[FPEnv] Add pragma FP_CONTRACT support under strict FP.

[pengfei]

Summary

Support pragma FP_CONTRACT under strict FP.

[pengfei]

Remove dead code.

[craig.topper]

The title of this review is misleading. It should at least mention FPEnv, constrained intrinsics, or strict fp or something. Right now it sounds like FP_CONTRACT isn't supported at all.

Can we split most of the X86 changes into a separate patch? Most of it can be tested with fneg and constrained.fma.

[craig.topper]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

7090

Can you make the SDValue Result an argument of this and only capture 'this'. I don't like depending on reassigning Result.

[pengfei]

Address review comments.

[craig.topper]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

7061

Why is Result a reference? It's not modified is it? Don't use auto for parameter types. llvm coding style prefers auto to only be used when the type is easily assumed by someone reading the code.

[pengfei]

Address review comment.

[andrew.w.kaylor]

clang/lib/CodeGen/CGExprScalar.cpp

3381

You shouldn't just assume that MulOp is a constrained intrinsic. Cast to ConstrainedFPIntrinsic and use ConstrainedFPIntrinsic::getRoundingMode() and ConstrainedFPIntrinsic::getExceptionBehavior(). The cast will effectively assert that MulOp is a constrained intrisic. I think that should always be true.

3423

I don't think we should ever non-constrained create FMul instructions if Builder is in FP constrained mode, but you should assert that somewhere above. Maybe move this block above line 3409 and add:

assert(LHSBinOp->getOpcode() != llvm::Instruction::FMul && RHSBinOp->getOpcode() != llvm::Instruction::FMul);

clang/test/CodeGen/constrained-math-builtins.c

160

I'd like to see a test that verifies the calls generated in the function and specifically a test that verifies that the constrained fneg is generated if needed.

llvm/docs/LangRef.rst

16094

s/specifie/specify

s/the exception behavior/the rounding mode and exception behavior

16104

missing metadata arguments

llvm/include/llvm/CodeGen/BasicTTIImpl.h

1515

I don't think that matters. The cost calculation here is a conservative estimate based on the cost if we are unable to generate an FMA instruction. So a constrained fmuladd that can't be lowered to FMA will be lower the same way a contrained mul followed by a constrained add would be.

llvm/include/llvm/CodeGen/ISDOpcodes.h

355

Something is wrong with this comment. I'm not sure what it's trying to say but the grammar is wrong.

After looking through the rest of the code, I think I understand what's going on. I think we need a verbose comment to explain it. Here's my suggestion

FMULADD/STRICT_FMULADD -- These are intermediate opcodes used to handle the constrained.fmuladd intrinsic. The FMULADD opcode only exists because it is required for correct macro expansion and default handling (which is never reached). There should never be a node with ISD::FMULADD. The STRICT_FMULADD opcode is used to allow selectionDAGBuilder::visitConstrainedFPIntrinsic to determine (based on TargetOptions and target cost information) whether the constrained.fmuladd intrinsic should be lowered to FMA or separate FMUL and FADD operations.

Having thought through that, however, it strikes me as a lot of overhead. Can we just add special handling for the constrained.fmuladd intrinsic and make the decision then to create either a STRICT_FMA node or separate STRICT_FMUL and STRICT_FADD?

The idea that ISD::FMULADD is going to exist as a defined opcode but we never intend to add any support for handling it is particularly bad.

[cameron.mcinally]

clang/lib/CodeGen/CGExprScalar.cpp

3436

I don't think it's safe to fuse a FMUL and FADD if the intermediate rounding isn't exactly the same as those individual operations. FMULADD doesn't guarantee that, does it?

[cameron.mcinally]

clang/lib/CodeGen/CGExprScalar.cpp

3436

To be clear, we could miss very-edge-case overflow/underflow exceptions.

[cameron.mcinally]

clang/lib/CodeGen/CGExprScalar.cpp

3436

Ah, but I see C/C++ FP_CONTRACT allows the exceptions to be optimized away. Sorry for the noise.

[andrew.w.kaylor]

clang/lib/CodeGen/CGExprScalar.cpp

3436

We've talked about this before but I don't think we ever documented a decision as to whether we want to allow constrained intrinsics and fast math flags to be combined. This patch moves that decision into clang's decision to generate this intrinsic or not.

I think it definitely makes sense in the case of fp contraction, because even if a user cares about value safety they might want FMA, which is theorectically more accurate than the separate values even though it produces a different value. This is consistent with gcc (which produces FMA under "-ffp-contract=fast -fno-fast-math") and icc (which produced FMA under "-fp-model strict -fma").

For the record, I also think it makes sense to use nnan, ninf, and nsz with constrained intrinsics.

[pengfei]

Address review comments.

[pengfei]

clang/lib/CodeGen/CGExprScalar.cpp

3381

I prefer to reuse the operands from the fmul intrinsic here. 1). fmuladd always has the same exception/rounding mode with fmul. 2). the function getRoundingMode/getExceptionBehavior just return a enum value. We need more code to turn them into Value type.

3423

Add assertion in line 3380. We only need to check once there.

llvm/test/TableGen/GlobalISelEmitter-input-discard.td

18 ↗

(On Diff #238769)

It's strange the number is affected. I haven't found any cause.

[pengfei]

Remove unnecessary comment.

[cameron.mcinally]

clang/lib/CodeGen/CGExprScalar.cpp

3436

You had me until:

For the record, I also think it makes sense to use nnan, ninf, and nsz with constrained intrinsics.

To be clear, we'd need them for the fast case, but I don't see a lot of value for the strict case.

We definitely want reassoc/recip/etc for the optimized but trap-safe case, so that's enough to require FMF flags on constrained intrinsics alone.

We should probably break this conversation out into an llvm-dev thread...

[rjmccall]

Remember that the design is that constrained intrinsics must be used whenever *any* code in the function is constrained. It is not unreasonable that part of the function might be constrained and the rest subject to fast-math; it'd be a shame if the intrinsics couldn't even express that.

[andrew.w.kaylor]

clang/lib/CodeGen/CGExprScalar.cpp

3436

I agree about starting an llvm-dev thread. I'll send something out unless you've already done so by the time I finish typing it.

[craig.topper]

clang/lib/CodeGen/CGExprScalar.cpp

3381

Doesn't this need to be CreateConstrainedFPCall so that the strictfp attribute is added? That will take care of adding the metadata operands too.

[kpn]

clang/lib/CodeGen/CGExprScalar.cpp

3381

Is this code tested? I ran into a bug yesterday where CreateCall was used with a constrained intrinsic and the Instruction class blew up because the function signature was wrong. I wasn't passing in the metadata arguments.

So, yes, it should be, and it would might make sense for the patch to have test coverage that catches any other cases of this.

[craig.topper]

clang/lib/CodeGen/CGExprScalar.cpp

3381

This code is copying the metadata arguments from the fmul intrinsic, MulOp. that’s the getOperand(2) and getOperand(3).

[kpn]

clang/lib/CodeGen/CGExprScalar.cpp

3381

Ah, yes, thanks. Your comment about the attribute is still valid, though. And, yes, using CreateConstrainedFPCall() is the easiest way to fix the attribute.

[pengfei]

Address review comment.

[pengfei]

clang/lib/CodeGen/CGExprScalar.cpp

3381

Yes, it's the best choice. Thanks!

[craig.topper]

LGTM

[jhenderson]

llvm/docs/LangRef.rst

16065

This underline isn't long enough and is breaking the sphinx build bot. Please fix.

[pengfei]

llvm/docs/LangRef.rst

16065

Thanks! I'll fix it soon.

[Allen]

clang/lib/CodeGen/CGExprScalar.cpp

3382

Sorry, I'm not familiar with the optimization of the clang front end.
I'd like to ask, is this optimization supposed to assume that all the backends have instructions like Fmuladd?

[pengfei]

clang/lib/CodeGen/CGExprScalar.cpp

3382

No, it is a flexible intrinsic, which allows backends to choose their best approach. It can be either interpretered as mul + add or fma. It represents user doesn't care the differece between them.

clang/lib/CodeGen/CGExprScalar.cpp

return CGF.EmitCheckedInBoundsGEP(pointer, index, isSigned, isSubtraction,
                                     op.E->getExprLoc(), "add.ptr");
 }
 
 // Construct an fmuladd intrinsic to represent a fused mul-add of MulOp and
 // Addend. Use negMul and negAdd to negate the first operand of the Mul or
 // the add operand respectively. This allows fmuladd to represent a*b-c, or
 // c-a*b. Patterns in LLVM should catch the negated forms and translate them to
 // efficient operations.
-static Value* buildFMulAdd(llvm::BinaryOperator *MulOp, Value *Addend,
+static Value* buildFMulAdd(llvm::Instruction *MulOp, Value *Addend,
                            const CodeGenFunction &CGF, CGBuilderTy &Builder,
                            bool negMul, bool negAdd) {
   assert(!(negMul && negAdd) && "Only one of negMul and negAdd should be set.");
 
   Value *MulOp0 = MulOp->getOperand(0);
   Value *MulOp1 = MulOp->getOperand(1);
   if (negMul)
     MulOp0 = Builder.CreateFNeg(MulOp0, "neg");
   if (negAdd)
     Addend = Builder.CreateFNeg(Addend, "neg");
 
-  Value *FMulAdd = Builder.CreateCall(
+  Value *FMulAdd = nullptr;
+  if (Builder.getIsFPConstrained())
+    FMulAdd = Builder.CreateCall(
+        CGF.CGM.getIntrinsic(llvm::Intrinsic::experimental_constrained_fmuladd,
+                             Addend->getType()),
+        {MulOp0, MulOp1, Addend, MulOp->getOperand(2), MulOp->getOperand(3)});
+  else
+    FMulAdd = Builder.CreateCall(
         CGF.CGM.getIntrinsic(llvm::Intrinsic::fmuladd, Addend->getType()),
         {MulOp0, MulOp1, Addend});
-   MulOp->eraseFromParent();
 
-   return FMulAdd;
 }
 
 // Check whether it would be legal to emit an fmuladd intrinsic call to
 // represent op and if so, build the fmuladd.
 //
 // Checks that (a) the operation is fusable, and (b) -ffp-contract=on.
 // Does NOT check the type of the operation - it's assumed that this function
 // will be called from contexts where it's known that the type is contractable.
if (LHSBinOp->getOpcode() == llvm::Instruction::FMul &&
       return buildFMulAdd(LHSBinOp, op.RHS, CGF, Builder, false, isSub);
   }
   if (auto *RHSBinOp = dyn_cast<llvm::BinaryOperator>(op.RHS)) {
     if (RHSBinOp->getOpcode() == llvm::Instruction::FMul &&
         RHSBinOp->use_empty())
       return buildFMulAdd(RHSBinOp, op.LHS, CGF, Builder, isSub, false);
   }
 
+  if (Builder.getIsFPConstrained()) {
+    if (auto *LHSBinOp = dyn_cast<llvm::CallBase>(op.LHS)) {
+      if (LHSBinOp->getIntrinsicID() ==
+              llvm::Intrinsic::experimental_constrained_fmul &&
+          LHSBinOp->use_empty())
+        return buildFMulAdd(LHSBinOp, op.RHS, CGF, Builder, false, isSub);
+    }
+    if (auto *RHSBinOp = dyn_cast<llvm::CallBase>(op.RHS)) {
+      if (RHSBinOp->getIntrinsicID() ==
+              llvm::Intrinsic::experimental_constrained_fmul &&
+          RHSBinOp->use_empty())
+        return buildFMulAdd(RHSBinOp, op.LHS, CGF, Builder, isSub, false);
+    }
+  }
+
   return nullptr;
 }
 
 Value *ScalarExprEmitter::EmitAdd(const BinOpInfo &op) {
   if (op.LHS->getType()->isPointerTy() ||
       op.RHS->getType()->isPointerTy())
     return emitPointerArithmetic(CGF, op, CodeGenFunction::NotSubtraction);
 

clang/test/CodeGen/constrained-math-builtins.c

// CHECK: declare x86_fp80 @llvm.experimental.constrained.sqrt.f80(x86_fp80, metadata, metadata)
 
   __builtin_trunc(f);      __builtin_truncf(f);     __builtin_truncl(f);
 
 // CHECK: declare double @llvm.experimental.constrained.trunc.f64(double, metadata)
 // CHECK: declare float @llvm.experimental.constrained.trunc.f32(float, metadata)
 // CHECK: declare x86_fp80 @llvm.experimental.constrained.trunc.f80(x86_fp80, metadata)
 };
 
+#pragma STDC FP_CONTRACT ON
+void bar(float f) {
+  f * f + f;
+  (double)f * f + f;
+  (long double)f * f + f;
+
+// CHECK: declare float @llvm.experimental.constrained.fmuladd.f32(float, float, float, metadata, metadata)
+// CHECK: declare double @llvm.experimental.constrained.fmuladd.f64(double, double, double, metadata, metadata)
+// CHECK: declare x86_fp80 @llvm.experimental.constrained.fmuladd.f80(x86_fp80, x86_fp80, x86_fp80, metadata, metadata)
+};

llvm/docs/LangRef.rst

 - "``uno``": yields ``true`` if either operand is a NAN.
 
 The quiet comparison operation performed by
 '``llvm.experimental.constrained.fcmp``' will only raise an exception
 if either operand is a SNAN.  The signaling comparison operation
 performed by '``llvm.experimental.constrained.fcmps``' will raise an
 exception if either operand is a NAN (QNAN or SNAN).
 
+'``llvm.experimental.constrained.fmuladd``' Intrinsic
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Syntax:
+"""""""
+
+::
+
+      declare <type>
+      @llvm.experimental.constrained.fmuladd(<type> <op1>, <type> <op2>,
+                                             <type> <op3>,
+                                             metadata <rounding mode>,
+                                             metadata <exception behavior>)
+
+Overview:
+"""""""""
+
+The '``llvm.experimental.constrained.fmuladd``' intrinsic represents
+multiply-add expressions that can be fused if the code generator determines
+that (a) the target instruction set has support for a fused operation,
+and (b) that the fused operation is more efficient than the equivalent,
+separate pair of mul and add instructions.
+
+Arguments:
+""""""""""
+
+The first three arguments to the '``llvm.experimental.constrained.fmuladd``'
+intrinsic must be floating-point or vector of floating-point values.
+All three arguments must have identical types.
+
+The fourth and fifth arguments specifie the exception behavior as described
+above.
+
+Semantics:
+""""""""""
+
+The expression:
+
+::
+
+      %0 = call float @llvm.experimental.constrained.fmuladd.f32(%a, %b, %c)
+
+is equivalent to the expression:
+
+::
+
+      %0 = call float @llvm.experimental.constrained.fmul.f32(%a, %b)
+      %1 = call float @llvm.experimental.constrained.fadd.f32(%0, %c)
+
+except that it is unspecified whether rounding will be performed between the
+multiplication and addition steps. Fusion is not guaranteed, even if the target
+platform supports it.
+If a fused multiply-add is required, the corresponding
+:ref:`llvm.experimental.constrained.fma <int_fma>` intrinsic function should be
+used instead.
+This never sets errno, just as '``llvm.experimental.constrained.fma.*``'.
+
 Constrained libm-equivalent Intrinsics
 --------------------------------------
 
 In addition to the basic floating-point operations for which constrained
 intrinsics are described above, there are constrained versions of various
 operations which provide equivalent behavior to a corresponding libm function.
 These intrinsics allow the precise behavior of these operations with respect to
 rounding mode and exception behavior to be controlled.

llvm/include/llvm/CodeGen/BasicTTIImpl.h

case Intrinsic::pow:
       ISDs.push_back(ISD::FPOW);
       break;
     case Intrinsic::fma:
       ISDs.push_back(ISD::FMA);
       break;
     case Intrinsic::fmuladd:
       ISDs.push_back(ISD::FMA);
       break;
+    case Intrinsic::experimental_constrained_fmuladd:
+      ISDs.push_back(ISD::STRICT_FMA);
+      break;
     // FIXME: We should return 0 whenever getIntrinsicCost == TCC_Free.
     case Intrinsic::lifetime_start:
     case Intrinsic::lifetime_end:
     case Intrinsic::sideeffect:
       return 0;
     case Intrinsic::masked_store:
       return ConcreteTTI->getMaskedMemoryOpCost(Instruction::Store, Tys[0], 0,
                                                 0);
unsigned getIntrinsicInstrCost(
     if (MinCustomCostI != CustomCost.end())
       return *MinCustomCostI;
 
     // If we can't lower fmuladd into an FMA estimate the cost as a floating
     // point mul followed by an add.
     if (IID == Intrinsic::fmuladd)
       return ConcreteTTI->getArithmeticInstrCost(BinaryOperator::FMul, RetTy) +
              ConcreteTTI->getArithmeticInstrCost(BinaryOperator::FAdd, RetTy);
+    // FIXME: Is constrained intrinsic' cost equal to it's no strict one?
+    if (IID == Intrinsic::experimental_constrained_fmuladd)
+      return ConcreteTTI->getIntrinsicCost(
+                 Intrinsic::experimental_constrained_fmul, RetTy, Tys,
+                 nullptr) +
+             ConcreteTTI->getIntrinsicCost(
+                 Intrinsic::experimental_constrained_fadd, RetTy, Tys, nullptr);
 
     // Else, assume that we need to scalarize this intrinsic. For math builtins
     // this will emit a costly libcall, adding call overhead and spills. Make it
     // very expensive.
     if (RetTy->isVectorTy()) {
       unsigned ScalarizationCost =
           ((ScalarizationCostPassed != std::numeric_limits<unsigned>::max())
                ? ScalarizationCostPassed

llvm/include/llvm/CodeGen/ISDOpcodes.h

enum NodeType {
     STRICT_FP_EXTEND,
 
     /// STRICT_FSETCC/STRICT_FSETCCS - Constrained versions of SETCC, used
     /// for floating-point operands only.  STRICT_FSETCC performs a quiet
     /// comparison operation, while STRICT_FSETCCS performs a signaling
     /// comparison operation.
     STRICT_FSETCC, STRICT_FSETCCS,
 
+    /// FMULADD/STRICT_FMULADD - A intermediate node, made functions handle
+    /// constrained fmuladd the same as other constrained intrinsics.
+    FMULADD, STRICT_FMULADD,
+
     /// FMA - Perform a * b + c with no intermediate rounding step.
     FMA,
 
     /// FMAD - Perform a * b + c, while getting the same result as the
     /// separately rounded operations.
     FMAD,
 
     /// FCOPYSIGN(X, Y) - Return the value of X with the sign of Y.  NOTE: This

llvm/include/llvm/IR/ConstrainedOps.def

 FUNCTION(pow,             2, 1, experimental_constrained_pow,        FPOW)
 FUNCTION(powi,            2, 1, experimental_constrained_powi,       FPOWI)
 FUNCTION(rint,            1, 1, experimental_constrained_rint,       FRINT)
 FUNCTION(round,           1, 0, experimental_constrained_round,      FROUND)
 FUNCTION(sin,             1, 1, experimental_constrained_sin,        FSIN)
 FUNCTION(sqrt,            1, 1, experimental_constrained_sqrt,       FSQRT)
 FUNCTION(trunc,           1, 0, experimental_constrained_trunc,      FTRUNC)
 
+// This is definition for fmuladd intrinsic function, that is converted into
+// constrained FMA or FMUL + FADD intrinsics.
+FUNCTION(fmuladd,         3, 1, experimental_constrained_fmuladd,    FMULADD)
+
 #undef INSTRUCTION
 #undef FUNCTION
 #undef CMP_INSTRUCTION

llvm/include/llvm/IR/Intrinsics.td

let IntrProperties = [IntrInaccessibleMemOnly, IntrWillReturn] in {
 
   def int_experimental_constrained_fma : Intrinsic<[ llvm_anyfloat_ty ],
                                                     [ LLVMMatchType<0>,
                                                       LLVMMatchType<0>,
                                                       LLVMMatchType<0>,
                                                       llvm_metadata_ty,
                                                       llvm_metadata_ty ]>;
 
+  def int_experimental_constrained_fmuladd : Intrinsic<[ llvm_anyfloat_ty ],
+                                                       [ LLVMMatchType<0>,
+                                                         LLVMMatchType<0>,
+                                                         LLVMMatchType<0>,
+                                                         llvm_metadata_ty,
+                                                         llvm_metadata_ty ]>;
+
   def int_experimental_constrained_fptosi : Intrinsic<[ llvm_anyint_ty ],
                                                     [ llvm_anyfloat_ty,
                                                       llvm_metadata_ty ]>;
 
   def int_experimental_constrained_fptoui : Intrinsic<[ llvm_anyint_ty ],
                                                     [ llvm_anyfloat_ty,
                                                       llvm_metadata_ty ]>;
 

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

#include "llvm/IR/ConstrainedOps.def"
   case ISD::STRICT_FSETCC:
   case ISD::STRICT_FSETCCS: {
     auto *FPCmp = dyn_cast<ConstrainedFPCmpIntrinsic>(&FPI);
     Opers.push_back(DAG.getCondCode(getFCmpCondCode(FPCmp->getPredicate())));
     break;
   }
   }
 
-  SDVTList VTs = DAG.getVTList(ValueVTs);
-  SDValue Result = DAG.getNode(Opcode, sdl, VTs, Opers);
-
+  auto pushOutChain = [this](auto &Result, auto EB) {
     assert(Result.getNode()->getNumValues() == 2);
 
     // Push node to the appropriate list so that future instructions can be
     // chained up correctly.
     SDValue OutChain = Result.getValue(1);
-  switch (FPI.getExceptionBehavior().getValue()) {
+    switch (EB) {
     case fp::ExceptionBehavior::ebIgnore:
       // The only reason why ebIgnore nodes still need to be chained is that
       // they might depend on the current rounding mode, and therefore must
       // not be moved across instruction that may change that mode.
       LLVM_FALLTHROUGH;
     case fp::ExceptionBehavior::ebMayTrap:
       // These must not be moved across calls or instructions that may change
       // floating-point exception masks.
       PendingConstrainedFP.push_back(OutChain);
       break;
     case fp::ExceptionBehavior::ebStrict:
       // These must not be moved across calls or instructions that may change
       // floating-point exception masks or read floating-point exception flags.
       // In addition, they cannot be optimized out even if unused.
       PendingConstrainedFPStrict.push_back(OutChain);
       break;
     }
+  };
+
+  SDVTList VTs = DAG.getVTList(ValueVTs);
+  fp::ExceptionBehavior EB = FPI.getExceptionBehavior().getValue();
+  SDValue Result;
+
+  if (Opcode == ISD::STRICT_FMULADD) {
+    Opcode = ISD::STRICT_FMA;
+    // Break fmuladd into fmul and fadd.
+    if (TM.Options.AllowFPOpFusion == FPOpFusion::Strict ||
+        !TLI.isFMAFasterThanFMulAndFAdd(DAG.getMachineFunction(),
+                                        ValueVTs[0])) {
+      Opers.pop_back();
+      Result = DAG.getNode(ISD::STRICT_FMUL, sdl, VTs, Opers);
+      pushOutChain(Result, EB);
+      Opcode = ISD::STRICT_FADD;
+      Opers.clear();
+      Opers.push_back(Result.getValue(1));
+      Opers.push_back(Result.getValue(0));
+      Opers.push_back(getValue(FPI.getArgOperand(2)));
+    }
+  }
+
+  Result = DAG.getNode(Opcode, sdl, VTs, Opers);
+  pushOutChain(Result, EB);
 
   SDValue FPResult = Result.getValue(0);
   setValue(&FPI, FPResult);
 }
 
 std::pair<SDValue, SDValue>
 SelectionDAGBuilder::lowerInvokable(TargetLowering::CallLoweringInfo &CLI,
                                     const BasicBlock *EHPadBB) {

llvm/test/CodeGen/X86/fp-intrinsics-fma.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -O3 -mtriple=x86_64-pc-linux < %s | FileCheck %s --check-prefixes=COMMON,NOFMA
 ; RUN: llc -O3 -mtriple=x86_64-pc-linux -mattr=+fma < %s | FileCheck %s --check-prefixes=COMMON,FMA
 ; RUN: llc -O3 -mtriple=x86_64-pc-linux -mattr=+avx512f < %s | FileCheck %s --check-prefixes=COMMON,FMA
 
+; Verify constrained fmul and fadd aren't fused.
+define float @f11(float %0, float %1, float %2) #0 {
+; NOFMA-LABEL: f11:
+; NOFMA:       # %bb.0: # %entry
+; NOFMA-NEXT:    mulss %xmm1, %xmm0
+; NOFMA-NEXT:    addss %xmm2, %xmm0
+; NOFMA-NEXT:    retq
+;
+; FMA-LABEL: f11:
+; FMA:       # %bb.0: # %entry
+; FMA-NEXT:    vmulss %xmm1, %xmm0, %xmm0
+; FMA-NEXT:    vaddss %xmm2, %xmm0, %xmm0
+; FMA-NEXT:    retq
+entry:
+  %3 = call float @llvm.experimental.constrained.fmul.f32(float %0, float %1,
+                                                          metadata !"round.dynamic",
+                                                          metadata !"fpexcept.strict") #0
+  %4 = call float @llvm.experimental.constrained.fadd.f32(float %3, float %2,
+                                                          metadata !"round.dynamic",
+                                                          metadata !"fpexcept.strict") #0
+  ret float %4
+}
+
+; Verify constrained fmul and fadd aren't fused.
+define double @f12(double %0, double %1, double %2) #0 {
+; NOFMA-LABEL: f12:
+; NOFMA:       # %bb.0: # %entry
+; NOFMA-NEXT:    mulsd %xmm1, %xmm0
+; NOFMA-NEXT:    addsd %xmm2, %xmm0
+; NOFMA-NEXT:    retq
+;
+; FMA-LABEL: f12:
+; FMA:       # %bb.0: # %entry
+; FMA-NEXT:    vmulsd %xmm1, %xmm0, %xmm0
+; FMA-NEXT:    vaddsd %xmm2, %xmm0, %xmm0
+; FMA-NEXT:    retq
+entry:
+  %3 = call double @llvm.experimental.constrained.fmul.f64(double %0, double %1,
+                                                           metadata !"round.dynamic",
+                                                           metadata !"fpexcept.strict") #0
+  %4 = call double @llvm.experimental.constrained.fadd.f64(double %3, double %2,
+                                                           metadata !"round.dynamic",
+                                                           metadata !"fpexcept.strict") #0
+  ret double %4
+}
+
+; Verify that fmuladd(3.5) isn't simplified when the rounding mode is
+; unknown.
+define float @f15() #0 {
+; NOFMA-LABEL: f15:
+; NOFMA:       # %bb.0: # %entry
+; NOFMA-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
+; NOFMA-NEXT:    movaps %xmm1, %xmm0
+; NOFMA-NEXT:    mulss %xmm1, %xmm0
+; NOFMA-NEXT:    addss %xmm1, %xmm0
+; NOFMA-NEXT:    retq
+;
+; FMA-LABEL: f15:
+; FMA:       # %bb.0: # %entry
+; FMA-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; FMA-NEXT:    vfmadd213ss {{.*#+}} xmm0 = (xmm0 * xmm0) + xmm0
+; FMA-NEXT:    retq
+entry:
+  %result = call float @llvm.experimental.constrained.fmuladd.f32(
+                                               float 3.5,
+                                               float 3.5,
+                                               float 3.5,
+                                               metadata !"round.dynamic",
+                                               metadata !"fpexcept.strict") #0
+  ret float %result
+}
+
+; Verify that fmuladd(42.1) isn't simplified when the rounding mode is
+; unknown.
+define double @f16() #0 {
+; NOFMA-LABEL: f16:
+; NOFMA:       # %bb.0: # %entry
+; NOFMA-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero
+; NOFMA-NEXT:    movapd %xmm1, %xmm0
+; NOFMA-NEXT:    mulsd %xmm1, %xmm0
+; NOFMA-NEXT:    addsd %xmm1, %xmm0
+; NOFMA-NEXT:    retq
+;
+; FMA-LABEL: f16:
+; FMA:       # %bb.0: # %entry
+; FMA-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
+; FMA-NEXT:    vfmadd213sd {{.*#+}} xmm0 = (xmm0 * xmm0) + xmm0
+; FMA-NEXT:    retq
+entry:
+  %result = call double @llvm.experimental.constrained.fmuladd.f64(
+                                               double 42.1,
+                                               double 42.1,
+                                               double 42.1,
+                                               metadata !"round.dynamic",
+                                               metadata !"fpexcept.strict") #0
+  ret double %result
+}
+
 ; Verify that fma(3.5) isn't simplified when the rounding mode is
 ; unknown.
 define float @f17() #0 {
 ; NOFMA-LABEL: f17:
 ; NOFMA:       # %bb.0: # %entry
 ; NOFMA-NEXT:    pushq %rax
 ; NOFMA-NEXT:    .cfi_def_cfa_offset 16
 ; NOFMA-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
%result = call double @llvm.experimental.constrained.fma.f64(
                                                double 42.1,
                                                metadata !"round.dynamic",
                                                metadata !"fpexcept.strict") #0
   ret double %result
 }
 
 attributes #0 = { strictfp }
 
+declare float @llvm.experimental.constrained.fmul.f32(float, float, metadata, metadata)
+declare float @llvm.experimental.constrained.fadd.f32(float, float, metadata, metadata)
+declare double @llvm.experimental.constrained.fmul.f64(double, double, metadata, metadata)
+declare double @llvm.experimental.constrained.fadd.f64(double, double, metadata, metadata)
 declare float @llvm.experimental.constrained.fma.f32(float, float, float, metadata, metadata)
 declare double @llvm.experimental.constrained.fma.f64(double, double, double, metadata, metadata)
+declare float @llvm.experimental.constrained.fmuladd.f32(float, float, float, metadata, metadata)
+declare double @llvm.experimental.constrained.fmuladd.f64(double, double, double, metadata, metadata)