemonly() inaccessiblemem_or_argmemonly
 ; CHECK: declare void @f.inaccessiblemem_or_argmemonly() #34
+declare void @f.strictfp() #35
 
 ; Functions -- section
 declare void @f.section() section "80"
@@ -1252,6 +1253,9 @@ exit:
   call void @f.nobuiltin() builtin
   ; CHECK: call void @f.nobuiltin() #42
 
+  call void @f.strictfp() strictfp
+  ; CHECK: call void @f.strictfp() #43
+
   call fastcc noalias i32* @f.noalias() noinline
   ; CHECK: call fastcc noalias i32* @f.noalias() #12
   tail call ghccc nonnull i32* @f.nonnull() minsize
@@ -1670,6 +1674,7 @@ define i8** @constexpr() {
 ; CHECK: attributes #40 = { writeonly }
 ; CHECK: attributes #41 = { speculatable }
 ; CHECK: attributes #42 = { builtin }
+; CHECK: attributes #43 = { strictfp }
 
 ;; Metadata
 

test/Transforms/DCE/calls-errno.ll

@@ -76,6 +76,10 @@ entry:
 ; CHECK-NEXT: %cos3 = call double @cos(double 0.000000e+00)
   %cos3 = call double @cos(double 0.000000e+00) nobuiltin
 
+; cos(1) strictfp sets FP status flags
+; CHECK-NEXT: %cos4 = call double @cos(double 1.000000e+00)
+  %cos4 = call double @cos(double 1.000000e+00) strictfp
+
 ; pow(0, 1) is 0
   %pow1 = call double @pow(double 0x7FF0000000000000, double 1.000000e+00)
 

test/Transforms/InstCombine/constant-fold-libfunc.ll

@@ -12,9 +12,20 @@ define double @test_simplify_acos() {
   ret double %pi
 }
 
+; Check that we don't constant fold builtin functions.
+
 define double @test_acos_nobuiltin() {
 ; CHECK-LABEL: @test_acos_nobuiltin
   %pi = call double @acos(double -1.000000e+00) nobuiltin 
 ; CHECK: call double @acos(double -1.000000e+00)
   ret double %pi
 }
+
+; Check that we don't constant fold strictfp results that require rounding.
+
+define double @test_acos_strictfp() {
+; CHECK-LABEL: @test_acos_strictfp
+  %pi = call double @acos(double -1.000000e+00) strictfp 
+; CHECK: call double @acos(double -1.000000e+00)
+  ret double %pi
+}

test/Transforms/InstCombine/memcpy-1.ll

@@ -17,3 +17,12 @@ define i8* @test_simplify1(i8* %mem1, i8* %mem2, i32 %size) {
   ret i8* %ret
 }
 
+; Verify that the strictfp attr doesn't block this optimization.
+
+define i8* @test_simplify2(i8* %mem1, i8* %mem2, i32 %size) {
+; CHECK-LABEL: @test_simplify2(
+  %ret = call i8* @memcpy(i8* %mem1, i8* %mem2, i32 %size) strictfp
+; CHECK: call void @llvm.memcpy
+  ret i8* %ret
+; CHECK: ret i8* %mem1
+}

utils/vim/syntax/llvm.vim

@@ -144,6 +144,7 @@ syn keyword llvmKeyword
       \ ssp
       \ sspreq
       \ sspstrong
+      \ strictfp
       \ swiftcc
       \ tail
       \ target