Implementation of '#pragma STDC FENV_ROUND'

[sepavloff]

Summary

This pragma is introduced by forthcoming C2x standard and can be used to
set particular rounding mode without need to call 'fesetmode' or accessing
control mode registers directly. Previously this pragma was implemented in
clang partially, only for the purpose of using in constant expressions and
making tests.

This change implements the pragma according to the standard draft. It sets
up dynamic rounding mode in the compound statement where the pragma acts.
This is inevitable for targets that set rounding mode by changing some
control register. Some targets however allow specifying rounding mode as a
field in instructions, in this case setting dynamic rounding mode is not
mandatory. However the standard draft requests that invocations of
functions shall be evaluated using dynamic rounding mode. So the dynamic
rounding mode is set even if actually later it is unused. Removal of
unneeded manipulation on dynamic rounding mode can be made by a separate
IR optimization.

The implementation uses intrinsic functions 'flt_rounds' and
'set_rounding' to save/restore dynamic rounding mode. It is not the best
way, because these functions need to extract rounding mode from the
content of some control register or pack into it. More efficient
implementation could save/restore entire value of the control register. It
requires functions 'fegetmode' and 'fesetmode' to be implemented as LLVM
intrinsic functions. When these functions will be implemented, the same
mechanism could be used in implementation of other similar fp pragmas.

[aaron.ballman]

Thank you for working on this (and thank you for your patience with how long it's taken me to get to this review)!

Adding another reviewer who has recently been poking at floating point environment pragmas in case Zahira has some thoughts here as well.

clang/lib/CodeGen/CGStmt.cpp

504

I was imagining that if the new rounding mode is dynamic, we still need to reset the rounding mode back to what it was when the thread was created or to the previous rounding mode set by a library call, but now I'm not certain. I commented on one of the test cases about what I was thinking.

clang/test/CodeGen/pragma-fenv_round.c

55–56

I was trying to figure out whether this was correct or not when I was looking at the codegen logic. C2x 7.6.2p3 says "... If no FENV_ROUND pragma is in effect, or the specified constant rounding mode is FE_DYNAMIC, rounding is according to the mode specified by the dynamic floating-point environment, which is the dynamic rounding mode that was established either at thread creation time or by a call to fesetround, fesetmode, fesetenv, or feupdateenv. ..." p2 says: "... When outside external declarations, the pragma takes effect from its occurrence until another FENV_ROUND pragma is encountered, or until the end of the translation unit. When inside a compound statement, the pragma takes effect from its first occurrence until another FENV_ROUND pragma is encountered (including within a nested compound statement), or until the end of the compound statement; ...". Finally, p4 says: "... Within the scope of an FENV_ROUND pragma establishing a mode other than FE_DYNAMIC, ... shall be evaluated according to the specified constant rounding mode (as though no constant mode was specified and the corresponding dynamic rounding mode had been established by a call to fesetround)."

I *think* this means that the rounding mode for result += z; should actually be FE_TOWARDZERO, but I'm not 100% sure. We enter the compound statement for the function body, then we enter a constant rounding mode of FE_TOWARDZERO. Then we get into a new nested compound statement, and we enter the rounding mode of FE_DYNAMIC, which says it should behave in the manner of the last call to fesetround, which was (as-if) called by the earlier pragma.

Thoughts?

I would be good to have a test for the behavior at file scope as well as at compound scope, e.g.,

float default_mode = 1.0f / 3.0f;

#pragma STDC FENV_ROUND FE_UPWARD  
float up = 1.0f / 3.0f;

#pragma STDC FENV_ROUND FE_DOWNWARD
float down = 1.0f / 3.0f;

#pragma STDC FENV_ROUND FE_DYNAMIC
float dynamic = 1.0f / 3.0f;

I'd expect the up and down cases to have different values, and I'd expect the dynamic case to either be the same as down or the same as default_mode.

72

I think it'd be nice to add a test case here like:

float func_21(float x, float y) {
  return x + y;
}

to demonstrate that FE_DOWNWARD is still in effect within this function too.

93

Another fun test case to add would be a C++ case:

#pragma STDC FENV_ROUND FE_DOWNWARD
float func_cxx(float x, float y, float z = 1.0f/3.0f) {
  return x + y + z;
}

int main() {
#pragma STDC FENV_ROUND FE_UPWARD
  func_cxx(1.0f, 2.0f);
}

is z evaluated as upward (because default arguments are evaluated at the call site IIRC), or downward (because of the rounding mode when defining the function)?

I suppose another interesting case is:

inline float func(float x, float y) {
#pragma STDC FENV_ROUND FE_DOWNWARD
  return x + y;
}

int main(void) {
  func(1.0f, 2.0f);
  float f = 1.0f / 3.0f;
}

Even if the call to func is truly inlined, I still wouldn't expect f to be calculated in downward (unless that's the default rounding mode). Is that how you read the standard as well?

[sepavloff]

Rebased and addressed reviever's comments

[sepavloff]

clang/lib/CodeGen/CGStmt.cpp

504

Yes, it is required by the standard draft (7.6.2p2):

... at the end of a compound statement the static rounding
mode is restored to its condition just before the compound statement.

Rounding mode is not set if FE_DYNAMIC is in effect, but it is restored upon leaving such region, see test func_4 in pragma-fenv_round.c.

clang/test/CodeGen/pragma-fenv_round.c

55–56

It is easier to rationalize effect of FENV_ROUND if target supports static rounding mode, like RISCV. On such target the rounding mode may be specified in every instruction that depends on it. Such instruction does not depend on FP environment, the rounding mode is known at compile-time, and it perfectly fits notion of constant rounding mode provided in the standard. There is also a special value of rounding mode that does not designates concrete rounding but instruct the processor to take rounding mode from a special register. This is dynamic rounding mode, In general it is unknown at compile-time and may be set, for example, by a call to fesetround. On such target FE_DYNAMIC means that instructions are created with dynamic rounding mode, while any other mode selected by FENV_ROUND results in concrete static rounding mode.

As for this example (func_3), the first pragma sets FE_TOWARDZERO mode. This is a static rounding mode, it is known in compile-time. In the nested compound statement pragma FENV_ROUND FE_DYNAMIC requires that FP operations use dynamic rounding mode. It means that actual rounding mode is known in runtime and the actual rounding mode must be set somewhere by a call to fesetround or some other manner. In this test no 'FENV_ACCESS` is specified, to compiler may assume the dynamic rounding mode is FE_TONEAREST. In the test func_4 there is pragma FENV_ACCESS ON and rounding mode specified in the FP operation is dynamic.

I would be good to have a test for the behavior at file scope as well as at compound scope, e.g.,

Added. FE_DYNAMIC in this case is identical to FE_TONEAREST, because initial dynamic rounding mode is such.

72

Added.

93

As FENV_ROUND is defined in C standard, its interaction with C++ is undefined. In this case it is reasonable to have "most obvious" behavior, although it is not well-defined without standard. In this case default argument initializer is specified in the region where #pragma STDC FENV_ROUND FE_DOWNWARD is in effect, so it seems natural to have the initialized calculated with this mode. Of course, this behavior may be changed. Added some relevant tests in pragma-fenv_round.cpp.

Inlining is made at llvm part, there are tests in https://github.com/llvm/llvm-project/blob/main/llvm/test/Transforms/Inline/inline-strictfp.ll.

clang/include/clang/Basic/DiagnosticParseKinds.td

def err_pragma_file_or_compound_scope : Error<
   "'#pragma %0' can only appear at file scope or at the start of a "
   "compound statement">;
 // - #pragma stdc unknown
 def ext_stdc_pragma_ignored : ExtWarn<"unknown pragma in STDC namespace">,
    InGroup<UnknownPragmas>;
 // The C standard 7.6.1p2 says "The [FENV_ACCESS] pragma shall occur either
 // outside external declarations or preceding all explicit declarations and
 // statements inside a compound statement.
-def warn_stdc_fenv_round_not_supported :
-   Warning<"pragma STDC FENV_ROUND is not supported">,
-   InGroup<UnknownPragmas>;
 def warn_stdc_unknown_rounding_mode : Warning<
   "invalid or unsupported rounding mode in '#pragma STDC FENV_ROUND' - ignored">,
   InGroup<IgnoredPragmas>;
 def warn_pragma_fp_ignored : Warning<
   "'#pragma %0' is not supported on this target - ignored">,
   InGroup<IgnoredPragmas>;
 // - #pragma comment
 def err_pragma_comment_malformed : Error<

clang/lib/CodeGen/CGStmt.cpp

case Stmt::CaseStmtClass:
     break;
   case Stmt::SEHLeaveStmtClass:
     EmitSEHLeaveStmt(cast<SEHLeaveStmt>(*S));
     break;
   }
   return true;
 }
 
+namespace {
+/// Cleanup action that restores floating-point control modes upon leaving
+/// a scope.
+struct FPControlModesCleanup final : EHScopeStack::Cleanup {
+  llvm::Value *PreviousModes;
+  FPControlModesCleanup(llvm::Value *M) : PreviousModes(M) {}
+  void Emit(CodeGenFunction &CGF, Flags flags) override {
+    // TODO: using 'setmode' must be more efficient.
+    CGF.Builder.CreateIntrinsic(llvm::Intrinsic::set_rounding, {},
+                                {PreviousModes});
+  }
+};
+} // namespace
+
+void CodeGenFunction::emitSetFPControlModes(const CompoundStmt &S) {
+  if (!S.hasStoredFPFeatures())
+    return;
+  FPOptionsOverride FPO = S.getStoredFPFeatures();
+  FPOptions FP = FPO.applyOverrides(CurFPFeatures);
+  llvm::RoundingMode OldRM = CurFPFeatures.getRoundingMode();
+  llvm::RoundingMode NewRM = FP.getRoundingMode();
+  if (OldRM == NewRM)
+    return;
+  llvm::Value *PreviousRM = nullptr;
+  if (OldRM == llvm::RoundingMode::Dynamic) {
+    llvm::Function *FGetRound = CGM.getIntrinsic(llvm::Intrinsic::flt_rounds);
+    PreviousRM = Builder.CreateCall(FGetRound);
+  } else {
+    PreviousRM = llvm::ConstantInt::get(Int32Ty, static_cast<uint64_t>(OldRM));
+  }
+  if (NewRM != llvm::RoundingMode::Dynamic)
+    Builder.CreateIntrinsic(
+        llvm::Intrinsic::set_rounding, {},
+        llvm::ConstantInt::get(Int32Ty, static_cast<uint64_t>(NewRM)));
+  EHStack.pushCleanup<FPControlModesCleanup>(NormalAndEHCleanup, PreviousRM);
+}
+
 /// EmitCompoundStmt - Emit a compound statement {..} node.  If GetLast is true,
 /// this captures the expression result of the last sub-statement and returns it
 /// (for use by the statement expression extension).
 Address CodeGenFunction::EmitCompoundStmt(const CompoundStmt &S, bool GetLast,
                                           AggValueSlot AggSlot) {
   PrettyStackTraceLoc CrashInfo(getContext().getSourceManager(),S.getLBracLoc(),
                              "LLVM IR generation of compound statement ('{}')");
 
   // Keep track of the current cleanup stack depth, including debug scopes.
   LexicalScope Scope(*this, S.getSourceRange());
 
   return EmitCompoundStmtWithoutScope(S, GetLast, AggSlot);
 }
 
 Address
 CodeGenFunction::EmitCompoundStmtWithoutScope(const CompoundStmt &S,
                                               bool GetLast,
                                               AggValueSlot AggSlot) {
 
   const Stmt *ExprResult = S.getStmtExprResult();
   assert((!GetLast || (GetLast && ExprResult)) &&
          "If GetLast is true then the CompoundStmt must have a StmtExprResult");
 
+  // Optionally set up the new FP environment, if the compound statement
+  // contains a pragma that modifies it.
+  emitSetFPControlModes(S);
   CGFPOptionsRAII SavedFPFeatues(*this, S.getNewFPOptions(CurFPFeatures));
 
   Address RetAlloca = Address::invalid();
 
   for (auto *CurStmt : S.body()) {
     if (GetLast && ExprResult == CurStmt) {
       // We have to special case labels here.  They are statements, but when put
       // at the end of a statement expression, they yield the value of their

clang/lib/CodeGen/CodeGenFunction.h

ComplexPairTy EmitComplexPrePostIncDec(const UnaryOperator *E, LValue LV,
                                          bool isInc, bool isPre);
 
   /// Converts Location to a DebugLoc, if debug information is enabled.
   llvm::DebugLoc SourceLocToDebugLoc(SourceLocation Location);
 
   /// Get the record field index as represented in debug info.
   unsigned getDebugInfoFIndex(const RecordDecl *Rec, unsigned FieldIndex);
 
+  /// Optionally emit code that sets required floating-point control modes (like
+  /// rounding direction). Also creates corresponding cleanup action if the code
+  /// is emitted.
+  void emitSetFPControlModes(const CompoundStmt &S);
 
   //===--------------------------------------------------------------------===//
   //                            Declaration Emission
   //===--------------------------------------------------------------------===//
 
   /// EmitDecl - Emit a declaration.
   ///
   /// This function can be called with a null (unreachable) insert point.

clang/lib/Parse/ParsePragma.cpp

void PragmaSTDC_FENV_ROUNDHandler::HandlePragma(Preprocessor &PP,
   PP.Lex(Tok);
 
   if (Tok.isNot(tok::eod)) {
     PP.Diag(Tok.getLocation(), diag::warn_pragma_extra_tokens_at_eol)
         << "STDC FENV_ROUND";
     return;
   }
 
-  // Until the pragma is fully implemented, issue a warning.
-  PP.Diag(Tok.getLocation(), diag::warn_stdc_fenv_round_not_supported);
-
   MutableArrayRef<Token> Toks(PP.getPreprocessorAllocator().Allocate<Token>(1),
                               1);
   Toks[0].startToken();
   Toks[0].setKind(tok::annot_pragma_fenv_round);
   Toks[0].setLocation(Tok.getLocation());
   Toks[0].setAnnotationEndLoc(Tok.getLocation());
   Toks[0].setAnnotationValue(
       reinterpret_cast<void *>(static_cast<uintptr_t>(RM)));

clang/test/CodeGen/complex-strictfp.c

 #pragma STDC FENV_ROUND FE_UPWARD
 
 _Complex double g1, g2;
 _Complex float cf;
 double D;
 
 // CHECK-LABEL: @test3a(
 // CHECK-NEXT:  entry:
+// CHECK-NEXT:    call void @llvm.set.rounding(i32 2) #[[ATTR2:[0-9]+]]
 // CHECK-NEXT:    [[TMP0:%.*]] = load double, double* @D, align 8
 // CHECK-NEXT:    [[CF_REAL:%.*]] = load float, float* getelementptr inbounds ({ float, float }, { float, float }* @cf, i32 0, i32 0), align 4
 // CHECK-NEXT:    [[CF_IMAG:%.*]] = load float, float* getelementptr inbounds ({ float, float }, { float, float }* @cf, i32 0, i32 1), align 4
-// CHECK-NEXT:    [[CONV:%.*]] = call double @llvm.experimental.constrained.fpext.f64.f32(float [[CF_REAL]], metadata !"fpexcept.strict") #[[ATTR2:[0-9]+]]
+// CHECK-NEXT:    [[CONV:%.*]] = call double @llvm.experimental.constrained.fpext.f64.f32(float [[CF_REAL]], metadata !"fpexcept.strict") #[[ATTR2]]
 // CHECK-NEXT:    [[CONV1:%.*]] = call double @llvm.experimental.constrained.fpext.f64.f32(float [[CF_IMAG]], metadata !"fpexcept.strict") #[[ATTR2]]
 // CHECK-NEXT:    [[ADD_R:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[CONV]], double [[TMP0]], metadata !"round.upward", metadata !"fpexcept.strict") #[[ATTR2]]
 // CHECK-NEXT:    [[CONV2:%.*]] = call float @llvm.experimental.constrained.fptrunc.f32.f64(double [[ADD_R]], metadata !"round.upward", metadata !"fpexcept.strict") #[[ATTR2]]
 // CHECK-NEXT:    [[CONV3:%.*]] = call float @llvm.experimental.constrained.fptrunc.f32.f64(double [[CONV1]], metadata !"round.upward", metadata !"fpexcept.strict") #[[ATTR2]]
 // CHECK-NEXT:    store float [[CONV2]], float* getelementptr inbounds ({ float, float }, { float, float }* @cf, i32 0, i32 0), align 4
 // CHECK-NEXT:    store float [[CONV3]], float* getelementptr inbounds ({ float, float }, { float, float }* @cf, i32 0, i32 1), align 4
+// CHECK-NEXT:    call void @llvm.set.rounding(i32 1) #[[ATTR2]]
 // CHECK-NEXT:    ret void
 //
 void test3a(void) {
   cf += D;
 }
 
 // CHECK-LABEL: @test3b(
 // CHECK-NEXT:  entry:
+// CHECK-NEXT:    call void @llvm.set.rounding(i32 2) #[[ATTR2]]
 // CHECK-NEXT:    [[CF_REAL:%.*]] = load float, float* getelementptr inbounds ({ float, float }, { float, float }* @cf, i32 0, i32 0), align 4
 // CHECK-NEXT:    [[CF_IMAG:%.*]] = load float, float* getelementptr inbounds ({ float, float }, { float, float }* @cf, i32 0, i32 1), align 4
 // CHECK-NEXT:    [[CONV:%.*]] = call double @llvm.experimental.constrained.fpext.f64.f32(float [[CF_REAL]], metadata !"fpexcept.strict") #[[ATTR2]]
 // CHECK-NEXT:    [[CONV1:%.*]] = call double @llvm.experimental.constrained.fpext.f64.f32(float [[CF_IMAG]], metadata !"fpexcept.strict") #[[ATTR2]]
 // CHECK-NEXT:    [[TMP0:%.*]] = load double, double* @D, align 8
 // CHECK-NEXT:    [[ADD_R:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[TMP0]], double [[CONV]], metadata !"round.upward", metadata !"fpexcept.strict") #[[ATTR2]]
 // CHECK-NEXT:    store double [[ADD_R]], double* @D, align 8
+// CHECK-NEXT:    call void @llvm.set.rounding(i32 1) #[[ATTR2]]
 // CHECK-NEXT:    ret void
 //
 void test3b(void) {
   D += cf;
 }
 
 // CHECK-LABEL: @test3c(
 // CHECK-NEXT:  entry:
+// CHECK-NEXT:    call void @llvm.set.rounding(i32 2) #[[ATTR2]]
 // CHECK-NEXT:    [[G1_REAL:%.*]] = load double, double* getelementptr inbounds ({ double, double }, { double, double }* @g1, i32 0, i32 0), align 8
 // CHECK-NEXT:    [[G1_IMAG:%.*]] = load double, double* getelementptr inbounds ({ double, double }, { double, double }* @g1, i32 0, i32 1), align 8
 // CHECK-NEXT:    [[CF_REAL:%.*]] = load float, float* getelementptr inbounds ({ float, float }, { float, float }* @cf, i32 0, i32 0), align 4
 // CHECK-NEXT:    [[CF_IMAG:%.*]] = load float, float* getelementptr inbounds ({ float, float }, { float, float }* @cf, i32 0, i32 1), align 4
 // CHECK-NEXT:    [[CONV:%.*]] = call double @llvm.experimental.constrained.fpext.f64.f32(float [[CF_REAL]], metadata !"fpexcept.strict") #[[ATTR2]]
 // CHECK-NEXT:    [[CONV1:%.*]] = call double @llvm.experimental.constrained.fpext.f64.f32(float [[CF_IMAG]], metadata !"fpexcept.strict") #[[ATTR2]]
 // CHECK-NEXT:    [[CALL:%.*]] = call { double, double } @__divdc3(double noundef [[CONV]], double noundef [[CONV1]], double noundef [[G1_REAL]], double noundef [[G1_IMAG]]) #[[ATTR3:[0-9]+]]
 // CHECK-NEXT:    [[TMP0:%.*]] = extractvalue { double, double } [[CALL]], 0
 // CHECK-NEXT:    [[TMP1:%.*]] = extractvalue { double, double } [[CALL]], 1
 // CHECK-NEXT:    [[CONV2:%.*]] = call float @llvm.experimental.constrained.fptrunc.f32.f64(double [[TMP0]], metadata !"round.upward", metadata !"fpexcept.strict") #[[ATTR2]]
 // CHECK-NEXT:    [[CONV3:%.*]] = call float @llvm.experimental.constrained.fptrunc.f32.f64(double [[TMP1]], metadata !"round.upward", metadata !"fpexcept.strict") #[[ATTR2]]
 // CHECK-NEXT:    store float [[CONV2]], float* getelementptr inbounds ({ float, float }, { float, float }* @cf, i32 0, i32 0), align 4
 // CHECK-NEXT:    store float [[CONV3]], float* getelementptr inbounds ({ float, float }, { float, float }* @cf, i32 0, i32 1), align 4
+// CHECK-NEXT:    call void @llvm.set.rounding(i32 1) #[[ATTR2]]
 // CHECK-NEXT:    ret void
 //
 void test3c(void) {
   cf /= g1;
 }
 
 // CHECK-LABEL: @test3d(
 // CHECK-NEXT:  entry:
+// CHECK-NEXT:    call void @llvm.set.rounding(i32 2) #[[ATTR2]]
 // CHECK-NEXT:    [[G1_REAL:%.*]] = load double, double* getelementptr inbounds ({ double, double }, { double, double }* @g1, i32 0, i32 0), align 8
 // CHECK-NEXT:    [[G1_IMAG:%.*]] = load double, double* getelementptr inbounds ({ double, double }, { double, double }* @g1, i32 0, i32 1), align 8
 // CHECK-NEXT:    [[TMP0:%.*]] = load double, double* @D, align 8
 // CHECK-NEXT:    [[ADD_R:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[G1_REAL]], double [[TMP0]], metadata !"round.upward", metadata !"fpexcept.strict") #[[ATTR2]]
 // CHECK-NEXT:    store double [[ADD_R]], double* getelementptr inbounds ({ double, double }, { double, double }* @g1, i32 0, i32 0), align 8
 // CHECK-NEXT:    store double [[G1_IMAG]], double* getelementptr inbounds ({ double, double }, { double, double }* @g1, i32 0, i32 1), align 8
+// CHECK-NEXT:    call void @llvm.set.rounding(i32 1) #[[ATTR2]]
 // CHECK-NEXT:    ret void
 //
 void test3d(void) {
   g1 = g1 + D;
 }
 
 // CHECK-LABEL: @test3e(
 // CHECK-NEXT:  entry:
+// CHECK-NEXT:    call void @llvm.set.rounding(i32 2) #[[ATTR2]]
 // CHECK-NEXT:    [[TMP0:%.*]] = load double, double* @D, align 8
 // CHECK-NEXT:    [[G1_REAL:%.*]] = load double, double* getelementptr inbounds ({ double, double }, { double, double }* @g1, i32 0, i32 0), align 8
 // CHECK-NEXT:    [[G1_IMAG:%.*]] = load double, double* getelementptr inbounds ({ double, double }, { double, double }* @g1, i32 0, i32 1), align 8
 // CHECK-NEXT:    [[ADD_R:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[TMP0]], double [[G1_REAL]], metadata !"round.upward", metadata !"fpexcept.strict") #[[ATTR2]]
 // CHECK-NEXT:    store double [[ADD_R]], double* getelementptr inbounds ({ double, double }, { double, double }* @g1, i32 0, i32 0), align 8
 // CHECK-NEXT:    store double [[G1_IMAG]], double* getelementptr inbounds ({ double, double }, { double, double }* @g1, i32 0, i32 1), align 8
+// CHECK-NEXT:    call void @llvm.set.rounding(i32 1) #[[ATTR2]]
 // CHECK-NEXT:    ret void
 //
 void test3e(void) {
   g1 = D + g1;
 }
 
 // CHECK-LABEL: @t1(
 // CHECK-NEXT:  entry:
+// CHECK-NEXT:    call void @llvm.set.rounding(i32 2) #[[ATTR2]]
 // CHECK-NEXT:    [[CONV:%.*]] = call float @llvm.experimental.constrained.fptrunc.f32.f64(double 4.000000e+00, metadata !"round.upward", metadata !"fpexcept.strict") #[[ATTR2]]
 // CHECK-NEXT:    store float [[CONV]], float* getelementptr inbounds ({ float, float }, { float, float }* @cf, i32 0, i32 0), align 4
+// CHECK-NEXT:    call void @llvm.set.rounding(i32 1) #[[ATTR2]]
 // CHECK-NEXT:    ret void
 //
 void t1(void) {
   (__real__ cf) = 4.0;
 }
 
 // CHECK-LABEL: @t2(
 // CHECK-NEXT:  entry:
+// CHECK-NEXT:    call void @llvm.set.rounding(i32 2) #[[ATTR2]]
 // CHECK-NEXT:    [[CONV:%.*]] = call float @llvm.experimental.constrained.fptrunc.f32.f64(double 4.000000e+00, metadata !"round.upward", metadata !"fpexcept.strict") #[[ATTR2]]
 // CHECK-NEXT:    store float [[CONV]], float* getelementptr inbounds ({ float, float }, { float, float }* @cf, i32 0, i32 1), align 4
+// CHECK-NEXT:    call void @llvm.set.rounding(i32 1) #[[ATTR2]]
 // CHECK-NEXT:    ret void
 //
 void t2(void) {
   (__imag__ cf) = 4.0;
 }
 
 // CHECK-LABEL: @t91(
 // CHECK-NEXT:  entry:
 // CHECK-NEXT:    [[C:%.*]] = alloca [0 x i8], align 1
+// CHECK-NEXT:    call void @llvm.set.rounding(i32 2) #[[ATTR2]]
 // CHECK-NEXT:    br i1 false, label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
 // CHECK:       cond.true:
 // CHECK-NEXT:    [[CONV:%.*]] = call double @llvm.experimental.constrained.fpext.f64.f32(float 2.000000e+00, metadata !"fpexcept.strict") #[[ATTR2]]
 // CHECK-NEXT:    br label [[COND_END:%.*]]
 // CHECK:       cond.false:
 // CHECK-NEXT:    [[CONV1:%.*]] = call double @llvm.experimental.constrained.fpext.f64.f32(float 2.000000e+00, metadata !"fpexcept.strict") #[[ATTR2]]
 // CHECK-NEXT:    br label [[COND_END]]
 // CHECK:       cond.end:
 // CHECK-NEXT:    [[COND_R:%.*]] = phi double [ [[CONV]], [[COND_TRUE]] ], [ [[CONV1]], [[COND_FALSE]] ]
 // CHECK-NEXT:    [[COND_I:%.*]] = phi double [ 0.000000e+00, [[COND_TRUE]] ], [ 0.000000e+00, [[COND_FALSE]] ]
+// CHECK-NEXT:    call void @llvm.set.rounding(i32 1) #[[ATTR2]]
 // CHECK-NEXT:    ret void
 //
 void t91(void) {
   // Check for proper type promotion of conditional expression
   char c[(int)(sizeof(typeof((0 ? 2.0f : (_Complex double) 2.0f))) - sizeof(_Complex double))];
   // Check for proper codegen
   (0 ? 2.0f : (_Complex double) 2.0f);
 }
 
 // CHECK-LABEL: @t92(
 // CHECK-NEXT:  entry:
 // CHECK-NEXT:    [[C:%.*]] = alloca [0 x i8], align 1
+// CHECK-NEXT:    call void @llvm.set.rounding(i32 2) #[[ATTR2]]
 // CHECK-NEXT:    br i1 false, label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
 // CHECK:       cond.true:
 // CHECK-NEXT:    [[CONV:%.*]] = call double @llvm.experimental.constrained.fpext.f64.f32(float 2.000000e+00, metadata !"fpexcept.strict") #[[ATTR2]]
 // CHECK-NEXT:    br label [[COND_END:%.*]]
 // CHECK:       cond.false:
 // CHECK-NEXT:    [[CONV1:%.*]] = call double @llvm.experimental.constrained.fpext.f64.f32(float 2.000000e+00, metadata !"fpexcept.strict") #[[ATTR2]]
 // CHECK-NEXT:    br label [[COND_END]]
 // CHECK:       cond.end:
 // CHECK-NEXT:    [[COND_R:%.*]] = phi double [ [[CONV]], [[COND_TRUE]] ], [ [[CONV1]], [[COND_FALSE]] ]
 // CHECK-NEXT:    [[COND_I:%.*]] = phi double [ 0.000000e+00, [[COND_TRUE]] ], [ 0.000000e+00, [[COND_FALSE]] ]
+// CHECK-NEXT:    call void @llvm.set.rounding(i32 1) #[[ATTR2]]
 // CHECK-NEXT:    ret void
 //
 void t92(void) {
   // Check for proper type promotion of conditional expression
   char c[(int)(sizeof(typeof((0 ? (_Complex double) 2.0f : 2.0f))) - sizeof(_Complex double))];
   // Check for proper codegen
   (0 ? (_Complex double) 2.0f : 2.0f);
 }
 

clang/test/CodeGen/pragma-fenv_round.c

+// RUN: %clang_cc1 -S -triple x86_64-linux-gnu -emit-llvm %s -o - | FileCheck %s
+
+float default_mode = 1.0f / 3.0f;
+// CHECK: @default_mode = global float 0x3FD5555560000000
+
+#pragma STDC FENV_ROUND FE_UPWARD
+float upward = 1.0f / 3.0f;
+// CHECK: @upward = global float 0x3FD5555560000000
+
+#pragma STDC FENV_ROUND FE_DOWNWARD
+float downward = 1.0f / 3.0f;
+// CHECK: @downward = global float 0x3FD5555540000000
+
+#pragma STDC FENV_ROUND FE_DYNAMIC
+float dynamic = 1.0f / 3.0f;
+// CHECK: @dynamic = global float 0x3FD5555560000000
+
+#pragma STDC FENV_ROUND FE_TONEAREST
+
+
+float func_1(float w, float x, float y, float z) {
+  #pragma STDC FENV_ROUND FE_TOWARDZERO
+  float result = x * y;
+  {
+    #pragma STDC FENV_ROUND FE_UPWARD
+    result += z;
+  }
+  return result - w;
+}
+
+// CHECK-LABEL: @func_1
+// CHECK:  call void @llvm.set.rounding(i32 0)
+// CHECK:  call float @llvm.experimental.constrained.fmul.f32({{.*}}, metadata !"round.towardzero", metadata !"fpexcept.ignore")
+// CHECK:  call void @llvm.set.rounding(i32 2)
+// CHECK:  call float @llvm.experimental.constrained.fadd.f32({{.*}}, metadata !"round.upward", metadata !"fpexcept.ignore")
+// CHECK:  call void @llvm.set.rounding(i32 0)
+// CHECK:  call float @llvm.experimental.constrained.fsub.f32({{.*}}, metadata !"round.towardzero", metadata !"fpexcept.ignore")
+// CHECK:  call void @llvm.set.rounding(i32 1)
+
+
+float func_2(float w, float x, float y, float z) {
+  #pragma STDC FENV_ROUND FE_TOWARDZERO
+  float result = x * y;
+  {
+    #pragma STDC FENV_ROUND FE_TOWARDZERO
+    result += z;
+  }
+  return result - w;
+}
+
+// CHECK-LABEL: @func_2
+// CHECK:  call void @llvm.set.rounding(i32 0)
+// CHECK:  call float @llvm.experimental.constrained.fmul.f32({{.*}}, metadata !"round.towardzero", metadata !"fpexcept.ignore")
+// CHECK:  call float @llvm.experimental.constrained.fadd.f32({{.*}}, metadata !"round.towardzero", metadata !"fpexcept.ignore")
+// CHECK:  call float @llvm.experimental.constrained.fsub.f32({{.*}}, metadata !"round.towardzero", metadata !"fpexcept.ignore")
+// CHECK:  call void @llvm.set.rounding(i32 1)
+
+
+float func_3(float w, float x, float y, float z) {
+  #pragma STDC FENV_ROUND FE_TOWARDZERO
+  float result = x * y;
+  {
+    #pragma STDC FENV_ROUND FE_DYNAMIC
+    result += z;
+  }
+  return result - w;
+}
+
+// CHECK-LABEL: @func_3
+// CHECK:  call void @llvm.set.rounding(i32 0)
+// CHECK:  call float @llvm.experimental.constrained.fmul.f32({{.*}}, metadata !"round.towardzero", metadata !"fpexcept.ignore")
+// CHECK:  call void @llvm.set.rounding(i32 1)
+// CHECK:  call float @llvm.experimental.constrained.fadd.f32({{.*}}, metadata !"round.tonearest", metadata !"fpexcept.ignore")
+// CHECK:  call void @llvm.set.rounding(i32 0)
+// CHECK:  call float @llvm.experimental.constrained.fsub.f32({{.*}}, metadata !"round.towardzero", metadata !"fpexcept.ignore")
+// CHECK:  call void @llvm.set.rounding(i32 1)
+
+
+float func_4(float w, float x, float y, float z) {
+  #pragma STDC FENV_ROUND FE_TOWARDZERO
+  float result = x * y;
+  {
+    #pragma STDC FENV_ROUND FE_DYNAMIC
+    #pragma STDC FENV_ACCESS ON
+    result += z;
+  }
+  return result - w;
+}
+
+// CHECK-LABEL: @func_4
+// CHECK:  call void @llvm.set.rounding(i32 0)
+// CHECK:  call float @llvm.experimental.constrained.fmul.f32({{.*}}, metadata !"round.towardzero", metadata !"fpexcept.ignore")
+// CHECK:  call float @llvm.experimental.constrained.fadd.f32({{.*}}, metadata !"round.dynamic", metadata !"fpexcept.strict")
+// CHECK:  call void @llvm.set.rounding(i32 0)
+// CHECK:  call float @llvm.experimental.constrained.fsub.f32({{.*}}, metadata !"round.towardzero", metadata !"fpexcept.ignore")
+// CHECK:  call void @llvm.set.rounding(i32 1)
+
+
+float func_5(float w, float x, float y, float z) {
+  #pragma STDC FENV_ROUND FE_DYNAMIC
+  #pragma STDC FENV_ACCESS ON
+  #pragma clang fp exceptions(ignore)
+  float result = x * y;
+  {
+    #pragma STDC FENV_ROUND FE_UPWARD
+    result += z;
+  }
+  return result - w;
+}
+
+// CHECK-LABEL: @func_5
+// CHECK:  call float @llvm.experimental.constrained.fmul.f32({{.*}}, metadata !"round.dynamic", metadata !"fpexcept.ignore")
+// CHECK:  [[RM:%[0-9]+]] = call i32 @llvm.flt.rounds()
+// CHECK:  call void @llvm.set.rounding(i32 2)
+// CHECK:  call float @llvm.experimental.constrained.fadd.f32({{.*}}, metadata !"round.upward", metadata !"fpexcept.ignore")
+// CHECK:  call void @llvm.set.rounding(i32 [[RM]])
+// CHECK:  call float @llvm.experimental.constrained.fsub.f32({{.*}}, metadata !"round.dynamic", metadata !"fpexcept.ignore")
+// CHECK:  call void @llvm.set.rounding(i32 1)
+
+
+float func_6(float w, float x, float y, float z) {
+  #pragma STDC FENV_ROUND FE_DYNAMIC
+  #pragma STDC FENV_ACCESS ON
+  #pragma clang fp exceptions(ignore)
+  float result = x * y;
+  {
+    #pragma STDC FENV_ROUND FE_DYNAMIC
+    result += z;
+  }
+  return result - w;
+}
+
+// CHECK-LABEL: @func_6
+// CHECK:  call float @llvm.experimental.constrained.fmul.f32({{.*}}, metadata !"round.dynamic", metadata !"fpexcept.ignore")
+// CHECK:  call float @llvm.experimental.constrained.fadd.f32({{.*}}, metadata !"round.dynamic", metadata !"fpexcept.ignore")
+// CHECK:  call float @llvm.experimental.constrained.fsub.f32({{.*}}, metadata !"round.dynamic", metadata !"fpexcept.ignore")
+// CHECK:  call void @llvm.set.rounding(i32 1)
+
+
+#pragma STDC FENV_ROUND FE_DOWNWARD
+float func_20(float x, float y) {
+  return x + y;
+}
+
+// CHECK-LABEL: func_20
+// CHECK:  call void @llvm.set.rounding(i32 3)
+// CHECK:  call float @llvm.experimental.constrained.fadd.f32({{.*}}, metadata !"round.downward", metadata !"fpexcept.ignore")
+// CHECK:  call void @llvm.set.rounding(i32 1)
+
+float func_21(float x, float y) {
+  return x - y;
+}
+
+// CHECK-LABEL: func_21
+// CHECK:  call void @llvm.set.rounding(i32 3)
+// CHECK:  call float @llvm.experimental.constrained.fsub.f32({{.*}}, metadata !"round.downward", metadata !"fpexcept.ignore")
+// CHECK:  call void @llvm.set.rounding(i32 1)
+
+#pragma STDC FENV_ROUND FE_DOWNWARD
+float func_22(int x) {
+  return x;
+}
+
+// CHECK-LABEL: @func_22
+// CHECK:  call void @llvm.set.rounding(i32 3)
+// CHECK:  call float @llvm.experimental.constrained.sitofp.f32.i32({{.*}}, metadata !"round.downward", metadata !"fpexcept.ignore")
+// CHECK:  call void @llvm.set.rounding(i32 1)
+
+
+#pragma STDC FENV_ROUND FE_DOWNWARD
+float func_23(float w, float x, float y, float z) {
+  #pragma STDC FENV_ROUND FE_TOWARDZERO
+  float result = x * y;
+  {
+    result += z;
+  }
+  return result - w;
+}
+
+// CHECK-LABEL: @func_23
+// CHECK:  call void @llvm.set.rounding(i32 0)
+// CHECK:  call float @llvm.experimental.constrained.fmul.f32({{.*}}, metadata !"round.towardzero", metadata !"fpexcept.ignore")
+// CHECK:  call float @llvm.experimental.constrained.fadd.f32({{.*}}, metadata !"round.towardzero", metadata !"fpexcept.ignore")
+// CHECK:  call float @llvm.experimental.constrained.fsub.f32({{.*}}, metadata !"round.towardzero", metadata !"fpexcept.ignore")
+// CHECK:  call void @llvm.set.rounding(i32 1)
+
+float func_24(float w, float x, float y, float z) {
+  #pragma STDC FENV_ACCESS ON
+  #pragma STDC FENV_ROUND FE_DYNAMIC
+  float result = x * y;
+  {
+    #pragma STDC FENV_ROUND FE_TOWARDZERO
+    result += z;
+  }
+  return result - w;
+}
+
+// CHECK-LABEL: func_24
+// CHECK: call float @llvm.experimental.constrained.fmul.f32({{.*}}, metadata !"round.dynamic", metadata !"fpexcept.strict")
+// CHECK: [[RM:%.*]] = call i32 @llvm.flt.rounds()
+// CHECK: call void @llvm.set.rounding(i32 0)
+// CHECK: call float @llvm.experimental.constrained.fadd.f32({{.*}}, metadata !"round.towardzero", metadata !"fpexcept.strict")
+// CHECK: call void @llvm.set.rounding(i32 [[RM]])
+// CHECK: call float @llvm.experimental.constrained.fsub.f32({{.*}}, metadata !"round.dynamic", metadata !"fpexcept.strict")
+// CHECK: call void @llvm.set.rounding(i32 1)

clang/test/CodeGen/pragma-fenv_round.cpp

+// RUN: %clang_cc1 -S -triple x86_64-linux-gnu -emit-llvm %s -o - | FileCheck %s
+
+#pragma STDC FENV_ROUND FE_DOWNWARD
+
+float func_01(float x, float y, float z = 1.0f/3.0f) {
+  return x + y + z;
+}
+// CHECK-LABEL: @_Z7func_01fff
+// CHECK: call void @llvm.set.rounding(i32 3)
+// CHECK: call float @llvm.experimental.constrained.fadd.f32({{.*}}, metadata !"round.downward", metadata !"fpexcept.ignore")
+// CHECK: call float @llvm.experimental.constrained.fadd.f32({{.*}}, metadata !"round.downward", metadata !"fpexcept.ignore")
+// CHECK: call void @llvm.set.rounding(i32 1)
+
+float func_02() {
+#pragma STDC FENV_ROUND FE_UPWARD
+  return func_01(1.0f, 2.0f);
+}
+// CHECK-LABEL: @_Z7func_02v
+// CHECK: call void @llvm.set.rounding(i32 2)
+// CHECK: call float @llvm.experimental.constrained.fdiv.f32(float 1.000000e+00, float 3.000000e+00, metadata !"round.downward", metadata !"fpexcept.ignore")
+// CHECK: call noundef float @_Z7func_01fff(float noundef 1.000000e+00, float noundef 2.000000e+00, float noundef {{.*}})
+// CHECK: call void @llvm.set.rounding(i32 1)
+
+float func_03(float x, float y, float z = 1.0f/3.0f) {
+#pragma STDC FENV_ROUND FE_TOWARDZERO
+  return x + y + z;
+}
+// CHECK-LABEL: @_Z7func_03fff
+// CHECK: call void @llvm.set.rounding(i32 0)
+// CHECK: call float @llvm.experimental.constrained.fadd.f32({{.*}}, metadata !"round.towardzero", metadata !"fpexcept.ignore")
+// CHECK: call float @llvm.experimental.constrained.fadd.f32({{.*}}, metadata !"round.towardzero", metadata !"fpexcept.ignore")
+// CHECK: call void @llvm.set.rounding(i32 1)
+
+float func_04() {
+#pragma STDC FENV_ROUND FE_UPWARD
+  return func_01(1.0f, 2.0f);
+}
+// CHECK-LABEL: @_Z7func_04v
+// CHECK: call void @llvm.set.rounding(i32 2)
+// CHECK: call float @llvm.experimental.constrained.fdiv.f32(float 1.000000e+00, float 3.000000e+00, metadata !"round.downward", metadata !"fpexcept.ignore")
+// CHECK: call noundef float @_Z7func_01fff(float noundef 1.000000e+00, float noundef 2.000000e+00, float noundef {{.*}})
+// CHECK: call void @llvm.set.rounding(i32 1)

clang/test/Parser/pragma-fenv_round.c

 // RUN: %clang_cc1 -fexperimental-strict-floating-point -fsyntax-only -Wignored-pragmas -verify %s
 
 #pragma STDC FENV_ROUND ON   // expected-warning {{invalid or unsupported rounding mode}}
 
 float func_01(int x, float y) {
   if (x)
     return y + 2;
   #pragma STDC FENV_ROUND FE_DOWNWARD // expected-error{{'#pragma STDC FENV_ROUND' can only appear at file scope or at the start of a compound statement}}
-                                      // expected-warning@-1{{pragma STDC FENV_ROUND is not supported}}
   return x + y;
 }