[X86] Custom lower ISD::FROUND with SSE4.1 to avoid a libcall.

[craig.topper]

Summary

ISD::FROUND is defined to round to nearest with ties rounding
away from 0. This mode isn't supported in hardware on X86.

But as long as we aren't compiling with trapping math, we can
emulate this with floor(X + copysign(nextafter(0.5, 0.0), X)).

We have to use nextafter to avoid some corner cases that adding
0.5 would have. For example, if X is nextafter(0.5, 0.0) it should
round to 0.0, but adding 0.5 would need one extra bit of mantissa
than can be stored so it rounds to 1.0. Adding nextafter(0.5, 0.0)
instead will just increase the exponent by 1 and leave the mantissa
as all 1s. This would be nextafter(1.0, 0.0) which will floor to 0.0.

Techically this requires -fno-trapping-math which isn't our default.
But if we care about exceptions we should be using constrained
intrinsics. Constrained intrinsics would use STRICT_FROUND which
won't go through this code.

Fixes PR42195.

[spatel]

LGTM

llvm/lib/Target/X86/X86ISelLowering.cpp

20460

Include at least part of the text from this patch description as a block comment for this function, so:

/// ISD::FROUND is defined to round to nearest with ties rounding
/// away from 0. This mode isn't supported in hardware on X86.
/// But as long as we aren't compiling with trapping math, we can
/// emulate this with floor(X + copysign(nextafter(0.5, 0.0), X)).
/// ...

llvm/test/CodeGen/X86/vec_round.ll

9

This test/file was added with:
rL188048
...to show we wouldn't crash?
But it doesn't add value now that we have more thorough tests. I'd delete it.

[rtrieu]

I ran across a llvm_unreachable that points to this commit. Repro instructions below:

test.ii

double compare1(double x, double y) { return ((int)x< y) ? x : y; }
double compare2(double x, double y) { return y != 0.0 ? y : x; }

int compareint(int x, int y, int z) { return (x < y) ? y : (z < x) ? z : x; }

class C {
public:
  C(double arg) { 
    constexpr int k1 = -(1 << 23); 
    constexpr int k2 = (1 << 23) - 1;
    array[0] = compareint(arg, k1, k2);
  }
  
  char array[3];
};
extern "C" double round(double);
constexpr double kEight = 8;

C create(int b) {
  double d1 = b * kEight;
  double d2 = round(d1);
  double d3 = compare1(0.0, d2);
  double d4 = compare2(0.0, d3);
  return C(d4);
}

void loop(int* b, C *ptr, long j) {
  for (int i = 0; i < j; ++i)
    ptr[i] = create(b[i]);
}

clang command:

clang \
"-cc1" \
"-triple" "x86_64-unknown-linux-gnu" \
"-emit-obj" \
"-target-cpu" "x86-64" \
"-target-feature" "+avx" \
"-target-feature" "+avx2" \
"-target-feature" "+avx512f" \
"-O3" \
"-vectorize-loops" \
"-x" "c++" "test.ii"

output:

PromoteIntegerResult #0: t177: v16i24 = X86ISD::VTRUNCS t216

Do not know how to promote this operator!
UNREACHABLE executed at llvm-project/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp:54!

0.	Program arguments: clang -cc1 -triple x86_64-unknown-linux-gnu -emit-obj -target-cpu x86-64 -target-feature +avx -target-feature +avx2 -target-feature +avx512f -O3 -vectorize-loops -x c++ test.ii
1.	<eof> parser at end of file
2.	Code generation
3.	Running pass 'Function Pass Manager' on module 'test.ii'.
4.	Running pass 'X86 DAG->DAG Instruction Selection' on function '@_Z4loopPiP1Cl'

[efriedma]

Looks like some combine is incorrectly producing X86ISD::VTRUNCS? Probably just exposed by this patch.

[craig.topper]

In D73607#2806917, @rtrieu wrote:

I ran across a llvm_unreachable that points to this commit. Repro instructions below:

test.ii

double compare1(double x, double y) { return ((int)x< y) ? x : y; }
double compare2(double x, double y) { return y != 0.0 ? y : x; }

int compareint(int x, int y, int z) { return (x < y) ? y : (z < x) ? z : x; }

class C {
public:
  C(double arg) { 
    constexpr int k1 = -(1 << 23); 
    constexpr int k2 = (1 << 23) - 1;
    array[0] = compareint(arg, k1, k2);
  }
  
  char array[3];
};
extern "C" double round(double);
constexpr double kEight = 8;

C create(int b) {
  double d1 = b * kEight;
  double d2 = round(d1);
  double d3 = compare1(0.0, d2);
  double d4 = compare2(0.0, d3);
  return C(d4);
}

void loop(int* b, C *ptr, long j) {
  for (int i = 0; i < j; ++i)
    ptr[i] = create(b[i]);
}

clang command:

clang \
"-cc1" \
"-triple" "x86_64-unknown-linux-gnu" \
"-emit-obj" \
"-target-cpu" "x86-64" \
"-target-feature" "+avx" \
"-target-feature" "+avx2" \
"-target-feature" "+avx512f" \
"-O3" \
"-vectorize-loops" \
"-x" "c++" "test.ii"

output:

PromoteIntegerResult #0: t177: v16i24 = X86ISD::VTRUNCS t216

Do not know how to promote this operator!
UNREACHABLE executed at llvm-project/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp:54!

0.	Program arguments: clang -cc1 -triple x86_64-unknown-linux-gnu -emit-obj -target-cpu x86-64 -target-feature +avx -target-feature +avx2 -target-feature +avx512f -O3 -vectorize-loops -x c++ test.ii
1.	<eof> parser at end of file
2.	Code generation
3.	Running pass 'Function Pass Manager' on module 'test.ii'.
4.	Running pass 'X86 DAG->DAG Instruction Selection' on function '@_Z4loopPiP1Cl'

This should fix it

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index b89e1674..90babf3 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -45273,7 +45273,8 @@ static SDValue combineTruncateWithSat(SDValue In, EVT VT, const SDLoc &DL,
 
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
   if (TLI.isTypeLegal(InVT) && InVT.isVector() && SVT != MVT::i1 &&
-      Subtarget.hasAVX512() && (InSVT != MVT::i16 || Subtarget.hasBWI())) {
+      Subtarget.hasAVX512() && (InSVT != MVT::i16 || Subtarget.hasBWI()) &&
+      (SVT == MVT::i32 || SVT == MVT::i16 || SVT == MVT::i8)) {
     unsigned TruncOpc = 0;
     SDValue SatVal;
     if (auto SSatVal = detectSSatPattern(In, VT)) {

Posted here https://reviews.llvm.org/D103940

[rtrieu]

Thanks for the quick fix. I've verified that it fixes my full test case.

llvm/lib/Target/X86/X86ISelLowering.cpp

for (MVT RoundedTy : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
       setOperationAction(ISD::FCEIL,             RoundedTy,  Legal);
       setOperationAction(ISD::STRICT_FCEIL,      RoundedTy,  Legal);
       setOperationAction(ISD::FTRUNC,            RoundedTy,  Legal);
       setOperationAction(ISD::STRICT_FTRUNC,     RoundedTy,  Legal);
       setOperationAction(ISD::FRINT,             RoundedTy,  Legal);
       setOperationAction(ISD::STRICT_FRINT,      RoundedTy,  Legal);
       setOperationAction(ISD::FNEARBYINT,        RoundedTy,  Legal);
       setOperationAction(ISD::STRICT_FNEARBYINT, RoundedTy,  Legal);
+
+      setOperationAction(ISD::FROUND,            RoundedTy,  Custom);
     }
 
     setOperationAction(ISD::SMAX,               MVT::v16i8, Legal);
     setOperationAction(ISD::SMAX,               MVT::v4i32, Legal);
     setOperationAction(ISD::UMAX,               MVT::v8i16, Legal);
     setOperationAction(ISD::UMAX,               MVT::v4i32, Legal);
     setOperationAction(ISD::SMIN,               MVT::v16i8, Legal);
     setOperationAction(ISD::SMIN,               MVT::v4i32, Legal);
for (auto VT : { MVT::v8f32, MVT::v4f64 }) {
       setOperationAction(ISD::FCEIL,             VT, Legal);
       setOperationAction(ISD::STRICT_FCEIL,      VT, Legal);
       setOperationAction(ISD::FTRUNC,            VT, Legal);
       setOperationAction(ISD::STRICT_FTRUNC,     VT, Legal);
       setOperationAction(ISD::FRINT,             VT, Legal);
       setOperationAction(ISD::STRICT_FRINT,      VT, Legal);
       setOperationAction(ISD::FNEARBYINT,        VT, Legal);
       setOperationAction(ISD::STRICT_FNEARBYINT, VT, Legal);
+
+      setOperationAction(ISD::FROUND,            VT, Custom);
+
       setOperationAction(ISD::FNEG,              VT, Custom);
       setOperationAction(ISD::FABS,              VT, Custom);
       setOperationAction(ISD::FCOPYSIGN,         VT, Custom);
     }
 
     // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
     // even though v8i16 is a legal type.
     setOperationPromotedToType(ISD::FP_TO_SINT,        MVT::v8i16, MVT::v8i32);
for (auto VT : { MVT::v16f32, MVT::v8f64 }) {
       setOperationAction(ISD::STRICT_FCEIL,      VT, Legal);
       setOperationAction(ISD::FTRUNC,            VT, Legal);
       setOperationAction(ISD::STRICT_FTRUNC,     VT, Legal);
       setOperationAction(ISD::FRINT,             VT, Legal);
       setOperationAction(ISD::STRICT_FRINT,      VT, Legal);
       setOperationAction(ISD::FNEARBYINT,        VT, Legal);
       setOperationAction(ISD::STRICT_FNEARBYINT, VT, Legal);
 
+      setOperationAction(ISD::FROUND,            VT, Custom);
+
       setOperationAction(ISD::SELECT,           VT, Custom);
     }
 
     // Without BWI we need to use custom lowering to handle MVT::v64i8 input.
     for (auto VT : {MVT::v16i32, MVT::v8i64, MVT::v64i8}) {
       setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Custom);
       setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Custom);
     }

 /// Depending on uarch and/or optimizing for size, we might prefer to use a
 /// vector operation in place of the typical scalar operation.
 SDValue X86TargetLowering::lowerFaddFsub(SDValue Op, SelectionDAG &DAG) const {
   assert((Op.getValueType() == MVT::f32 || Op.getValueType() == MVT::f64) &&
          "Only expecting float/double");
   return lowerAddSubToHorizontalOp(Op, DAG, Subtarget);
 }
 
+static SDValue LowerFROUND(SDValue Op, SelectionDAG &DAG) {
+  SDValue N0 = Op.getOperand(0);
+  SDLoc dl(Op);
+  MVT VT = Op.getSimpleValueType();
+
+  // N0 += copysign(nextafter(0.5, 0.0), N0)
+  const fltSemantics &Sem = SelectionDAG::EVTToAPFloatSemantics(VT);
+  bool Ignored;
+  APFloat Point5Pred = APFloat(0.5f);
+  Point5Pred.convert(Sem, APFloat::rmNearestTiesToEven, &Ignored);
+  Point5Pred.next(/*nextDown*/true);
+
+  SDValue Adder = DAG.getNode(ISD::FCOPYSIGN, dl, VT,
+                              DAG.getConstantFP(Point5Pred, dl, VT), N0);
+  N0 = DAG.getNode(ISD::FADD, dl, VT, N0, Adder);
+
+  // Truncate the result to remove fraction.
+  return DAG.getNode(ISD::FTRUNC, dl, VT, N0);
+}
+
 /// The only differences between FABS and FNEG are the mask and the logic op.
 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
   assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
          "Wrong opcode for lowering FABS or FNEG.");
 
   bool IsFABS = (Op.getOpcode() == ISD::FABS);
 
SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   case ISD::FP_EXTEND:
   case ISD::STRICT_FP_EXTEND:   return LowerFP_EXTEND(Op, DAG);
   case ISD::FP_ROUND:
   case ISD::STRICT_FP_ROUND:    return LowerFP_ROUND(Op, DAG);
   case ISD::LOAD:               return LowerLoad(Op, Subtarget, DAG);
   case ISD::STORE:              return LowerStore(Op, Subtarget, DAG);
   case ISD::FADD:
   case ISD::FSUB:               return lowerFaddFsub(Op, DAG);
+  case ISD::FROUND:             return LowerFROUND(Op, DAG);
   case ISD::FABS:
   case ISD::FNEG:               return LowerFABSorFNEG(Op, DAG);
   case ISD::FCOPYSIGN:          return LowerFCOPYSIGN(Op, DAG);
   case ISD::FGETSIGN:           return LowerFGETSIGN(Op, DAG);
   case ISD::SETCC:
   case ISD::STRICT_FSETCC:
   case ISD::STRICT_FSETCCS:     return LowerSETCC(Op, DAG);
   case ISD::SETCCCARRY:         return LowerSETCCCARRY(Op, DAG);

llvm/test/CodeGen/X86/extractelement-fp.ll

; X86-NEXT:    retl
   %v = call <4 x double> @llvm.nearbyint.v4f64(<4 x double> %x)
   %r = extractelement <4 x double> %v, i32 0
   ret double %r
 }
 
 define float @round_v4f32(<4 x float> %x) nounwind {
 ; X64-LABEL: round_v4f32:
 ; X64:       # %bb.0:
-; X64-NEXT:    jmp roundf # TAILCALL
+; X64-NEXT:    vbroadcastss {{.*#+}} xmm1 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; X64-NEXT:    vandps %xmm1, %xmm0, %xmm1
+; X64-NEXT:    vbroadcastss {{.*#+}} xmm2 = [4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1]
+; X64-NEXT:    vorps %xmm1, %xmm2, %xmm1
+; X64-NEXT:    vaddss %xmm1, %xmm0, %xmm0
+; X64-NEXT:    vroundss $11, %xmm0, %xmm0, %xmm0
+; X64-NEXT:    retq
 ;
 ; X86-LABEL: round_v4f32:
 ; X86:       # %bb.0:
 ; X86-NEXT:    pushl %eax
+; X86-NEXT:    vbroadcastss {{.*#+}} xmm1 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; X86-NEXT:    vandps %xmm1, %xmm0, %xmm1
+; X86-NEXT:    vbroadcastss {{.*#+}} xmm2 = [4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1]
+; X86-NEXT:    vorps %xmm1, %xmm2, %xmm1
+; X86-NEXT:    vaddss %xmm1, %xmm0, %xmm0
+; X86-NEXT:    vroundss $11, %xmm0, %xmm0, %xmm0
 ; X86-NEXT:    vmovss %xmm0, (%esp)
-; X86-NEXT:    calll roundf
+; X86-NEXT:    flds (%esp)
 ; X86-NEXT:    popl %eax
 ; X86-NEXT:    retl
   %v = call <4 x float> @llvm.round.v4f32(<4 x float> %x)
   %r = extractelement <4 x float> %v, i32 0
   ret float %r
 }
 
 define double @round_v4f64(<4 x double> %x) nounwind {
 ; X64-LABEL: round_v4f64:
 ; X64:       # %bb.0:
-; X64-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; X64-NEXT:    vandpd {{.*}}(%rip), %xmm0, %xmm1
+; X64-NEXT:    vmovddup {{.*#+}} xmm2 = [4.9999999999999994E-1,4.9999999999999994E-1]
+; X64-NEXT:    # xmm2 = mem[0,0]
+; X64-NEXT:    vorpd %xmm1, %xmm2, %xmm1
+; X64-NEXT:    vaddsd %xmm1, %xmm0, %xmm0
+; X64-NEXT:    vroundsd $11, %xmm0, %xmm0, %xmm0
 ; X64-NEXT:    vzeroupper
-; X64-NEXT:    jmp round # TAILCALL
+; X64-NEXT:    retq
 ;
 ; X86-LABEL: round_v4f64:
 ; X86:       # %bb.0:
+; X86-NEXT:    pushl %ebp
+; X86-NEXT:    movl %esp, %ebp
+; X86-NEXT:    andl $-8, %esp
 ; X86-NEXT:    subl $8, %esp
-; X86-NEXT:    vmovlps %xmm0, (%esp)
+; X86-NEXT:    vandpd {{\.LCPI.*}}, %xmm0, %xmm1
+; X86-NEXT:    vmovddup {{.*#+}} xmm2 = [4.9999999999999994E-1,4.9999999999999994E-1]
+; X86-NEXT:    # xmm2 = mem[0,0]
+; X86-NEXT:    vorpd %xmm1, %xmm2, %xmm1
+; X86-NEXT:    vaddsd %xmm1, %xmm0, %xmm0
+; X86-NEXT:    vroundsd $11, %xmm0, %xmm0, %xmm0
+; X86-NEXT:    vmovsd %xmm0, (%esp)
+; X86-NEXT:    fldl (%esp)
+; X86-NEXT:    movl %ebp, %esp
+; X86-NEXT:    popl %ebp
 ; X86-NEXT:    vzeroupper
-; X86-NEXT:    calll round
-; X86-NEXT:    addl $8, %esp
 ; X86-NEXT:    retl
   %v = call <4 x double> @llvm.round.v4f64(<4 x double> %x)
   %r = extractelement <4 x double> %v, i32 0
   ret double %r
 }
 
 define float @rcp_v4f32(<4 x float> %x) nounwind {
 ; X64-LABEL: rcp_v4f32:

llvm/test/CodeGen/X86/fp-round.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+sse2 | FileCheck %s --check-prefix=SSE2
+; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+sse4.1 | FileCheck %s --check-prefixes=SSE41
+; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX1
+; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx512f | FileCheck %s --check-prefixes=AVX,AVX512
+
+define float @round_f32(float %x) {
+; SSE2-LABEL: round_f32:
+; SSE2:       ## %bb.0:
+; SSE2-NEXT:    jmp _roundf ## TAILCALL
+;
+; SSE41-LABEL: round_f32:
+; SSE41:       ## %bb.0:
+; SSE41-NEXT:    movaps {{.*#+}} xmm1 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; SSE41-NEXT:    andps %xmm0, %xmm1
+; SSE41-NEXT:    movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
+; SSE41-NEXT:    orps %xmm1, %xmm2
+; SSE41-NEXT:    addss %xmm0, %xmm2
+; SSE41-NEXT:    xorps %xmm0, %xmm0
+; SSE41-NEXT:    roundss $11, %xmm2, %xmm0
+; SSE41-NEXT:    retq
+;
+; AVX1-LABEL: round_f32:
+; AVX1:       ## %bb.0:
+; AVX1-NEXT:    vandps {{.*}}(%rip), %xmm0, %xmm1
+; AVX1-NEXT:    vbroadcastss {{.*#+}} xmm2 = [4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1]
+; AVX1-NEXT:    vorps %xmm1, %xmm2, %xmm1
+; AVX1-NEXT:    vaddss %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vroundss $11, %xmm0, %xmm0, %xmm0
+; AVX1-NEXT:    retq
+;
+; AVX512-LABEL: round_f32:
+; AVX512:       ## %bb.0:
+; AVX512-NEXT:    vbroadcastss {{.*#+}} xmm1 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; AVX512-NEXT:    vandps %xmm1, %xmm0, %xmm1
+; AVX512-NEXT:    vbroadcastss {{.*#+}} xmm2 = [4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1]
+; AVX512-NEXT:    vorps %xmm1, %xmm2, %xmm1
+; AVX512-NEXT:    vaddss %xmm1, %xmm0, %xmm0
+; AVX512-NEXT:    vroundss $11, %xmm0, %xmm0, %xmm0
+; AVX512-NEXT:    retq
+  %a = call float @llvm.round.f32(float %x)
+  ret float %a
+}
+
+define double @round_f64(double %x) {
+; SSE2-LABEL: round_f64:
+; SSE2:       ## %bb.0:
+; SSE2-NEXT:    jmp _round ## TAILCALL
+;
+; SSE41-LABEL: round_f64:
+; SSE41:       ## %bb.0:
+; SSE41-NEXT:    movapd {{.*#+}} xmm1 = [-0.0E+0,-0.0E+0]
+; SSE41-NEXT:    andpd %xmm0, %xmm1
+; SSE41-NEXT:    movsd {{.*#+}} xmm2 = mem[0],zero
+; SSE41-NEXT:    orpd %xmm1, %xmm2
+; SSE41-NEXT:    addsd %xmm0, %xmm2
+; SSE41-NEXT:    xorps %xmm0, %xmm0
+; SSE41-NEXT:    roundsd $11, %xmm2, %xmm0
+; SSE41-NEXT:    retq
+;
+; AVX-LABEL: round_f64:
+; AVX:       ## %bb.0:
+; AVX-NEXT:    vandpd {{.*}}(%rip), %xmm0, %xmm1
+; AVX-NEXT:    vmovddup {{.*#+}} xmm2 = [4.9999999999999994E-1,4.9999999999999994E-1]
+; AVX-NEXT:    ## xmm2 = mem[0,0]
+; AVX-NEXT:    vorpd %xmm1, %xmm2, %xmm1
+; AVX-NEXT:    vaddsd %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    vroundsd $11, %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    retq
+  %a = call double @llvm.round.f64(double %x)
+  ret double %a
+}
+
+define <4 x float> @round_v4f32(<4 x float> %x) {
+; SSE2-LABEL: round_v4f32:
+; SSE2:       ## %bb.0:
+; SSE2-NEXT:    subq $56, %rsp
+; SSE2-NEXT:    .cfi_def_cfa_offset 64
+; SSE2-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,1,2,3]
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    unpcklps (%rsp), %xmm0 ## 16-byte Folded Reload
+; SSE2-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; SSE2-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,1,2,3]
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
+; SSE2-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
+; SSE2-NEXT:    unpcklpd (%rsp), %xmm1 ## 16-byte Folded Reload
+; SSE2-NEXT:    ## xmm1 = xmm1[0],mem[0]
+; SSE2-NEXT:    movaps %xmm1, %xmm0
+; SSE2-NEXT:    addq $56, %rsp
+; SSE2-NEXT:    retq
+;
+; SSE41-LABEL: round_v4f32:
+; SSE41:       ## %bb.0:
+; SSE41-NEXT:    movaps {{.*#+}} xmm1 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; SSE41-NEXT:    andps %xmm0, %xmm1
+; SSE41-NEXT:    orps {{.*}}(%rip), %xmm1
+; SSE41-NEXT:    addps %xmm0, %xmm1
+; SSE41-NEXT:    roundps $11, %xmm1, %xmm0
+; SSE41-NEXT:    retq
+;
+; AVX1-LABEL: round_v4f32:
+; AVX1:       ## %bb.0:
+; AVX1-NEXT:    vandps {{.*}}(%rip), %xmm0, %xmm1
+; AVX1-NEXT:    vorps {{.*}}(%rip), %xmm1, %xmm1
+; AVX1-NEXT:    vaddps %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vroundps $11, %xmm0, %xmm0
+; AVX1-NEXT:    retq
+;
+; AVX512-LABEL: round_v4f32:
+; AVX512:       ## %bb.0:
+; AVX512-NEXT:    vbroadcastss {{.*#+}} xmm1 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; AVX512-NEXT:    vandps %xmm1, %xmm0, %xmm1
+; AVX512-NEXT:    vbroadcastss {{.*#+}} xmm2 = [4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1]
+; AVX512-NEXT:    vorps %xmm1, %xmm2, %xmm1
+; AVX512-NEXT:    vaddps %xmm1, %xmm0, %xmm0
+; AVX512-NEXT:    vroundps $11, %xmm0, %xmm0
+; AVX512-NEXT:    retq
+  %a = call <4 x float> @llvm.round.v4f32(<4 x float> %x)
+  ret <4 x float> %a
+}
+
+define <2 x double> @round_v2f64(<2 x double> %x) {
+; SSE2-LABEL: round_v2f64:
+; SSE2:       ## %bb.0:
+; SSE2-NEXT:    subq $40, %rsp
+; SSE2-NEXT:    .cfi_def_cfa_offset 48
+; SSE2-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
+; SSE2-NEXT:    callq _round
+; SSE2-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps (%rsp), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
+; SSE2-NEXT:    callq _round
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
+; SSE2-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
+; SSE2-NEXT:    movaps %xmm1, %xmm0
+; SSE2-NEXT:    addq $40, %rsp
+; SSE2-NEXT:    retq
+;
+; SSE41-LABEL: round_v2f64:
+; SSE41:       ## %bb.0:
+; SSE41-NEXT:    movapd {{.*#+}} xmm1 = [-0.0E+0,-0.0E+0]
+; SSE41-NEXT:    andpd %xmm0, %xmm1
+; SSE41-NEXT:    orpd {{.*}}(%rip), %xmm1
+; SSE41-NEXT:    addpd %xmm0, %xmm1
+; SSE41-NEXT:    roundpd $11, %xmm1, %xmm0
+; SSE41-NEXT:    retq
+;
+; AVX-LABEL: round_v2f64:
+; AVX:       ## %bb.0:
+; AVX-NEXT:    vandpd {{.*}}(%rip), %xmm0, %xmm1
+; AVX-NEXT:    vorpd {{.*}}(%rip), %xmm1, %xmm1
+; AVX-NEXT:    vaddpd %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    vroundpd $11, %xmm0, %xmm0
+; AVX-NEXT:    retq
+  %a = call <2 x double> @llvm.round.v2f64(<2 x double> %x)
+  ret <2 x double> %a
+}
+
+define <8 x float> @round_v8f32(<8 x float> %x) {
+; SSE2-LABEL: round_v8f32:
+; SSE2:       ## %bb.0:
+; SSE2-NEXT:    subq $72, %rsp
+; SSE2-NEXT:    .cfi_def_cfa_offset 80
+; SSE2-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,1,2,3]
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps (%rsp), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Folded Reload
+; SSE2-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; SSE2-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps (%rsp), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps (%rsp), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,1,2,3]
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
+; SSE2-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
+; SSE2-NEXT:    unpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Folded Reload
+; SSE2-NEXT:    ## xmm1 = xmm1[0],mem[0]
+; SSE2-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,1,2,3]
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    unpcklps (%rsp), %xmm0 ## 16-byte Folded Reload
+; SSE2-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; SSE2-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,1,2,3]
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
+; SSE2-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
+; SSE2-NEXT:    unpcklpd (%rsp), %xmm1 ## 16-byte Folded Reload
+; SSE2-NEXT:    ## xmm1 = xmm1[0],mem[0]
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    addq $72, %rsp
+; SSE2-NEXT:    retq
+;
+; SSE41-LABEL: round_v8f32:
+; SSE41:       ## %bb.0:
+; SSE41-NEXT:    movaps {{.*#+}} xmm2 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; SSE41-NEXT:    movaps %xmm0, %xmm3
+; SSE41-NEXT:    andps %xmm2, %xmm3
+; SSE41-NEXT:    movaps {{.*#+}} xmm4 = [4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1]
+; SSE41-NEXT:    orps %xmm4, %xmm3
+; SSE41-NEXT:    addps %xmm0, %xmm3
+; SSE41-NEXT:    roundps $11, %xmm3, %xmm0
+; SSE41-NEXT:    andps %xmm1, %xmm2
+; SSE41-NEXT:    orps %xmm4, %xmm2
+; SSE41-NEXT:    addps %xmm1, %xmm2
+; SSE41-NEXT:    roundps $11, %xmm2, %xmm1
+; SSE41-NEXT:    retq
+;
+; AVX1-LABEL: round_v8f32:
+; AVX1:       ## %bb.0:
+; AVX1-NEXT:    vandps {{.*}}(%rip), %ymm0, %ymm1
+; AVX1-NEXT:    vorps {{.*}}(%rip), %ymm1, %ymm1
+; AVX1-NEXT:    vaddps %ymm1, %ymm0, %ymm0
+; AVX1-NEXT:    vroundps $11, %ymm0, %ymm0
+; AVX1-NEXT:    retq
+;
+; AVX512-LABEL: round_v8f32:
+; AVX512:       ## %bb.0:
+; AVX512-NEXT:    vbroadcastss {{.*#+}} ymm1 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; AVX512-NEXT:    vandps %ymm1, %ymm0, %ymm1
+; AVX512-NEXT:    vbroadcastss {{.*#+}} ymm2 = [4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1]
+; AVX512-NEXT:    vorps %ymm1, %ymm2, %ymm1
+; AVX512-NEXT:    vaddps %ymm1, %ymm0, %ymm0
+; AVX512-NEXT:    vroundps $11, %ymm0, %ymm0
+; AVX512-NEXT:    retq
+  %a = call <8 x float> @llvm.round.v8f32(<8 x float> %x)
+  ret <8 x float> %a
+}
+
+define <4 x double> @round_v4f64(<4 x double> %x) {
+; SSE2-LABEL: round_v4f64:
+; SSE2:       ## %bb.0:
+; SSE2-NEXT:    subq $56, %rsp
+; SSE2-NEXT:    .cfi_def_cfa_offset 64
+; SSE2-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
+; SSE2-NEXT:    callq _round
+; SSE2-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps (%rsp), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
+; SSE2-NEXT:    callq _round
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
+; SSE2-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
+; SSE2-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    callq _round
+; SSE2-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
+; SSE2-NEXT:    callq _round
+; SSE2-NEXT:    movaps (%rsp), %xmm1 ## 16-byte Reload
+; SSE2-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    addq $56, %rsp
+; SSE2-NEXT:    retq
+;
+; SSE41-LABEL: round_v4f64:
+; SSE41:       ## %bb.0:
+; SSE41-NEXT:    movapd {{.*#+}} xmm2 = [-0.0E+0,-0.0E+0]
+; SSE41-NEXT:    movapd %xmm0, %xmm3
+; SSE41-NEXT:    andpd %xmm2, %xmm3
+; SSE41-NEXT:    movapd {{.*#+}} xmm4 = [4.9999999999999994E-1,4.9999999999999994E-1]
+; SSE41-NEXT:    orpd %xmm4, %xmm3
+; SSE41-NEXT:    addpd %xmm0, %xmm3
+; SSE41-NEXT:    roundpd $11, %xmm3, %xmm0
+; SSE41-NEXT:    andpd %xmm1, %xmm2
+; SSE41-NEXT:    orpd %xmm4, %xmm2
+; SSE41-NEXT:    addpd %xmm1, %xmm2
+; SSE41-NEXT:    roundpd $11, %xmm2, %xmm1
+; SSE41-NEXT:    retq
+;
+; AVX1-LABEL: round_v4f64:
+; AVX1:       ## %bb.0:
+; AVX1-NEXT:    vandpd {{.*}}(%rip), %ymm0, %ymm1
+; AVX1-NEXT:    vorpd {{.*}}(%rip), %ymm1, %ymm1
+; AVX1-NEXT:    vaddpd %ymm1, %ymm0, %ymm0
+; AVX1-NEXT:    vroundpd $11, %ymm0, %ymm0
+; AVX1-NEXT:    retq
+;
+; AVX512-LABEL: round_v4f64:
+; AVX512:       ## %bb.0:
+; AVX512-NEXT:    vbroadcastsd {{.*#+}} ymm1 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; AVX512-NEXT:    vandpd %ymm1, %ymm0, %ymm1
+; AVX512-NEXT:    vbroadcastsd {{.*#+}} ymm2 = [4.9999999999999994E-1,4.9999999999999994E-1,4.9999999999999994E-1,4.9999999999999994E-1]
+; AVX512-NEXT:    vorpd %ymm1, %ymm2, %ymm1
+; AVX512-NEXT:    vaddpd %ymm1, %ymm0, %ymm0
+; AVX512-NEXT:    vroundpd $11, %ymm0, %ymm0
+; AVX512-NEXT:    retq
+  %a = call <4 x double> @llvm.round.v4f64(<4 x double> %x)
+  ret <4 x double> %a
+}
+
+define <16 x float> @round_v16f32(<16 x float> %x) {
+; SSE2-LABEL: round_v16f32:
+; SSE2:       ## %bb.0:
+; SSE2-NEXT:    subq $104, %rsp
+; SSE2-NEXT:    .cfi_def_cfa_offset 112
+; SSE2-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps %xmm1, (%rsp) ## 16-byte Spill
+; SSE2-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,1,2,3]
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Folded Reload
+; SSE2-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; SSE2-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,1,2,3]
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
+; SSE2-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
+; SSE2-NEXT:    unpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Folded Reload
+; SSE2-NEXT:    ## xmm1 = xmm1[0],mem[0]
+; SSE2-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps (%rsp), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,1,2,3]
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps (%rsp), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Folded Reload
+; SSE2-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; SSE2-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps (%rsp), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps (%rsp), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,1,2,3]
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
+; SSE2-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
+; SSE2-NEXT:    unpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Folded Reload
+; SSE2-NEXT:    ## xmm1 = xmm1[0],mem[0]
+; SSE2-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,1,2,3]
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    unpcklps (%rsp), %xmm0 ## 16-byte Folded Reload
+; SSE2-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; SSE2-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,1,2,3]
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    movaps (%rsp), %xmm1 ## 16-byte Reload
+; SSE2-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
+; SSE2-NEXT:    unpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Folded Reload
+; SSE2-NEXT:    ## xmm1 = xmm1[0],mem[0]
+; SSE2-NEXT:    movaps %xmm1, (%rsp) ## 16-byte Spill
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,1,2,3]
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Folded Reload
+; SSE2-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; SSE2-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,1,2,3]
+; SSE2-NEXT:    callq _roundf
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 ## 16-byte Reload
+; SSE2-NEXT:    unpcklps {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]
+; SSE2-NEXT:    unpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 ## 16-byte Folded Reload
+; SSE2-NEXT:    ## xmm3 = xmm3[0],mem[0]
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
+; SSE2-NEXT:    movaps (%rsp), %xmm2 ## 16-byte Reload
+; SSE2-NEXT:    addq $104, %rsp
+; SSE2-NEXT:    retq
+;
+; SSE41-LABEL: round_v16f32:
+; SSE41:       ## %bb.0:
+; SSE41-NEXT:    movaps {{.*#+}} xmm4 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; SSE41-NEXT:    movaps %xmm0, %xmm5
+; SSE41-NEXT:    andps %xmm4, %xmm5
+; SSE41-NEXT:    movaps {{.*#+}} xmm6 = [4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1]
+; SSE41-NEXT:    orps %xmm6, %xmm5
+; SSE41-NEXT:    addps %xmm0, %xmm5
+; SSE41-NEXT:    roundps $11, %xmm5, %xmm0
+; SSE41-NEXT:    movaps %xmm1, %xmm5
+; SSE41-NEXT:    andps %xmm4, %xmm5
+; SSE41-NEXT:    orps %xmm6, %xmm5
+; SSE41-NEXT:    addps %xmm1, %xmm5
+; SSE41-NEXT:    roundps $11, %xmm5, %xmm1
+; SSE41-NEXT:    movaps %xmm2, %xmm5
+; SSE41-NEXT:    andps %xmm4, %xmm5
+; SSE41-NEXT:    orps %xmm6, %xmm5
+; SSE41-NEXT:    addps %xmm2, %xmm5
+; SSE41-NEXT:    roundps $11, %xmm5, %xmm2
+; SSE41-NEXT:    andps %xmm3, %xmm4
+; SSE41-NEXT:    orps %xmm6, %xmm4
+; SSE41-NEXT:    addps %xmm3, %xmm4
+; SSE41-NEXT:    roundps $11, %xmm4, %xmm3
+; SSE41-NEXT:    retq
+;
+; AVX1-LABEL: round_v16f32:
+; AVX1:       ## %bb.0:
+; AVX1-NEXT:    vmovaps {{.*#+}} ymm2 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; AVX1-NEXT:    vandps %ymm2, %ymm0, %ymm3
+; AVX1-NEXT:    vmovaps {{.*#+}} ymm4 = [4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1]
+; AVX1-NEXT:    vorps %ymm3, %ymm4, %ymm3
+; AVX1-NEXT:    vaddps %ymm3, %ymm0, %ymm0
+; AVX1-NEXT:    vroundps $11, %ymm0, %ymm0
+; AVX1-NEXT:    vandps %ymm2, %ymm1, %ymm2
+; AVX1-NEXT:    vorps %ymm2, %ymm4, %ymm2
+; AVX1-NEXT:    vaddps %ymm2, %ymm1, %ymm1
+; AVX1-NEXT:    vroundps $11, %ymm1, %ymm1
+; AVX1-NEXT:    retq
+;
+; AVX512-LABEL: round_v16f32:
+; AVX512:       ## %bb.0:
+; AVX512-NEXT:    vpandd {{.*}}(%rip){1to16}, %zmm0, %zmm1
+; AVX512-NEXT:    vpord {{.*}}(%rip){1to16}, %zmm1, %zmm1
+; AVX512-NEXT:    vaddps %zmm1, %zmm0, %zmm0
+; AVX512-NEXT:    vrndscaleps $11, %zmm0, %zmm0
+; AVX512-NEXT:    retq
+  %a = call <16 x float> @llvm.round.v16f32(<16 x float> %x)
+  ret <16 x float> %a
+}
+
+define <8 x double> @round_v8f64(<8 x double> %x) {
+; SSE2-LABEL: round_v8f64:
+; SSE2:       ## %bb.0:
+; SSE2-NEXT:    subq $88, %rsp
+; SSE2-NEXT:    .cfi_def_cfa_offset 96
+; SSE2-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps %xmm1, (%rsp) ## 16-byte Spill
+; SSE2-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    callq _round
+; SSE2-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
+; SSE2-NEXT:    callq _round
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
+; SSE2-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
+; SSE2-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps (%rsp), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    callq _round
+; SSE2-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps (%rsp), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
+; SSE2-NEXT:    callq _round
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
+; SSE2-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
+; SSE2-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    callq _round
+; SSE2-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
+; SSE2-NEXT:    callq _round
+; SSE2-NEXT:    movaps (%rsp), %xmm1 ## 16-byte Reload
+; SSE2-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
+; SSE2-NEXT:    movaps %xmm1, (%rsp) ## 16-byte Spill
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    callq _round
+; SSE2-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
+; SSE2-NEXT:    callq _round
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 ## 16-byte Reload
+; SSE2-NEXT:    movlhps {{.*#+}} xmm3 = xmm3[0],xmm0[0]
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
+; SSE2-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
+; SSE2-NEXT:    movaps (%rsp), %xmm2 ## 16-byte Reload
+; SSE2-NEXT:    addq $88, %rsp
+; SSE2-NEXT:    retq
+;
+; SSE41-LABEL: round_v8f64:
+; SSE41:       ## %bb.0:
+; SSE41-NEXT:    movapd {{.*#+}} xmm4 = [-0.0E+0,-0.0E+0]
+; SSE41-NEXT:    movapd %xmm0, %xmm5
+; SSE41-NEXT:    andpd %xmm4, %xmm5
+; SSE41-NEXT:    movapd {{.*#+}} xmm6 = [4.9999999999999994E-1,4.9999999999999994E-1]
+; SSE41-NEXT:    orpd %xmm6, %xmm5
+; SSE41-NEXT:    addpd %xmm0, %xmm5
+; SSE41-NEXT:    roundpd $11, %xmm5, %xmm0
+; SSE41-NEXT:    movapd %xmm1, %xmm5
+; SSE41-NEXT:    andpd %xmm4, %xmm5
+; SSE41-NEXT:    orpd %xmm6, %xmm5
+; SSE41-NEXT:    addpd %xmm1, %xmm5
+; SSE41-NEXT:    roundpd $11, %xmm5, %xmm1
+; SSE41-NEXT:    movapd %xmm2, %xmm5
+; SSE41-NEXT:    andpd %xmm4, %xmm5
+; SSE41-NEXT:    orpd %xmm6, %xmm5
+; SSE41-NEXT:    addpd %xmm2, %xmm5
+; SSE41-NEXT:    roundpd $11, %xmm5, %xmm2
+; SSE41-NEXT:    andpd %xmm3, %xmm4
+; SSE41-NEXT:    orpd %xmm6, %xmm4
+; SSE41-NEXT:    addpd %xmm3, %xmm4
+; SSE41-NEXT:    roundpd $11, %xmm4, %xmm3
+; SSE41-NEXT:    retq
+;
+; AVX1-LABEL: round_v8f64:
+; AVX1:       ## %bb.0:
+; AVX1-NEXT:    vmovapd {{.*#+}} ymm2 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; AVX1-NEXT:    vandpd %ymm2, %ymm0, %ymm3
+; AVX1-NEXT:    vmovapd {{.*#+}} ymm4 = [4.9999999999999994E-1,4.9999999999999994E-1,4.9999999999999994E-1,4.9999999999999994E-1]
+; AVX1-NEXT:    vorpd %ymm3, %ymm4, %ymm3
+; AVX1-NEXT:    vaddpd %ymm3, %ymm0, %ymm0
+; AVX1-NEXT:    vroundpd $11, %ymm0, %ymm0
+; AVX1-NEXT:    vandpd %ymm2, %ymm1, %ymm2
+; AVX1-NEXT:    vorpd %ymm2, %ymm4, %ymm2
+; AVX1-NEXT:    vaddpd %ymm2, %ymm1, %ymm1
+; AVX1-NEXT:    vroundpd $11, %ymm1, %ymm1
+; AVX1-NEXT:    retq
+;
+; AVX512-LABEL: round_v8f64:
+; AVX512:       ## %bb.0:
+; AVX512-NEXT:    vpandq {{.*}}(%rip){1to8}, %zmm0, %zmm1
+; AVX512-NEXT:    vporq {{.*}}(%rip){1to8}, %zmm1, %zmm1
+; AVX512-NEXT:    vaddpd %zmm1, %zmm0, %zmm0
+; AVX512-NEXT:    vrndscalepd $11, %zmm0, %zmm0
+; AVX512-NEXT:    retq
+  %a = call <8 x double> @llvm.round.v8f64(<8 x double> %x)
+  ret <8 x double> %a
+}
+
+declare float @llvm.round.f32(float)
+declare double @llvm.round.f64(double)
+declare <4 x float> @llvm.round.v4f32(<4 x float>)
+declare <2 x double> @llvm.round.v2f64(<2 x double>)
+declare <8 x float> @llvm.round.v8f32(<8 x float>)
+declare <4 x double> @llvm.round.v4f64(<4 x double>)
+declare <16 x float> @llvm.round.v16f32(<16 x float>)
+declare <8 x double> @llvm.round.v8f64(<8 x double>)

llvm/test/CodeGen/X86/vec-libcalls.ll

 ; CHECK-NEXT:    retq
   %r = call <2 x float> @llvm.rint.v2f32(<2 x float> %x)
   ret <2 x float> %r
 }
 
 define <2 x float> @round_v2f32(<2 x float> %x) nounwind {
 ; CHECK-LABEL: round_v2f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    subq $40, %rsp
-; CHECK-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
-; CHECK-NEXT:    callq roundf
-; CHECK-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; CHECK-NEXT:    vmovshdup (%rsp), %xmm0 # 16-byte Folded Reload
-; CHECK-NEXT:    # xmm0 = mem[1,1,3,3]
-; CHECK-NEXT:    callq roundf
-; CHECK-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; CHECK-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
-; CHECK-NEXT:    addq $40, %rsp
+; CHECK-NEXT:    vandps {{.*}}(%rip), %xmm0, %xmm1
+; CHECK-NEXT:    vorps {{.*}}(%rip), %xmm1, %xmm1
+; CHECK-NEXT:    vaddps %xmm1, %xmm0, %xmm0
+; CHECK-NEXT:    vroundps $11, %xmm0, %xmm0
 ; CHECK-NEXT:    retq
   %r = call <2 x float> @llvm.round.v2f32(<2 x float> %x)
   ret <2 x float> %r
 }
 
 define <2 x float> @sqrt_v2f32(<2 x float> %x) nounwind {
 ; CHECK-LABEL: sqrt_v2f32:
 ; CHECK:       # %bb.0:

llvm/test/CodeGen/X86/vec_round.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -mcpu=nehalem -mtriple=x86_64-unknown-linux-gnu < %s | FileCheck %s
 target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-linux-gnu"
 
 declare void @use(<2 x double>)
 
 ; Function Attrs: nounwind uwtable
 define void @test() {
 ; CHECK-LABEL: test:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    pushq %rax
 ; CHECK-NEXT:    .cfi_def_cfa_offset 16
-; CHECK-NEXT:    callq round
-; CHECK-NEXT:    movddup {{.*#+}} xmm0 = xmm0[0,0]
+; CHECK-NEXT:    movaps {{.*#+}} xmm0 = [NaN,NaN]
 ; CHECK-NEXT:    callq use
 ; CHECK-NEXT:    popq %rax
 ; CHECK-NEXT:    .cfi_def_cfa_offset 8
 ; CHECK-NEXT:    retq
 entry:
   %tmp = call <2 x double> @llvm.round.v2f64(<2 x double> undef)
   call void @use(<2 x double> %tmp)
   ret void
 }
 
 ; Function Attrs: nounwind readonly
 declare <2 x double> @llvm.round.v2f64(<2 x double>) #0
 
 attributes #0 = { nounwind readonly }