[DAGCombiner] restrict (float)((int) f) --> ftrunc with no-signed-zeros

[spatel]

Summary

As noted in the D44909 review, the transform from (fptosi+sitofp) to ftrunc can produce -0.0 where the original code does not:

#include <stdio.h>
  
int main(int argc) {
  float x;
  x = -0.8 * argc;
  printf("%f\n", (float)((int)x));
  return 0;
}

$ clang -O0 -mavx fp.c ; ./a.out 
0.000000
$ clang -O1 -mavx fp.c ; ./a.out 
-0.000000

Ideally, we'd use IR/node flags to predicate the transform, but the IR parser doesn't currently allow fast-math-flags on the cast instructions. So for now, just use the function attribute that corresponds to clang's "-fno-signed-zeros" option.

[lebedev.ri]

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

11049–11050 ↗

(On Diff #150979)

Couldn't we do and (ftrunc %x), (-1 >> 1), i.e. unset the sign bit?
Or is that worse than not transforming to ftrunc in the first place?

[spatel]

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

11049–11050 ↗

(On Diff #150979)

You'd want to keep this in FP, so I think that would be:
fabs (ftrunc x)

And that could be better than the alternative in some cases. See D44909 for what this looks like without using ftrunc. If the sequence is more than 2 converts, fabs is likely a winner.

But given the fallout from the original patch, I expect this is the patch most people would prefer. :)
We'll avoid the UB headaches for most code, and code that cares about FP perf likely has some loosened FP constraints anyway.

[lebedev.ri]

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

11049–11050 ↗

(On Diff #150979)

You'd want to keep this in FP, so I think that would be:
fabs (ftrunc x)

Right, right. Otherwise you might need to leave fp domain, etc etc.

But given the fallout from the original patch, I expect this is the patch most people would prefer. :)

That cat is already out of the bag though, and modulo this edge-case, the transform is valid i think.
Backtracking like this, while not for no reason, may be viewed as acceptance of views of some who
think UB is bad and lack of UB [in code] should not be assumed by compilers. But this is purely my opinion.

[spatel]

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

11049–11050 ↗

(On Diff #150979)

Depends on your perspective. For brave (much appreciated!) projects that track LLVM trunk, yes they've already suffered/benefited. But most potential customers won't see this change until the next major release.

But this discussion is independent of this particular patch. We have a miscompile that must be fixed. Salvaging perf is secondary, and I don't think we can do it universally - ie, converting 2 casts to fabs+ftrunc or select+fcmp+fabs+ftrunc requires a TLI hook, so that's definitely a bigger patch than what we have here currently.

[spatel]

Ping.

[lebedev.ri]

Should this adjust the ReleaseNotes?

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

11049–11050 ↗

(On Diff #150979)

Could you please at least add a FIXME? :)

[spatel]

In D48085#1136860, @lebedev.ri wrote:

Should this adjust the ReleaseNotes?

The docs provide extra warning/workaround, and this change doesn't affect that language IMO (there's less chance we're going to break code, but we don't have to disclose that).

For reference, here are links to the docs that we added with the previous patches (last updated with D46236 I think):
https://clang.llvm.org/docs/ReleaseNotes.html#new-compiler-flags
https://clang.llvm.org/docs/UsersManual.html#controlling-code-generation
http://llvm.org/docs/ReleaseNotes.html#non-comprehensive-list-of-changes-in-this-release

Let me know if you see anything that can be improved.

[spatel]

Patch updated:
Added a 'TODO' comment about using a FABS-based sequence if we can't ignore -0.0.

[lebedev.ri]

In D48085#1136900, @spatel wrote:

In D48085#1136860, @lebedev.ri wrote:

Should this adjust the ReleaseNotes?

The docs provide extra warning/workaround, and this change doesn't affect that language IMO (there's less chance we're going to break code, but we don't have to disclose that).

For reference, here are links to the docs that we added with the previous patches (last updated with D46236 I think):
https://clang.llvm.org/docs/ReleaseNotes.html#new-compiler-flags
https://clang.llvm.org/docs/UsersManual.html#controlling-code-generation
http://llvm.org/docs/ReleaseNotes.html#non-comprehensive-list-of-changes-in-this-release

Let me know if you see anything that can be improved.

The text in first and last links suggests that such an optimization is always being done.
But now it will only be done in presence of no-signed-zeros-fp-math attribute.
Which is controlled by -ffast-math (or maybe some more fine-grained option, too?)

[spatel]

In D48085#1136910, @lebedev.ri wrote:

In D48085#1136900, @spatel wrote:

In D48085#1136860, @lebedev.ri wrote:

Should this adjust the ReleaseNotes?

The docs provide extra warning/workaround, and this change doesn't affect that language IMO (there's less chance we're going to break code, but we don't have to disclose that).

For reference, here are links to the docs that we added with the previous patches (last updated with D46236 I think):
https://clang.llvm.org/docs/ReleaseNotes.html#new-compiler-flags
https://clang.llvm.org/docs/UsersManual.html#controlling-code-generation
http://llvm.org/docs/ReleaseNotes.html#non-comprehensive-list-of-changes-in-this-release

Let me know if you see anything that can be improved.

The text in first and last links suggests that such an optimization is always being done.
But now it will only be done in presence of no-signed-zeros-fp-math attribute.
Which is controlled by -ffast-math (or maybe some more fine-grained option, too?)

Right - it should still be allowed with "-fno-signed-zeros" at a minimum.

If we change the docs to provide the exact details of the potential optimization, then we have to adjust the docs anytime that changes (eg, if we implement the fabs trick for some subset of data types on some subset targets). So my preference is to not document the implementation at that level, but if you think that the text is misleading, I'll add a note.

[spatel]

In D48085#1136910, @lebedev.ri wrote:

In D48085#1136900, @spatel wrote:

In D48085#1136860, @lebedev.ri wrote:

Should this adjust the ReleaseNotes?

The docs provide extra warning/workaround, and this change doesn't affect that language IMO (there's less chance we're going to break code, but we don't have to disclose that).

For reference, here are links to the docs that we added with the previous patches (last updated with D46236 I think):
https://clang.llvm.org/docs/ReleaseNotes.html#new-compiler-flags
https://clang.llvm.org/docs/UsersManual.html#controlling-code-generation
http://llvm.org/docs/ReleaseNotes.html#non-comprehensive-list-of-changes-in-this-release

Let me know if you see anything that can be improved.

The text in first and last links suggests that such an optimization is always being done.

One more point here: if it sounds like it's always done, then that's already wrong.

We only do the transformation when the target has a legal FTRUNC op, so some fraction of targets are never affected by this transform regardless of fast-math. Eg, x86 doesn't have roundXX before SSE4.1.

[spatel]

Ping * 2.

[lebedev.ri]

I still think the docs need some adjustment.
And not for all arches the there are full set of tests (i.e. one with "no-signed-zeros-fp-math"="true", one without).
But this is a regression, so i'd guess that can be iterated later on.

It would be best if someone else could agree/disagree, but other than that, LG.

[spatel]

In D48085#1143602, @lebedev.ri wrote:

I still think the docs need some adjustment.
And not for all arches the there are full set of tests (i.e. one with "no-signed-zeros-fp-math"="true", one without).
But this is a regression, so i'd guess that can be iterated later on.

It would be best if someone else could agree/disagree, but other than that, LG.

Thanks! I'm not opposed to rewording the docs, but I'm also not sure how to adjust them to make it better rather than worse.

If anyone would like to take a shot at that, I'm happy to review (and as I said, I think any text tweaking should be independent of this code change...though others may disagree with that assessment).

llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

static SDValue foldFPToIntToFP(SDNode *N, SelectionDAG &DAG,
   // unexpected results. Ie, programs may be relying on the platform-specific
   // undefined behavior when the float-to-int conversion overflows.
   const Function &F = DAG.getMachineFunction().getFunction();
   Attribute StrictOverflow = F.getFnAttribute("strict-float-cast-overflow");
   if (StrictOverflow.getValueAsString().equals("false"))
     return SDValue();
 
   // We only do this if the target has legal ftrunc. Otherwise, we'd likely be
-  // replacing casts with a libcall.
+  // replacing casts with a libcall. We also must be allowed to ignore -0.0
+  // because FTRUNC will return -0.0 for (-1.0, -0.0), but using integer
+  // conversions would return +0.0.
+  // FIXME: We should be able to use node-level FMF here.
+  // TODO: If strict math, should we use FABS (+ range check for signed cast)?
   EVT VT = N->getValueType(0);
-  if (!TLI.isOperationLegal(ISD::FTRUNC, VT))
+  if (!TLI.isOperationLegal(ISD::FTRUNC, VT) ||
+      !DAG.getTarget().Options.NoSignedZerosFPMath)
     return SDValue();
 
   // fptosi/fptoui round towards zero, so converting from FP to integer and
   // back is the same as an 'ftrunc': [us]itofp (fpto[us]i X) --> ftrunc X
   SDValue N0 = N->getOperand(0);
   if (N->getOpcode() == ISD::SINT_TO_FP && N0.getOpcode() == ISD::FP_TO_SINT &&
       N0.getOperand(0).getValueType() == VT)
     return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0.getOperand(0));

llvm/trunk/test/CodeGen/AArch64/ftrunc.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -mtriple=aarch64-unknown-unknown < %s | FileCheck %s
 
-define float @trunc_unsigned_f32(float %x) {
+define float @trunc_unsigned_f32(float %x) #0 {
 ; CHECK-LABEL: trunc_unsigned_f32:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    frintz s0, s0
 ; CHECK-NEXT:    ret
   %i = fptoui float %x to i32
   %r = uitofp i32 %i to float
   ret float %r
 }
 
-define double @trunc_unsigned_f64(double %x) {
+define double @trunc_unsigned_f64(double %x) #0 {
 ; CHECK-LABEL: trunc_unsigned_f64:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    frintz d0, d0
 ; CHECK-NEXT:    ret
   %i = fptoui double %x to i64
   %r = uitofp i64 %i to double
   ret double %r
 }
 
-define float @trunc_signed_f32(float %x) {
+define float @trunc_signed_f32(float %x) #0 {
 ; CHECK-LABEL: trunc_signed_f32:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    frintz s0, s0
 ; CHECK-NEXT:    ret
   %i = fptosi float %x to i32
   %r = sitofp i32 %i to float
   ret float %r
 }
 
-define double @trunc_signed_f64(double %x) {
+define double @trunc_signed_f64(double %x) #0 {
 ; CHECK-LABEL: trunc_signed_f64:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    frintz d0, d0
 ; CHECK-NEXT:    ret
   %i = fptosi double %x to i64
   %r = sitofp i64 %i to double
   ret double %r
 }
 
+attributes #0 = { "no-signed-zeros-fp-math"="true" }
+

llvm/trunk/test/CodeGen/PowerPC/fp-int128-fp-combine.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -O0 -mtriple=powerpc64le-unknown-linux-gnu < %s | FileCheck %s
 
 ; xscvdpsxds should NOT be emitted, since it saturates the result down to i64.
+; We can't use friz here because it may return -0.0 where the original code doesn't.
+
 define float @f_i128_f(float %v) {
 ; CHECK-LABEL: f_i128_f:
 ; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    mflr 0
+; CHECK-NEXT:    std 0, 16(1)
+; CHECK-NEXT:    stdu 1, -32(1)
+; CHECK-NEXT:    .cfi_def_cfa_offset 32
+; CHECK-NEXT:    .cfi_offset lr, 16
+; CHECK-NEXT:    bl __fixsfti
+; CHECK-NEXT:    nop
+; CHECK-NEXT:    bl __floattisf
+; CHECK-NEXT:    nop
+; CHECK-NEXT:    addi 1, 1, 32
+; CHECK-NEXT:    ld 0, 16(1)
+; CHECK-NEXT:    mtlr 0
+; CHECK-NEXT:    blr
+entry:
+  %a = fptosi float %v to i128
+  %b = sitofp i128 %a to float
+  ret float %b
+}
+
+; NSZ, so it's safe to friz.
+
+define float @f_i128_fi_nsz(float %v) #0 {
+; CHECK-LABEL: f_i128_fi_nsz:
+; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    friz 1, 1
 ; CHECK-NEXT:    blr
 entry:
   %a = fptosi float %v to i128
   %b = sitofp i128 %a to float
   ret float %b
 }
+
+attributes #0 = { "no-signed-zeros-fp-math"="true" }
+

llvm/trunk/test/CodeGen/PowerPC/fp-to-int-to-fp.ll

entry:
   %conv1 = uitofp i64 %conv to double
   ret double %conv1
 
 ; FPCVT-LABEL: @fooudl
 ; FPCVT: friz 1, 1
 ; FPCVT: blr
 }
 
-attributes #0 = { nounwind readnone }
+attributes #0 = { nounwind readnone "no-signed-zeros-fp-math"="true" }
 

llvm/trunk/test/CodeGen/PowerPC/ftrunc-vec.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -mcpu=pwr8 -mtriple=powerpc64le-unknown-unknown -verify-machineinstrs < %s | FileCheck %s
 
-define <4 x float> @truncf32(<4 x float> %a) {
+define <4 x float> @truncf32(<4 x float> %a) #0 {
 ; CHECK-LABEL: truncf32:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    xvrspiz 34, 34
 ; CHECK-NEXT:    blr
   %t0 = fptosi <4 x float> %a to <4 x i32>
   %t1 = sitofp <4 x i32> %t0 to <4 x float>
   ret <4 x float> %t1
 }
 
-define <2 x double> @truncf64(<2 x double> %a) {
+define <2 x double> @truncf64(<2 x double> %a) #0 {
 ; CHECK-LABEL: truncf64:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    xvrdpiz 34, 34
 ; CHECK-NEXT:    blr
   %t0 = fptosi <2 x double> %a to <2 x i64>
   %t1 = sitofp <2 x i64> %t0 to <2 x double>
   ret <2 x double> %t1
 }
 
-define <4 x float> @truncf32u(<4 x float> %a) {
+define <4 x float> @truncf32u(<4 x float> %a) #0 {
 ; CHECK-LABEL: truncf32u:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    xvrspiz 34, 34
 ; CHECK-NEXT:    blr
   %t0 = fptoui <4 x float> %a to <4 x i32>
   %t1 = uitofp <4 x i32> %t0 to <4 x float>
   ret <4 x float> %t1
 }
 
-define <2 x double> @truncf64u(<2 x double> %a) {
+define <2 x double> @truncf64u(<2 x double> %a) #0 {
 ; CHECK-LABEL: truncf64u:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    xvrdpiz 34, 34
 ; CHECK-NEXT:    blr
   %t0 = fptoui <2 x double> %a to <2 x i64>
   %t1 = uitofp <2 x i64> %t0 to <2 x double>
   ret <2 x double> %t1
 }
 
+attributes #0 = { "no-signed-zeros-fp-math"="true" }
+

llvm/trunk/test/CodeGen/PowerPC/no-extra-fp-conv-ldst.ll

entry:
   %conv1 = uitofp i32 %conv to double
   ret double %conv1
 
 ; CHECK-LABEL: @fooud
 ; CHECK: friz 1, 1
 ; CHECK: blr
 }
 
-attributes #0 = { nounwind readonly }
+attributes #0 = { nounwind readonly "no-signed-zeros-fp-math"="true" }
 

llvm/trunk/test/CodeGen/X86/2011-10-19-widen_vselect.ll

 }
 
 define void @full_test() {
 ; X32-LABEL: full_test:
 ; X32:       # %bb.0: # %entry
 ; X32-NEXT:    subl $60, %esp
 ; X32-NEXT:    .cfi_def_cfa_offset 64
 ; X32-NEXT:    movsd {{.*#+}} xmm2 = mem[0],zero
-; X32-NEXT:    roundps $11, %xmm2, %xmm1
+; X32-NEXT:    cvttps2dq %xmm2, %xmm0
+; X32-NEXT:    cvtdq2ps %xmm0, %xmm1
 ; X32-NEXT:    xorps %xmm0, %xmm0
 ; X32-NEXT:    cmpltps %xmm2, %xmm0
 ; X32-NEXT:    movaps {{.*#+}} xmm3 = <1,1,u,u>
 ; X32-NEXT:    addps %xmm1, %xmm3
 ; X32-NEXT:    movaps %xmm1, %xmm4
 ; X32-NEXT:    blendvps %xmm0, %xmm3, %xmm4
 ; X32-NEXT:    cmpeqps %xmm2, %xmm1
 ; X32-NEXT:    movaps %xmm1, %xmm0
 ; X32-NEXT:    blendvps %xmm0, %xmm2, %xmm4
 ; X32-NEXT:    movss %xmm4, {{[0-9]+}}(%esp)
 ; X32-NEXT:    movshdup {{.*#+}} xmm0 = xmm4[1,1,3,3]
 ; X32-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
 ; X32-NEXT:    movss %xmm4, {{[0-9]+}}(%esp)
 ; X32-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
 ; X32-NEXT:    addl $60, %esp
 ; X32-NEXT:    .cfi_def_cfa_offset 4
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: full_test:
 ; X64:       # %bb.0: # %entry
 ; X64-NEXT:    movsd {{.*#+}} xmm2 = mem[0],zero
-; X64-NEXT:    roundps $11, %xmm2, %xmm1
+; X64-NEXT:    cvttps2dq %xmm2, %xmm0
+; X64-NEXT:    cvtdq2ps %xmm0, %xmm1
 ; X64-NEXT:    xorps %xmm0, %xmm0
 ; X64-NEXT:    cmpltps %xmm2, %xmm0
 ; X64-NEXT:    movaps {{.*#+}} xmm3 = <1,1,u,u>
 ; X64-NEXT:    addps %xmm1, %xmm3
 ; X64-NEXT:    movaps %xmm1, %xmm4
 ; X64-NEXT:    blendvps %xmm0, %xmm3, %xmm4
 ; X64-NEXT:    cmpeqps %xmm2, %xmm1
 ; X64-NEXT:    movaps %xmm1, %xmm0

llvm/trunk/test/CodeGen/X86/avx-cvttp2si.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-- -mattr=avx              | FileCheck %s --check-prefixes=AVX,AVX1
 ; RUN: llc < %s -mtriple=x86_64-- -mattr=avx512f,avx512vl | FileCheck %s --check-prefixes=AVX,AVX512
 
 ; PR37751 - https://bugs.llvm.org/show_bug.cgi?id=37751
 ; We can't combine into 'round' instructions because the behavior is different for out-of-range values.
 
 declare <8 x i32> @llvm.x86.avx.cvtt.ps2dq.256(<8 x float>)
 declare <4 x i32> @llvm.x86.avx.cvtt.pd2dq.256(<4 x double>)
 
-define <8 x float> @float_to_int_to_float_mem_v8f32(<8 x float>* %p) {
+define <8 x float> @float_to_int_to_float_mem_v8f32(<8 x float>* %p) #0 {
 ; AVX-LABEL: float_to_int_to_float_mem_v8f32:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vcvttps2dq (%rdi), %ymm0
 ; AVX-NEXT:    vcvtdq2ps %ymm0, %ymm0
 ; AVX-NEXT:    retq
   %x = load <8 x float>, <8 x float>* %p, align 16
   %fptosi = tail call <8 x i32> @llvm.x86.avx.cvtt.ps2dq.256(<8 x float> %x)
   %sitofp = sitofp <8 x i32> %fptosi to <8 x float>
   ret <8 x float> %sitofp
 }
 
-define <8 x float> @float_to_int_to_float_reg_v8f32(<8 x float> %x) {
+define <8 x float> @float_to_int_to_float_reg_v8f32(<8 x float> %x) #0 {
 ; AVX-LABEL: float_to_int_to_float_reg_v8f32:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vcvttps2dq %ymm0, %ymm0
 ; AVX-NEXT:    vcvtdq2ps %ymm0, %ymm0
 ; AVX-NEXT:    retq
   %fptosi = tail call <8 x i32> @llvm.x86.avx.cvtt.ps2dq.256(<8 x float> %x)
   %sitofp = sitofp <8 x i32> %fptosi to <8 x float>
   ret <8 x float> %sitofp
 }
 
-define <4 x double> @float_to_int_to_float_mem_v4f64(<4 x double>* %p) {
+define <4 x double> @float_to_int_to_float_mem_v4f64(<4 x double>* %p) #0 {
 ; AVX-LABEL: float_to_int_to_float_mem_v4f64:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vcvttpd2dqy (%rdi), %xmm0
 ; AVX-NEXT:    vcvtdq2pd %xmm0, %ymm0
 ; AVX-NEXT:    retq
   %x = load <4 x double>, <4 x double>* %p, align 16
   %fptosi = tail call <4 x i32> @llvm.x86.avx.cvtt.pd2dq.256(<4 x double> %x)
   %sitofp = sitofp <4 x i32> %fptosi to <4 x double>
   ret <4 x double> %sitofp
 }
 
-define <4 x double> @float_to_int_to_float_reg_v4f64(<4 x double> %x) {
+define <4 x double> @float_to_int_to_float_reg_v4f64(<4 x double> %x) #0 {
 ; AVX-LABEL: float_to_int_to_float_reg_v4f64:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vcvttpd2dq %ymm0, %xmm0
 ; AVX-NEXT:    vcvtdq2pd %xmm0, %ymm0
 ; AVX-NEXT:    retq
   %fptosi = tail call <4 x i32> @llvm.x86.avx.cvtt.pd2dq.256(<4 x double> %x)
   %sitofp = sitofp <4 x i32> %fptosi to <4 x double>
   ret <4 x double> %sitofp
 }
 
+attributes #0 = { "no-signed-zeros-fp-math"="true" }
+

llvm/trunk/test/CodeGen/X86/ftrunc.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2    | FileCheck %s --check-prefix=SSE2
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.1  | FileCheck %s --check-prefix=SSE41
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx     | FileCheck %s --check-prefix=AVX1
 
-define float @trunc_unsigned_f32(float %x) nounwind {
+define float @trunc_unsigned_f32(float %x) #0 {
 ; SSE2-LABEL: trunc_unsigned_f32:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    cvttss2si %xmm0, %rax
 ; SSE2-NEXT:    movl %eax, %eax
 ; SSE2-NEXT:    xorps %xmm0, %xmm0
 ; SSE2-NEXT:    cvtsi2ssq %rax, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE41-LABEL: trunc_unsigned_f32:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    roundss $11, %xmm0, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: trunc_unsigned_f32:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vroundss $11, %xmm0, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
   %i = fptoui float %x to i32
   %r = uitofp i32 %i to float
   ret float %r
 }
 
-define double @trunc_unsigned_f64(double %x) nounwind {
+define double @trunc_unsigned_f64(double %x) #0 {
 ; SSE2-LABEL: trunc_unsigned_f64:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero
 ; SSE2-NEXT:    movapd %xmm0, %xmm2
 ; SSE2-NEXT:    subsd %xmm1, %xmm2
 ; SSE2-NEXT:    cvttsd2si %xmm2, %rax
 ; SSE2-NEXT:    movabsq $-9223372036854775808, %rcx # imm = 0x8000000000000000
 ; SSE2-NEXT:    xorq %rax, %rcx

 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vroundsd $11, %xmm0, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
   %i = fptoui double %x to i64
   %r = uitofp i64 %i to double
   ret double %r
 }
 
-define <4 x float> @trunc_unsigned_v4f32(<4 x float> %x) nounwind {
+define <4 x float> @trunc_unsigned_v4f32(<4 x float> %x) #0 {
 ; SSE2-LABEL: trunc_unsigned_v4f32:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movaps %xmm0, %xmm1
 ; SSE2-NEXT:    shufps {{.*#+}} xmm1 = xmm1[3,1],xmm0[2,3]
 ; SSE2-NEXT:    cvttss2si %xmm1, %rax
 ; SSE2-NEXT:    movd %eax, %xmm1
 ; SSE2-NEXT:    movaps %xmm0, %xmm2
 ; SSE2-NEXT:    movhlps {{.*#+}} xmm2 = xmm0[1],xmm2[1]

 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vroundps $11, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
   %i = fptoui <4 x float> %x to <4 x i32>
   %r = uitofp <4 x i32> %i to <4 x float>
   ret <4 x float> %r
 }
 
-define <2 x double> @trunc_unsigned_v2f64(<2 x double> %x) nounwind {
+define <2 x double> @trunc_unsigned_v2f64(<2 x double> %x) #0 {
 ; SSE2-LABEL: trunc_unsigned_v2f64:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movaps %xmm0, %xmm1
 ; SSE2-NEXT:    movhlps {{.*#+}} xmm1 = xmm0[1],xmm1[1]
 ; SSE2-NEXT:    movsd {{.*#+}} xmm2 = mem[0],zero
 ; SSE2-NEXT:    movaps %xmm1, %xmm3
 ; SSE2-NEXT:    subsd %xmm2, %xmm3
 ; SSE2-NEXT:    cvttsd2si %xmm3, %rax

 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vroundpd $11, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
   %i = fptoui <2 x double> %x to <2 x i64>
   %r = uitofp <2 x i64> %i to <2 x double>
   ret <2 x double> %r
 }
 
-define <4 x double> @trunc_unsigned_v4f64(<4 x double> %x) nounwind {
+define <4 x double> @trunc_unsigned_v4f64(<4 x double> %x) #0 {
 ; SSE2-LABEL: trunc_unsigned_v4f64:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movaps %xmm1, %xmm3
 ; SSE2-NEXT:    movhlps {{.*#+}} xmm3 = xmm1[1],xmm3[1]
 ; SSE2-NEXT:    movsd {{.*#+}} xmm2 = mem[0],zero
 ; SSE2-NEXT:    movaps %xmm3, %xmm4
 ; SSE2-NEXT:    subsd %xmm2, %xmm4
 ; SSE2-NEXT:    cvttsd2si %xmm4, %rcx

 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vroundpd $11, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
   %i = fptoui <4 x double> %x to <4 x i64>
   %r = uitofp <4 x i64> %i to <4 x double>
   ret <4 x double> %r
 }
 
-define float @trunc_signed_f32(float %x) nounwind {
+define float @trunc_signed_f32(float %x) #0 {
 ; SSE2-LABEL: trunc_signed_f32:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    cvttss2si %xmm0, %eax
 ; SSE2-NEXT:    xorps %xmm0, %xmm0
 ; SSE2-NEXT:    cvtsi2ssl %eax, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE41-LABEL: trunc_signed_f32:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    roundss $11, %xmm0, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: trunc_signed_f32:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vroundss $11, %xmm0, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
   %i = fptosi float %x to i32
   %r = sitofp i32 %i to float
   ret float %r
 }
 
-define double @trunc_signed_f64(double %x) nounwind {
+define double @trunc_signed_f64(double %x) #0 {
 ; SSE2-LABEL: trunc_signed_f64:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    cvttsd2si %xmm0, %rax
 ; SSE2-NEXT:    xorps %xmm0, %xmm0
 ; SSE2-NEXT:    cvtsi2sdq %rax, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE41-LABEL: trunc_signed_f64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    roundsd $11, %xmm0, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: trunc_signed_f64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vroundsd $11, %xmm0, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
   %i = fptosi double %x to i64
   %r = sitofp i64 %i to double
   ret double %r
 }
 
-define <4 x float> @trunc_signed_v4f32(<4 x float> %x) nounwind {
+define <4 x float> @trunc_signed_v4f32(<4 x float> %x) #0 {
 ; SSE2-LABEL: trunc_signed_v4f32:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    cvttps2dq %xmm0, %xmm0
 ; SSE2-NEXT:    cvtdq2ps %xmm0, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE41-LABEL: trunc_signed_v4f32:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    roundps $11, %xmm0, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: trunc_signed_v4f32:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vroundps $11, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
   %i = fptosi <4 x float> %x to <4 x i32>
   %r = sitofp <4 x i32> %i to <4 x float>
   ret <4 x float> %r
 }
 
-define <2 x double> @trunc_signed_v2f64(<2 x double> %x) nounwind {
+define <2 x double> @trunc_signed_v2f64(<2 x double> %x) #0 {
 ; SSE2-LABEL: trunc_signed_v2f64:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    cvttsd2si %xmm0, %rax
 ; SSE2-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
 ; SSE2-NEXT:    cvttsd2si %xmm0, %rcx
 ; SSE2-NEXT:    xorps %xmm0, %xmm0
 ; SSE2-NEXT:    cvtsi2sdq %rax, %xmm0
 ; SSE2-NEXT:    cvtsi2sdq %rcx, %xmm1

 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vroundpd $11, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
   %i = fptosi <2 x double> %x to <2 x i64>
   %r = sitofp <2 x i64> %i to <2 x double>
   ret <2 x double> %r
 }
 
-define <4 x double> @trunc_signed_v4f64(<4 x double> %x) nounwind {
+define <4 x double> @trunc_signed_v4f64(<4 x double> %x) #0 {
 ; SSE2-LABEL: trunc_signed_v4f64:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    cvttsd2si %xmm1, %rax
 ; SSE2-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1]
 ; SSE2-NEXT:    cvttsd2si %xmm1, %rcx
 ; SSE2-NEXT:    cvttsd2si %xmm0, %rdx
 ; SSE2-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
 ; SSE2-NEXT:    cvttsd2si %xmm0, %rsi

 ; AVX1-NEXT:    vcvttsd2si %xmm0, %rax
 ; AVX1-NEXT:    vcvtsi2sdq %rax, %xmm1, %xmm0
 ; AVX1-NEXT:    retq
   %i = fptosi double %x to i64
   %r = sitofp i64 %i to double
   ret double %r
 }
 
-attributes #1 = { nounwind "strict-float-cast-overflow"="false" }
+attributes #0 = { nounwind "no-signed-zeros-fp-math"="true" }
+attributes #1 = { nounwind "no-signed-zeros-fp-math"="true" "strict-float-cast-overflow"="false" }
 

llvm/trunk/test/CodeGen/X86/sse-cvttp2si.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-- -mattr=sse4.1           | FileCheck %s --check-prefixes=SSE
 ; RUN: llc < %s -mtriple=x86_64-- -mattr=avx              | FileCheck %s --check-prefixes=AVX,AVX1
 ; RUN: llc < %s -mtriple=x86_64-- -mattr=avx512f,avx512vl | FileCheck %s --check-prefixes=AVX,AVX512
 
 ; PR37751 - https://bugs.llvm.org/show_bug.cgi?id=37751
 ; We can't combine into 'round' instructions because the behavior is different for out-of-range values.
 
 declare <4 x i32> @llvm.x86.sse2.cvttps2dq(<4 x float>)
 declare <4 x i32> @llvm.x86.sse2.cvttpd2dq(<2 x double>)
 declare i32 @llvm.x86.sse.cvttss2si(<4 x float>)
 declare i64 @llvm.x86.sse.cvttss2si64(<4 x float>)
 declare i32 @llvm.x86.sse2.cvttsd2si(<2 x double>)
 declare i64 @llvm.x86.sse2.cvttsd2si64(<2 x double>)
 
-define float @float_to_int_to_float_mem_f32_i32(<4 x float>* %p) {
+define float @float_to_int_to_float_mem_f32_i32(<4 x float>* %p) #0 {
 ; SSE-LABEL: float_to_int_to_float_mem_f32_i32:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    cvttss2si (%rdi), %eax
 ; SSE-NEXT:    cvtsi2ssl %eax, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: float_to_int_to_float_mem_f32_i32:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vcvttss2si (%rdi), %eax
 ; AVX-NEXT:    vcvtsi2ssl %eax, %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %x = load <4 x float>, <4 x float>* %p, align 16
   %fptosi = tail call i32 @llvm.x86.sse.cvttss2si(<4 x float> %x)
   %sitofp = sitofp i32 %fptosi to float
   ret float %sitofp
 }
 
-define float @float_to_int_to_float_reg_f32_i32(<4 x float> %x) {
+define float @float_to_int_to_float_reg_f32_i32(<4 x float> %x) #0 {
 ; SSE-LABEL: float_to_int_to_float_reg_f32_i32:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    cvttss2si %xmm0, %eax
 ; SSE-NEXT:    xorps %xmm0, %xmm0
 ; SSE-NEXT:    cvtsi2ssl %eax, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: float_to_int_to_float_reg_f32_i32:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vcvttss2si %xmm0, %eax
 ; AVX-NEXT:    vcvtsi2ssl %eax, %xmm1, %xmm0
 ; AVX-NEXT:    retq
   %fptosi = tail call i32 @llvm.x86.sse.cvttss2si(<4 x float> %x)
   %sitofp = sitofp i32 %fptosi to float
   ret float %sitofp
 }
 
-define float @float_to_int_to_float_mem_f32_i64(<4 x float>* %p) {
+define float @float_to_int_to_float_mem_f32_i64(<4 x float>* %p) #0 {
 ; SSE-LABEL: float_to_int_to_float_mem_f32_i64:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    cvttss2si (%rdi), %rax
 ; SSE-NEXT:    cvtsi2ssq %rax, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: float_to_int_to_float_mem_f32_i64:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vcvttss2si (%rdi), %rax
 ; AVX-NEXT:    vcvtsi2ssq %rax, %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %x = load <4 x float>, <4 x float>* %p, align 16
   %fptosi = tail call i64 @llvm.x86.sse.cvttss2si64(<4 x float> %x)
   %sitofp = sitofp i64 %fptosi to float
   ret float %sitofp
 }
 
-define float @float_to_int_to_float_reg_f32_i64(<4 x float> %x) {
+define float @float_to_int_to_float_reg_f32_i64(<4 x float> %x) #0 {
 ; SSE-LABEL: float_to_int_to_float_reg_f32_i64:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    cvttss2si %xmm0, %rax
 ; SSE-NEXT:    xorps %xmm0, %xmm0
 ; SSE-NEXT:    cvtsi2ssq %rax, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: float_to_int_to_float_reg_f32_i64:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vcvttss2si %xmm0, %rax
 ; AVX-NEXT:    vcvtsi2ssq %rax, %xmm1, %xmm0
 ; AVX-NEXT:    retq
   %fptosi = tail call i64 @llvm.x86.sse.cvttss2si64(<4 x float> %x)
   %sitofp = sitofp i64 %fptosi to float
   ret float %sitofp
 }
 
-define double @float_to_int_to_float_mem_f64_i32(<2 x double>* %p) {
+define double @float_to_int_to_float_mem_f64_i32(<2 x double>* %p) #0 {
 ; SSE-LABEL: float_to_int_to_float_mem_f64_i32:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    cvttsd2si (%rdi), %eax
 ; SSE-NEXT:    cvtsi2sdl %eax, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: float_to_int_to_float_mem_f64_i32:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vcvttsd2si (%rdi), %eax
 ; AVX-NEXT:    vcvtsi2sdl %eax, %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %x = load <2 x double>, <2 x double>* %p, align 16
   %fptosi = tail call i32 @llvm.x86.sse2.cvttsd2si(<2 x double> %x)
   %sitofp = sitofp i32 %fptosi to double
   ret double %sitofp
 }
 
-define double @float_to_int_to_float_reg_f64_i32(<2 x double> %x) {
+define double @float_to_int_to_float_reg_f64_i32(<2 x double> %x) #0 {
 ; SSE-LABEL: float_to_int_to_float_reg_f64_i32:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    cvttsd2si %xmm0, %eax
 ; SSE-NEXT:    xorps %xmm0, %xmm0
 ; SSE-NEXT:    cvtsi2sdl %eax, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: float_to_int_to_float_reg_f64_i32:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vcvttsd2si %xmm0, %eax
 ; AVX-NEXT:    vcvtsi2sdl %eax, %xmm1, %xmm0
 ; AVX-NEXT:    retq
   %fptosi = tail call i32 @llvm.x86.sse2.cvttsd2si(<2 x double> %x)
   %sitofp = sitofp i32 %fptosi to double
   ret double %sitofp
 }
 
-define double @float_to_int_to_float_mem_f64_i64(<2 x double>* %p) {
+define double @float_to_int_to_float_mem_f64_i64(<2 x double>* %p) #0 {
 ; SSE-LABEL: float_to_int_to_float_mem_f64_i64:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    cvttsd2si (%rdi), %rax
 ; SSE-NEXT:    cvtsi2sdq %rax, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: float_to_int_to_float_mem_f64_i64:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vcvttsd2si (%rdi), %rax
 ; AVX-NEXT:    vcvtsi2sdq %rax, %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %x = load <2 x double>, <2 x double>* %p, align 16
   %fptosi = tail call i64 @llvm.x86.sse2.cvttsd2si64(<2 x double> %x)
   %sitofp = sitofp i64 %fptosi to double
   ret double %sitofp
 }
 
-define double @float_to_int_to_float_reg_f64_i64(<2 x double> %x) {
+define double @float_to_int_to_float_reg_f64_i64(<2 x double> %x) #0 {
 ; SSE-LABEL: float_to_int_to_float_reg_f64_i64:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    cvttsd2si %xmm0, %rax
 ; SSE-NEXT:    xorps %xmm0, %xmm0
 ; SSE-NEXT:    cvtsi2sdq %rax, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: float_to_int_to_float_reg_f64_i64:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vcvttsd2si %xmm0, %rax
 ; AVX-NEXT:    vcvtsi2sdq %rax, %xmm1, %xmm0
 ; AVX-NEXT:    retq
   %fptosi = tail call i64 @llvm.x86.sse2.cvttsd2si64(<2 x double> %x)
   %sitofp = sitofp i64 %fptosi to double
   ret double %sitofp
 }
 
-define <4 x float> @float_to_int_to_float_mem_v4f32(<4 x float>* %p) {
+define <4 x float> @float_to_int_to_float_mem_v4f32(<4 x float>* %p) #0 {
 ; SSE-LABEL: float_to_int_to_float_mem_v4f32:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    cvttps2dq (%rdi), %xmm0
 ; SSE-NEXT:    cvtdq2ps %xmm0, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: float_to_int_to_float_mem_v4f32:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vcvttps2dq (%rdi), %xmm0
 ; AVX-NEXT:    vcvtdq2ps %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %x = load <4 x float>, <4 x float>* %p, align 16
   %fptosi = tail call <4 x i32> @llvm.x86.sse2.cvttps2dq(<4 x float> %x)
   %sitofp = sitofp <4 x i32> %fptosi to <4 x float>
   ret <4 x float> %sitofp
 }
 
-define <4 x float> @float_to_int_to_float_reg_v4f32(<4 x float> %x) {
+define <4 x float> @float_to_int_to_float_reg_v4f32(<4 x float> %x) #0 {
 ; SSE-LABEL: float_to_int_to_float_reg_v4f32:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    cvttps2dq %xmm0, %xmm0
 ; SSE-NEXT:    cvtdq2ps %xmm0, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: float_to_int_to_float_reg_v4f32:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vcvttps2dq %xmm0, %xmm0
 ; AVX-NEXT:    vcvtdq2ps %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %fptosi = tail call <4 x i32> @llvm.x86.sse2.cvttps2dq(<4 x float> %x)
   %sitofp = sitofp <4 x i32> %fptosi to <4 x float>
   ret <4 x float> %sitofp
 }
 
-define <2 x double> @float_to_int_to_float_mem_v2f64(<2 x double>* %p) {
+define <2 x double> @float_to_int_to_float_mem_v2f64(<2 x double>* %p) #0 {
 ; SSE-LABEL: float_to_int_to_float_mem_v2f64:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    cvttpd2dq (%rdi), %xmm0
 ; SSE-NEXT:    cvtdq2pd %xmm0, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: float_to_int_to_float_mem_v2f64:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vcvttpd2dqx (%rdi), %xmm0
 ; AVX-NEXT:    vcvtdq2pd %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %x = load <2 x double>, <2 x double>* %p, align 16
   %fptosi = tail call <4 x i32> @llvm.x86.sse2.cvttpd2dq(<2 x double> %x)
   %concat = shufflevector <4 x i32> %fptosi, <4 x i32> undef, <2 x i32> <i32 0, i32 1>
   %sitofp = sitofp <2 x i32> %concat to <2 x double>
   ret <2 x double> %sitofp
 }
 
-define <2 x double> @float_to_int_to_float_reg_v2f64(<2 x double> %x) {
+define <2 x double> @float_to_int_to_float_reg_v2f64(<2 x double> %x) #0 {
 ; SSE-LABEL: float_to_int_to_float_reg_v2f64:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    cvttpd2dq %xmm0, %xmm0
 ; SSE-NEXT:    cvtdq2pd %xmm0, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: float_to_int_to_float_reg_v2f64:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vcvttpd2dq %xmm0, %xmm0
 ; AVX-NEXT:    vcvtdq2pd %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %fptosi = tail call <4 x i32> @llvm.x86.sse2.cvttpd2dq(<2 x double> %x)
   %concat = shufflevector <4 x i32> %fptosi, <4 x i32> undef, <2 x i32> <i32 0, i32 1>
   %sitofp = sitofp <2 x i32> %concat to <2 x double>
   ret <2 x double> %sitofp
 }
 
+attributes #0 = { "no-signed-zeros-fp-math"="true" }
+