Add constrained fptrunc and fpext intrinsics

[kpn]

Summary

This ticket was split off from D43515. It contains just the experimental constrained fptrunc and fpext intrinsics plus related changes like documentation. These two intrinsics are simpler to implement so I don't see a reason they need to wait on the rest of D43515.

[kpn]

I just realized I missed a couple of changes. Let me work on those and I'll update later.

[kpn]

It was pointed out before I split this ticket out that I wasn't properly passing the correct operand in ExpandNode() when handling the two new STRICT nodes. I've corrected that.

[cwabbott]

docs/LangRef.rst

14529

Hi,

I've been working on implementing a Vulkan extension which allows the user to specifiy different rounding modes for the AMDGPU backend. I'm not sure how this works in C/C++, but we're required to support floating-point truncation with non-standard rounding modes. Is there a reason the rounding mode isn't an argument here?

[kpn]

docs/LangRef.rst

14529

Going back and rereading D43515, I don't see an explicit reason given back then. And I can't find anything in the C99 or IEEE 754 standards, or in the LLVM documentation, that would mandate any particular rounding mode. So I'm open to adding a rounding mode argument to the constrained fptrunc.

Andrew? What do you think of making constrained fptrunc go back to taking a rounding mode argument?

Having said that, the constrained FP intrinsics are to avoid optimizations that change program behavior taking traps into account. Is this the behavior you need for Vulcan?

[cwabbott]

docs/LangRef.rst

14529

Vulkan doesn't support trapping floating-point exceptions, so we don't have to worry about that. However, my understanding is that we still need to communicate the rounding mode to LLVM, to prevent it from constant folding floating-point operations with the wrong rounding mode, so we still need to use the intrinsics.

There's also the complication that we may need to emit some "internal" floating-point operations which need to have some defined rounding mode different from what the user specified. The backend has total control over the control register that specifies the rounding mode, and in some cases (e.g. fp32 -> fp16 truncation) the rounding mode is actually specified statically by the instruction itself rather than the control register, so my thought was that we could make AMDGPU just always use the rounding specified by the argument to the constrained intrinsic, emitting changes to the control register if necessary. I'm not sure if the CodeGen infrastructure is set up to do that.

[cwabbott]

docs/LangRef.rst

14529

Oh, and I forgot another thing: the extension also adds support for letting the user either flush or preserve denormalized values. However, this is per-source-module, and sometimes we need to stitch together multiple source modules which have different rounding needs, emitting an instruction in between them to change the denorm flushing and/or rounding mode. So it seems we really do need to use the constrained intrinsics, to prevent code motion of floating-point operations around that register setting.

[kpn]

OK, I'm working on it now.

[kpn]

Add back a rounding mode argument to constrained fptrunc as requested by Connor Abbott.

[andrew.w.kaylor]

docs/LangRef.rst

14529

We should definitely have a rounding mode argument for fptrunc. I think the reason we missed that the first time around is probably due to the unfortunate naming of this operation (i.e. it isn't actually truncating) and the confusion with ISD::ROUND.

[andrew.w.kaylor]

This is looking pretty good. I don't think I know the Selection DAG well enough to offer a proper review of that. I'll see if I can get Craig's attention on it.

docs/index.rst

194

This seems a bit too prominently placed. Most people don't care about these intrinsics. I would recommend sinking this down into the subsytem documentation section. There is no clear organization there, so it's hard to say where it should go. Maybe just above or below the exception handling section (just based on my perception of the generality of each).

lib/IR/Verifier.cpp

4659

The formatting is non-standard and inconsistent in this section.

4674

You could combine these two lines as:

if (auto *VecTy = dyn_cast<VectorType>(OperandTy))

4685

Redefining "Operand" here makes the code confusing. I'd rather see Operand and Result as separate variables and make local variables for their types. I would also make the Assert statements more specific to what they are actually checking. For instance,

if (OperandTy->isVectorTy()) {

Assert(ResultTy->isVectorTy(), ...
Assert(OperandTy->getVectorNumElements() == ResultTy->getVectorNumElements(), ...

}

Is there a reason that these vector checks are specific to fptrunc and fpext? Is it just because they don't have a "same type" restriction in the intrinsic definition?

The floating point type assertions could be simplified as

Assert(OperandTy->isFPorFPVectorTy(),...
Assert(ResultTy->isFPorFPVectorTy(),...

You should also be checking that OperandTy->getScalarSizeInBits() > ResultTy->getScalarSizeInBits for fptrunc and vice versa for fpext.

[craig.topper]

lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp

711

What makes us reach this case? I would expect we'd scalarize based on the result type before we got to the operand type.

3059

We can't really widen this can we? Won't that put garbage in the upper elements?

[craig.topper]

test/CodeGen/X86/vector-constrained-fp-intrinsics.ll

2566

Please add an AVX command line so v4f64 will be a legal type.

[kpn]

lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp

711

Test case CodeGen/AArch64/neon-fpround_f128.ll says this code is needed. That case goes through the non-STRICT version of this function. This STRICT function was copied from the non-STRICT function and called in the appropriate places alongside that function. And a trivial conversion of the test case to use the constrained intrinsics does indeed go through this STRICT function.

I did want to try to keep the functions unified, but sometimes the result was too ugly to live.

3059

Is that what getUNDEF() does? Give llvm license to put garbage in registers? My assumption is that the code is fine because this function is a copy of WidenVecRes_Convert() with the needed changes for the strict node being chained. If if there's a problem here there's also a problem in that function.

lib/IR/Verifier.cpp

4674

The rewrite moots this point.

4685

I've rewritten this code and with your suggestions it does look much nicer.

I think I did put the vector checks are for fptrunc and fpext because they aren't checked earlier.

I've also added the checks for the appropriate changes in ScalarSizeInBits.

[kpn]

Address review comments.

[craig.topper]

lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp

3059

That is what the undef means. The existing code isn't required to be exception safe so garbage is fine. The constrained intrinsics have to be exception safe. This is why the implementation of WidenVecRes_StrictFP is different than the non-trapping case in WidenVecRes_BinaryCanTrap. I believe FADD/FSUB/FDIV/FMUL are considered non-trapping on most targets.

[kpn]

I'm seeing a regression in WebAssembly/PR40267.ll that I need to look into. If anyone else is seeing this failure let me know.

lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp

3059

It looks like this "WidenNumElts % InVTNumElts" block can just be eliminated. The fallback at the end of the function should handle the case without any undefs.

[cameron.mcinally]

lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

2846

Would this lowering to EmitStackConvert(...) discard the rounding mode? I'm not familiar with this code, but I *think* it would.

E.g. STRICT_FP_ROUND would lower to a truncating store.

[cameron.mcinally]

lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp

711

Have you considered scalarizing the operands with a generic function? Something like ScalarizeVecRes_StrictFPOp(...)? I suspect we'll see reuse of this code.

@craig.topper

What makes us reach this case? I would expect we'd scalarize based on the result type before we got to the operand type.

The different operand and result types for these operations are probably why ScalarizeVecRes_StrictFPOp(...) didn't trip. Just guessing though...

[kpn]

lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

2846

Yes it would, assuming EmitStackConvert() works correctly for floating point types. A 'make check' of llvm with the default targets enabled doesn't trigger the call to DAG.getTruncStore(), so I'm not actually sure it works.

But none of the other intrinsics handle rounding modes currently, so I think we should leave this to future work.

lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp

711

You know, I thought about it. But trying to generalize it would hide the parallel with ScalarizeVecRes_FP_ROUND(). Right now the connection between the two is obvious. I'm not sure that we'd gain much at this point in time by generalizing and losing that bit of readability.

If we do end up needing to generalize this code in the future then we can do it then.

[kpn]

Address review comments.

Fix regression on WebAssembly accidentally caused by incorrect svn update.

[kpn]

Rebase. Ping.

[craig.topper]

docs/LangRef.rst

14511

This should be in a separate patch.

include/llvm/IR/Intrinsics.td

592

This line [ should probably be lined up with the line above. Same with fpext.

lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

1784

This line doesn't make sense to me. This is replacing the users of result 1 of the load you just created with Chain. But no one has seen this load yet so how can its result 1 have any users?

2831

Shouldn't you be passing the input chain operand of Node into the last argument? Right now it looks like you're using the Chain result from Node itself. But we want to delete Node.

lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp

713

Assert that OpNo is 1.

1530

use a temporary for the OpNo instead of of repeatedly calling N->isStrictFPOpcode()

1933

This comment must have been copy pasted. We aren't operating on a "load" here

4024

Add curly braces to the else and fix the indentation

lib/CodeGen/SelectionDAG/SelectionDAG.cpp

7682

Can this be Node->getValueType(0)? Accessing ValueList directly seems pretty unusual.

lib/IR/Verifier.cpp

4630–4712

Most of this looks like it belongs in a separate patch. This patch shoudl focus on adding the new intrinsics. Gaps in old intrinsics should be separated.

test/CodeGen/X86/vector-constrained-fp-intrinsics.ll

3

Do this as a pre-commit?

[kpn]

docs/LangRef.rst

14511

What should be in a separate patch? The new AddingConstrainedIntrinsics.rst file?

[kpn]

lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

1784

Agreed. I'm fixing it now.

2831

I'm now changing EmitStackConvert to splice itself into the chain. So passing Node is about to become correct since I need both ends of the chain to do the splicing.

[kpn]

Address review comments.

I did find that I didn't need a new, special version of ReplaceNode(). One of the existing ones appears to do the job. That meant I could simplify the changes to EmitStackConvert() and avoid chain splicing there. This, in turn, means I only needed to pass in one end of the chain to EmitStackConvert(). I hope the result is more readable.

[andrew.w.kaylor]

lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

2829

Can this cause the wrong rounding mode to be used? For X86 targets I would guess it will result in an instruction that uses the runtime rounding mode, but I'm not sure we can count on that for all targets. Also, if the value being converted is a constant, might this get folded using the default rounding mode? And if it is a constant and we knew the rounding mode based on an argument to the intrinsic, we might want to fold it (though ideally that would have happened before this).

[kpn]

lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

2829

Isn't this an argument for a strict load node types and maybe a strict store type as well?

I'd be surprised if we were doing constant folding of a store+load combination, but if we are then either a strict load or store node would disable that folding.

Strictly speaking, using the runtime rounding mode would be incorrect if that wasn't what was specified in the rounding mode field of the intrinsic. So even X86 could be wrong.

We aren't doing anything with the rounding and exception arguments yet for any of the new intrinsics. So how about I file a bug noting that fact and noting that we need to not forget about this case? Then we can move on and come back later.

[andrew.w.kaylor]

lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

2829

Actually, if the runtime rounding mode doesn't match the rounding argument for non-dynamic cases that's a user error or a bug somewhere upstream in the compiler. The rounding mode argument is supposed to tell us what the rounding mode is at this point in the program. It is not supposed to control the rounding mode. The non-dynamic rounding modes are essentially an "assume" kind of directive.

I see that the constant folding case I'm concerned about will actually be covered by one of your test cases, so I think we can safely move forward with this the way you have it.

lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp

1530

I'd prefer to see this as:

unsigned OpNo = N->isStrictFPOpcode() ? 1 : 0;

1921

Again, the use of a boolean result as an index is awkward.

test/CodeGen/X86/fp-intrinsics.ll

303

This comment is wrong.

test/CodeGen/X86/vector-constrained-fp-intrinsics.ll

3

Expanding on Craig's comment here. I think he is suggesting that you add the AVX run line and all of the associated new checks as a separate patch before the fpext/fptrrunc patch lands.

[kpn]

Address review comments.

[andrew.w.kaylor]

lib/CodeGen/SelectionDAG/SelectionDAG.cpp

7664

It might be helpful to have a comment here explaining why STRICT_FP_ROUND isn't unary. That can be done in a separate change if everything else here is ready to commit. It all looks good to me, but I'm hoping Craig can give your widen vector changes one last look to make sure that's doing what he expected.

[craig.topper]

lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

2835

This break is unreachable.

2850

This break is unreachable

lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp

3042

I'm not sure this is safe. There's no way of knowing how the input was widened. I don't think you can access more than the original vector width worth of elements.

3054

I don't think this is safe either.

3995

Again I don't think this is safe. We don't what's in widened elements of the input.

test/CodeGen/X86/vector-constrained-fp-intrinsics.ll

2487

I think this is wrong. It's reading 4 elements, but we don't know what is in the 4th element.

2558

This is reading 4 float elements from memory. We should only be reading 2.

2584

This reads 4 floats from memory, but we should only read 3.

[cameron.mcinally]

lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

2829

Can this cause the wrong rounding mode to be used? For X86 targets I would guess it will result in an instruction that uses the runtime rounding mode, but I'm not sure we can count on that for all targets.

I think it's more severe than that. The details are not fresh in my mind, but IIRC EmitStackConvert(...) emits a straight truncate on X86 (well, at least for one case I looked at. Could be wrong too.). We probably shouldn't be calling it for a strict round, and rather go to an instruction that honors rounding mode.

[kpn]

lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

2829

My impression is that EmitStackConvert() is one of those functions that is only used as a last resort when an ISA lacks a good reg+reg instruction. So this is probably already handled for most cases.

If you know of a way to write a strict FP test case that uses it I'd love to see it. And if you have one then we should probably turn it into a library call to honor the rounding mode.

lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp

3042

I believe you are correct. I'm looking at this now.

[andrew.w.kaylor]

lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

2829

@cameron.mcinally , what do you mean by "a straight truncate"?

[cameron.mcinally]

lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

2829

Nope, I made a mistake. EmitStackConvert(...) ends up creating a truncstore, which could end up as an FST. I had mistakenly thought that this rounded to nearest, but it looks like it uses the FPU control word.

Assuming that fenv.h sets the FPU control word correctly (I didn't check), it should be fine.

[andrew.w.kaylor]

lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

2829

Regarding fenv.h, if you use the fesetround() function from fenv.h to set the rounding mode, it will update both MXCSR and FPCW. If you use an intrinsic or inline assembly or whatever to set one of them without the other, you're on your own (and fegetround() will fail if they have different settings).

[kpn]

Address review comments: Add a comment as requested. Remove bogus optimizations to let the base case work. Update tests.

[kpn]

Why does Verifier::visitIntrinsicCall() not choke when given an intrinsic it doesn't know about?

I added back the lines accidentally lost so the new intrinsics do now get checked.

[craig.topper]

include/llvm/IR/Intrinsics.td

672

A lot of this FIXME still applies doesn't it?

lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp

3032

I think this is an unused variable

lib/IR/Verifier.cpp

4627

Indented too far

4629

Indented too far

[cameron.mcinally]

include/llvm/IR/Intrinsics.td

672

We definitely need FCMP. I attempted a patch for this, D54649, but FCMP is Custom lowered on X86. We still need a good way to handle STRICT nodes that need to be Custom lowered.

Also FYI that the line below can be removed. We decided not to implement constrained FABS and FCOPYSIGN... unless a problem is found.

[kpn]

Address review comments.

[craig.topper]

LGTM with that one remaining unused variable fixed.

lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp

3031

Is InWidenVT also unused?

[kpn]

Excellent. Thank you for all the reviews! I will commit this on Monday morning so I can keep an eye on the bots.

docs/AddingConstrainedIntrinsics.rst

+==================================================
+How To Add A Constrained Floating-Point Intrinsic
+==================================================
+
+.. contents::
+   :local:
+
+.. warning::
+  This is a work in progress.
+
+Add the intrinsic
+=================
+
+Multiple files need to be updated when adding a new constrained intrinsic.
+
+Add the new intrinsic to the table of intrinsics.::
+
+  include/llvm/IR/Intrinsics.td
+
+Update class ConstrainedFPIntrinsic to know about the intrinsics.::
+
+  include/llvm/IR/IntrinsicInst.h
+
+Functions like ConstrainedFPIntrinsic::isUnaryOp() or
+ConstrainedFPIntrinsic::isTernaryOp() may need to know about the new
+intrinsic.::
+
+  lib/IR/IntrinsicInst.cpp
+
+Update the IR verifier::
+
+  lib/IR/Verifier.cpp
+
+Add SelectionDAG node types
+===========================
+
+Add the new STRICT version of the node type to the ISD::NodeType enum.::
+
+  include/llvm/CodeGen/ISDOpcodes.h
+
+In class SDNode update isStrictFPOpcode()::
+
+  include/llvm/CodeGen/SelectionDAGNodes.h
+
+A mapping from the STRICT SDnode type to the non-STRICT is done in
+TargetLoweringBase::getStrictFPOperationAction(). This allows STRICT
+nodes to be legalized similarly to the non-STRICT node type.::
+
+  include/llvm/CodeGen/TargetLowering.h
+
+Building the SelectionDAG
+-------------------------
+
+The switch statement in SelectionDAGBuilder::visitIntrinsicCall() needs
+to be updated to call SelectionDAGBuilder::visitConstrainedFPIntrinsic().
+That function, in turn, needs to be updated to know how to create the
+SDNode for the intrinsic. The new STRICT node will eventually be converted
+to the matching non-STRICT node. For this reason it _must_ have the same
+operands and values as the non-STRICT version in case the non-STRICT
+version's default lowering is used. This means that if the non-STRICT
+version of the node does not use the chain then the STRICT node cannot
+either.::
+
+  lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
+
+Most of the STRICT nodes get legalized the same as their matching non-STRICT
+counterparts. A new STRICT node with this property must get added to the
+switch in SelectionDAGLegalize::LegalizeOp().::
+
+  lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
+
+Other parts of the legalizer may need to be updated as well. Look for
+places where the non-STRICT counterpart is legalized and update as needed.
+Be careful of the chain since STRICT nodes use it but their counterparts
+often don't.::
+
+The code to do the conversion or mutation of the STRICT node to a non-STRICT
+version of the node happens in SelectionDAG::mutateStrictFPToFP(). Be
+careful updating this function since some nodes are always chained and
+some are not. Some nodes have the same return type as their input operand,
+but some are different. Both of these points must be properly handled.::
+
+  lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+
+To make debug logs readable it is helpful to update the SelectionDAG's
+debug logger:::
+
+  lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp
+
+Add documentation and tests
+===========================
+
+::
+
+  docs/LangRef.rst

docs/LangRef.rst

 """"""""""
 
 The first three arguments to the '``llvm.experimental.constrained.fma``'
 intrinsic must be :ref:`floating-point <t_floating>` or :ref:`vector
 <t_vector>` of floating-point values. All arguments must have identical types.
 
 The fourth and fifth arguments specify the rounding mode and exception behavior
 as described above.
 
 Semantics:
 """"""""""
 
 The result produced is the product of the first two operands added to the third
 operand computed with infinite precision, and then rounded to the target
 precision.
 
+'``llvm.experimental.constrained.fptrunc``' Intrinsic
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Syntax:
+"""""""
+
+::
+
+      declare <ty2>
+      @llvm.experimental.constrained.fptrunc(<type> <value>,
+                                          metadata <exception behavior>)
+
+Overview:
+"""""""""
+
+The '``llvm.experimental.constrained.fptrunc``' intrinsic truncates ``value``
+to type ``ty2``.
+
+Arguments:
+""""""""""
+
+The first argument to the '``llvm.experimental.constrained.fptrunc``'
+intrinsic must be :ref:`floating point <t_floating>` or :ref:`vector
+<t_vector>` of floating point values. This argument must be larger in size
+than the result.
+
+The second argument specifies the exception behavior as described above.
+
+Semantics:
+""""""""""
+
+The result produced is a floating point value truncated to be smaller in size
+than the operand.
+
+'``llvm.experimental.constrained.fpext``' Intrinsic
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Syntax:
+"""""""
+
+::
+
+      declare <ty2>
+      @llvm.experimental.constrained.fpext(<type> <value>,
+                                          metadata <exception behavior>)
+
+Overview:
+"""""""""
+
+The '``llvm.experimental.constrained.fpext``' intrinsic extends a 
+floating-point ``value`` to a larger floating-point value.
+
+Arguments:
+""""""""""
+
+The first argument to the '``llvm.experimental.constrained.fpext``'
+intrinsic must be :ref:`floating point <t_floating>` or :ref:`vector
+<t_vector>` of floating point values. This argument must be smaller in size
+than the result.
+
+The second argument specifies the exception behavior as described above.
+
+Semantics:
+""""""""""
+
+The result produced is a floating point value extended to be larger in size
+than the operand. All restrictions that apply to the fpext instruction also
+apply to this intrinsic.
+
 Constrained libm-equivalent Intrinsics
 --------------------------------------
 
 In addition to the basic floating-point operations for which constrained
 intrinsics are described above, there are constrained versions of various
 operations which provide equivalent behavior to a corresponding libm function.
 These intrinsics allow the precise behavior of these operations with respect to
 rounding mode and exception behavior to be controlled.

docs/index.rst

 .. toctree::
    :hidden:
 
    Atomics
    CodingStandards
    CommandLine
    CompilerWriterInfo
    ExtendingLLVM
+   AddingConstrainedIntrinsics
    HowToSetUpLLVMStyleRTTI
    ProgrammersManual
    Extensions
    LibFuzzer
    FuzzingLLVM
    ScudoHardenedAllocator
    OptBisect
 

 
 :doc:`HowToSetUpLLVMStyleRTTI`
   How to make ``isa<>``, ``dyn_cast<>``, etc. available for clients of your
   class hierarchy.
 
 :doc:`ExtendingLLVM`
   Look here to see how to add instructions and intrinsics to LLVM.
 
+:doc:`AddingConstrainedIntrinsics`
+   Gives the steps necessary when adding a new constrained math intrinsic
+   to LLVM.
+
 `Doxygen generated documentation <http://llvm.org/doxygen/>`_
   (`classes <http://llvm.org/doxygen/inherits.html>`_)
 
 `Documentation for Go bindings <http://godoc.org/llvm.org/llvm/bindings/go/llvm>`_
 
 `Github Source Repository Browser <http://github.com/llvm/llvm-project//>`_
    ..
 

include/llvm/CodeGen/ISDOpcodes.h

enum NodeType {
     /// These will be lowered to the equivalent non-constrained pseudo-op
     /// (or expanded to the equivalent library call) before final selection.
     /// They are used to limit optimizations while the DAG is being optimized.
     STRICT_FSQRT, STRICT_FPOW, STRICT_FPOWI, STRICT_FSIN, STRICT_FCOS,
     STRICT_FEXP, STRICT_FEXP2, STRICT_FLOG, STRICT_FLOG10, STRICT_FLOG2,
     STRICT_FRINT, STRICT_FNEARBYINT, STRICT_FMAXNUM, STRICT_FMINNUM,
     STRICT_FCEIL, STRICT_FFLOOR, STRICT_FROUND, STRICT_FTRUNC,
 
+    /// X = STRICT_FP_ROUND(Y, TRUNC) - Rounding 'Y' from a larger floating 
+    /// point type down to the precision of the destination VT.  TRUNC is a 
+    /// flag, which is always an integer that is zero or one.  If TRUNC is 0,
+    /// this is a normal rounding, if it is 1, this FP_ROUND is known to not
+    /// change the value of Y.
+    ///
+    /// The TRUNC = 1 case is used in cases where we know that the value will
+    /// not be modified by the node, because Y is not using any of the extra
+    /// precision of source type.  This allows certain transformations like
+    /// STRICT_FP_EXTEND(STRICT_FP_ROUND(X,1)) -> X which are not safe for
+    /// STRICT_FP_EXTEND(STRICT_FP_ROUND(X,0)) because the extra bits aren't
+    /// removed.
+    /// It is used to limit optimizations while the DAG is being optimized.
+    STRICT_FP_ROUND,
+
+    /// X = STRICT_FP_EXTEND(Y) - Extend a smaller FP type into a larger FP
+    /// type.
+    /// It is used to limit optimizations while the DAG is being optimized.
+    STRICT_FP_EXTEND,
+
     /// FMA - Perform a * b + c with no intermediate rounding step.
     FMA,
 
     /// FMAD - Perform a * b + c, while getting the same result as the
     /// separately rounded operations.
     FMAD,
 
     /// FCOPYSIGN(X, Y) - Return the value of X with the sign of Y.  NOTE: This

include/llvm/CodeGen/SelectionDAGNodes.h

switch (NodeType) {
       case ISD::STRICT_FRINT:
       case ISD::STRICT_FNEARBYINT:
       case ISD::STRICT_FMAXNUM:
       case ISD::STRICT_FMINNUM:
       case ISD::STRICT_FCEIL:
       case ISD::STRICT_FFLOOR:
       case ISD::STRICT_FROUND:
       case ISD::STRICT_FTRUNC:
+      case ISD::STRICT_FP_ROUND:
+      case ISD::STRICT_FP_EXTEND:
         return true;
     }
   }
 
   /// Test if this node has a post-isel opcode, directly
   /// corresponding to a MachineInstr opcode.
   bool isMachineOpcode() const { return NodeType < 0; }
 

include/llvm/CodeGen/TargetLowering.h

switch (Op) {
       case ISD::STRICT_FRINT: EqOpc = ISD::FRINT; break;
       case ISD::STRICT_FNEARBYINT: EqOpc = ISD::FNEARBYINT; break;
       case ISD::STRICT_FMAXNUM: EqOpc = ISD::FMAXNUM; break;
       case ISD::STRICT_FMINNUM: EqOpc = ISD::FMINNUM; break;
       case ISD::STRICT_FCEIL: EqOpc = ISD::FCEIL; break;
       case ISD::STRICT_FFLOOR: EqOpc = ISD::FFLOOR; break;
       case ISD::STRICT_FROUND: EqOpc = ISD::FROUND; break;
       case ISD::STRICT_FTRUNC: EqOpc = ISD::FTRUNC; break;
+      case ISD::STRICT_FP_ROUND: EqOpc = ISD::FP_ROUND; break;
+      case ISD::STRICT_FP_EXTEND: EqOpc = ISD::FP_EXTEND; break;
     }
 
     auto Action = getOperationAction(EqOpc, VT);
 
     // We don't currently handle Custom or Promote for strict FP pseudo-ops.
     // For now, we just expand for those cases.
     if (Action != Legal)
       Action = Expand;

include/llvm/IR/IntrinsicInst.h

public:
     static bool classof(const IntrinsicInst *I) {
       switch (I->getIntrinsicID()) {
       case Intrinsic::experimental_constrained_fadd:
       case Intrinsic::experimental_constrained_fsub:
       case Intrinsic::experimental_constrained_fmul:
       case Intrinsic::experimental_constrained_fdiv:
       case Intrinsic::experimental_constrained_frem:
       case Intrinsic::experimental_constrained_fma:
+      case Intrinsic::experimental_constrained_fptrunc:
+      case Intrinsic::experimental_constrained_fpext:
       case Intrinsic::experimental_constrained_sqrt:
       case Intrinsic::experimental_constrained_pow:
       case Intrinsic::experimental_constrained_powi:
       case Intrinsic::experimental_constrained_sin:
       case Intrinsic::experimental_constrained_cos:
       case Intrinsic::experimental_constrained_exp:
       case Intrinsic::experimental_constrained_exp2:
       case Intrinsic::experimental_constrained_log:

include/llvm/IR/Intrinsics.td

let IntrProperties = [IntrInaccessibleMemOnly] in {
 
   def int_experimental_constrained_fma : Intrinsic<[ llvm_anyfloat_ty ],
                                                     [ LLVMMatchType<0>,
                                                       LLVMMatchType<0>,
                                                       LLVMMatchType<0>,
                                                       llvm_metadata_ty,
                                                       llvm_metadata_ty ]>;
 
+  def int_experimental_constrained_fptrunc : Intrinsic<[ llvm_anyfloat_ty ],
+                                                    [ llvm_anyfloat_ty,
+                                                      llvm_metadata_ty ]>;
+
+  def int_experimental_constrained_fpext : Intrinsic<[ llvm_anyfloat_ty ],
+                                                    [ llvm_anyfloat_ty,
+                                                      llvm_metadata_ty ]>;
+
   // These intrinsics are sensitive to the rounding mode so we need constrained
   // versions of each of them.  When strict rounding and exception control are
   // not required the non-constrained versions of these intrinsics should be
   // used.
   def int_experimental_constrained_sqrt : Intrinsic<[ llvm_anyfloat_ty ],
                                                     [ LLVMMatchType<0>,
                                                       llvm_metadata_ty,
                                                       llvm_metadata_ty ]>;
def int_experimental_constrained_round : Intrinsic<[ llvm_anyfloat_ty ],
                                                      [ LLVMMatchType<0>,
                                                       llvm_metadata_ty,
                                                       llvm_metadata_ty ]>;
   def int_experimental_constrained_trunc : Intrinsic<[ llvm_anyfloat_ty ],
                                                      [ LLVMMatchType<0>,
                                                        llvm_metadata_ty,
                                                        llvm_metadata_ty ]>;
 }
-// FIXME: Add intrinsics for fcmp, fptrunc, fpext, fptoui and fptosi.
-// FIXME: Add intrinsics for fabs and copysign? 
 
 
 //===------------------------- Expect Intrinsics --------------------------===//
 //
 def int_expect : Intrinsic<[llvm_anyint_ty], [LLVMMatchType<0>,
                                               LLVMMatchType<0>], [IntrNoMem]>;
 
 //===-------------------- Bit Manipulation Intrinsics ---------------------===//

lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

SDValue ExpandIntLibCall(SDNode *Node, bool isSigned,
                            RTLIB::Libcall Call_I16,
                            RTLIB::Libcall Call_I32,
                            RTLIB::Libcall Call_I64,
                            RTLIB::Libcall Call_I128);
   void ExpandDivRemLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
   void ExpandSinCosLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
 
   SDValue EmitStackConvert(SDValue SrcOp, EVT SlotVT, EVT DestVT,
-                           const SDLoc &dl);
+                           const SDLoc &dl, SDNode *OldChain = nullptr);
   SDValue ExpandBUILD_VECTOR(SDNode *Node);
   SDValue ExpandSCALAR_TO_VECTOR(SDNode *Node);
   void ExpandDYNAMIC_STACKALLOC(SDNode *Node,
                                 SmallVectorImpl<SDValue> &Results);
   void getSignAsIntValue(FloatSignAsInt &State, const SDLoc &DL,
                          SDValue Value) const;
   SDValue modifySignAsInt(const FloatSignAsInt &State, const SDLoc &DL,
                           SDValue NewIntValue) const;
void ReplaceNode(SDNode *Old, const SDValue *New) {
     for (unsigned i = 0, e = Old->getNumValues(); i != e; ++i) {
       LLVM_DEBUG(dbgs() << (i == 0 ? "     with:      " : "      and:      ");
                  New[i]->dump(&DAG));
       if (UpdatedNodes)
         UpdatedNodes->insert(New[i].getNode());
     }
     ReplacedNode(Old);
   }
+
+  void ReplaceNodeIgnoreChain(SDNode *Old, SDNode *New) {
+    LLVM_DEBUG(dbgs() << " ... replacing: "; Old->dump(&DAG);
+               dbgs() << "     with:      "; New->dump(&DAG));
+
+    DAG.ReplaceAllUsesOfValueWith(SDValue(Old, 0), SDValue(New, 0));
+    if (UpdatedNodes)
+      UpdatedNodes->insert(New);
+    ReplacedNode(Old);
+  }
 };
 
 } // end anonymous namespace
 
 /// Return a vector shuffle operation which
 /// performs the same shuffle in terms of order or result bytes, but on a type
 /// whose vector element type is narrower than the original shuffle type.
 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
#endif
   case ISD::STRICT_FRINT:
   case ISD::STRICT_FNEARBYINT:
   case ISD::STRICT_FMAXNUM:
   case ISD::STRICT_FMINNUM:
   case ISD::STRICT_FCEIL:
   case ISD::STRICT_FFLOOR:
   case ISD::STRICT_FROUND:
   case ISD::STRICT_FTRUNC:
+  case ISD::STRICT_FP_ROUND:
+  case ISD::STRICT_FP_EXTEND:
     // These pseudo-ops get legalized as if they were their non-strict
     // equivalent.  For instance, if ISD::FSQRT is legal then ISD::STRICT_FSQRT
     // is also legal, but if ISD::FSQRT requires expansion then so does
     // ISD::STRICT_FSQRT.
     Action = TLI.getStrictFPOperationAction(Node->getOpcode(),
                                             Node->getValueType(0));
     break;
   case ISD::SADDSAT:
bool SelectionDAGLegalize::LegalizeSetCCCondCode(EVT VT, SDValue &LHS,
   return false;
 }
 
 /// Emit a store/load combination to the stack.  This stores
 /// SrcOp to a stack slot of type SlotVT, truncating it if needed.  It then does
 /// a load from the stack slot to DestVT, extending it if needed.
 /// The resultant code need not be legal.
 SDValue SelectionDAGLegalize::EmitStackConvert(SDValue SrcOp, EVT SlotVT,
-                                               EVT DestVT, const SDLoc &dl) {
+                                               EVT DestVT, const SDLoc &dl,
+                                               SDNode *OldChain) {
   // Create the stack frame object.
   unsigned SrcAlign = DAG.getDataLayout().getPrefTypeAlignment(
       SrcOp.getValueType().getTypeForEVT(*DAG.getContext()));
   SDValue FIPtr = DAG.CreateStackTemporary(SlotVT, SrcAlign);
 
   FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(FIPtr);
   int SPFI = StackPtrFI->getIndex();
   MachinePointerInfo PtrInfo =
       MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SPFI);
 
   unsigned SrcSize = SrcOp.getValueSizeInBits();
   unsigned SlotSize = SlotVT.getSizeInBits();
   unsigned DestSize = DestVT.getSizeInBits();
   Type *DestType = DestVT.getTypeForEVT(*DAG.getContext());
   unsigned DestAlign = DAG.getDataLayout().getPrefTypeAlignment(DestType);
 
   // Emit a store to the stack slot.  Use a truncstore if the input value is
   // later than DestVT.
   SDValue Store;
+  SDValue Chain = SDValue(OldChain, 1);
+  if (!OldChain)
+    Chain = DAG.getEntryNode();
 
   if (SrcSize > SlotSize)
-    Store = DAG.getTruncStore(DAG.getEntryNode(), dl, SrcOp, FIPtr, PtrInfo,
+    Store = DAG.getTruncStore(Chain, dl, SrcOp, FIPtr, PtrInfo,
                               SlotVT, SrcAlign);
   else {
     assert(SrcSize == SlotSize && "Invalid store");
     Store =
-        DAG.getStore(DAG.getEntryNode(), dl, SrcOp, FIPtr, PtrInfo, SrcAlign);
+        DAG.getStore(Chain, dl, SrcOp, FIPtr, PtrInfo, SrcAlign);
   }
 
-  // Result is a load from the stack slot.
-  if (SlotSize == DestSize)
-    return DAG.getLoad(DestVT, dl, Store, FIPtr, PtrInfo, DestAlign);
+  SDValue Load;
 
+  // Result is a load from the stack slot.
+  if (SlotSize == DestSize) {
+    Load = DAG.getLoad(DestVT, dl, Store, FIPtr, PtrInfo, DestAlign);
+  }
+  else {
     assert(SlotSize < DestSize && "Unknown extension!");
-  return DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT, Store, FIPtr, PtrInfo, SlotVT,
-                        DestAlign);
+    Load = DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT, Store, FIPtr, 
+                          PtrInfo, SlotVT, DestAlign);
+  }
+
+  if (OldChain) {
+    DAG.ReplaceAllUsesOfValueWith(SDValue(Load.getNode(), 1), Chain);
+  }
+
+  return Load;
 }
 
 SDValue SelectionDAGLegalize::ExpandSCALAR_TO_VECTOR(SDNode *Node) {
   SDLoc dl(Node);
   // Create a vector sized/aligned stack slot, store the value to element #0,
   // then load the whole vector back out.
   SDValue StackPtr = DAG.CreateStackTemporary(Node->getValueType(0));
 
case ISD::UNDEF: {
     if (VT.isInteger())
       Results.push_back(DAG.getConstant(0, dl, VT));
     else {
       assert(VT.isFloatingPoint() && "Unknown value type!");
       Results.push_back(DAG.getConstantFP(0, dl, VT));
     }
     break;
   }
+  case ISD::STRICT_FP_ROUND:
+    Tmp1 = EmitStackConvert(Node->getOperand(1), 
+                            Node->getValueType(0),
+                            Node->getValueType(0), dl, Node);
+    ReplaceNodeIgnoreChain(Node, Tmp1.getNode());
+    LLVM_DEBUG(dbgs() << "Successfully expanded STRICT_FP_ROUND node\n");
+    return true;
+    break;
   case ISD::FP_ROUND:
   case ISD::BITCAST:
-    Tmp1 = EmitStackConvert(Node->getOperand(0), Node->getValueType(0),
+    Tmp1 = EmitStackConvert(Node->getOperand(0), 
+                            Node->getValueType(0),
                             Node->getValueType(0), dl);
     Results.push_back(Tmp1);
     break;
+  case ISD::STRICT_FP_EXTEND:
+    Tmp1 = EmitStackConvert(Node->getOperand(1),
+                            Node->getOperand(1).getValueType(),
+                            Node->getValueType(0), dl, Node);
+    ReplaceNodeIgnoreChain(Node, Tmp1.getNode());
+    LLVM_DEBUG(dbgs() << "Successfully expanded STRICT_FP_EXTEND node\n");
+    return true;
+    break;
   case ISD::FP_EXTEND:
     Tmp1 = EmitStackConvert(Node->getOperand(0),
                             Node->getOperand(0).getValueType(),
                             Node->getValueType(0), dl);
     Results.push_back(Tmp1);
     break;
   case ISD::SIGN_EXTEND_INREG: {
     EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();

lib/CodeGen/SelectionDAG/LegalizeTypes.h

private:
   SDValue ScalarizeVecRes_StrictFPOp(SDNode *N);
   SDValue ScalarizeVecRes_InregOp(SDNode *N);
   SDValue ScalarizeVecRes_VecInregOp(SDNode *N);
 
   SDValue ScalarizeVecRes_BITCAST(SDNode *N);
   SDValue ScalarizeVecRes_BUILD_VECTOR(SDNode *N);
   SDValue ScalarizeVecRes_EXTRACT_SUBVECTOR(SDNode *N);
   SDValue ScalarizeVecRes_FP_ROUND(SDNode *N);
+  SDValue ScalarizeVecRes_STRICT_FP_ROUND(SDNode *N);
   SDValue ScalarizeVecRes_FPOWI(SDNode *N);
   SDValue ScalarizeVecRes_INSERT_VECTOR_ELT(SDNode *N);
   SDValue ScalarizeVecRes_LOAD(LoadSDNode *N);
   SDValue ScalarizeVecRes_SCALAR_TO_VECTOR(SDNode *N);
   SDValue ScalarizeVecRes_VSELECT(SDNode *N);
   SDValue ScalarizeVecRes_SELECT(SDNode *N);
   SDValue ScalarizeVecRes_SELECT_CC(SDNode *N);
   SDValue ScalarizeVecRes_SETCC(SDNode *N);
   SDValue ScalarizeVecRes_UNDEF(SDNode *N);
   SDValue ScalarizeVecRes_VECTOR_SHUFFLE(SDNode *N);
 
   SDValue ScalarizeVecRes_SMULFIX(SDNode *N);
 
   // Vector Operand Scalarization: <1 x ty> -> ty.
   bool ScalarizeVectorOperand(SDNode *N, unsigned OpNo);
   SDValue ScalarizeVecOp_BITCAST(SDNode *N);
   SDValue ScalarizeVecOp_UnaryOp(SDNode *N);
   SDValue ScalarizeVecOp_CONCAT_VECTORS(SDNode *N);
   SDValue ScalarizeVecOp_EXTRACT_VECTOR_ELT(SDNode *N);
   SDValue ScalarizeVecOp_VSELECT(SDNode *N);
   SDValue ScalarizeVecOp_VSETCC(SDNode *N);
   SDValue ScalarizeVecOp_STORE(StoreSDNode *N, unsigned OpNo);
   SDValue ScalarizeVecOp_FP_ROUND(SDNode *N, unsigned OpNo);
+  SDValue ScalarizeVecOp_STRICT_FP_ROUND(SDNode *N, unsigned OpNo);
 
   //===--------------------------------------------------------------------===//
   // Vector Splitting Support: LegalizeVectorTypes.cpp
   //===--------------------------------------------------------------------===//
 
   /// Given a processed vector Op which was split into vectors of half the size,
   /// this method returns the halves. The first elements of Op coincide with the
   /// elements of Lo; the remaining elements of Op coincide with the elements of
private:
   SDValue WidenVecRes_UNDEF(SDNode *N);
   SDValue WidenVecRes_VECTOR_SHUFFLE(ShuffleVectorSDNode *N);
 
   SDValue WidenVecRes_Ternary(SDNode *N);
   SDValue WidenVecRes_Binary(SDNode *N);
   SDValue WidenVecRes_BinaryCanTrap(SDNode *N);
   SDValue WidenVecRes_StrictFP(SDNode *N);
   SDValue WidenVecRes_Convert(SDNode *N);
+  SDValue WidenVecRes_Convert_StrictFP(SDNode *N);
   SDValue WidenVecRes_FCOPYSIGN(SDNode *N);
   SDValue WidenVecRes_POWI(SDNode *N);
   SDValue WidenVecRes_Shift(SDNode *N);
   SDValue WidenVecRes_Unary(SDNode *N);
   SDValue WidenVecRes_InregOp(SDNode *N);
 
   // Widen Vector Operand.
   bool WidenVectorOperand(SDNode *N, unsigned OpNo);

lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp

SDValue VectorLegalizer::LegalizeOp(SDValue Op) {
   case ISD::STRICT_FRINT:
   case ISD::STRICT_FNEARBYINT:
   case ISD::STRICT_FMAXNUM:
   case ISD::STRICT_FMINNUM:
   case ISD::STRICT_FCEIL:
   case ISD::STRICT_FFLOOR:
   case ISD::STRICT_FROUND:
   case ISD::STRICT_FTRUNC:
+  case ISD::STRICT_FP_ROUND:
+  case ISD::STRICT_FP_EXTEND:
     // These pseudo-ops get legalized as if they were their non-strict
     // equivalent.  For instance, if ISD::FSQRT is legal then ISD::STRICT_FSQRT
     // is also legal, but if ISD::FSQRT requires expansion then so does
     // ISD::STRICT_FSQRT.
     Action = TLI.getStrictFPOperationAction(Node->getOpcode(),
                                             Node->getValueType(0));
     break;
   case ISD::ADD:
for (unsigned i = 0; i < NumElems; ++i) {
 
     // Now process the remaining operands.
     for (unsigned j = 1; j < NumOpers; ++j) {
       SDValue Oper = Op.getOperand(j);
       EVT OperVT = Oper.getValueType();
 
       if (OperVT.isVector())
         Oper = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
-                           EltVT, Oper, Idx);
+                           OperVT.getVectorElementType(), Oper, Idx);
 
       Opers.push_back(Oper);
     }
 
     SDValue ScalarOp = DAG.getNode(Op->getOpcode(), dl, ValueVTs, Opers);
 
     OpValues.push_back(ScalarOp.getValue(0));
     OpChains.push_back(ScalarOp.getValue(1));

lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp

 #endif
     report_fatal_error("Do not know how to scalarize the result of this "
                        "operator!\n");
 
   case ISD::MERGE_VALUES:      R = ScalarizeVecRes_MERGE_VALUES(N, ResNo);break;
   case ISD::BITCAST:           R = ScalarizeVecRes_BITCAST(N); break;
   case ISD::BUILD_VECTOR:      R = ScalarizeVecRes_BUILD_VECTOR(N); break;
   case ISD::EXTRACT_SUBVECTOR: R = ScalarizeVecRes_EXTRACT_SUBVECTOR(N); break;
+  case ISD::STRICT_FP_ROUND:   R = ScalarizeVecRes_STRICT_FP_ROUND(N); break;
   case ISD::FP_ROUND:          R = ScalarizeVecRes_FP_ROUND(N); break;
   case ISD::FP_ROUND_INREG:    R = ScalarizeVecRes_InregOp(N); break;
   case ISD::FPOWI:             R = ScalarizeVecRes_FPOWI(N); break;
   case ISD::INSERT_VECTOR_ELT: R = ScalarizeVecRes_INSERT_VECTOR_ELT(N); break;
   case ISD::LOAD:           R = ScalarizeVecRes_LOAD(cast<LoadSDNode>(N));break;
   case ISD::SCALAR_TO_VECTOR:  R = ScalarizeVecRes_SCALAR_TO_VECTOR(N); break;
   case ISD::SIGN_EXTEND_INREG: R = ScalarizeVecRes_InregOp(N); break;
   case ISD::VSELECT:           R = ScalarizeVecRes_VSELECT(N); break;
#endif
   case ISD::STRICT_FRINT:
   case ISD::STRICT_FNEARBYINT:
   case ISD::STRICT_FMAXNUM:
   case ISD::STRICT_FMINNUM:
   case ISD::STRICT_FCEIL:
   case ISD::STRICT_FFLOOR:
   case ISD::STRICT_FROUND:
   case ISD::STRICT_FTRUNC:
+  case ISD::STRICT_FP_EXTEND:
     R = ScalarizeVecRes_StrictFPOp(N);
     break;
   case ISD::SMULFIX:
     R = ScalarizeVecRes_SMULFIX(N);
     break;
   }
 
   // If R is null, the sub-method took care of registering the result.

 
 SDValue DAGTypeLegalizer::ScalarizeVecRes_FP_ROUND(SDNode *N) {
   EVT NewVT = N->getValueType(0).getVectorElementType();
   SDValue Op = GetScalarizedVector(N->getOperand(0));
   return DAG.getNode(ISD::FP_ROUND, SDLoc(N),
                      NewVT, Op, N->getOperand(1));
 }
 
+SDValue DAGTypeLegalizer::ScalarizeVecRes_STRICT_FP_ROUND(SDNode *N) {
+  EVT NewVT = N->getValueType(0).getVectorElementType();
+  SDValue Op = GetScalarizedVector(N->getOperand(1));
+  SDValue Res = DAG.getNode(ISD::STRICT_FP_ROUND, SDLoc(N),
+                            { NewVT, MVT::Other }, 
+                            { N->getOperand(0), Op, N->getOperand(2) });
+  // Legalize the chain result - switch anything that used the old chain to
+  // use the new one.
+  ReplaceValueWith(SDValue(N, 1), Res.getValue(1));
+  return Res;
+}
+
 SDValue DAGTypeLegalizer::ScalarizeVecRes_FPOWI(SDNode *N) {
   SDValue Op = GetScalarizedVector(N->getOperand(0));
   return DAG.getNode(ISD::FPOWI, SDLoc(N),
                      Op.getValueType(), Op, N->getOperand(1));
 }
 
 SDValue DAGTypeLegalizer::ScalarizeVecRes_INSERT_VECTOR_ELT(SDNode *N) {
   // The value to insert may have a wider type than the vector element type,
case ISD::VSELECT:
       Res = ScalarizeVecOp_VSELECT(N);
       break;
     case ISD::SETCC:
       Res = ScalarizeVecOp_VSETCC(N);
       break;
     case ISD::STORE:
       Res = ScalarizeVecOp_STORE(cast<StoreSDNode>(N), OpNo);
       break;
+    case ISD::STRICT_FP_ROUND:
+      Res = ScalarizeVecOp_STRICT_FP_ROUND(N, OpNo);
+      break;
     case ISD::FP_ROUND:
       Res = ScalarizeVecOp_FP_ROUND(N, OpNo);
       break;
     }
   }
 
   // If the result is null, the sub-method took care of registering results etc.
   if (!Res.getNode()) return false;

 SDValue DAGTypeLegalizer::ScalarizeVecOp_FP_ROUND(SDNode *N, unsigned OpNo) {
   SDValue Elt = GetScalarizedVector(N->getOperand(0));
   SDValue Res = DAG.getNode(ISD::FP_ROUND, SDLoc(N),
                             N->getValueType(0).getVectorElementType(), Elt,
                             N->getOperand(1));
   return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(N), N->getValueType(0), Res);
 }
 
+SDValue DAGTypeLegalizer::ScalarizeVecOp_STRICT_FP_ROUND(SDNode *N, 
+                                                         unsigned OpNo) {
+  SDValue Elt = GetScalarizedVector(N->getOperand(1));
+  SDValue Res = DAG.getNode(ISD::STRICT_FP_ROUND, SDLoc(N),
+                            { N->getValueType(0).getVectorElementType(), MVT::Other },
+                            { N->getOperand(0), Elt, N->getOperand(2) });
+  // Legalize the chain result - switch anything that used the old chain to
+  // use the new one.
+  ReplaceValueWith(SDValue(N, 1), Res.getValue(1));
+  return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(N), N->getValueType(0), Res);
+} 
+
 //===----------------------------------------------------------------------===//
 //  Result Vector Splitting
 //===----------------------------------------------------------------------===//
 
 /// This method is called when the specified result of the specified node is
 /// found to need vector splitting. At this point, the node may also have
 /// invalid operands or may have other results that need legalization, we just
 /// know that (at least) one result needs vector splitting.
#endif
   case ISD::FEXP2:
   case ISD::FFLOOR:
   case ISD::FLOG:
   case ISD::FLOG10:
   case ISD::FLOG2:
   case ISD::FNEARBYINT:
   case ISD::FNEG:
   case ISD::FP_EXTEND:
+  case ISD::STRICT_FP_EXTEND:
   case ISD::FP_ROUND:
+  case ISD::STRICT_FP_ROUND:
   case ISD::FP_TO_SINT:
   case ISD::FP_TO_UINT:
   case ISD::FRINT:
   case ISD::FROUND:
   case ISD::FSIN:
   case ISD::FSQRT:
   case ISD::FTRUNC:
   case ISD::SINT_TO_FP:
void DAGTypeLegalizer::SplitVecRes_UnaryOp(SDNode *N, SDValue &Lo,
                                            SDValue &Hi) {
   // Get the dest types - they may not match the input types, e.g. int_to_fp.
   EVT LoVT, HiVT;
   SDLoc dl(N);
   std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
 
   // If the input also splits, handle it directly for a compile time speedup.
   // Otherwise split it by hand.
-  EVT InVT = N->getOperand(0).getValueType();
+  EVT InVT = N->getOperand(N->isStrictFPOpcode()).getValueType();
   if (getTypeAction(InVT) == TargetLowering::TypeSplitVector)
-    GetSplitVector(N->getOperand(0), Lo, Hi);
+    GetSplitVector(N->getOperand(N->isStrictFPOpcode()), Lo, Hi);
   else
-    std::tie(Lo, Hi) = DAG.SplitVectorOperand(N, 0);
+    std::tie(Lo, Hi) = DAG.SplitVectorOperand(N, N->isStrictFPOpcode());
 
   if (N->getOpcode() == ISD::FP_ROUND) {
     Lo = DAG.getNode(N->getOpcode(), dl, LoVT, Lo, N->getOperand(1));
     Hi = DAG.getNode(N->getOpcode(), dl, HiVT, Hi, N->getOperand(1));
+  } else if (N->getOpcode() == ISD::STRICT_FP_ROUND) {
+    Lo = DAG.getNode(N->getOpcode(), dl, { LoVT, MVT::Other }, 
+                     { N->getOperand(0), Lo, N->getOperand(2) });
+    Hi = DAG.getNode(N->getOpcode(), dl, { HiVT, MVT::Other }, 
+                     { N->getOperand(0), Hi, N->getOperand(2) });
+    SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 
+                                   Lo.getValue(1), Hi.getValue(1));
+    ReplaceValueWith(SDValue(N, 1), NewChain);
+  } else if (N->isStrictFPOpcode()) {
+    Lo = DAG.getNode(N->getOpcode(), dl, { LoVT, MVT::Other }, 
+                     { N->getOperand(0), Lo });
+    Hi = DAG.getNode(N->getOpcode(), dl, { HiVT, MVT::Other }, 
+                     { N->getOperand(0), Hi });
+    // Legalize the chain result - switch anything that used the old chain to
+    // use the new one.
+    SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 
+                                   Lo.getValue(1), Hi.getValue(1));
+    ReplaceValueWith(SDValue(N, 1), NewChain);
   } else {
     Lo = DAG.getNode(N->getOpcode(), dl, LoVT, Lo);
     Hi = DAG.getNode(N->getOpcode(), dl, HiVT, Hi);
   }
 }
 
 void DAGTypeLegalizer::SplitVecRes_ExtendOp(SDNode *N, SDValue &Lo,
                                             SDValue &Hi) {
#endif
     case ISD::SETCC:             Res = SplitVecOp_VSETCC(N); break;
     case ISD::BITCAST:           Res = SplitVecOp_BITCAST(N); break;
     case ISD::EXTRACT_SUBVECTOR: Res = SplitVecOp_EXTRACT_SUBVECTOR(N); break;
     case ISD::EXTRACT_VECTOR_ELT:Res = SplitVecOp_EXTRACT_VECTOR_ELT(N); break;
     case ISD::CONCAT_VECTORS:    Res = SplitVecOp_CONCAT_VECTORS(N); break;
     case ISD::TRUNCATE:
       Res = SplitVecOp_TruncateHelper(N);
       break;
+    case ISD::STRICT_FP_ROUND:
     case ISD::FP_ROUND:          Res = SplitVecOp_FP_ROUND(N); break;
     case ISD::FCOPYSIGN:         Res = SplitVecOp_FCOPYSIGN(N); break;
     case ISD::STORE:
       Res = SplitVecOp_STORE(cast<StoreSDNode>(N), OpNo);
       break;
     case ISD::MSTORE:
       Res = SplitVecOp_MSTORE(cast<MaskedStoreSDNode>(N), OpNo);
       break;
case ISD::UINT_TO_FP:
       else
         Res = SplitVecOp_UnaryOp(N);
       break;
     case ISD::FP_TO_SINT:
     case ISD::FP_TO_UINT:
     case ISD::CTTZ:
     case ISD::CTLZ:
     case ISD::CTPOP:
+    case ISD::STRICT_FP_EXTEND:
     case ISD::FP_EXTEND:
     case ISD::SIGN_EXTEND:
     case ISD::ZERO_EXTEND:
     case ISD::ANY_EXTEND:
     case ISD::FTRUNC:
     case ISD::FCANONICALIZE:
       Res = SplitVecOp_UnaryOp(N);
       break;
#endif
   // If the result is null, the sub-method took care of registering results etc.
   if (!Res.getNode()) return false;
 
   // If the result is N, the sub-method updated N in place.  Tell the legalizer
   // core about this.
   if (Res.getNode() == N)
     return true;
 
+  if (N->isStrictFPOpcode())
+    assert(Res.getValueType() == N->getValueType(0) && N->getNumValues() == 2 &&
+           "Invalid operand expansion");
+  else
     assert(Res.getValueType() == N->getValueType(0) && N->getNumValues() == 1 &&
          "Invalid operand expansion");
 
   ReplaceValueWith(SDValue(N, 0), Res);
   return false;
 }
 
 SDValue DAGTypeLegalizer::SplitVecOp_VSELECT(SDNode *N, unsigned OpNo) {
   // The only possibility for an illegal operand is the mask, since result type
SDValue DAGTypeLegalizer::SplitVecOp_VECREDUCE(SDNode *N, unsigned OpNo) {
   return DAG.getNode(N->getOpcode(), dl, ResVT, Partial, N->getFlags());
 }
 
 SDValue DAGTypeLegalizer::SplitVecOp_UnaryOp(SDNode *N) {
   // The result has a legal vector type, but the input needs splitting.
   EVT ResVT = N->getValueType(0);
   SDValue Lo, Hi;
   SDLoc dl(N);
-  GetSplitVector(N->getOperand(0), Lo, Hi);
+  GetSplitVector(N->getOperand(N->isStrictFPOpcode()), Lo, Hi);
   EVT InVT = Lo.getValueType();
 
   EVT OutVT = EVT::getVectorVT(*DAG.getContext(), ResVT.getVectorElementType(),
                                InVT.getVectorNumElements());
 
+  if (N->isStrictFPOpcode()) {
+    Lo = DAG.getNode(N->getOpcode(), dl, { OutVT, MVT::Other }, 
+                     { N->getOperand(0), Lo });
+    Hi = DAG.getNode(N->getOpcode(), dl, { OutVT, MVT::Other }, 
+                     { N->getOperand(0), Hi });
+
+    // Build a factor node to remember that this load is independent of the
+    // other one.
+    SDValue Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
+                             Hi.getValue(1));
+  
+    // Legalize the chain result - switch anything that used the old chain to
+    // use the new one.
+    ReplaceValueWith(SDValue(N, 1), Ch);
+  } else {
     Lo = DAG.getNode(N->getOpcode(), dl, OutVT, Lo);
     Hi = DAG.getNode(N->getOpcode(), dl, OutVT, Hi);
+  }
 
   return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
 }
 
 SDValue DAGTypeLegalizer::SplitVecOp_BITCAST(SDNode *N) {
   // For example, i64 = BITCAST v4i16 on alpha.  Typically the vector will
   // end up being split all the way down to individual components.  Convert the
   // split pieces into integers and reassemble.

 }
 
 
 SDValue DAGTypeLegalizer::SplitVecOp_FP_ROUND(SDNode *N) {
   // The result has a legal vector type, but the input needs splitting.
   EVT ResVT = N->getValueType(0);
   SDValue Lo, Hi;
   SDLoc DL(N);
-  GetSplitVector(N->getOperand(0), Lo, Hi);
+  GetSplitVector(N->getOperand(N->isStrictFPOpcode()), Lo, Hi);
   EVT InVT = Lo.getValueType();
 
   EVT OutVT = EVT::getVectorVT(*DAG.getContext(), ResVT.getVectorElementType(),
                                InVT.getVectorNumElements());
 
+  if (N->isStrictFPOpcode()) {
+    Lo = DAG.getNode(N->getOpcode(), DL, { OutVT, MVT::Other }, 
+                     { N->getOperand(0), Lo, N->getOperand(2) });
+    Hi = DAG.getNode(N->getOpcode(), DL, { OutVT, MVT::Other }, 
+                     { N->getOperand(0), Hi, N->getOperand(2) });
+    // Legalize the chain result - switch anything that used the old chain to
+    // use the new one.
+    SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, 
+                                   Lo.getValue(1), Hi.getValue(1));
+    ReplaceValueWith(SDValue(N, 1), NewChain);
+  } else {
     Lo = DAG.getNode(ISD::FP_ROUND, DL, OutVT, Lo, N->getOperand(1));
     Hi = DAG.getNode(ISD::FP_ROUND, DL, OutVT, Hi, N->getOperand(1));
+  }
 
   return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResVT, Lo, Hi);
 }
 
 SDValue DAGTypeLegalizer::SplitVecOp_FCOPYSIGN(SDNode *N) {
   // The result (and the first input) has a legal vector type, but the second
   // input needs splitting.
   return DAG.UnrollVectorOp(N, N->getValueType(0).getVectorNumElements());
#endif
   case ISD::SIGN_EXTEND:
   case ISD::SINT_TO_FP:
   case ISD::TRUNCATE:
   case ISD::UINT_TO_FP:
   case ISD::ZERO_EXTEND:
     Res = WidenVecRes_Convert(N);
     break;
 
+  case ISD::STRICT_FP_EXTEND:
+  case ISD::STRICT_FP_ROUND:
+    Res = WidenVecRes_Convert_StrictFP(N);
+    break;
+
   case ISD::FABS:
   case ISD::FCEIL:
   case ISD::FCOS:
   case ISD::FEXP:
   case ISD::FEXP2:
   case ISD::FFLOOR:
   case ISD::FLOG:
   case ISD::FLOG10:
if (N->getNumOperands() == 1)
       Ops[i] = DAG.getNode(Opcode, DL, EltVT, Val);
     else
       Ops[i] = DAG.getNode(Opcode, DL, EltVT, Val, N->getOperand(1), Flags);
   }
 
   return DAG.getBuildVector(WidenVT, DL, Ops);
 }
 
+SDValue DAGTypeLegalizer::WidenVecRes_Convert_StrictFP(SDNode *N) {
+  SDValue InOp = N->getOperand(1);
+  SDLoc DL(N);
+  SmallVector<SDValue, 4> NewOps(N->op_begin(), N->op_end());
+
+  EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+  unsigned WidenNumElts = WidenVT.getVectorNumElements();
+  SmallVector<EVT, 2> WidenVTs = { WidenVT, MVT::Other };
+
+  EVT InVT = InOp.getValueType();
+  EVT InEltVT = InVT.getVectorElementType();
+  EVT InWidenVT = EVT::getVectorVT(*DAG.getContext(), InEltVT, WidenNumElts);
+  SmallVector<EVT, 2> InWidenVTs = { InWidenVT, MVT::Other };
+
+  unsigned Opcode = N->getOpcode();
+  unsigned InVTNumElts = InVT.getVectorNumElements();
+  if (getTypeAction(InVT) == TargetLowering::TypeWidenVector) {
+    InOp = GetWidenedVector(N->getOperand(1));
+    InVT = InOp.getValueType();
+    InVTNumElts = InVT.getVectorNumElements();
+    if (InVTNumElts == WidenNumElts) {
+      NewOps[1] = InOp;
+      SDValue Result = DAG.getNode(Opcode, DL, WidenVTs, NewOps);
+      ReplaceValueWith(SDValue(N, 1), Result.getValue(1)); // Chain
+      return Result;
+    }
+  }
+
+  if (TLI.isTypeLegal(InWidenVT)) {
+    // Because the result and the input are different vector types, widening
+    // the result could create a legal type but widening the input might make
+    // it an illegal type that might lead to repeatedly splitting the input
+    // and then widening it. To avoid this, we widen the input only if
+    // it results in a legal type.
+    if (WidenNumElts % InVTNumElts == 0) {
+      // Widen the input and call convert on the widened input vector.
+      unsigned NumConcat = WidenNumElts/InVTNumElts;
+      SmallVector<SDValue, 16> Ops(NumConcat, DAG.getUNDEF(InVT));
+      Ops[0] = InOp;
+      SDValue InVec = DAG.getNode(ISD::CONCAT_VECTORS, DL, InWidenVT, Ops);
+
+      NewOps[1] = InVec;
+      SDValue Result = DAG.getNode(Opcode, DL, WidenVTs, NewOps);
+      ReplaceValueWith(SDValue(N, 1), Result.getValue(1)); // Chain
+      return Result;
+    }
+
+    if (InVTNumElts % WidenNumElts == 0) {
+      NewOps[1] = DAG.getNode(
+          ISD::EXTRACT_SUBVECTOR, DL, InWidenVT, InOp,
+          DAG.getConstant(0, DL, TLI.getVectorIdxTy(DAG.getDataLayout())));
+      // Extract the input and convert the shorten input vector.
+      SDValue Result = DAG.getNode(Opcode, DL, WidenVTs, NewOps);
+      ReplaceValueWith(SDValue(N, 1), Result.getValue(1)); // Chain
+      return Result;
+    }
+  }
+
+  // Otherwise unroll into some nasty scalar code and rebuild the vector.
+  EVT EltVT = WidenVT.getVectorElementType();
+  SmallVector<EVT, 2> EltVTs = { EltVT, MVT::Other };
+  SmallVector<SDValue, 16> Ops(WidenNumElts, DAG.getUNDEF(EltVT));
+  SmallVector<SDValue, 32> OpChains;
+  // Use the original element count so we don't do more scalar opts than
+  // necessary.
+  unsigned MinElts = N->getValueType(0).getVectorNumElements();
+  for (unsigned i=0; i < MinElts; ++i) {
+    NewOps[1] = DAG.getNode(
+        ISD::EXTRACT_VECTOR_ELT, DL, InEltVT, InOp,
+        DAG.getConstant(i, DL, TLI.getVectorIdxTy(DAG.getDataLayout())));
+    Ops[i] = DAG.getNode(Opcode, DL, EltVTs, NewOps);
+    OpChains.push_back(Ops[i].getValue(1));
+  }
+  SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, OpChains);
+  ReplaceValueWith(SDValue(N, 1), NewChain);
+
+  return DAG.getBuildVector(WidenVT, DL, Ops);
+}
+
 SDValue DAGTypeLegalizer::WidenVecRes_EXTEND_VECTOR_INREG(SDNode *N) {
   unsigned Opcode = N->getOpcode();
   SDValue InOp = N->getOperand(0);
   SDLoc DL(N);
 
   EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
   EVT WidenSVT = WidenVT.getVectorElementType();
   unsigned WidenNumElts = WidenVT.getVectorNumElements();
#endif
 
   case ISD::ANY_EXTEND:
   case ISD::SIGN_EXTEND:
   case ISD::ZERO_EXTEND:
     Res = WidenVecOp_EXTEND(N);
     break;
 
   case ISD::FP_EXTEND:
+  case ISD::STRICT_FP_EXTEND:
   case ISD::FP_TO_SINT:
   case ISD::FP_TO_UINT:
   case ISD::SINT_TO_FP:
   case ISD::UINT_TO_FP:
   case ISD::TRUNCATE:
     Res = WidenVecOp_Convert(N);
     break;
   }
 
   // If Res is null, the sub-method took care of registering the result.
   if (!Res.getNode()) return false;
 
   // If the result is N, the sub-method updated N in place.  Tell the legalizer
   // core about this.
   if (Res.getNode() == N)
     return true;
 
 
+  if (N->isStrictFPOpcode())
+    assert(Res.getValueType() == N->getValueType(0) && N->getNumValues() == 2 &&
+           "Invalid operand expansion");
+  else
     assert(Res.getValueType() == N->getValueType(0) && N->getNumValues() == 1 &&
            "Invalid operand expansion");
 
   ReplaceValueWith(SDValue(N, 0), Res);
   return false;
 }
 
 SDValue DAGTypeLegalizer::WidenVecOp_EXTEND(SDNode *N) {
   SDLoc DL(N);
   EVT VT = N->getValueType(0);

 }
 
 SDValue DAGTypeLegalizer::WidenVecOp_Convert(SDNode *N) {
   // Since the result is legal and the input is illegal.
   EVT VT = N->getValueType(0);
   EVT EltVT = VT.getVectorElementType();
   SDLoc dl(N);
   unsigned NumElts = VT.getVectorNumElements();
-  SDValue InOp = N->getOperand(0);
+  SDValue InOp = N->getOperand(N->isStrictFPOpcode());
   assert(getTypeAction(InOp.getValueType()) ==
              TargetLowering::TypeWidenVector &&
          "Unexpected type action");
   InOp = GetWidenedVector(InOp);
   EVT InVT = InOp.getValueType();
   unsigned Opcode = N->getOpcode();
 
   // See if a widened result type would be legal, if so widen the node.
   EVT WideVT = EVT::getVectorVT(*DAG.getContext(), EltVT,
                                 InVT.getVectorNumElements());
   if (TLI.isTypeLegal(WideVT)) {
-    SDValue Res = DAG.getNode(Opcode, dl, WideVT, InOp);
+    SDValue Res;
+    if (N->isStrictFPOpcode()) {
+      Res = DAG.getNode(Opcode, dl, { WideVT, MVT::Other }, 
+                        { N->getOperand(0), InOp });
+      // Legalize the chain result - switch anything that used the old chain to
+      // use the new one.
+      ReplaceValueWith(SDValue(N, 1), Res.getValue(1));
+    } else
+      Res = DAG.getNode(Opcode, dl, WideVT, InOp);
     return DAG.getNode(
         ISD::EXTRACT_SUBVECTOR, dl, VT, Res,
         DAG.getConstant(0, dl, TLI.getVectorIdxTy(DAG.getDataLayout())));
   }
 
   EVT InEltVT = InVT.getVectorElementType();
 
   // Unroll the convert into some scalar code and create a nasty build vector.
   SmallVector<SDValue, 16> Ops(NumElts);
+  if (N->isStrictFPOpcode()) {
+    SmallVector<SDValue, 4> NewOps(N->op_begin(), N->op_end());
+    SmallVector<SDValue, 32> OpChains;
+    for (unsigned i=0; i < NumElts; ++i) {
+      NewOps[1] = DAG.getNode(
+          ISD::EXTRACT_VECTOR_ELT, dl, InEltVT, InOp,
+          DAG.getConstant(i, dl, TLI.getVectorIdxTy(DAG.getDataLayout())));
+      Ops[i] = DAG.getNode(Opcode, dl, { EltVT, MVT::Other }, NewOps);
+      OpChains.push_back(Ops[i].getValue(1));
+    }
+    SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OpChains);
+    ReplaceValueWith(SDValue(N, 1), NewChain);
+  } else
   for (unsigned i=0; i < NumElts; ++i)
     Ops[i] = DAG.getNode(
         Opcode, dl, EltVT,
         DAG.getNode(
             ISD::EXTRACT_VECTOR_ELT, dl, InEltVT, InOp,
             DAG.getConstant(i, dl, TLI.getVectorIdxTy(DAG.getDataLayout()))));
 
   return DAG.getBuildVector(VT, dl, Ops);
 }

lib/CodeGen/SelectionDAG/SelectionDAG.cpp

case ISD::STRICT_FNEARBYINT:
     IsUnary = true;
     break;
   case ISD::STRICT_FMAXNUM: NewOpc = ISD::FMAXNUM; break;
   case ISD::STRICT_FMINNUM: NewOpc = ISD::FMINNUM; break;
   case ISD::STRICT_FCEIL: NewOpc = ISD::FCEIL; IsUnary = true; break;
   case ISD::STRICT_FFLOOR: NewOpc = ISD::FFLOOR; IsUnary = true; break;
   case ISD::STRICT_FROUND: NewOpc = ISD::FROUND; IsUnary = true; break;
   case ISD::STRICT_FTRUNC: NewOpc = ISD::FTRUNC; IsUnary = true; break;
+  case ISD::STRICT_FP_ROUND: NewOpc = ISD::FP_ROUND; break;
+  case ISD::STRICT_FP_EXTEND: NewOpc = ISD::FP_EXTEND; IsUnary = true; break;
   }
 
   // We're taking this node out of the chain, so we need to re-link things.
   SDValue InputChain = Node->getOperand(0);
   SDValue OutputChain = SDValue(Node, 1);
   ReplaceAllUsesOfValueWith(OutputChain, InputChain);
 
-  SDVTList VTs = getVTList(Node->getOperand(1).getValueType());
+  SDVTList VTs;
   SDNode *Res = nullptr;
+
+  switch (OrigOpc) {
+  default:
+    VTs = getVTList(Node->getOperand(1).getValueType());
+    break;
+  case ISD::STRICT_FP_ROUND:
+  case ISD::STRICT_FP_EXTEND:
+    VTs = getVTList(Node->ValueList[0]);
+    break;
+  }
+
   if (IsUnary)
     Res = MorphNodeTo(Node, NewOpc, VTs, { Node->getOperand(1) });
   else if (IsTernary)
     Res = MorphNodeTo(Node, NewOpc, VTs, { Node->getOperand(1),
                                            Node->getOperand(2),
                                            Node->getOperand(3)});
   else
     Res = MorphNodeTo(Node, NewOpc, VTs, { Node->getOperand(1),

lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

setValue(&I, DAG.getNode(ISD::FMA, sdl,
                              getValue(I.getArgOperand(2))));
     return nullptr;
   case Intrinsic::experimental_constrained_fadd:
   case Intrinsic::experimental_constrained_fsub:
   case Intrinsic::experimental_constrained_fmul:
   case Intrinsic::experimental_constrained_fdiv:
   case Intrinsic::experimental_constrained_frem:
   case Intrinsic::experimental_constrained_fma:
+  case Intrinsic::experimental_constrained_fptrunc:
+  case Intrinsic::experimental_constrained_fpext:
   case Intrinsic::experimental_constrained_sqrt:
   case Intrinsic::experimental_constrained_pow:
   case Intrinsic::experimental_constrained_powi:
   case Intrinsic::experimental_constrained_sin:
   case Intrinsic::experimental_constrained_cos:
   case Intrinsic::experimental_constrained_exp:
   case Intrinsic::experimental_constrained_exp2:
   case Intrinsic::experimental_constrained_log:
case Intrinsic::experimental_constrained_fdiv:
     Opcode = ISD::STRICT_FDIV;
     break;
   case Intrinsic::experimental_constrained_frem:
     Opcode = ISD::STRICT_FREM;
     break;
   case Intrinsic::experimental_constrained_fma:
     Opcode = ISD::STRICT_FMA;
     break;
+  case Intrinsic::experimental_constrained_fptrunc:
+    Opcode = ISD::STRICT_FP_ROUND;
+    break;
+  case Intrinsic::experimental_constrained_fpext:
+    Opcode = ISD::STRICT_FP_EXTEND;
+    break;
   case Intrinsic::experimental_constrained_sqrt:
     Opcode = ISD::STRICT_FSQRT;
     break;
   case Intrinsic::experimental_constrained_pow:
     Opcode = ISD::STRICT_FPOW;
     break;
   case Intrinsic::experimental_constrained_powi:
     Opcode = ISD::STRICT_FPOWI;
void SelectionDAGBuilder::visitConstrainedFPIntrinsic(
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
   SDValue Chain = getRoot();
   SmallVector<EVT, 4> ValueVTs;
   ComputeValueVTs(TLI, DAG.getDataLayout(), FPI.getType(), ValueVTs);
   ValueVTs.push_back(MVT::Other); // Out chain
 
   SDVTList VTs = DAG.getVTList(ValueVTs);
   SDValue Result;
-  if (FPI.isUnaryOp())
+  if (Opcode == ISD::STRICT_FP_ROUND)
+    Result = DAG.getNode(Opcode, sdl, VTs, 
+                          { Chain, getValue(FPI.getArgOperand(0)), 
+                               DAG.getTargetConstant(0, sdl,
+                               TLI.getPointerTy(DAG.getDataLayout())) });
+  else if (FPI.isUnaryOp())
     Result = DAG.getNode(Opcode, sdl, VTs,
                          { Chain, getValue(FPI.getArgOperand(0)) });
   else if (FPI.isTernaryOp())
     Result = DAG.getNode(Opcode, sdl, VTs,
                          { Chain, getValue(FPI.getArgOperand(0)),
                                   getValue(FPI.getArgOperand(1)),
                                   getValue(FPI.getArgOperand(2)) });
   else

lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp

#endif
   case ISD::ZERO_EXTEND:                return "zero_extend";
   case ISD::ANY_EXTEND:                 return "any_extend";
   case ISD::SIGN_EXTEND_INREG:          return "sign_extend_inreg";
   case ISD::ANY_EXTEND_VECTOR_INREG:    return "any_extend_vector_inreg";
   case ISD::SIGN_EXTEND_VECTOR_INREG:   return "sign_extend_vector_inreg";
   case ISD::ZERO_EXTEND_VECTOR_INREG:   return "zero_extend_vector_inreg";
   case ISD::TRUNCATE:                   return "truncate";
   case ISD::FP_ROUND:                   return "fp_round";
+  case ISD::STRICT_FP_ROUND:            return "strict_fp_round";
   case ISD::FLT_ROUNDS_:                return "flt_rounds";
   case ISD::FP_ROUND_INREG:             return "fp_round_inreg";
   case ISD::FP_EXTEND:                  return "fp_extend";
+  case ISD::STRICT_FP_EXTEND:           return "strict_fp_extend";
 
   case ISD::SINT_TO_FP:                 return "sint_to_fp";
   case ISD::UINT_TO_FP:                 return "uint_to_fp";
   case ISD::FP_TO_SINT:                 return "fp_to_sint";
   case ISD::FP_TO_UINT:                 return "fp_to_uint";
   case ISD::BITCAST:                    return "bitcast";
   case ISD::ADDRSPACECAST:              return "addrspacecast";
   case ISD::FP16_TO_FP:                 return "fp16_to_fp";

lib/IR/IntrinsicInst.cpp

return StringSwitch<ExceptionBehavior>(ExceptionArg)
     .Case("fpexcept.strict",  ebStrict)
     .Default(ebInvalid);
 }
 
 bool ConstrainedFPIntrinsic::isUnaryOp() const {
   switch (getIntrinsicID()) {
     default:
       return false;
+    case Intrinsic::experimental_constrained_fptrunc:
+    case Intrinsic::experimental_constrained_fpext:
     case Intrinsic::experimental_constrained_sqrt:
     case Intrinsic::experimental_constrained_sin:
     case Intrinsic::experimental_constrained_cos:
     case Intrinsic::experimental_constrained_exp:
     case Intrinsic::experimental_constrained_exp2:
     case Intrinsic::experimental_constrained_log:
     case Intrinsic::experimental_constrained_log10:
     case Intrinsic::experimental_constrained_log2:

lib/IR/Verifier.cpp

Assert(isa<ConstantInt>(Call.getArgOperand(1)),
            Call);
     break;
   case Intrinsic::experimental_constrained_fadd:
   case Intrinsic::experimental_constrained_fsub:
   case Intrinsic::experimental_constrained_fmul:
   case Intrinsic::experimental_constrained_fdiv:
   case Intrinsic::experimental_constrained_frem:
   case Intrinsic::experimental_constrained_fma:
+  case Intrinsic::experimental_constrained_fptrunc:
+  case Intrinsic::experimental_constrained_fpext:
   case Intrinsic::experimental_constrained_sqrt:
   case Intrinsic::experimental_constrained_pow:
   case Intrinsic::experimental_constrained_powi:
   case Intrinsic::experimental_constrained_sin:
   case Intrinsic::experimental_constrained_cos:
   case Intrinsic::experimental_constrained_exp:
   case Intrinsic::experimental_constrained_exp2:
   case Intrinsic::experimental_constrained_log:
static DISubprogram *getSubprogram(Metadata *LocalScope) {
 
   // Just return null; broken scope chains are checked elsewhere.
   assert(!isa<DILocalScope>(LocalScope) && "Unknown type of local scope");
   return nullptr;
 }
 
 void Verifier::visitConstrainedFPIntrinsic(ConstrainedFPIntrinsic &FPI) {
   unsigned NumOperands = FPI.getNumArgOperands();
-  Assert(((NumOperands == 5 && FPI.isTernaryOp()) ||
-          (NumOperands == 3 && FPI.isUnaryOp()) || (NumOperands == 4)),
+  bool HasExceptionMD = false;
+  bool HasRoundingMD = false;
+  switch (FPI.getIntrinsicID())
+  {
+  case Intrinsic::experimental_constrained_sqrt:
+  case Intrinsic::experimental_constrained_sin:
+  case Intrinsic::experimental_constrained_cos:
+  case Intrinsic::experimental_constrained_exp:
+  case Intrinsic::experimental_constrained_exp2:
+  case Intrinsic::experimental_constrained_log:
+  case Intrinsic::experimental_constrained_log10:
+  case Intrinsic::experimental_constrained_log2:
+  case Intrinsic::experimental_constrained_rint:
+  case Intrinsic::experimental_constrained_nearbyint:
+  case Intrinsic::experimental_constrained_ceil:
+  case Intrinsic::experimental_constrained_floor:
+  case Intrinsic::experimental_constrained_round:
+  case Intrinsic::experimental_constrained_trunc:
+    Assert((NumOperands == 3),
            "invalid arguments for constrained FP intrinsic", &FPI);
+    HasExceptionMD = true;
+    HasRoundingMD = true;
+    break;
+
+  case Intrinsic::experimental_constrained_fma:
+    Assert((NumOperands == 5),
+           "invalid arguments for constrained FP intrinsic", &FPI);
+    HasExceptionMD = true;
+    HasRoundingMD = true;
+    break;
+
+  case Intrinsic::experimental_constrained_fadd:
+  case Intrinsic::experimental_constrained_fsub:
+  case Intrinsic::experimental_constrained_fmul:
+  case Intrinsic::experimental_constrained_fdiv:
+  case Intrinsic::experimental_constrained_frem:
+  case Intrinsic::experimental_constrained_pow:
+  case Intrinsic::experimental_constrained_powi:
+  case Intrinsic::experimental_constrained_maxnum:
+  case Intrinsic::experimental_constrained_minnum:
+    Assert((NumOperands == 4),
+           "invalid arguments for constrained FP intrinsic", &FPI);
+    HasExceptionMD = true;
+    HasRoundingMD = true;
+    break;
+
+  case Intrinsic::experimental_constrained_fptrunc:
+  case Intrinsic::experimental_constrained_fpext: {
+    Assert((NumOperands == 2),
+           "invalid arguments for constrained FP intrinsic", &FPI);
+    HasExceptionMD = true;
+
+    Value *Operand = FPI.getArgOperand(0);
+    uint64_t NumSrcElem = 0;
+    if (Operand->getType()->isVectorTy()) {
+      auto *OperandT = cast<VectorType>(Operand->getType());
+      NumSrcElem = OperandT->getNumElements();
+      Assert(OperandT->getVectorElementType()->isFloatingPointTy(),
+             "Intrinsic first argument vector must be floating point", 
+             &FPI);
+    }
+    else
+      Assert(Operand->getType()->isFloatingPointTy(),
+             "Intrinsic first argument must be floating point", 
+             &FPI);
+
+    Operand = &FPI;
+    Assert((NumSrcElem > 0) == Operand->getType()->isVectorTy(),
+           "Intrinsic first argument and result disagree on vector use",
+           &FPI);
+    if (Operand->getType()->isVectorTy()) {
+      auto *OperandT = cast<VectorType>(Operand->getType());
+      Assert(NumSrcElem == OperandT->getNumElements(),
+             "Intrinsic first argument and result vector lengths must be equal",
+             &FPI);
+      Assert(OperandT->getVectorElementType()->isFloatingPointTy(),
+             "Intrinsic result vector must be floating point", 
+             &FPI);
+    }
+    else
+       Assert(Operand->getType()->isFloatingPointTy(),
+              "Intrinsic result must be an floating point", 
+             &FPI);
+  }
+    break;
+
+  default:
+    llvm_unreachable("Invalid constrained FP intrinsic!");
+  }
+
+  if (HasExceptionMD) {
     Assert(isa<MetadataAsValue>(FPI.getArgOperand(NumOperands-1)),
            "invalid exception behavior argument", &FPI);
-  Assert(isa<MetadataAsValue>(FPI.getArgOperand(NumOperands-2)),
+    Assert(FPI.getExceptionBehavior() != ConstrainedFPIntrinsic::ebInvalid,
+           "invalid exception behavior argument", &FPI);
+  }
+  if (HasRoundingMD) {
+    int RoundingIdx = (HasExceptionMD ? NumOperands - 2 : NumOperands - 1);
+    Assert(isa<MetadataAsValue>(FPI.getArgOperand(RoundingIdx)),
            "invalid rounding mode argument", &FPI);
     Assert(FPI.getRoundingMode() != ConstrainedFPIntrinsic::rmInvalid,
            "invalid rounding mode argument", &FPI);
-  Assert(FPI.getExceptionBehavior() != ConstrainedFPIntrinsic::ebInvalid,
-         "invalid exception behavior argument", &FPI);
+  }
 }
 
 void Verifier::visitDbgIntrinsic(StringRef Kind, DbgVariableIntrinsic &DII) {
   auto *MD = cast<MetadataAsValue>(DII.getArgOperand(0))->getMetadata();
   AssertDI(isa<ValueAsMetadata>(MD) ||
              (isa<MDNode>(MD) && !cast<MDNode>(MD)->getNumOperands()),
          "invalid llvm.dbg." + Kind + " intrinsic address/value", &DII, MD);
   AssertDI(isa<DILocalVariable>(DII.getRawVariable()),

test/CodeGen/X86/fp-intrinsics.ll

entry:
   %rem = call double @llvm.experimental.constrained.frem.f64(
                                                double 1.000000e+00,
                                                double 1.000000e+01,
                                                metadata !"round.dynamic",
                                                metadata !"fpexcept.strict")
   ret double %rem
 }
 
+; Verify that round(42.1) isn't simplified when the rounding mode is
+; unknown.
+; Verify that no gross errors happen.
+; CHECK-LABEL: @f21
+; COMMON: cvtsd2ss
+define float @f21() {
+entry:
+  %result = call float @llvm.experimental.constrained.fptrunc.f32.f64(
+                                               double 42.1,
+                                               metadata !"fpexcept.strict")
+  ret float %result
+}
+
+; Verify that fpext(42.1) isn't simplified when the rounding mode is
+; unknown.
+; Verify that no gross errors happen.
+; CHECK-LABEL: @f22
+; COMMON: cvtss2sd
+define double @f22(float %x) {
+entry:
+  %result = call double @llvm.experimental.constrained.fpext.f64.f32(float %x,
+                                               metadata !"fpexcept.strict")
+  ret double %result
+}
+
 @llvm.fp.env = thread_local global i8 zeroinitializer, section "llvm.metadata"
 declare double @llvm.experimental.constrained.fadd.f64(double, double, metadata, metadata)
 declare double @llvm.experimental.constrained.fsub.f64(double, double, metadata, metadata)
 declare double @llvm.experimental.constrained.fmul.f64(double, double, metadata, metadata)
 declare double @llvm.experimental.constrained.fdiv.f64(double, double, metadata, metadata)
 declare double @llvm.experimental.constrained.frem.f64(double, double, metadata, metadata)
 declare double @llvm.experimental.constrained.sqrt.f64(double, metadata, metadata)
 declare double @llvm.experimental.constrained.pow.f64(double, double, metadata, metadata)
 declare double @llvm.experimental.constrained.powi.f64(double, i32, metadata, metadata)
 declare double @llvm.experimental.constrained.sin.f64(double, metadata, metadata)
 declare double @llvm.experimental.constrained.cos.f64(double, metadata, metadata)
 declare double @llvm.experimental.constrained.exp.f64(double, metadata, metadata)
 declare double @llvm.experimental.constrained.exp2.f64(double, metadata, metadata)
 declare double @llvm.experimental.constrained.log.f64(double, metadata, metadata)
 declare double @llvm.experimental.constrained.log10.f64(double, metadata, metadata)
 declare double @llvm.experimental.constrained.log2.f64(double, metadata, metadata)
 declare double @llvm.experimental.constrained.rint.f64(double, metadata, metadata)
 declare double @llvm.experimental.constrained.nearbyint.f64(double, metadata, metadata)
 declare float @llvm.experimental.constrained.fma.f32(float, float, float, metadata, metadata)
 declare double @llvm.experimental.constrained.fma.f64(double, double, double, metadata, metadata)
+declare float @llvm.experimental.constrained.fptrunc.f32.f64(double, metadata)
+declare double @llvm.experimental.constrained.fpext.f64.f32(float, metadata)
+

test/CodeGen/X86/vector-constrained-fp-intrinsics.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -O3 -mtriple=x86_64-pc-linux < %s | FileCheck %s
 
 define <1 x float> @constrained_vector_fdiv_v1f32() {
 ; CHECK-LABEL: constrained_vector_fdiv_v1f32:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
 ; CHECK-NEXT:    divss {{.*}}(%rip), %xmm0
 ; CHECK-NEXT:    retq
 entry:
   %div = call <1 x float> @llvm.experimental.constrained.fdiv.v1f32(
%min = call <4 x double> @llvm.experimental.constrained.minnum.v4f64(
                                               double 46.0, double 47.0>,
                                 <4 x double> <double 40.0, double 41.0,
                                               double 42.0, double 43.0>,
                                 metadata !"round.dynamic",
                                 metadata !"fpexcept.strict")
   ret <4 x double> %min
 }
 
+define <1 x float> @constrained_vector_fptrunc_v1f64() {
+; CHECK-LABEL: constrained_vector_fptrunc_v1f64:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
+; CHECK-NEXT:    cvtsd2ss %xmm0, %xmm0
+; CHECK-NEXT:    retq
+entry:
+  %result = call <1 x float> @llvm.experimental.constrained.fptrunc.v1f32.v1f64(
+                                <1 x double><double 42.1>,
+                                metadata !"fpexcept.strict")
+  ret <1 x float> %result
+}
+
+define <2 x float> @constrained_vector_fptrunc_v2f64() {
+; CHECK-LABEL: constrained_vector_fptrunc_v2f64:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
+; CHECK-NEXT:    cvtsd2ss %xmm0, %xmm1
+; CHECK-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
+; CHECK-NEXT:    cvtsd2ss %xmm0, %xmm0
+; CHECK-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
+; CHECK-NEXT:    retq
+entry:
+  %result = call <2 x float> @llvm.experimental.constrained.fptrunc.v2f32.v2f64(
+                                <2 x double><double 42.1, double 42.2>,
+                                metadata !"fpexcept.strict")
+  ret <2 x float> %result
+}
+
+define <3 x float> @constrained_vector_fptrunc_v3f64() {
+; CHECK-LABEL: constrained_vector_fptrunc_v3f64:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
+; CHECK-NEXT:    cvtsd2ss %xmm0, %xmm1
+; CHECK-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
+; CHECK-NEXT:    cvtsd2ss %xmm0, %xmm0
+; CHECK-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
+; CHECK-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero
+; CHECK-NEXT:    cvtsd2ss %xmm1, %xmm1
+; CHECK-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; CHECK-NEXT:    retq
+entry:
+  %result = call <3 x float> @llvm.experimental.constrained.fptrunc.v3f32.v3f64(
+                                <3 x double><double 42.1, double 42.2,
+                                             double 42.3>,
+                                metadata !"fpexcept.strict")
+  ret <3 x float> %result
+}
+
+define <4 x float> @constrained_vector_fptrunc_v4f64() {
+; CHECK-LABEL: constrained_vector_fptrunc_v4f64:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
+; CHECK-NEXT:    cvtsd2ss %xmm0, %xmm0
+; CHECK-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero
+; CHECK-NEXT:    cvtsd2ss %xmm1, %xmm1
+; CHECK-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
+; CHECK-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
+; CHECK-NEXT:    cvtsd2ss %xmm0, %xmm2
+; CHECK-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
+; CHECK-NEXT:    cvtsd2ss %xmm0, %xmm0
+; CHECK-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
+; CHECK-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; CHECK-NEXT:    retq
+entry:
+  %result = call <4 x float> @llvm.experimental.constrained.fptrunc.v4f32.v4f64(
+                                <4 x double><double 42.1, double 42.2,
+                                             double 42.3, double 42.4>,
+                                metadata !"fpexcept.strict")
+  ret <4 x float> %result
+}
+
+define <1 x double> @constrained_vector_fpext_v1f32() {
+; CHECK-LABEL: constrained_vector_fpext_v1f32:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; CHECK-NEXT:    cvtss2sd %xmm0, %xmm0
+; CHECK-NEXT:    retq
+entry:
+  %result = call <1 x double> @llvm.experimental.constrained.fpext.v1f64.v1f32(
+                                <1 x float><float 42.0>,
+                                metadata !"fpexcept.strict")
+  ret <1 x double> %result
+}
+
+define <2 x double> @constrained_vector_fpext_v2f32() {
+; CHECK-LABEL: constrained_vector_fpext_v2f32:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; CHECK-NEXT:    cvtss2sd %xmm0, %xmm1
+; CHECK-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; CHECK-NEXT:    cvtss2sd %xmm0, %xmm0
+; CHECK-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; CHECK-NEXT:    retq
+entry:
+  %result = call <2 x double> @llvm.experimental.constrained.fpext.v2f64.v2f32(
+                                <2 x float><float 42.0, float 43.0>,
+                                metadata !"fpexcept.strict")
+  ret <2 x double> %result
+}
+
+define <3 x double> @constrained_vector_fpext_v3f32() {
+; CHECK-LABEL: constrained_vector_fpext_v3f32:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; CHECK-NEXT:    cvtss2sd %xmm0, %xmm0
+; CHECK-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
+; CHECK-NEXT:    cvtss2sd %xmm1, %xmm1
+; CHECK-NEXT:    movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
+; CHECK-NEXT:    cvtss2sd %xmm2, %xmm2
+; CHECK-NEXT:    movsd %xmm2, -{{[0-9]+}}(%rsp)
+; CHECK-NEXT:    fldl -{{[0-9]+}}(%rsp)
+; CHECK-NEXT:    retq
+entry:
+  %result = call <3 x double> @llvm.experimental.constrained.fpext.v3f64.v3f32(
+                                <3 x float><float 42.0, float 43.0,
+                                            float 44.0>,
+                                metadata !"fpexcept.strict")
+  ret <3 x double> %result
+}
+
+define <4 x double> @constrained_vector_fpext_v4f32() {
+; CHECK-LABEL: constrained_vector_fpext_v4f32:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; CHECK-NEXT:    cvtss2sd %xmm0, %xmm1
+; CHECK-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; CHECK-NEXT:    cvtss2sd %xmm0, %xmm0
+; CHECK-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; CHECK-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
+; CHECK-NEXT:    cvtss2sd %xmm1, %xmm2
+; CHECK-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
+; CHECK-NEXT:    cvtss2sd %xmm1, %xmm1
+; CHECK-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
+; CHECK-NEXT:    retq
+entry:
+  %result = call <4 x double> @llvm.experimental.constrained.fpext.v4f64.v4f32(
+                                <4 x float><float 42.0, float 43.0,
+                                            float 44.0, float 45.0>,
+                                metadata !"fpexcept.strict")
+  ret <4 x double> %result
+}
+
 define <1 x float> @constrained_vector_ceil_v1f32() {
 ; CHECK-LABEL: constrained_vector_ceil_v1f32:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    pushq %rax
 ; CHECK-NEXT:    .cfi_def_cfa_offset 16
 ; CHECK-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
 ; CHECK-NEXT:    callq ceilf
 ; CHECK-NEXT:    popq %rax
 ; CHECK-NEXT:    .cfi_def_cfa_offset 8
 ; CHECK-NEXT:    retq
 entry:
   %ceil = call <1 x float> @llvm.experimental.constrained.ceil.v1f32(
                                <1 x float> <float 1.5>,
                                metadata !"round.dynamic",
                                metadata !"fpexcept.strict")
   ret <1 x float> %ceil
 }
 
 define <2 x double> @constrained_vector_ceil_v2f64() {
 ; CHECK-LABEL: constrained_vector_ceil_v2f64:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    subq $24, %rsp
 ; CHECK-NEXT:    .cfi_def_cfa_offset 32
 ; CHECK-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero

 declare <2 x double> @llvm.experimental.constrained.exp2.v2f64(<2 x double>, metadata, metadata)
 declare <2 x double> @llvm.experimental.constrained.log.v2f64(<2 x double>, metadata, metadata)
 declare <2 x double> @llvm.experimental.constrained.log10.v2f64(<2 x double>, metadata, metadata)
 declare <2 x double> @llvm.experimental.constrained.log2.v2f64(<2 x double>, metadata, metadata)
 declare <2 x double> @llvm.experimental.constrained.rint.v2f64(<2 x double>, metadata, metadata)
 declare <2 x double> @llvm.experimental.constrained.nearbyint.v2f64(<2 x double>, metadata, metadata)
 declare <2 x double> @llvm.experimental.constrained.maxnum.v2f64(<2 x double>, <2 x double>, metadata, metadata)
 declare <2 x double> @llvm.experimental.constrained.minnum.v2f64(<2 x double>, <2 x double>, metadata, metadata)
+declare <2 x float> @llvm.experimental.constrained.fptrunc.v2f32.v2f64(<2 x double>, metadata)
+declare <2 x double> @llvm.experimental.constrained.fpext.v2f64.v2f32(<2 x float>, metadata)
 declare <2 x double> @llvm.experimental.constrained.ceil.v2f64(<2 x double>, metadata, metadata)
 declare <2 x double> @llvm.experimental.constrained.floor.v2f64(<2 x double>, metadata, metadata)
 declare <2 x double> @llvm.experimental.constrained.round.v2f64(<2 x double>, metadata, metadata)
 declare <2 x double> @llvm.experimental.constrained.trunc.v2f64(<2 x double>, metadata, metadata)
 
 ; Scalar width declarations
 declare <1 x float> @llvm.experimental.constrained.fadd.v1f32(<1 x float>, <1 x float>, metadata, metadata)
 declare <1 x float> @llvm.experimental.constrained.fsub.v1f32(<1 x float>, <1 x float>, metadata, metadata)

 declare <1 x float> @llvm.experimental.constrained.exp2.v1f32(<1 x float>, metadata, metadata)
 declare <1 x float> @llvm.experimental.constrained.log.v1f32(<1 x float>, metadata, metadata)
 declare <1 x float> @llvm.experimental.constrained.log10.v1f32(<1 x float>, metadata, metadata)
 declare <1 x float> @llvm.experimental.constrained.log2.v1f32(<1 x float>, metadata, metadata)
 declare <1 x float> @llvm.experimental.constrained.rint.v1f32(<1 x float>, metadata, metadata)
 declare <1 x float> @llvm.experimental.constrained.nearbyint.v1f32(<1 x float>, metadata, metadata)
 declare <1 x float> @llvm.experimental.constrained.maxnum.v1f32(<1 x float>, <1 x float>, metadata, metadata)
 declare <1 x float> @llvm.experimental.constrained.minnum.v1f32(<1 x float>, <1 x float>, metadata, metadata)
+declare <1 x float> @llvm.experimental.constrained.fptrunc.v1f32.v1f64(<1 x double>, metadata)
+declare <1 x double> @llvm.experimental.constrained.fpext.v1f64.v1f32(<1 x float>, metadata)
 declare <1 x float> @llvm.experimental.constrained.ceil.v1f32(<1 x float>, metadata, metadata)
 declare <1 x float> @llvm.experimental.constrained.floor.v1f32(<1 x float>, metadata, metadata)
 declare <1 x float> @llvm.experimental.constrained.round.v1f32(<1 x float>, metadata, metadata)
 declare <1 x float> @llvm.experimental.constrained.trunc.v1f32(<1 x float>, metadata, metadata)
 
 ; Illegal width declarations
 declare <3 x float> @llvm.experimental.constrained.fadd.v3f32(<3 x float>, <3 x float>, metadata, metadata)
 declare <3 x double> @llvm.experimental.constrained.fadd.v3f64(<3 x double>, <3 x double>, metadata, metadata)

 declare <3 x float> @llvm.experimental.constrained.rint.v3f32(<3 x float>, metadata, metadata)
 declare <3 x double> @llvm.experimental.constrained.rint.v3f64(<3 x double>, metadata, metadata)
 declare <3 x float> @llvm.experimental.constrained.nearbyint.v3f32(<3 x float>, metadata, metadata)
 declare <3 x double> @llvm.experimental.constrained.nearbyint.v3f64(<3 x double>, metadata, metadata)
 declare <3 x float> @llvm.experimental.constrained.maxnum.v3f32(<3 x float>, <3 x float>, metadata, metadata)
 declare <3 x double> @llvm.experimental.constrained.maxnum.v3f64(<3 x double>, <3 x double>, metadata, metadata)
 declare <3 x float> @llvm.experimental.constrained.minnum.v3f32(<3 x float>, <3 x float>, metadata, metadata)
 declare <3 x double> @llvm.experimental.constrained.minnum.v3f64(<3 x double>, <3 x double>, metadata, metadata)
+declare <3 x float> @llvm.experimental.constrained.fptrunc.v3f32.v3f64(<3 x double>, metadata)
+declare <3 x double> @llvm.experimental.constrained.fpext.v3f64.v3f32(<3 x float>, metadata)
 declare <3 x float> @llvm.experimental.constrained.ceil.v3f32(<3 x float>, metadata, metadata)
 declare <3 x double> @llvm.experimental.constrained.ceil.v3f64(<3 x double>, metadata, metadata)
 declare <3 x float> @llvm.experimental.constrained.floor.v3f32(<3 x float>, metadata, metadata)
 declare <3 x double> @llvm.experimental.constrained.floor.v3f64(<3 x double>, metadata, metadata)
 declare <3 x float> @llvm.experimental.constrained.round.v3f32(<3 x float>, metadata, metadata)
 declare <3 x double> @llvm.experimental.constrained.round.v3f64(<3 x double>, metadata, metadata)
 declare <3 x float> @llvm.experimental.constrained.trunc.v3f32(<3 x float>, metadata, metadata)
 declare <3 x double> @llvm.experimental.constrained.trunc.v3f64(<3 x double>, metadata, metadata)

 declare <4 x double> @llvm.experimental.constrained.exp2.v4f64(<4 x double>, metadata, metadata)
 declare <4 x double> @llvm.experimental.constrained.log.v4f64(<4 x double>, metadata, metadata)
 declare <4 x double> @llvm.experimental.constrained.log10.v4f64(<4 x double>, metadata, metadata)
 declare <4 x double> @llvm.experimental.constrained.log2.v4f64(<4 x double>, metadata, metadata)
 declare <4 x double> @llvm.experimental.constrained.rint.v4f64(<4 x double>, metadata, metadata)
 declare <4 x double> @llvm.experimental.constrained.nearbyint.v4f64(<4 x double>, metadata, metadata)
 declare <4 x double> @llvm.experimental.constrained.maxnum.v4f64(<4 x double>, <4 x double>, metadata, metadata)
 declare <4 x double> @llvm.experimental.constrained.minnum.v4f64(<4 x double>, <4 x double>, metadata, metadata)
+declare <4 x float> @llvm.experimental.constrained.fptrunc.v4f32.v4f64(<4 x double>, metadata)
+declare <4 x double> @llvm.experimental.constrained.fpext.v4f64.v4f32(<4 x float>, metadata)
 declare <4 x double> @llvm.experimental.constrained.ceil.v4f64(<4 x double>, metadata, metadata)
 declare <4 x double> @llvm.experimental.constrained.floor.v4f64(<4 x double>, metadata, metadata)
 declare <4 x double> @llvm.experimental.constrained.round.v4f64(<4 x double>, metadata, metadata)
 declare <4 x double> @llvm.experimental.constrained.trunc.v4f64(<4 x double>, metadata, metadata)
-

test/Feature/fp-intrinsics.ll

 define double @f17() {
 entry:
   %result = call double @llvm.experimental.constrained.fma.f64(double 42.1, double 42.1, double 42.1,
                                                metadata !"round.dynamic",
                                                metadata !"fpexcept.strict")
   ret double %result
 }
 
+; Verify that fptrunc(42.1) isn't simplified when the rounding mode is
+; unknown.
+; CHECK-LABEL: f20
+; CHECK: call float @llvm.experimental.constrained.fptrunc
+define float @f20() {
+entry:
+  %result = call float @llvm.experimental.constrained.fptrunc.f32.f64(
+                                               double 42.1,
+                                               metadata !"fpexcept.strict")
+  ret float %result
+}
+
+; Verify that fpext(42.1) isn't simplified when the rounding mode is
+; unknown.
+; CHECK-LABEL: f21
+; CHECK: call double @llvm.experimental.constrained.fpext
+define double @f21() {
+entry:
+  %result = call double @llvm.experimental.constrained.fpext.f64.f32(float 42.0,
+                                               metadata !"fpexcept.strict")
+  ret double %result
+}
+
 @llvm.fp.env = thread_local global i8 zeroinitializer, section "llvm.metadata"
 declare double @llvm.experimental.constrained.fdiv.f64(double, double, metadata, metadata)
 declare double @llvm.experimental.constrained.fmul.f64(double, double, metadata, metadata)
 declare double @llvm.experimental.constrained.fadd.f64(double, double, metadata, metadata)
 declare double @llvm.experimental.constrained.fsub.f64(double, double, metadata, metadata)
 declare double @llvm.experimental.constrained.sqrt.f64(double, metadata, metadata)
 declare double @llvm.experimental.constrained.pow.f64(double, double, metadata, metadata)
 declare double @llvm.experimental.constrained.powi.f64(double, i32, metadata, metadata)
 declare double @llvm.experimental.constrained.sin.f64(double, metadata, metadata)
 declare double @llvm.experimental.constrained.cos.f64(double, metadata, metadata)
 declare double @llvm.experimental.constrained.exp.f64(double, metadata, metadata)
 declare double @llvm.experimental.constrained.exp2.f64(double, metadata, metadata)
 declare double @llvm.experimental.constrained.log.f64(double, metadata, metadata)
 declare double @llvm.experimental.constrained.log10.f64(double, metadata, metadata)
 declare double @llvm.experimental.constrained.log2.f64(double, metadata, metadata)
 declare double @llvm.experimental.constrained.rint.f64(double, metadata, metadata)
 declare double @llvm.experimental.constrained.nearbyint.f64(double, metadata, metadata)
 declare double @llvm.experimental.constrained.fma.f64(double, double, double, metadata, metadata)
+declare float @llvm.experimental.constrained.fptrunc.f32.f64(double, metadata)
+declare double @llvm.experimental.constrained.fpext.f64.f32(float, metadata)