Added intrinsics for access to FP environment

[sepavloff]

Summary

The change implements intrinsics 'get_fpenv', 'set_fpenv' and 'reset_fpenv'.
They are used to read floating-point environment, set it or reset to
some default state. They do the same actions as C library functions
'fegetenv' and 'fesetenv'. By default these intrinsics are lowered to calls
to these functions.

The new intrinsics specify FP environment as a value of integer type, it
is convenient of most targets where the FP state is a content of some
register. Some targets however use long representations. On X86 the size
of FP environment is 256 bits, and even half of this size is not a legal
ibteger type. To facilitate legalization in such cases, two sets of DAG
nodes is used. Nodes GET_FPENV and SET_FPENV are used when FP
environment may be represented by a legal integer type. Nodes
GET_FPENV_MEM and SET_FPENV_MEM consider FP environment as a region in
memory, much like fesetenv and fegetenv do. They are used when
target has long representation for floationg-point state.

[kpn]

I don't see the need. Changing the FP environment in a mixed environment program is the responsibility of the programmer, and standard calls already exist for this.

[efriedma]

It is expected that targets will implement custom lowering to proper machine instructions for better performance.

Do you have any performance numbers?

[sepavloff]

In D71742#1793243, @kpn wrote:

I don't see the need. Changing the FP environment in a mixed environment program is the responsibility of the programmer, and standard calls already exist for this.

This is about inlining. In the code like this:

double f1(double x, double y) {
  return x + y;
}
double f2(double x, double y) {
  #pragma STDC FENV_ACCESS ON
  ...
  return f1(x, y);
}

compiler might inline call to f1 in f2. However the inlined function f1 expects default FP environment but is called in some other one.

In D71742#1793447, @efriedma wrote:

It is expected that targets will implement custom lowering to proper machine instructions for better performance.

Do you have any performance numbers?

I don't have them. But there are some considerations why optimization may be desirable:

• For some targets FP environment is a content of 1 or 2 registers. It this case custom lowering can store the environment in registers and avoid expenses for function calls and memory operations.
• Some targets encode FP environment in instructions. In this case these intrinsics shoul not produce any code.
• Probably we need to reset FP environment before calls to external functions in strictfp function and restore it upon return. In this case number of calls to such intrinsics may be large enough.

[kpn]

In D71742#1794638, @sepavloff wrote:

In D71742#1793243, @kpn wrote:

I don't see the need. Changing the FP environment in a mixed environment program is the responsibility of the programmer, and standard calls already exist for this.

This is about inlining. In the code like this:

double f1(double x, double y) {
  return x + y;
}
double f2(double x, double y) {
  #pragma STDC FENV_ACCESS ON
  ...
  return f1(x, y);
}

compiler might inline call to f1 in f2. However the inlined function f1 expects default FP environment but is called in some other one.

Nothing here changes my statement. The compiler does _not_ change the FP environment because of the #pragma. So f1() here would behave the same whether it was inlined or not.

[sepavloff]

In D71742#1794901, @kpn wrote:

In D71742#1794638, @sepavloff wrote:

In D71742#1793243, @kpn wrote:

I don't see the need. Changing the FP environment in a mixed environment program is the responsibility of the programmer, and standard calls already exist for this.

This is about inlining. In the code like this:

double f1(double x, double y) {
  return x + y;
}
double f2(double x, double y) {
  #pragma STDC FENV_ACCESS ON
  ...
  return f1(x, y);
}

compiler might inline call to f1 in f2. However the inlined function f1 expects default FP environment but is called in some other one.

Nothing here changes my statement. The compiler does _not_ change the FP environment because of the #pragma. So f1() here would behave the same whether it was inlined or not.

When f1 is defined, no pragma is in act, so its body is executed in default FP environment. f2 contains #pragma, so FP environment in its body may differ from the default. When f1 is inlined into f2, the body of f1 becomes a part of the body of f2. Basic blocks of f1 would be executed in the environment, set in f2. To keep semantics of f1 we must execute its BBs in the default environment.

[craig.topper]

In D71742#1795013, @sepavloff wrote:

In D71742#1794901, @kpn wrote:

In D71742#1794638, @sepavloff wrote:

In D71742#1793243, @kpn wrote:

I don't see the need. Changing the FP environment in a mixed environment program is the responsibility of the programmer, and standard calls already exist for this.

This is about inlining. In the code like this:

double f1(double x, double y) {
  return x + y;
}
double f2(double x, double y) {
  #pragma STDC FENV_ACCESS ON
  ...
  return f1(x, y);
}

compiler might inline call to f1 in f2. However the inlined function f1 expects default FP environment but is called in some other one.

Nothing here changes my statement. The compiler does _not_ change the FP environment because of the #pragma. So f1() here would behave the same whether it was inlined or not.

When f1 is defined, no pragma is in act, so its body is executed in default FP environment. f2 contains #pragma, so FP environment in its body may differ from the default. When f1 is inlined into f2, the body of f1 becomes a part of the body of f2. Basic blocks of f1 would be executed in the environment, set in f2. To keep semantics of f1 we must execute its BBs in the default environment.

I don’t see how inlining is any different than f2 calling f1 without it being inlined. f1 would execute with the f2 environment in that case too.

[arsenm]

llvm/docs/LangRef.rst

25370

Why a pointer? Why not an i64 argument? The lowering of this through memory seems problematic for any target that doesn't implement this as the fe libcalls

[uweigand]

In D71742#1795014, @craig.topper wrote:

In D71742#1795013, @sepavloff wrote:

In D71742#1794901, @kpn wrote:

In D71742#1794638, @sepavloff wrote:

In D71742#1793243, @kpn wrote:

I don't see the need. Changing the FP environment in a mixed environment program is the responsibility of the programmer, and standard calls already exist for this.

This is about inlining. In the code like this:

double f1(double x, double y) {
  return x + y;
}
double f2(double x, double y) {
  #pragma STDC FENV_ACCESS ON
  ...
  return f1(x, y);
}

compiler might inline call to f1 in f2. However the inlined function f1 expects default FP environment but is called in some other one.

Nothing here changes my statement. The compiler does _not_ change the FP environment because of the #pragma. So f1() here would behave the same whether it was inlined or not.

When f1 is defined, no pragma is in act, so its body is executed in default FP environment. f2 contains #pragma, so FP environment in its body may differ from the default. When f1 is inlined into f2, the body of f1 becomes a part of the body of f2. Basic blocks of f1 would be executed in the environment, set in f2. To keep semantics of f1 we must execute its BBs in the default environment.

I don’t see how inlining is any different than f2 calling f1 without it being inlined. f1 would execute with the f2 environment in that case too.

Exactly. It is the responsibility of the programmer to ensure f1 is being called with the correct environment set. C11 7.6.1.2 explicitly says:

If part of a program [...] runs under non-default mode settings, but was translated with the state for the FENV_ACCESS pragma ‘‘off’’, the behavior is undefined.

The only issue for the compiler w.r.t. inlining is the case where the programmer writes correct code in f2, i.e. resets the environment to the default state before calling f1, but then f1 is inlined into f2. In this scenario, the compiler must ensure to preserve correctness by not scheduling any part of the inlined f1 to before the instruction in f2 that resets the environment. This can be achieved e.g. by the inliner replacing all regular FP instructions with constrained intrinsics (which may specify the default mode).

[sepavloff]

Updated patch

• Rebased,
• Used more consistent naming.

[jdoerfert]

llvm/include/llvm/IR/Intrinsics.td

1074

[Drive by][Unrelated] nosync, nofree missing. Should be readonly or writeonly I suspect. Arguments can be nocapture.

[arsenm]

llvm/include/llvm/IR/Intrinsics.td

1070–1072

Using a pointer for this is problematic, and one with a hardcoded 0 address space doubly so

[sepavloff]

Updated patch

• fixed some formatting issues,
• added missed attrubutes to the new intrinsics,
• pointer arguments support address spaces.

[sepavloff]

llvm/include/llvm/IR/Intrinsics.td

1070–1072

Changed type to llvm_anyptr_ty, which must support arbitrary address spaces.

1074

Fixed, thank you.

[arsenm]

llvm/include/llvm/IR/Intrinsics.td

1070–1072

Accepting any address space is only a half-fix. Why not make this return llvm_anyint_ty, and define it as zext or truncated to the expected target size in the backend? This wouldn't require lowering to introduce stack usage for example for something that's usually read directly out of a register

[sepavloff]

llvm/include/llvm/IR/Intrinsics.td

1070–1072

Why not make this return llvm_anyint_ty, and define it as zext or truncated to the expected target size in the backend?

In this case X86 represents FP environment as i256. It is not clear how to legalize this scalar type. Passing FPEnv through memory allows to avoid issues in legalization. For targets that get/set environment by simple register moves the intermediate stack slot is eliminated (see D81843). Well, almost eliminated, only write to stack slot remains, but it is DCE deficiency.

[sepavloff]

Removed dependency on DataLayout

For the task of the current patch availability of FPEnv size in
DataLayout is not strictly necessary.

[arsenm]

llvm/include/llvm/IR/Intrinsics.td

1070–1072

I'm not sure what you mean. You could legalize i256 by storing to the stack and reloading parts for all operations? This probably already works.

[sepavloff]

llvm/include/llvm/IR/Intrinsics.td

1070–1072

I tried to implement a solution in which get.fpenv returns integer value. The solution seems to work, however it has its own drawbacks:

• In the case of x86 get.fpenv returns i256. This is illegal type, so we cannot set custom lowering through setOperationAction. TLI.getOperationAction always returns Expand as the type is extended. To cope with this difficulty, DAGTypeLegalizer::ExpandIntegerResult has to call TLI.ReplaceNodeResults directly.

• It is necessary to use special node class to represent GET_FPENV, otherwise TLI.ReplaceNodeResults has no access to the chain token argument.

• The code created for the source:

	define void @func_01(i256* %fpenv) {
	entry:
	  %fpe = call i256 @llvm.get.fpenv.i256()
	  store i256 %fpe, i256* %fpenv
	  ret void
	}

uses stack variable, although it should use pointer argument directly. Optimization that would eliminate it is now absent.

This way is possible, but it requires more efforts to implement. There must be serious reasons why we prefer this way over using pointer argument.

Could you please explain the concern of using pointer in the intrinsic? What is wrong with the implementation used in D81843? Why it cannot be used by targets where FP environment is simply a content of a register?

[arsenm]

llvm/include/llvm/IR/Intrinsics.td

1070–1072

It could be, but it forces the lowering to introduce stack usage. Now an alloca has to stick around for the scratch pad until codegen, and SROA can't eliminate it. For AMDGPU for example, we change ABI lowering and other optimization strategies based on whether there is stack usage in the incoming IR. We don't really have other intrinsics that force you into this situation

I also noticed we have llvm.flt.rounds already, so I'm confused why this is different

[sepavloff]

llvm/include/llvm/IR/Intrinsics.td

1070–1072

Thank you for the explanation. I'll try to elaborate the other way.

I also noticed we have llvm.flt.rounds already, so I'm confused why this is different

llvm.flt.rounds only reads the current rounding mode. These intrinsics save/restore the entire FP environment.

[sepavloff]

Intrinsics now does not use pointers

[arsenm]

Is this still relevant?

[sepavloff]

Updated the patch

[sepavloff]

Updated patch

• Use special DAG nodes if FP environment is passed in memory,
• Various cleanups.

[sepavloff]

Updated patch after test started using autogenerated assertions in another commit.

[sepavloff]

Add memoperand to GET_FPENV_MEM and SET_FPENV_MEM

[arsenm]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

6551–6553

Can this just go into an Expand action like normal? We randomly expand certain things in the DAG builder but the inconsistency is annoying

[arsenm]

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

12327

If not getting it from the actual declaration, should probably use the default program address space (don't really care if you fix it here, 90% of the uses of this are wrong as it is)

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

6591

This should use the alignment for EnvVT, not the datalayout's stack alignment

[sepavloff]

Updated patch

• Use type alignment rather than stack one,
• Use getPointerMemTy instead of getPointerTy.

[sepavloff]

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

12327

IIRC default program address space is for function code. I put getPointerMemTy instead of getPointerTy, although now it behave in the same way. I don't know to put correct address space here, as declaration of the function is unavailable.

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

6551–6553

This code does the same job as other cases in this function, - it creates relevant DAG nodes. Their expansion occurs in SelectionDAG::Legalize, as for other nodes. The only difference is using two kings of nodes, register and memory-based ones.

Using only one kind of node could simplify code in this function but complicates code in other parts. On some targets FP environment is long and corresponding integer type is illegal (i256 on x86). In such case DAG nodes have to be expanded early, when DAG legalizes types. If FP environment can be represented by a legal type, the expansion occurs late, in SelectionDAG::Legalize. There were two points when the expansion occurred. This approach was used in previous versions of this patch, it complicated the implementation and made it fragile.

Using separate nodes for operations that involves memory allows to treat all FP environment operations in uniform manner, at the cost of small complication of this function.

6591

Fixed.

[sepavloff]

Ping.

[arsenm]

LGTM with the getPointerMemTy thing reverted

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

12327

pointer mem ty makes less sense

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

6551–6553

This hack doesn't require repeating in globalisel

6561

CreateStackTemporary should be fixed to directly take Align input

llvm/docs/LangRef.rst

 """"""""""
 
 The '``llvm.set.rounding``' intrinsic sets the current rounding mode. It is
 similar to C library function 'fesetround', however this intrinsic does not
 return any value and uses platform-independent representation of IEEE rounding
 modes.
 
 
+'``llvm.get.fpenv``' Intrinsic
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Syntax:
+"""""""
+
+::
+
+      declare <integer_type> @llvm.get.fpenv()
+
+Overview:
+"""""""""
+
+The '``llvm.get.fpenv``' intrinsic returns bits of the current floating-point
+environment. The return value type is platform-specific.
+
+Semantics:
+""""""""""
+
+The '``llvm.get.fpenv``' intrinsic reads the current floating-point environment
+and returns it as an integer value.
+
+
+'``llvm.set.fpenv``' Intrinsic
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Syntax:
+"""""""
+
+::
+
+      declare void @llvm.set.fpenv(<integer_type> <val>)
+
+Overview:
+"""""""""
+
+The '``llvm.set.fpenv``' intrinsic sets the current floating-point environment.
+
+Arguments:
+""""""""""
+
+The argument is an integer representing the new floating-point environment. The
+integer type is platform-specific.
+
+Semantics:
+""""""""""
+
+The '``llvm.set.fpenv``' intrinsic sets the current floating-point environment
+to the state specified by the argument. The state may be previously obtained by a
+call to '``llvm.get.fpenv``' or synthesised in a platform-dependent way.
+
+
+'``llvm.reset.fpenv``' Intrinsic
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Syntax:
+"""""""
+
+::
+
+      declare void @llvm.reset.fpenv()
+
+Overview:
+"""""""""
+
+The '``llvm.reset.fpenv``' intrinsic sets the default floating-point environment.
+
+Semantics:
+""""""""""
+
+The '``llvm.reset.fpenv``' intrinsic sets the current floating-point environment
+to default state. It is similar to the call 'fesetenv(FE_DFL_ENV)', except it
+does not return any value.
+
+
 Floating-Point Test Intrinsics
 ------------------------------
 
 These functions get properties of floating-point values.
 
 
 .. _llvm.is.fpclass:
 

 
 Syntax:
 """""""
 
 ::
 
       declare void @llvm.var.annotation(ptr <val>, ptr <str>, ptr <str>, i32  <int>)
 
 Overview:
 """""""""
 
 The '``llvm.var.annotation``' intrinsic.
 
 Arguments:
 """"""""""
 
 The first argument is a pointer to a value, the second is a pointer to a

llvm/include/llvm/CodeGen/ISDOpcodes.h

enum NodeType {
   /// as less than 0.0. While FMINNUM_IEEE/FMAXNUM_IEEE follow IEEE 754-2008
   /// semantics, FMINIMUM/FMAXIMUM follow IEEE 754-2018 draft semantics.
   FMINIMUM,
   FMAXIMUM,
 
   /// FSINCOS - Compute both fsin and fcos as a single operation.
   FSINCOS,
 
+  /// Gets the current floating-point environment. The first operand is a token
+  /// chain. The results are FP environment, represented by an integer value,
+  /// and a token chain.
+  GET_FPENV,
+
+  /// Sets the current floating-point environment. The first operand is a token
+  /// chain, the second is FP environment, represented by an integer value. The
+  /// result is a token chain.
+  SET_FPENV,
+
+  /// Set floating-point environment to default state. The first operand and the
+  /// result are token chains.
+  RESET_FPENV,
+
+  /// Gets the current floating-point environment. The first operand is a token
+  /// chain, the second is a pointer to memory, where FP environment is stored
+  /// to. The result is a token chain.
+  GET_FPENV_MEM,
+
+  /// Sets the current floating point environment. The first operand is a token
+  /// chain, the second is a pointer to memory, where FP environment is loaded
+  /// from. The result is a token chain.
+  SET_FPENV_MEM,
+
   /// LOAD and STORE have token chains as their first operand, then the same
   /// operands as an LLVM load/store instruction, then an offset node that
   /// is added / subtracted from the base pointer to form the address (for
   /// indexed memory ops).
   LOAD,
   STORE,
 
   /// DYNAMIC_STACKALLOC - Allocate some number of bytes on the stack aligned

llvm/include/llvm/CodeGen/SelectionDAG.h

bool isSafeToSpeculativelyExecute(unsigned Opcode) const {
     case ISD::UREM:
     case ISD::UDIVREM:
       return false;
     default:
       return true;
     }
   }
 
+  SDValue makeStateFunctionCall(unsigned LibFunc, SDValue Ptr, SDValue InChain,
+                                const SDLoc &DLoc);
+
 private:
   void InsertNode(SDNode *N);
   bool RemoveNodeFromCSEMaps(SDNode *N);
   void AddModifiedNodeToCSEMaps(SDNode *N);
   SDNode *FindModifiedNodeSlot(SDNode *N, SDValue Op, void *&InsertPos);
   SDNode *FindModifiedNodeSlot(SDNode *N, SDValue Op1, SDValue Op2,
                                void *&InsertPos);
   SDNode *FindModifiedNodeSlot(SDNode *N, ArrayRef<SDValue> Ops,

llvm/include/llvm/IR/Intrinsics.td

def int_objectsize : DefaultAttrsIntrinsic<[llvm_anyint_ty],
                                 ImmArg<ArgIndex<3>>]>,
                                ClangBuiltin<"__builtin_object_size">;
 
 //===--------------- Access to Floating Point Environment -----------------===//
 //
 
 let IntrProperties = [IntrInaccessibleMemOnly, IntrWillReturn] in {
   def int_get_rounding  : DefaultAttrsIntrinsic<[llvm_i32_ty], []>;
   def int_set_rounding  : DefaultAttrsIntrinsic<[], [llvm_i32_ty]>;
+  def int_get_fpenv     : DefaultAttrsIntrinsic<[llvm_anyint_ty], []>;
+  def int_set_fpenv     : DefaultAttrsIntrinsic<[], [llvm_anyint_ty]>;
+  def int_reset_fpenv   : DefaultAttrsIntrinsic<[], []>;
 }
 
 //===--------------- Floating Point Properties ----------------------------===//
 //
 
 def int_is_fpclass
     : DefaultAttrsIntrinsic<[LLVMScalarOrSameVectorWidth<0, llvm_i1_ty>],
                             [llvm_anyfloat_ty, llvm_i32_ty],
                             [IntrNoMem, IntrSpeculatable, ImmArg<ArgIndex<1>>]>;

llvm/include/llvm/IR/RuntimeLibcalls.def

 HANDLE_LIBCALL(LRINT_F128, "lrintl")
 HANDLE_LIBCALL(LRINT_PPCF128, "lrintl")
 HANDLE_LIBCALL(LLRINT_F32, "llrintf")
 HANDLE_LIBCALL(LLRINT_F64, "llrint")
 HANDLE_LIBCALL(LLRINT_F80, "llrintl")
 HANDLE_LIBCALL(LLRINT_F128, "llrintl")
 HANDLE_LIBCALL(LLRINT_PPCF128, "llrintl")
 
+// Floating point environment
+HANDLE_LIBCALL(FEGETENV, "fegetenv")
+HANDLE_LIBCALL(FESETENV, "fesetenv")
+
 // Conversion
 HANDLE_LIBCALL(FPEXT_F32_PPCF128, "__gcc_stoq")
 HANDLE_LIBCALL(FPEXT_F64_PPCF128, "__gcc_dtoq")
 HANDLE_LIBCALL(FPEXT_F80_F128, "__extendxftf2")
 HANDLE_LIBCALL(FPEXT_F64_F128, "__extenddftf2")
 HANDLE_LIBCALL(FPEXT_F32_F128, "__extendsftf2")
 HANDLE_LIBCALL(FPEXT_F16_F128, "__extendhftf2")
 HANDLE_LIBCALL(FPEXT_F16_F80, "__extendhfxf2")

llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

Action = TLI.getOperationAction(Node->getOpcode(),
                                     Node->getValueType(0));
     break;
   case ISD::VAARG:
     Action = TLI.getOperationAction(Node->getOpcode(),
                                     Node->getValueType(0));
     if (Action != TargetLowering::Promote)
       Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
     break;
+  case ISD::SET_FPENV:
+    Action = TLI.getOperationAction(Node->getOpcode(),
+                                    Node->getOperand(1).getValueType());
+    break;
   case ISD::FP_TO_FP16:
   case ISD::FP_TO_BF16:
   case ISD::SINT_TO_FP:
   case ISD::UINT_TO_FP:
   case ISD::EXTRACT_VECTOR_ELT:
   case ISD::LROUND:
   case ISD::LLROUND:
   case ISD::LRINT:
case ISD::CTLZ_ZERO_UNDEF:
     case MVT::i64:
       Results.push_back(ExpandLibCall(RTLIB::CTLZ_I64, Node, false));
       break;
     case MVT::i128:
       Results.push_back(ExpandLibCall(RTLIB::CTLZ_I128, Node, false));
       break;
     }
     break;
+  case ISD::RESET_FPENV: {
+    // It is legalized to call 'fesetenv(FE_DFL_ENV)'. On most targets
+    // FE_DFL_ENV is defined as '((const fenv_t *) -1)' in glibc.
+    SDValue Ptr = DAG.getIntPtrConstant(-1LL, dl);
+    SDValue Chain = Node->getOperand(0);
+    Results.push_back(
+        DAG.makeStateFunctionCall(RTLIB::FESETENV, Ptr, Chain, dl));
+    break;
+  }
+  case ISD::GET_FPENV_MEM: {
+    SDValue Chain = Node->getOperand(0);
+    SDValue EnvPtr = Node->getOperand(1);
+    Results.push_back(
+        DAG.makeStateFunctionCall(RTLIB::FEGETENV, EnvPtr, Chain, dl));
+    break;
+  }
+  case ISD::SET_FPENV_MEM: {
+    SDValue Chain = Node->getOperand(0);
+    SDValue EnvPtr = Node->getOperand(1);
+    Results.push_back(
+        DAG.makeStateFunctionCall(RTLIB::FESETENV, EnvPtr, Chain, dl));
+    break;
+  }
   }
 
   // Replace the original node with the legalized result.
   if (!Results.empty()) {
     LLVM_DEBUG(dbgs() << "Successfully converted node to libcall\n");
     ReplaceNode(Node, Results.data());
   } else
     LLVM_DEBUG(dbgs() << "Could not convert node to libcall\n");

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

case ISD::FMAXNUM: {
     if (Opcode == ISD::FMAXNUM)
       NeutralAF.changeSign();
 
     return getConstantFP(NeutralAF, DL, VT);
   }
   }
 }
 
+/// Helper used to make a call to a library function that has one argument of
+/// pointer type.
+///
+/// Such functions include 'fegetmode', 'fesetenv' and some others, which are
+/// used to get or set floating-point state. They have one argument of pointer
+/// type, which points to the memory region containing bits of the
+/// floating-point state. The value returned by such function is ignored in the
+/// created call.
+///
+/// \param LibFunc Reference to library function (value of RTLIB::Libcall).
+/// \param Ptr Pointer used to save/load state.
+/// \param InChain Ingoing token chain.
+/// \returns Outgoing chain token.
+SDValue SelectionDAG::makeStateFunctionCall(unsigned LibFunc, SDValue Ptr,
+                                            SDValue InChain,
+                                            const SDLoc &DLoc) {
+  assert(InChain.getValueType() == MVT::Other && "Expected token chain");
+  TargetLowering::ArgListTy Args;
+  TargetLowering::ArgListEntry Entry;
+  Entry.Node = Ptr;
+  Entry.Ty = Ptr.getValueType().getTypeForEVT(*getContext());
+  Args.push_back(Entry);
+  RTLIB::Libcall LC = static_cast<RTLIB::Libcall>(LibFunc);
+  SDValue Callee = getExternalSymbol(TLI->getLibcallName(LC),
+                                     TLI->getPointerTy(getDataLayout()));
+  TargetLowering::CallLoweringInfo CLI(*this);
+  CLI.setDebugLoc(DLoc).setChain(InChain).setLibCallee(
+      TLI->getLibcallCallingConv(LC), Type::getVoidTy(*getContext()), Callee,
+      std::move(Args));
+  return TLI->LowerCallTo(CLI).second;
+}
+
 void SelectionDAG::copyExtraInfo(SDNode *From, SDNode *To) {
   assert(From && To && "Invalid SDNode; empty source SDValue?");
   auto I = SDEI.find(From);
   if (I == SDEI.end())
     return;
 
   // Use of operator[] on the DenseMap may cause an insertion, which invalidates
   // the iterator, hence the need to make a copy to prevent a use-after-free.

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

if (!TLI.isOperationLegalOrCustom(ISD::IS_FPCLASS, ArgVT)) {
       return;
     }
 
     SDValue Check = DAG.getTargetConstant(Test, sdl, MVT::i32);
     SDValue V = DAG.getNode(ISD::IS_FPCLASS, sdl, DestVT, {Op, Check}, Flags);
     setValue(&I, V);
     return;
   }
+  case Intrinsic::get_fpenv: {
+    const DataLayout DLayout = DAG.getDataLayout();
+    EVT EnvVT = TLI.getValueType(DLayout, I.getType());
+    // Use GET_FPENV if it is legal or custom. Otherwise use memory-based node
+    // and temporary storage in stack.
+    if (TLI.isOperationLegalOrCustom(ISD::SET_FPENV, EnvVT)) {
+      Res = DAG.getNode(
+          ISD::GET_FPENV, sdl,
+          DAG.getVTList(TLI.getValueType(DAG.getDataLayout(), I.getType()),
+                        MVT::Other),
+          DAG.getRoot());
+    } else {
+      SDValue Temp = DAG.CreateStackTemporary(EnvVT);
+      SDValue Chain = DAG.getRoot();
+      Chain = DAG.getNode(ISD::GET_FPENV_MEM, sdl, MVT::Other, Chain, Temp);
+      int SPFI = cast<FrameIndexSDNode>(Temp.getNode())->getIndex();
+      auto MPI =
+          MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SPFI);
+      Res = DAG.getLoad(EnvVT, sdl, Chain, Temp, MPI);
+    }
+    setValue(&I, Res);
+    DAG.setRoot(Res.getValue(1));
+    return;
+  }
+  case Intrinsic::set_fpenv: {
+    const DataLayout DLayout = DAG.getDataLayout();
+    SDValue Env = getValue(I.getArgOperand(0));
+    SDValue Chain = getRoot();
+    EVT EnvVT = Env.getValueType();
+    // If SET_FPENV is custom or legal, use it. Otherwise use loading
+    // environment from memory.
+    if (TLI.isOperationLegalOrCustom(ISD::SET_FPENV, EnvVT)) {
+      DAG.setRoot(DAG.getNode(ISD::SET_FPENV, sdl, MVT::Other, Chain, Env));
+    } else {
+      // Allocate space in stack, copy environment bits into it and use this
+      // memory in SET_FPENV_MEM.
+      SDValue Ptr = DAG.CreateStackTemporary(EnvVT);
+      Ptr = DAG.CreateStackTemporary(EnvVT);
+      int SPFI = cast<FrameIndexSDNode>(Ptr.getNode())->getIndex();
+      auto MPI =
+          MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SPFI);
+      Chain =
+          DAG.getStore(Chain, sdl, Env, Ptr, MPI, DLayout.getStackAlignment(),
+                       MachineMemOperand::MOStore);
+      Chain = DAG.getNode(ISD::SET_FPENV_MEM, sdl, MVT::Other, Chain, Ptr);
+      DAG.setRoot(Chain);
+    }
+    return;
+  }
+  case Intrinsic::reset_fpenv:
+    DAG.setRoot(DAG.getNode(ISD::RESET_FPENV, sdl, MVT::Other, getRoot()));
+    return;
   case Intrinsic::pcmarker: {
     SDValue Tmp = getValue(I.getArgOperand(0));
     DAG.setRoot(DAG.getNode(ISD::PCMARKER, sdl, MVT::Other, getRoot(), Tmp));
     return;
   }
   case Intrinsic::readcyclecounter: {
     SDValue Op = getRoot();
     Res = DAG.getNode(ISD::READCYCLECOUNTER, sdl,

llvm/lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp

#endif
   case ISD::PREALLOCATED_SETUP:
     return "call_setup";
   case ISD::PREALLOCATED_ARG:
     return "call_alloc";
 
   // Floating point environment manipulation
   case ISD::GET_ROUNDING:               return "get_rounding";
   case ISD::SET_ROUNDING:               return "set_rounding";
+  case ISD::GET_FPENV:                  return "get_fpenv";
+  case ISD::SET_FPENV:                  return "set_fpenv";
+  case ISD::RESET_FPENV:                return "reset_fpenv";
+  case ISD::GET_FPENV_MEM:              return "get_fpenv_mem";
+  case ISD::SET_FPENV_MEM:              return "set_fpenv_mem";
 
   // Bit manipulation
   case ISD::ABS:                        return "abs";
   case ISD::BITREVERSE:                 return "bitreverse";
   case ISD::BSWAP:                      return "bswap";
   case ISD::CTPOP:                      return "ctpop";
   case ISD::CTTZ:                       return "cttz";
   case ISD::CTTZ_ZERO_UNDEF:            return "cttz_zero_undef";

llvm/lib/CodeGen/TargetLoweringBase.cpp

#include "llvm/IR/ConstrainedOps.def"
 
     // Named vector shuffles default to expand.
     setOperationAction(ISD::VECTOR_SPLICE, VT, Expand);
 
     // VP operations default to expand.
 #define BEGIN_REGISTER_VP_SDNODE(SDOPC, ...)                                   \
     setOperationAction(ISD::SDOPC, VT, Expand);
 #include "llvm/IR/VPIntrinsics.def"
+
+    // FP environment operations default to expand.
+    setOperationAction(ISD::GET_FPENV, VT, Expand);
+    setOperationAction(ISD::SET_FPENV, VT, Expand);
+    setOperationAction(ISD::RESET_FPENV, VT, Expand);
   }
 
   // Most targets ignore the @llvm.prefetch intrinsic.
   setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
 
   // Most targets also ignore the @llvm.readcyclecounter intrinsic.
   setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand);
 
#include "llvm/IR/VPIntrinsics.def"
   // Default ISD::TRAP to expand (which turns it into abort).
   setOperationAction(ISD::TRAP, MVT::Other, Expand);
 
   // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand"
   // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP.
   setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand);
 
   setOperationAction(ISD::UBSANTRAP, MVT::Other, Expand);
+
+  setOperationAction(ISD::GET_FPENV_MEM, MVT::Other, Expand);
+  setOperationAction(ISD::SET_FPENV_MEM, MVT::Other, Expand);
 }
 
 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL,
                                                EVT) const {
   return MVT::getIntegerVT(DL.getPointerSizeInBits(0));
 }
 
 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, const DataLayout &DL,

llvm/test/CodeGen/ARM/fpenv.ll

 ; CHECK-NEXT:    bic r0, r0, #12582912
 ; CHECK-NEXT:    orr r0, r0, #8388608
 ; CHECK-NEXT:    vmsr fpscr, r0
 ; CHECK-NEXT:    mov pc, lr
   call void @llvm.set.rounding(i32 3)
   ret void
 }
 
+define i32 @fpenv_01() #0 {
+; CHECK-LABEL: fpenv_01:
+; CHECK:       @ %bb.0: @ %entry
+; CHECK-NEXT:    .save {r11, lr}
+; CHECK-NEXT:    push {r11, lr}
+; CHECK-NEXT:    .pad #8
+; CHECK-NEXT:    sub sp, sp, #8
+; CHECK-NEXT:    add r0, sp, #4
+; CHECK-NEXT:    bl fegetenv
+; CHECK-NEXT:    ldr r0, [sp, #4]
+; CHECK-NEXT:    add sp, sp, #8
+; CHECK-NEXT:    pop {r11, lr}
+; CHECK-NEXT:    mov pc, lr
+entry:
+  %fpenv = call i32 @llvm.get.fpenv.i32()
+  ret i32 %fpenv
+}
+
+define void @fpenv_02(i32 %fpenv) #0 {
+; CHECK-LABEL: fpenv_02:
+; CHECK:       @ %bb.0: @ %entry
+; CHECK-NEXT:    .save {r11, lr}
+; CHECK-NEXT:    push {r11, lr}
+; CHECK-NEXT:    .pad #8
+; CHECK-NEXT:    sub sp, sp, #8
+; CHECK-NEXT:    str r0, [sp]
+; CHECK-NEXT:    mov r0, sp
+; CHECK-NEXT:    bl fesetenv
+; CHECK-NEXT:    add sp, sp, #8
+; CHECK-NEXT:    pop {r11, lr}
+; CHECK-NEXT:    mov pc, lr
+entry:
+  call void @llvm.set.fpenv.i32(i32 %fpenv)
+  ret void
+}
+
+define void @fpenv_03() #0 {
+; CHECK-LABEL: fpenv_03:
+; CHECK:       @ %bb.0: @ %entry
+; CHECK-NEXT:    .save {r11, lr}
+; CHECK-NEXT:    push {r11, lr}
+; CHECK-NEXT:    mvn r0, #0
+; CHECK-NEXT:    bl fesetenv
+; CHECK-NEXT:    pop {r11, lr}
+; CHECK-NEXT:    mov pc, lr
+entry:
+  call void @llvm.reset.fpenv()
+  ret void
+}
+
+attributes #0 = { nounwind "use-soft-float"="true" }
 
 declare void @llvm.set.rounding(i32)
+declare i32 @llvm.get.fpenv.i32()
+declare void @llvm.set.fpenv.i32(i32 %fpenv)
+declare void @llvm.reset.fpenv()

llvm/test/CodeGen/X86/fpenv.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=-sse -verify-machineinstrs < %s | FileCheck %s -check-prefix=X86-NOSSE
 ; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=+sse -verify-machineinstrs < %s | FileCheck %s -check-prefix=X86-SSE
 ; RUN: llc -mtriple=x86_64-unknown-linux-gnu -verify-machineinstrs < %s | FileCheck %s -check-prefix=X64
 
 declare void @llvm.set.rounding(i32 %x)
+declare i256 @llvm.get.fpenv.i256()
+declare void @llvm.set.fpenv.i256(i256 %fpenv)
+declare void @llvm.reset.fpenv()
 
 define void @func_01() nounwind {
 ; X86-NOSSE-LABEL: func_01:
 ; X86-NOSSE:       # %bb.0:
 ; X86-NOSSE-NEXT:    pushl %eax
 ; X86-NOSSE-NEXT:    fnstcw (%esp)
 ; X86-NOSSE-NEXT:    orb $12, {{[0-9]+}}(%esp)
 ; X86-NOSSE-NEXT:    fldcw (%esp)

 ; X64-NEXT:    andl -{{[0-9]+}}(%rsp), %ecx
 ; X64-NEXT:    leal (%rcx,%rax,8), %eax
 ; X64-NEXT:    movl %eax, -{{[0-9]+}}(%rsp)
 ; X64-NEXT:    ldmxcsr -{{[0-9]+}}(%rsp)
 ; X64-NEXT:    retq
   call void @llvm.set.rounding(i32 %x)  ; Downward
   ret void
 }
+
+define void @fpenv_01(ptr %ptr) #0 {
+; X86-NOSSE-LABEL: fpenv_01:
+; X86-NOSSE:       # %bb.0: # %entry
+; X86-NOSSE-NEXT:    pushl %ebp
+; X86-NOSSE-NEXT:    pushl %ebx
+; X86-NOSSE-NEXT:    pushl %edi
+; X86-NOSSE-NEXT:    pushl %esi
+; X86-NOSSE-NEXT:    subl $60, %esp
+; X86-NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; X86-NOSSE-NEXT:    leal {{[0-9]+}}(%esp), %eax
+; X86-NOSSE-NEXT:    movl %eax, (%esp)
+; X86-NOSSE-NEXT:    calll fegetenv
+; X86-NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-NOSSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X86-NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-NOSSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X86-NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X86-NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %edi
+; X86-NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %ebx
+; X86-NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %ebp
+; X86-NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X86-NOSSE-NEXT:    movl %ecx, 24(%esi)
+; X86-NOSSE-NEXT:    movl %eax, 28(%esi)
+; X86-NOSSE-NEXT:    movl %ebp, 16(%esi)
+; X86-NOSSE-NEXT:    movl %ebx, 20(%esi)
+; X86-NOSSE-NEXT:    movl %edi, 8(%esi)
+; X86-NOSSE-NEXT:    movl %edx, 12(%esi)
+; X86-NOSSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X86-NOSSE-NEXT:    movl %eax, (%esi)
+; X86-NOSSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X86-NOSSE-NEXT:    movl %eax, 4(%esi)
+; X86-NOSSE-NEXT:    addl $60, %esp
+; X86-NOSSE-NEXT:    popl %esi
+; X86-NOSSE-NEXT:    popl %edi
+; X86-NOSSE-NEXT:    popl %ebx
+; X86-NOSSE-NEXT:    popl %ebp
+; X86-NOSSE-NEXT:    retl
+;
+; X86-SSE-LABEL: fpenv_01:
+; X86-SSE:       # %bb.0: # %entry
+; X86-SSE-NEXT:    pushl %ebp
+; X86-SSE-NEXT:    pushl %ebx
+; X86-SSE-NEXT:    pushl %edi
+; X86-SSE-NEXT:    pushl %esi
+; X86-SSE-NEXT:    subl $60, %esp
+; X86-SSE-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; X86-SSE-NEXT:    leal {{[0-9]+}}(%esp), %eax
+; X86-SSE-NEXT:    movl %eax, (%esp)
+; X86-SSE-NEXT:    calll fegetenv
+; X86-SSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-SSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X86-SSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-SSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X86-SSE-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X86-SSE-NEXT:    movl {{[0-9]+}}(%esp), %edi
+; X86-SSE-NEXT:    movl {{[0-9]+}}(%esp), %ebx
+; X86-SSE-NEXT:    movl {{[0-9]+}}(%esp), %ebp
+; X86-SSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-SSE-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X86-SSE-NEXT:    movl %ecx, 24(%esi)
+; X86-SSE-NEXT:    movl %eax, 28(%esi)
+; X86-SSE-NEXT:    movl %ebp, 16(%esi)
+; X86-SSE-NEXT:    movl %ebx, 20(%esi)
+; X86-SSE-NEXT:    movl %edi, 8(%esi)
+; X86-SSE-NEXT:    movl %edx, 12(%esi)
+; X86-SSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X86-SSE-NEXT:    movl %eax, (%esi)
+; X86-SSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X86-SSE-NEXT:    movl %eax, 4(%esi)
+; X86-SSE-NEXT:    addl $60, %esp
+; X86-SSE-NEXT:    popl %esi
+; X86-SSE-NEXT:    popl %edi
+; X86-SSE-NEXT:    popl %ebx
+; X86-SSE-NEXT:    popl %ebp
+; X86-SSE-NEXT:    retl
+;
+; X64-LABEL: fpenv_01:
+; X64:       # %bb.0: # %entry
+; X64-NEXT:    pushq %rbx
+; X64-NEXT:    subq $32, %rsp
+; X64-NEXT:    movq %rdi, %rbx
+; X64-NEXT:    movq %rsp, %rdi
+; X64-NEXT:    callq fegetenv@PLT
+; X64-NEXT:    movq (%rsp), %rax
+; X64-NEXT:    movq {{[0-9]+}}(%rsp), %rcx
+; X64-NEXT:    movq {{[0-9]+}}(%rsp), %rdx
+; X64-NEXT:    movq {{[0-9]+}}(%rsp), %rsi
+; X64-NEXT:    movq %rsi, 16(%rbx)
+; X64-NEXT:    movq %rdx, 24(%rbx)
+; X64-NEXT:    movq %rax, (%rbx)
+; X64-NEXT:    movq %rcx, 8(%rbx)
+; X64-NEXT:    addq $32, %rsp
+; X64-NEXT:    popq %rbx
+; X64-NEXT:    retq
+entry:
+  %env = call i256 @llvm.get.fpenv.i256()
+  store i256 %env, ptr %ptr
+  ret void
+}
+
+define void @fpenv_02(ptr %ptr) #0 {
+; X86-NOSSE-LABEL: fpenv_02:
+; X86-NOSSE:       # %bb.0: # %entry
+; X86-NOSSE-NEXT:    pushl %ebp
+; X86-NOSSE-NEXT:    pushl %ebx
+; X86-NOSSE-NEXT:    pushl %edi
+; X86-NOSSE-NEXT:    pushl %esi
+; X86-NOSSE-NEXT:    subl $76, %esp
+; X86-NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-NOSSE-NEXT:    movl (%eax), %ecx
+; X86-NOSSE-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X86-NOSSE-NEXT:    movl 4(%eax), %edx
+; X86-NOSSE-NEXT:    movl 12(%eax), %esi
+; X86-NOSSE-NEXT:    movl 8(%eax), %edi
+; X86-NOSSE-NEXT:    movl 20(%eax), %ebx
+; X86-NOSSE-NEXT:    movl 16(%eax), %ebp
+; X86-NOSSE-NEXT:    movl 28(%eax), %ecx
+; X86-NOSSE-NEXT:    movl 24(%eax), %eax
+; X86-NOSSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; X86-NOSSE-NEXT:    movl %ecx, {{[0-9]+}}(%esp)
+; X86-NOSSE-NEXT:    movl %ebp, {{[0-9]+}}(%esp)
+; X86-NOSSE-NEXT:    movl %ebx, {{[0-9]+}}(%esp)
+; X86-NOSSE-NEXT:    movl %edi, {{[0-9]+}}(%esp)
+; X86-NOSSE-NEXT:    movl %esi, {{[0-9]+}}(%esp)
+; X86-NOSSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X86-NOSSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; X86-NOSSE-NEXT:    movl %edx, {{[0-9]+}}(%esp)
+; X86-NOSSE-NEXT:    leal {{[0-9]+}}(%esp), %eax
+; X86-NOSSE-NEXT:    movl %eax, (%esp)
+; X86-NOSSE-NEXT:    calll fesetenv
+; X86-NOSSE-NEXT:    addl $76, %esp
+; X86-NOSSE-NEXT:    popl %esi
+; X86-NOSSE-NEXT:    popl %edi
+; X86-NOSSE-NEXT:    popl %ebx
+; X86-NOSSE-NEXT:    popl %ebp
+; X86-NOSSE-NEXT:    retl
+;
+; X86-SSE-LABEL: fpenv_02:
+; X86-SSE:       # %bb.0: # %entry
+; X86-SSE-NEXT:    pushl %ebp
+; X86-SSE-NEXT:    pushl %ebx
+; X86-SSE-NEXT:    pushl %edi
+; X86-SSE-NEXT:    pushl %esi
+; X86-SSE-NEXT:    subl $76, %esp
+; X86-SSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-SSE-NEXT:    movl (%eax), %ecx
+; X86-SSE-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X86-SSE-NEXT:    movl 4(%eax), %edx
+; X86-SSE-NEXT:    movl 12(%eax), %esi
+; X86-SSE-NEXT:    movl 8(%eax), %edi
+; X86-SSE-NEXT:    movl 20(%eax), %ebx
+; X86-SSE-NEXT:    movl 16(%eax), %ebp
+; X86-SSE-NEXT:    movl 28(%eax), %ecx
+; X86-SSE-NEXT:    movl 24(%eax), %eax
+; X86-SSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; X86-SSE-NEXT:    movl %ecx, {{[0-9]+}}(%esp)
+; X86-SSE-NEXT:    movl %ebp, {{[0-9]+}}(%esp)
+; X86-SSE-NEXT:    movl %ebx, {{[0-9]+}}(%esp)
+; X86-SSE-NEXT:    movl %edi, {{[0-9]+}}(%esp)
+; X86-SSE-NEXT:    movl %esi, {{[0-9]+}}(%esp)
+; X86-SSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X86-SSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; X86-SSE-NEXT:    movl %edx, {{[0-9]+}}(%esp)
+; X86-SSE-NEXT:    leal {{[0-9]+}}(%esp), %eax
+; X86-SSE-NEXT:    movl %eax, (%esp)
+; X86-SSE-NEXT:    calll fesetenv
+; X86-SSE-NEXT:    addl $76, %esp
+; X86-SSE-NEXT:    popl %esi
+; X86-SSE-NEXT:    popl %edi
+; X86-SSE-NEXT:    popl %ebx
+; X86-SSE-NEXT:    popl %ebp
+; X86-SSE-NEXT:    retl
+;
+; X64-LABEL: fpenv_02:
+; X64:       # %bb.0: # %entry
+; X64-NEXT:    subq $72, %rsp
+; X64-NEXT:    movq (%rdi), %rax
+; X64-NEXT:    movq 8(%rdi), %rcx
+; X64-NEXT:    movq 24(%rdi), %rdx
+; X64-NEXT:    movq 16(%rdi), %rsi
+; X64-NEXT:    movq %rsi, {{[0-9]+}}(%rsp)
+; X64-NEXT:    movq %rdx, {{[0-9]+}}(%rsp)
+; X64-NEXT:    movq %rax, {{[0-9]+}}(%rsp)
+; X64-NEXT:    movq %rcx, {{[0-9]+}}(%rsp)
+; X64-NEXT:    leaq {{[0-9]+}}(%rsp), %rdi
+; X64-NEXT:    callq fesetenv@PLT
+; X64-NEXT:    addq $72, %rsp
+; X64-NEXT:    retq
+entry:
+  %env = load i256, ptr %ptr
+  call void @llvm.set.fpenv.i256(i256 %env)
+  ret void
+}
+
+
+define void @fpenv_03() #0 {
+; X86-NOSSE-LABEL: fpenv_03:
+; X86-NOSSE:       # %bb.0: # %entry
+; X86-NOSSE-NEXT:    subl $12, %esp
+; X86-NOSSE-NEXT:    movl $-1, (%esp)
+; X86-NOSSE-NEXT:    calll fesetenv
+; X86-NOSSE-NEXT:    addl $12, %esp
+; X86-NOSSE-NEXT:    retl
+;
+; X86-SSE-LABEL: fpenv_03:
+; X86-SSE:       # %bb.0: # %entry
+; X86-SSE-NEXT:    subl $12, %esp
+; X86-SSE-NEXT:    movl $-1, (%esp)
+; X86-SSE-NEXT:    calll fesetenv
+; X86-SSE-NEXT:    addl $12, %esp
+; X86-SSE-NEXT:    retl
+;
+; X64-LABEL: fpenv_03:
+; X64:       # %bb.0: # %entry
+; X64-NEXT:    pushq %rax
+; X64-NEXT:    movq $-1, %rdi
+; X64-NEXT:    callq fesetenv@PLT
+; X64-NEXT:    popq %rax
+; X64-NEXT:    retq
+entry:
+  call void @llvm.reset.fpenv()
+  ret void
+}
+
+attributes #0 = { nounwind "use-soft-float"="true" }