Add support for fast-math flags to the FCmp instruction.

[jmolloy]

Summary

FCmp behaves a lot like a floating-point binary operator in many ways,
and can benefit from fast-math information. Flags such as nsz and nnan
can affect if this fcmp (in combination with a select) can be treated
as a fminnum/fmaxnum operation.

This adds backwards-compatible bitcode support, IR parsing and writing,
LangRef changes and IRBuilder changes. I'll need to audit InstSimplify
and InstCombine in a followup to find places where flags should be
copied.

[majnemer]

docs/LangRef.rst

6978–6981

We really need to say more here about the semantics of fast math flags in this position.

For example, does arcp on an fcmp imply that i can take:

%div = fdiv double 1.000000e+00, %a
%div1 = fdiv double 1.000000e+00, %b
%cmp = fcmp arcp olt double %div, %div1

and replace it with:

%cmp = fcmp ogt double %a, %b

?

[hfinkel]

docs/LangRef.rst

6978–6981

I agree. FWIW, I'd say the answer should be no. We should require arcp on the fdivs for that.

[jmolloy]

Hi Hal, David,

I've updated the patch to explicitly say that the only flags that should change semantics are those that allow assumptions about input arguments (not transforms).

Sorry about the long round-trip on this, I went on vacation for 2 weeks!

Cheers,

James

[jmolloy]

Ping!

[jmolloy]

Ping*2!

[hfinkel]

LGTM.

docs/LangRef.rst

6984

I'd rather word this as, "that allow assumptions to be made about the values of the input arguments", because implying that there is a mathematical 'domain' that includes NaN might be confusing.

6988

Why are you deleting this statement in this patch? Is this a separate cleanup?

[jmolloy]

Thanks Hal - I've committed this (without the extra text removal; I'll add that as a separate cleanup as you suggest) in r241901.

docs/LangRef.rst

 
 .. _fastmath:
 
 Fast-Math Flags
 ---------------
 
 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
-:ref:`frem <i_frem>`) have the following flags that can be set to enable
-otherwise unsafe floating point operations
+:ref:`frem <i_frem>`, :ref:`fcmp <i_fcmp>`) have the following flags that can
+be set to enable otherwise unsafe floating point operations
 
 ``nnan``
    No NaNs - Allow optimizations to assume the arguments and result are not
    NaN. Such optimizations are required to retain defined behavior over
    NaNs, but the value of the result is undefined.
 
 ``ninf``
    No Infs - Allow optimizations to assume the arguments and result are not

 '``fcmp``' Instruction
 ^^^^^^^^^^^^^^^^^^^^^^
 
 Syntax:
 """""""
 
 ::
 
-      <result> = fcmp <cond> <ty> <op1>, <op2>     ; yields i1 or <N x i1>:result
+      <result> = fcmp [fast-math flags]* <cond> <ty> <op1>, <op2>     ; yields i1 or <N x i1>:result
 
 Overview:
 """""""""
 
 The '``fcmp``' instruction returns a boolean value or vector of boolean
 values based on comparison of its operands.
 
 If the operands are floating point scalars, then the result type is a
#. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
    less than ``op2``.
 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
    less than or equal to ``op2``.
 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
    not equal to ``op2``.
 #. ``uno``: yields ``true`` if either operand is a QNAN.
 #. ``true``: always yields ``true``, regardless of operands.
 
+The ``fcmp`` instruction can also optionally take any number of
+:ref:`fast-math flags <fastmath>`, which are optimization hints to enable
+otherwise unsafe floating point optimizations.
+
+Any set of fast-math flags are legal on an ``fcmp`` instruction, but the
+only flags that have any effect on its semantics are those that allow
+assumptions to be made about the domain of input arguments; namely
+``nnan``, ``ninf``, and ``nsz``. See :ref:`fastmath` for more information.
+
 Example:
 """"""""
 
 .. code-block:: llvm
 
       <result> = fcmp oeq float 4.0, 5.0    ; yields: result=false
       <result> = fcmp one float 4.0, 5.0    ; yields: result=true
       <result> = fcmp olt float 4.0, 5.0    ; yields: result=true
       <result> = fcmp ueq double 1.0, 2.0   ; yields: result=false
 
-Note that the code generator does not yet support vector types with the
-``fcmp`` instruction.
-
 .. _i_phi:
 
 '``phi``' Instruction
 ^^^^^^^^^^^^^^^^^^^^^
 
 Syntax:
 """""""
 

include/llvm/IR/IRBuilder.h

public:
   }
   Value *CreateICmpSLT(Value *LHS, Value *RHS, const Twine &Name = "") {
     return CreateICmp(ICmpInst::ICMP_SLT, LHS, RHS, Name);
   }
   Value *CreateICmpSLE(Value *LHS, Value *RHS, const Twine &Name = "") {
     return CreateICmp(ICmpInst::ICMP_SLE, LHS, RHS, Name);
   }
 
-  Value *CreateFCmpOEQ(Value *LHS, Value *RHS, const Twine &Name = "") {
-    return CreateFCmp(FCmpInst::FCMP_OEQ, LHS, RHS, Name);
+  Value *CreateFCmpOEQ(Value *LHS, Value *RHS, const Twine &Name = "",
+                       MDNode *FPMathTag = nullptr) {
+    return CreateFCmp(FCmpInst::FCMP_OEQ, LHS, RHS, Name, FPMathTag);
   }
-  Value *CreateFCmpOGT(Value *LHS, Value *RHS, const Twine &Name = "") {
-    return CreateFCmp(FCmpInst::FCMP_OGT, LHS, RHS, Name);
+  Value *CreateFCmpOGT(Value *LHS, Value *RHS, const Twine &Name = "",
+                       MDNode *FPMathTag = nullptr) {
+    return CreateFCmp(FCmpInst::FCMP_OGT, LHS, RHS, Name, FPMathTag);
   }
-  Value *CreateFCmpOGE(Value *LHS, Value *RHS, const Twine &Name = "") {
-    return CreateFCmp(FCmpInst::FCMP_OGE, LHS, RHS, Name);
+  Value *CreateFCmpOGE(Value *LHS, Value *RHS, const Twine &Name = "",
+                       MDNode *FPMathTag = nullptr) {
+    return CreateFCmp(FCmpInst::FCMP_OGE, LHS, RHS, Name, FPMathTag);
   }
-  Value *CreateFCmpOLT(Value *LHS, Value *RHS, const Twine &Name = "") {
-    return CreateFCmp(FCmpInst::FCMP_OLT, LHS, RHS, Name);
+  Value *CreateFCmpOLT(Value *LHS, Value *RHS, const Twine &Name = "",
+                       MDNode *FPMathTag = nullptr) {
+    return CreateFCmp(FCmpInst::FCMP_OLT, LHS, RHS, Name, FPMathTag);
   }
-  Value *CreateFCmpOLE(Value *LHS, Value *RHS, const Twine &Name = "") {
-    return CreateFCmp(FCmpInst::FCMP_OLE, LHS, RHS, Name);
+  Value *CreateFCmpOLE(Value *LHS, Value *RHS, const Twine &Name = "",
+                       MDNode *FPMathTag = nullptr) {
+    return CreateFCmp(FCmpInst::FCMP_OLE, LHS, RHS, Name, FPMathTag);
   }
-  Value *CreateFCmpONE(Value *LHS, Value *RHS, const Twine &Name = "") {
-    return CreateFCmp(FCmpInst::FCMP_ONE, LHS, RHS, Name);
+  Value *CreateFCmpONE(Value *LHS, Value *RHS, const Twine &Name = "",
+                       MDNode *FPMathTag = nullptr) {
+    return CreateFCmp(FCmpInst::FCMP_ONE, LHS, RHS, Name, FPMathTag);
   }
-  Value *CreateFCmpORD(Value *LHS, Value *RHS, const Twine &Name = "") {
-    return CreateFCmp(FCmpInst::FCMP_ORD, LHS, RHS, Name);
+  Value *CreateFCmpORD(Value *LHS, Value *RHS, const Twine &Name = "",
+                       MDNode *FPMathTag = nullptr) {
+    return CreateFCmp(FCmpInst::FCMP_ORD, LHS, RHS, Name, FPMathTag);
   }
-  Value *CreateFCmpUNO(Value *LHS, Value *RHS, const Twine &Name = "") {
-    return CreateFCmp(FCmpInst::FCMP_UNO, LHS, RHS, Name);
+  Value *CreateFCmpUNO(Value *LHS, Value *RHS, const Twine &Name = "",
+                       MDNode *FPMathTag = nullptr) {
+    return CreateFCmp(FCmpInst::FCMP_UNO, LHS, RHS, Name, FPMathTag);
   }
-  Value *CreateFCmpUEQ(Value *LHS, Value *RHS, const Twine &Name = "") {
-    return CreateFCmp(FCmpInst::FCMP_UEQ, LHS, RHS, Name);
+  Value *CreateFCmpUEQ(Value *LHS, Value *RHS, const Twine &Name = "",
+                       MDNode *FPMathTag = nullptr) {
+    return CreateFCmp(FCmpInst::FCMP_UEQ, LHS, RHS, Name, FPMathTag);
   }
-  Value *CreateFCmpUGT(Value *LHS, Value *RHS, const Twine &Name = "") {
-    return CreateFCmp(FCmpInst::FCMP_UGT, LHS, RHS, Name);
+  Value *CreateFCmpUGT(Value *LHS, Value *RHS, const Twine &Name = "",
+                       MDNode *FPMathTag = nullptr) {
+    return CreateFCmp(FCmpInst::FCMP_UGT, LHS, RHS, Name, FPMathTag);
   }
-  Value *CreateFCmpUGE(Value *LHS, Value *RHS, const Twine &Name = "") {
-    return CreateFCmp(FCmpInst::FCMP_UGE, LHS, RHS, Name);
+  Value *CreateFCmpUGE(Value *LHS, Value *RHS, const Twine &Name = "",
+                       MDNode *FPMathTag = nullptr) {
+    return CreateFCmp(FCmpInst::FCMP_UGE, LHS, RHS, Name, FPMathTag);
   }
-  Value *CreateFCmpULT(Value *LHS, Value *RHS, const Twine &Name = "") {
-    return CreateFCmp(FCmpInst::FCMP_ULT, LHS, RHS, Name);
+  Value *CreateFCmpULT(Value *LHS, Value *RHS, const Twine &Name = "",
+                       MDNode *FPMathTag = nullptr) {
+    return CreateFCmp(FCmpInst::FCMP_ULT, LHS, RHS, Name, FPMathTag);
   }
-  Value *CreateFCmpULE(Value *LHS, Value *RHS, const Twine &Name = "") {
-    return CreateFCmp(FCmpInst::FCMP_ULE, LHS, RHS, Name);
+  Value *CreateFCmpULE(Value *LHS, Value *RHS, const Twine &Name = "",
+                       MDNode *FPMathTag = nullptr) {
+    return CreateFCmp(FCmpInst::FCMP_ULE, LHS, RHS, Name, FPMathTag);
   }
-  Value *CreateFCmpUNE(Value *LHS, Value *RHS, const Twine &Name = "") {
-    return CreateFCmp(FCmpInst::FCMP_UNE, LHS, RHS, Name);
+  Value *CreateFCmpUNE(Value *LHS, Value *RHS, const Twine &Name = "",
+                       MDNode *FPMathTag = nullptr) {
+    return CreateFCmp(FCmpInst::FCMP_UNE, LHS, RHS, Name, FPMathTag);
   }
 
   Value *CreateICmp(CmpInst::Predicate P, Value *LHS, Value *RHS,
                     const Twine &Name = "") {
     if (Constant *LC = dyn_cast<Constant>(LHS))
       if (Constant *RC = dyn_cast<Constant>(RHS))
         return Insert(Folder.CreateICmp(P, LC, RC), Name);
     return Insert(new ICmpInst(P, LHS, RHS), Name);
   }
   Value *CreateFCmp(CmpInst::Predicate P, Value *LHS, Value *RHS,
-                    const Twine &Name = "") {
+                    const Twine &Name = "", MDNode *FPMathTag = nullptr) {
     if (Constant *LC = dyn_cast<Constant>(LHS))
       if (Constant *RC = dyn_cast<Constant>(RHS))
         return Insert(Folder.CreateFCmp(P, LC, RC), Name);
-    return Insert(new FCmpInst(P, LHS, RHS), Name);
+    return Insert(AddFPMathAttributes(new FCmpInst(P, LHS, RHS),
+                                      FPMathTag, FMF), Name);
   }
 
   //===--------------------------------------------------------------------===//
   // Instruction creation methods: Other Instructions
   //===--------------------------------------------------------------------===//
 
   PHINode *CreatePHI(Type *Ty, unsigned NumReservedValues,
                      const Twine &Name = "") {

include/llvm/IR/Operator.h

public:
   }
 
   /// \brief Get the maximum error permitted by this operation in ULPs.  An
   /// accuracy of 0.0 means that the operation should be performed with the
   /// default precision.
   float getFPAccuracy() const;
 
   static inline bool classof(const Instruction *I) {
-    return I->getType()->isFPOrFPVectorTy();
+    return I->getType()->isFPOrFPVectorTy() ||
+      I->getOpcode() == Instruction::FCmp;
   }
   static inline bool classof(const Value *V) {
     return isa<Instruction>(V) && classof(cast<Instruction>(V));
   }
 };
 
 
 /// A helper template for defining operators for individual opcodes.

lib/AsmParser/LLParser.cpp

case lltok::kw_ashr: {
     return false;
   }
 
   case lltok::kw_urem:
   case lltok::kw_srem:   return ParseArithmetic(Inst, PFS, KeywordVal, 1);
   case lltok::kw_and:
   case lltok::kw_or:
   case lltok::kw_xor:    return ParseLogical(Inst, PFS, KeywordVal);
-  case lltok::kw_icmp:
-  case lltok::kw_fcmp:   return ParseCompare(Inst, PFS, KeywordVal);
+  case lltok::kw_icmp:   return ParseCompare(Inst, PFS, KeywordVal);
+  case lltok::kw_fcmp: {
+    FastMathFlags FMF = EatFastMathFlagsIfPresent();
+    int Res = ParseCompare(Inst, PFS, KeywordVal);
+    if (Res != 0)
+      return Res;
+    if (FMF.any())
+      Inst->setFastMathFlags(FMF);
+    return 0;
+  }
+
   // Casts.
   case lltok::kw_trunc:
   case lltok::kw_zext:
   case lltok::kw_sext:
   case lltok::kw_fptrunc:
   case lltok::kw_fpext:
   case lltok::kw_bitcast:
   case lltok::kw_addrspacecast:

lib/Bitcode/Reader/BitcodeReader.cpp

case bitc::COMDAT_SELECTION_KIND_LARGEST:
     return Comdat::Largest;
   case bitc::COMDAT_SELECTION_KIND_NO_DUPLICATES:
     return Comdat::NoDuplicates;
   case bitc::COMDAT_SELECTION_KIND_SAME_SIZE:
     return Comdat::SameSize;
   }
 }
 
+static FastMathFlags getDecodedFastMathFlags(unsigned Val) {
+  FastMathFlags FMF;
+  if (0 != (Val & FastMathFlags::UnsafeAlgebra))
+    FMF.setUnsafeAlgebra();
+  if (0 != (Val & FastMathFlags::NoNaNs))
+    FMF.setNoNaNs();
+  if (0 != (Val & FastMathFlags::NoInfs))
+    FMF.setNoInfs();
+  if (0 != (Val & FastMathFlags::NoSignedZeros))
+    FMF.setNoSignedZeros();
+  if (0 != (Val & FastMathFlags::AllowReciprocal))
+    FMF.setAllowReciprocal();
+  return FMF;
+}
+
 static void UpgradeDLLImportExportLinkage(llvm::GlobalValue *GV, unsigned Val) {
   switch (Val) {
   case 5: GV->setDLLStorageClass(GlobalValue::DLLImportStorageClass); break;
   case 6: GV->setDLLStorageClass(GlobalValue::DLLExportStorageClass); break;
   }
 }
 
 namespace llvm {
case bitc::FUNC_CODE_INST_BINOP: {    // BINOP: [opval, ty, opval, opcode]
             cast<BinaryOperator>(I)->setHasNoUnsignedWrap(true);
         } else if (Opc == Instruction::SDiv ||
                    Opc == Instruction::UDiv ||
                    Opc == Instruction::LShr ||
                    Opc == Instruction::AShr) {
           if (Record[OpNum] & (1 << bitc::PEO_EXACT))
             cast<BinaryOperator>(I)->setIsExact(true);
         } else if (isa<FPMathOperator>(I)) {
-          FastMathFlags FMF;
-          if (0 != (Record[OpNum] & FastMathFlags::UnsafeAlgebra))
-            FMF.setUnsafeAlgebra();
-          if (0 != (Record[OpNum] & FastMathFlags::NoNaNs))
-            FMF.setNoNaNs();
-          if (0 != (Record[OpNum] & FastMathFlags::NoInfs))
-            FMF.setNoInfs();
-          if (0 != (Record[OpNum] & FastMathFlags::NoSignedZeros))
-            FMF.setNoSignedZeros();
-          if (0 != (Record[OpNum] & FastMathFlags::AllowReciprocal))
-            FMF.setAllowReciprocal();
+          FastMathFlags FMF = getDecodedFastMathFlags(Record[OpNum]);
           if (FMF.any())
             I->setFastMathFlags(FMF);
         }
 
       }
       break;
     }
     case bitc::FUNC_CODE_INST_CAST: {    // CAST: [opval, opty, destty, castopc]
case bitc::FUNC_CODE_INST_CMP:   // CMP: [opty, opval, opval, pred]
       // Existed to differentiate between icmp/fcmp and vicmp/vfcmp which were
       // both legal on vectors but had different behaviour.
     case bitc::FUNC_CODE_INST_CMP2: { // CMP2: [opty, opval, opval, pred]
       // FCmp/ICmp returning bool or vector of bool
 
       unsigned OpNum = 0;
       Value *LHS, *RHS;
       if (getValueTypePair(Record, OpNum, NextValueNo, LHS) ||
-          popValue(Record, OpNum, NextValueNo, LHS->getType(), RHS) ||
-          OpNum+1 != Record.size())
+          popValue(Record, OpNum, NextValueNo, LHS->getType(), RHS))
+        return Error("Invalid record");
+
+      unsigned PredVal = Record[OpNum];
+      bool IsFP = LHS->getType()->isFPOrFPVectorTy();
+      FastMathFlags FMF;
+      if (IsFP && Record.size() > OpNum+1)
+        FMF = getDecodedFastMathFlags(Record[++OpNum]);
+
+      if (OpNum+1 != Record.size())
         return Error("Invalid record");
 
       if (LHS->getType()->isFPOrFPVectorTy())
-        I = new FCmpInst((FCmpInst::Predicate)Record[OpNum], LHS, RHS);
+        I = new FCmpInst((FCmpInst::Predicate)PredVal, LHS, RHS);
       else
-        I = new ICmpInst((ICmpInst::Predicate)Record[OpNum], LHS, RHS);
+        I = new ICmpInst((ICmpInst::Predicate)PredVal, LHS, RHS);
+
+      if (FMF.any())
+        I->setFastMathFlags(FMF);
       InstructionList.push_back(I);
       break;
     }
 
     case bitc::FUNC_CODE_INST_RET: // RET: [opty,opval<optional>]
       {
         unsigned Size = Record.size();
         if (Size == 0) {

lib/Bitcode/Writer/BitcodeWriter.cpp

case Instruction::InsertElement:
     break;
   case Instruction::ShuffleVector:
     Code = bitc::FUNC_CODE_INST_SHUFFLEVEC;
     PushValueAndType(I.getOperand(0), InstID, Vals, VE);
     pushValue(I.getOperand(1), InstID, Vals, VE);
     pushValue(I.getOperand(2), InstID, Vals, VE);
     break;
   case Instruction::ICmp:
-  case Instruction::FCmp:
+  case Instruction::FCmp: {
     // compare returning Int1Ty or vector of Int1Ty
     Code = bitc::FUNC_CODE_INST_CMP2;
     PushValueAndType(I.getOperand(0), InstID, Vals, VE);
     pushValue(I.getOperand(1), InstID, Vals, VE);
     Vals.push_back(cast<CmpInst>(I).getPredicate());
+    uint64_t Flags = GetOptimizationFlags(&I);
+    if (Flags != 0)
+      Vals.push_back(Flags);
     break;
+  }
 
   case Instruction::Ret:
     {
       Code = bitc::FUNC_CODE_INST_RET;
       unsigned NumOperands = I.getNumOperands();
       if (NumOperands == 0)
         AbbrevToUse = FUNCTION_INST_RET_VOID_ABBREV;
       else if (NumOperands == 1) {

test/Bitcode/fcmp-fast.ll

+; RUN: llvm-as < %s | llvm-dis > %t0
+; RUN: opt -S < %s > %t1
+; RUN: diff %t0 %t1
+; RUN: FileCheck < %t1 %s
+
+; Make sure flags on fcmp instructions are serialized/deserialized properly.
+
+define i1 @foo(float %a, float %b, double %c, double %d) {
+  ; CHECK:   %plain = fcmp ueq float %a, %b
+  %plain = fcmp ueq float %a, %b
+  ; CHECK:   %fast = fcmp fast olt float %a, %b
+  %fast = fcmp fast olt float %a, %b
+  ; CHECK:   %nsz = fcmp nsz uge float %a, %b
+  %nsz = fcmp nsz uge float %a, %b
+  ; CHECK:   %nnan = fcmp nnan nsz oge double %c, %d
+  %nnan = fcmp nnan nsz oge double %c, %d
+
+  %dce1 = or i1 %plain, %fast
+  %dce2 = or i1 %dce1, %nsz
+  %dce3 = or i1 %dce2, %nnan
+
+  ret i1 %dce3
+}

unittests/IR/IRBuilderTest.cpp

TEST_F(IRBuilderTest, GetIntTy) {
   IntegerType *IntPtrTy = Builder.getIntPtrTy(*DL);
   unsigned IntPtrBitSize =  DL->getPointerSizeInBits(0);
   EXPECT_EQ(IntPtrTy, IntegerType::get(Ctx, IntPtrBitSize));
   delete DL;
 }
 
 TEST_F(IRBuilderTest, FastMathFlags) {
   IRBuilder<> Builder(BB);
-  Value *F;
-  Instruction *FDiv, *FAdd;
+  Value *F, *FC;
+  Instruction *FDiv, *FAdd, *FCmp;
 
   F = Builder.CreateLoad(GV);
   F = Builder.CreateFAdd(F, F);
 
   EXPECT_FALSE(Builder.getFastMathFlags().any());
   ASSERT_TRUE(isa<Instruction>(F));
   FAdd = cast<Instruction>(F);
   EXPECT_FALSE(FAdd->hasNoNaNs());
TEST_F(IRBuilderTest, FastMathFlags) {
   EXPECT_TRUE(Builder.getFastMathFlags().any());
   EXPECT_TRUE(Builder.getFastMathFlags().AllowReciprocal);
   ASSERT_TRUE(isa<Instruction>(F));
   FDiv = cast<Instruction>(F);
   EXPECT_TRUE(FDiv->hasAllowReciprocal());
 
   Builder.clearFastMathFlags();
 
+  FC = Builder.CreateFCmpOEQ(F, F);
+  ASSERT_TRUE(isa<Instruction>(FC));
+  FCmp = cast<Instruction>(FC);
+  EXPECT_FALSE(FCmp->hasAllowReciprocal());
+
+  FMF.clear();
+  FMF.setAllowReciprocal();
+  Builder.SetFastMathFlags(FMF);
+
+  FC = Builder.CreateFCmpOEQ(F, F);
+  EXPECT_TRUE(Builder.getFastMathFlags().any());
+  EXPECT_TRUE(Builder.getFastMathFlags().AllowReciprocal);
+  ASSERT_TRUE(isa<Instruction>(FC));
+  FCmp = cast<Instruction>(FC);
+  EXPECT_TRUE(FCmp->hasAllowReciprocal());
+
+  Builder.clearFastMathFlags();
+
   // To test a copy, make sure that a '0' and a '1' change state. 
   F = Builder.CreateFDiv(F, F);
   ASSERT_TRUE(isa<Instruction>(F));
   FDiv = cast<Instruction>(F);
   EXPECT_FALSE(FDiv->getFastMathFlags().any());
   FDiv->setHasAllowReciprocal(true);
   FAdd->setHasAllowReciprocal(false);
   FDiv->copyFastMathFlags(FAdd);