Added llvm.is.constant intrinsic

[jyknight]

Summary

Add support for llvm.is.constant intrinsic (PR4898)

This adds llvm-side support for post-inline evaluation of the __builtin_constant_p GCC intrinsic.

[rengolin]

I'd add a bit more of tests, especially those that inspired this change...

--renato

docs/LangRef.rst

7178 ↗

(On Diff #10791)

... to the best of the compiler's knowledge ... :)

[• janusz]

Added a test case for inlining @llvm.is.constant

[• janusz]

Uploading again :)

[rengolin]

Thanks Janusz,

I'm adding Owen to review this patch, since it's not my area, but overall, the patch looks good.

About the tests, though the inline check is important, I was thinking something more elaborate.

Some examples that come to mind:

• static initializer
• generic function arguments
• function arguments where the function is static and only called one with a constant argument
• function arguments where the function is static and only called one with a variable argument
• expressions that will be simplified as constant because all variables were initialised with constants
• expressions that will NOT be simplified as constant because NOT all variables were initialised with constants

Those examples that *should* be constant and are not, should be marked as FIXME. Those examples that *should* NOT be constant and are, should be fixed before commit.

cheers,
--renato

test/CodeGen/Generic/inline-is-constant.ll

4 ↗

(On Diff #10830)

a more robust way would be to have two run lines, one with O0, another with O3, and use FileCheck for the True/False results.

[• janusz]

I agree that we should have more tests. I'd prefer to have an agreement that we want to add llvm.is.constant intrinsic before spending too much time on writing tests.

A follow up work is to add support for the new intrinsic in clang (__builtin_constant_p -> llvm.is.constant) and write more tests in C.

I have a few questions.

1. Do we want the code to promote or expand llvm.is.constant? On one hand it would make the code more complete but on the other it is not really required because zext() and trunc() don't change the 'constantness'

	%a = zext i1 %b to i32
	%v = call i1 @llvm.is.constant.i32(i32 %a)

and

	%a = trunc i256 %b to i32
	%v = call i1 @llvm.is.constant.i32(i32 %a)

will do the job. Is there a preference in the llvm community?

2. In my implementation llvm.is.constant is lowered to true/false rather late. Would it help in some cases if we added an early check in InstCombine? InstCombine would lower llvm.is.constant to true if the argument is known to be constant. Otherwise InstCombine would leave it unmodified.

[• janusz]

test/CodeGen/Generic/inline-is-constant.ll

4 ↗

(On Diff #10830)

If using lli is frowned upon, then the resolution is to call 'llc -debug' and search for the debug strings with FileCheck

[• rafael]

Ccing nick since it seems related to the enable_if attribute.

[• janusz]

Moved lowering llvm.is.constant to False from LegalizeDAG to SelectionDAG. Code to promote or expand llvm.is.constant is not required any more and the intrinsic works fine with different input integer types.

[rengolin]

docs/LangRef.rst

7164 ↗

(On Diff #10884)

As I understand, __builtin_constant_p can be used with any valid expression, and "constant" is the C semantics, so a numeric expression is constant if all its arguments are, string literals are always constant, but pointers are a bit more complicated.

Since this is a GNU extension and we're trying to implement it, it would be wise to add a FIXME and some comments on this documentation to that effect, if our intentions is to not expand to full expresisons, or to do so in the future.

[rengolin]

In D4276#14, @janusz wrote:

1. Do we want the code to promote or expand llvm.is.constant? On one hand it would make the code more complete but on the other it is not really required because zext() and trunc() don't change the 'constantness'

Depends on how hard would it be to infer constantness at a lower level with the added complexity. I'd say, as a general idea, leaving the IR untouched and adding logic to infer that later would be better, but I'm not sure of all the cases.

2. In my implementation llvm.is.constant is lowered to true/false rather late. Would it help in some cases if we added an early check in InstCombine? InstCombine would lower llvm.is.constant to true if the argument is known to be constant. Otherwise InstCombine would leave it unmodified.

Yes, that'd be great. I think we could teach even Clang about that, so that pre-processor constantness is explited at the front-end level. Though, I understand if those things take their time to get implemented.

[rengolin]

Any progress in this? If not, can we close this review?

[jyknight]

Taking this over.

[jyknight]

Rewrote patch significantly.

• Added support for other types.
• Deleted SelectionDAG-side IS_CONSTANT as unneeded.
• Added support in CodeGenPrepare to lower to false, as ISel cannot eliminate unreachable blocks. (similar to llvm.objectsize)

Updated and added more tests;

[efriedma]

It seems like you're waiting to fold llvm.is.constant until really late in the optimization pipeline; we probably want to fold it sometime in the "middle", so we get better optimization. Maybe just after function simplification passes; at that point, we're unlikely to get any more useful information about whether the argument is constant, and we want to simplify the code as much as possible before we run transforms like loop vectorization.

llvm/lib/Analysis/ConstantFolding.cpp

1607

Probably doesn't make sense to check for ConstantExpr here; if isManifestConstant is true for all the operands of a ConstantExpr, it should get constant-folded.

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

5901

Indentation.

[jyknight]

In D4276#1067497, @efriedma wrote:

It seems like you're waiting to fold llvm.is.constant until really late in the optimization pipeline; we probably want to fold it sometime in the "middle", so we get better optimization. Maybe just after function simplification passes; at that point, we're unlikely to get any more useful information about whether the argument is constant, and we want to simplify the code as much as possible before we run transforms like loop vectorization.

The most important constraint is that it needs to occur after inlining.

However, we already have llvm.objectsize intrinsic, and I wanted to handle the final failure in the same place, so that __builtin_constant_p works when given __builtin_object_size as an argument, even when the __builtin_object_size is unevaluable (and returns -1).

E.g. in GCC, this code:

int f1(char* x) {
    if (__builtin_constant_p(__builtin_object_size(x, 0))) {
        return __builtin_object_size(x, 0);
    }
    return -9999;
}

int f2() {
    char x[1000];
    return f1(x);
}

compiles such that:

f1 -> returns (constant) -1.
f2 -> returns (constant) 1000.

It's possible that both of them should move somewhere else, though...

llvm/lib/Analysis/ConstantFolding.cpp

1607

If I remove that, then this clause fails in the test case:

@llvm.is.constant.p0i64(i64* inttoptr (i32 42 to i64*))

because there's no way to express a pointer with a constant value directly.

[efriedma]

I guess lowering llvm.objectsize and llvm.is.constant at the same time makes sense; okay. They should probably both be lowered earlier, though.

llvm/lib/Analysis/ConstantFolding.cpp

1607

Okay.

[void]

llvm/docs/LangRef.rst

15454–15455

This is true only for expressions that evaluate to constants not immediates, right?

Should you also mention that if a function with llvm.is.constant() is inlined, any constants fed into the inlined function may cause llvm.is.constant() to return true? (or something to that affect)

llvm/include/llvm/IR/Intrinsics.td

904

80-columns?

llvm/lib/Transforms/Scalar/SCCP.cpp

1233

This seems unrelated. Was it meant to be in this patch? If so, could you comment that it's specifically related to llvm.is.constant()?

[jyknight]

Ran clang-format.
Expanded on docs.

llvm/docs/LangRef.rst

15454–15455

If you build with -O0, constant folding is not done, so even the @test_constant() function will return false.

(Of course, at the C level, the __builtin_constant_p intrinsic will still return true at -O0 when given a frontend constant).

Updated the docs to mention inlining explicitly.

llvm/include/llvm/IR/Intrinsics.td

904

Fixed.

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

5901

Fixed.

llvm/lib/Transforms/Scalar/SCCP.cpp

1233

It's due to llvm.is.constant being able to take a struct type as an argument. This code gets confused by seeing a call for which is canConstantFoldCallTo() returns true, but which can take a struct as an argument.

I'll add a bit to the commit message remarking why the change was made, but I don't think the code here really need mention llvm.is.constant.

[void]

llvm/docs/LangRef.rst

15454–15455

My comment is for documenting the llvm.is.constant intrinsic, not __builtin_constant_p(). :-) (People can write their own .ll files.) I just wanted to make sure that people know that this will evaluate to "true" even without constant folding:

%0 = i1 llvm.is.constant.i32(i32 37)

llvm/lib/Transforms/Scalar/SCCP.cpp

1233

I'm still a bit confused. This change will affect all call sites, not just those to llvm.is.constant. Or are you saying that the only time this situation will come up is when it involves the llvm.is.constant call?

[jyknight]

llvm/docs/LangRef.rst

15454–15455

Except that that it really _doesn't_ evaluate to true. :)

The only thing that turns a call to llvm.is.constant into "true" is the constant folding code, which doesn't get run in O0.

llvm/lib/Transforms/Scalar/SCCP.cpp

1233

This code cannot handle calls with a struct as an argument, it hits an assertion failure about not expecting a struct.

But it only ran in the first place when canConstantFoldCallTo returns true (see up a few lines at 1171), so it didn't actually come up before, because the other functions don't take structs as arguments.

[void]

llvm/docs/LangRef.rst

15454–15455

But why couldn't it evaluate to true at O0? Is there a valid reason why not? I mean, the intrinsic has to be lowered somehow at O0, so why not just evaluate it to true in cases like this?

llvm/lib/Transforms/Scalar/SCCP.cpp

1233

Ah! I see. Thanks.

[void]

Friendly ping? :-)

[jyknight]

llvm/docs/LangRef.rst

15454–15455

Yes, all the places it's lowered to false now in O0 (GlobalISel/IRTranslator.cpp, SelectionDAG/FastISel.cpp, SelectionDAG/SelectionDAGBuilder.cpp) could check if it can be folded to true first. It feels somewhat ugly to have to do that, though.

Of course, it's ugly that we have to do this last-chance lowering in 3 places in the first place. Same issue exists for llvm.objectsize too. Like efriedma suggested earlier, they should probably be both be lowered somewhere else...I'm not really sure where the best elsewhere to move it would be, though.

[void]

I have no more concerns with this patch.

Eli, Are you okay to approve this?

llvm/docs/LangRef.rst

15454–15455

It certainly is ugly that we have this in three places instead of one. Anyway, I suppose that as long as we document its behavior correctly, people won't complain. (And it's at -O0, so speed isn't the issue I suppose.)

[efriedma]

Needs a testcase for the SCCP fix. Otherwise LGTM.

[void]

Friendly ping. :-)

[Bharathi]

This feature would be very useful to us and testing this patch against our code base gave us an interesting case where llvm.is.constant (used in a conditional) is lowered to false in CodeGenPrepare and the dead branch that arises out of this lowering does not get eliminated. The code that gets generated after CodeGenPrepare pass has "br i1 false .." with both true and false conditions preserved and this propagates further and remains the same in the final assembly code.

In CodeGenPrepare::runOnFunction, ConstantFoldTerminator (which handles the br i1 false condition) is called only once and if after the transformation of ConstantFoldTerminator() and DeleteDeadBlock(), if we end up with code like "br i1 false", there is no further opportunity to clean them up. So calling the code under ' if (!DisableBranchOpts)' in a loop that is conditioned on MadeChange (similar to other places), fixes the issue. Is this reasonable ?

I wrote a patch to test this and I have one regression with llvm/test/CodeGen/AMDGPU/nested-loop-conditions.ll. I'm having some difficulty to determine if this warrants rewriting the test case to adjust with the new results or if this is a real issue.

Any pointers would be greatly appreciated. My simple fix (without any indentation changes) is:

--- a/llvm/lib/CodeGen/CodeGenPrepare.cpp
+++ b/llvm/lib/CodeGen/CodeGenPrepare.cpp
@@ -316,7 +316,9 @@ bool CodeGenPrepare::runOnFunction(Function &F) {
 
   SunkAddrs.clear();
 
+  MadeChange = true;
   if (!DisableBranchOpts) {
+  while (MadeChange) {
     MadeChange = false;
     SmallPtrSet<BasicBlock*, 8> WorkList;
     for (BasicBlock &BB : F) {
@@ -352,6 +354,7 @@ bool CodeGenPrepare::runOnFunction(Function &F) {
 
     EverMadeChange |= MadeChange;
   }
+ }
 
   if (!DisableGCOpts) {
     SmallVector<Instruction *, 2> Statepoints;

Thanks.

[void]

What's the status of this patch? I'm running into a situation where evaluating __builtin_constant_p through inlining may be useful.

[jyknight]

Rebased onto current head, no changes.

[jyknight]

Picking this back up again...

I'm ready to actually submit this now that Bill has the clang-side code that'll use it almost in place. Woo! :)

In D4276#1073957, @efriedma wrote:

Needs a testcase for the SCCP fix. Otherwise LGTM.

The SCCP bug is already be exhibited by the tests in test/CodeGen/Generic/is-constant.ll which pass a struct. Since I believe llvm.is.constant is the only thing we can constant-fold which takes struct args at the moment anyhow, that's really the only viable test. So I think there's nothing else needed here.

In D4276#1135976, @Bharathi wrote:

This feature would be very useful to us and testing this patch against our code base gave us an interesting case where llvm.is.constant (used in a conditional) is lowered to false in CodeGenPrepare and the dead branch that arises out of this lowering does not get eliminated. The code that gets generated after CodeGenPrepare pass has "br i1 false .." with both true and false conditions preserved and this propagates further and remains the same in the final assembly code.

It may make sense to iterate, as you say, but that can/should be a different patch.

[efriedma]

The SCCP bug is already be exhibited by the tests in test/CodeGen/Generic/is-constant.ll which pass a struct.

I'd prefer a testcase that explicitly calls "opt -sccp", as opposed to implicitly relying on the fact that opt -O2 includes SCCP. (You can just add a small test to test/Transforms/SCCP/ipsccp-basic.ll .)

[george.burgess.iv]

It may make sense to iterate, as you say, but that can/should be a different patch.

D49103 should take care of that, FWIW. Nothing to block this patch on (or vice-versa) IMO

[jyknight]

Added sccp testcase.

[jyknight]

In D4276#1289427, @efriedma wrote:

The SCCP bug is already be exhibited by the tests in test/CodeGen/Generic/is-constant.ll which pass a struct.

I'd prefer a testcase that explicitly calls "opt -sccp", as opposed to implicitly relying on the fact that opt -O2 includes SCCP. (You can just add a small test to test/Transforms/SCCP/ipsccp-basic.ll .)

Done. And verified that without the SCCP change, the test fails, and with it, passes.

[efriedma]

LGTM

llvm/docs/LangRef.rst

 None.
 
 Semantics:
 """"""""""
 
 This intrinsic actually does nothing, but optimizers must assume that it
 has externally observable side effects.
 
+'``llvm.is.constant.*``' Intrinsic
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Syntax:
+"""""""
+
+This is an overloaded intrinsic. You can use llvm.is.constant with any argument type.
+
+::
+
+      declare i1 @llvm.is.constant.i32(i32 %operand) nounwind readnone
+      declare i1 @llvm.is.constant.f32(float %operand) nounwind readnone
+      declare i1 @llvm.is.constant.TYPENAME(TYPE %operand) nounwind readnone
+
+Overview:
+"""""""""
+
+The '``llvm.is.constant``' intrinsic will return true if the argument
+is known to be a manifest compile-time constant. It is guaranteed to
+fold to either true or false before generating machine code.
+
+Semantics:
+""""""""""
+
+This intrinsic generates no code. If its argument is known to be a
+manifest compile-time constant value, then the intrinsic will be
+converted to a constant true value. Otherwise, it will be converted to
+a constant false value.
+
+In particular, note that if the argument is a constant expression
+which refers to a global (the address of which _is_ a constant, but
+not manifest during the compile), then the intrinsic evaluates to
+false.
+
+The result also intentionally depends on the result of optimization
+passes -- e.g., the result can change depending on whether a
+function gets inlined or not. A function's parameters are
+obviously not constant. However, a call like
+``llvm.is.constant.i32(i32 %param)`` *can* return true after the
+function is inlined, if the value passed to the function parameter was
+a constant.
+
+On the other hand, if constant folding is not run, it will never
+evaluate to true, even in simple cases.
+
 Stack Map Intrinsics
 --------------------
 
 LLVM provides experimental intrinsics to support runtime patching
 mechanisms commonly desired in dynamic language JITs. These intrinsics
 are described in :doc:`StackMaps`.
 
 Element Wise Atomic Memory Intrinsics

llvm/include/llvm/Analysis/TargetTransformInfoImpl.h

unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
     case Intrinsic::sideeffect:
     case Intrinsic::dbg_declare:
     case Intrinsic::dbg_value:
     case Intrinsic::dbg_label:
     case Intrinsic::invariant_start:
     case Intrinsic::invariant_end:
     case Intrinsic::launder_invariant_group:
     case Intrinsic::strip_invariant_group:
+    case Intrinsic::is_constant:
     case Intrinsic::lifetime_start:
     case Intrinsic::lifetime_end:
     case Intrinsic::objectsize:
     case Intrinsic::ptr_annotation:
     case Intrinsic::var_annotation:
     case Intrinsic::experimental_gc_result:
     case Intrinsic::experimental_gc_relocate:
     case Intrinsic::coro_alloc:

llvm/include/llvm/IR/Intrinsics.td

 def int_convert_from_fp16 : Intrinsic<[llvm_anyfloat_ty], [llvm_i16_ty]>;
 }
 
 // Clear cache intrinsic, default to ignore (ie. emit nothing)
 // maps to void __clear_cache() on supporting platforms
 def int_clear_cache : Intrinsic<[], [llvm_ptr_ty, llvm_ptr_ty],
                                 [], "llvm.clear_cache">;
 
+// Intrinsic to detect whether its argument is a constant.
+def int_is_constant : Intrinsic<[llvm_i1_ty], [llvm_any_ty], [IntrNoMem], "llvm.is.constant">;
+
+
 //===-------------------------- Masked Intrinsics -------------------------===//
 //
 def int_masked_store : Intrinsic<[], [llvm_anyvector_ty,
                                       LLVMAnyPointerType<LLVMMatchType<0>>,
                                       llvm_i32_ty,
                                       LLVMVectorSameWidth<0, llvm_i1_ty>],
                                  [IntrArgMemOnly]>;
 

llvm/lib/Analysis/ConstantFolding.cpp

bool llvm::canConstantFoldCallTo(ImmutableCallSite CS, const Function *F) {
   case Intrinsic::x86_avx512_vcvtss2usi32:
   case Intrinsic::x86_avx512_vcvtss2usi64:
   case Intrinsic::x86_avx512_cvttss2usi:
   case Intrinsic::x86_avx512_cvttss2usi64:
   case Intrinsic::x86_avx512_vcvtsd2usi32:
   case Intrinsic::x86_avx512_vcvtsd2usi64:
   case Intrinsic::x86_avx512_cvttsd2usi:
   case Intrinsic::x86_avx512_cvttsd2usi64:
+  case Intrinsic::is_constant:
     return true;
   default:
     return false;
   case Intrinsic::not_intrinsic: break;
   }
 
   if (!F->hasName())
     return false;
if (Ty->isDoubleTy())
     return Op->getValueAPF().convertToDouble();
 
   bool unused;
   APFloat APF = Op->getValueAPF();
   APF.convert(APFloat::IEEEdouble(), APFloat::rmNearestTiesToEven, &unused);
   return APF.convertToDouble();
 }
 
+static bool isManifestConstant(const Constant *c) {
+  if (isa<ConstantData>(c)) {
+    return true;
+  } else if (isa<ConstantAggregate>(c) || isa<ConstantExpr>(c)) {
+    for (const Value *subc : c->operand_values()) {
+      if (!isManifestConstant(cast<Constant>(subc)))
+        return false;
+    }
+    return true;
+  }
+  return false;
+}
+
 Constant *ConstantFoldScalarCall(StringRef Name, unsigned IntrinsicID, Type *Ty,
                                  ArrayRef<Constant *> Operands,
                                  const TargetLibraryInfo *TLI,
                                  ImmutableCallSite CS) {
   if (Operands.size() == 1) {
+    if (IntrinsicID == Intrinsic::is_constant) {
+      // We know we have a "Constant" argument. But we want to only
+      // return true for manifest constants, not those that depend on
+      // constants with unknowable values, e.g. GlobalValue or BlockAddress.
+      if (isManifestConstant(Operands[0]))
+        return ConstantInt::getTrue(Ty->getContext());
+      return nullptr;
+    }
     if (isa<UndefValue>(Operands[0])) {
       // cosine(arg) is between -1 and 1. cosine(invalid arg) is NaN
       if (IntrinsicID == Intrinsic::cos)
         return Constant::getNullValue(Ty);
       if (IntrinsicID == Intrinsic::bswap ||
           IntrinsicID == Intrinsic::bitreverse ||
           IntrinsicID == Intrinsic::launder_invariant_group ||
           IntrinsicID == Intrinsic::strip_invariant_group)

llvm/lib/CodeGen/CodeGenPrepare.cpp

void getAnalysisUsage(AnalysisUsage &AU) const override {
       // FIXME: When we can selectively preserve passes, preserve the domtree.
       AU.addRequired<ProfileSummaryInfoWrapperPass>();
       AU.addRequired<TargetLibraryInfoWrapperPass>();
       AU.addRequired<TargetTransformInfoWrapperPass>();
       AU.addRequired<LoopInfoWrapperPass>();
     }
 
   private:
+    template <typename F>
+    void resetIteratorIfInvalidatedWhileCalling(BasicBlock *BB, F f) {
+      // Substituting can cause recursive simplifications, which can invalidate
+      // our iterator.  Use a WeakTrackingVH to hold onto it in case this
+      // happens.
+      Value *CurValue = &*CurInstIterator;
+      WeakTrackingVH IterHandle(CurValue);
+
+      f();
+
+      // If the iterator instruction was recursively deleted, start over at the
+      // start of the block.
+      if (IterHandle != CurValue) {
+        CurInstIterator = BB->begin();
+        SunkAddrs.clear();
+      }
+    }
+
     bool eliminateFallThrough(Function &F);
     bool eliminateMostlyEmptyBlocks(Function &F);
     BasicBlock *findDestBlockOfMergeableEmptyBlock(BasicBlock *BB);
     bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
     void eliminateMostlyEmptyBlock(BasicBlock *BB);
     bool isMergingEmptyBlockProfitable(BasicBlock *BB, BasicBlock *DestBB,
                                        bool isPreheader);
     bool optimizeBlock(BasicBlock &BB, bool &ModifiedDT);
bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool &ModifiedDT) {
   IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
   if (II) {
     switch (II->getIntrinsicID()) {
     default: break;
     case Intrinsic::objectsize: {
       // Lower all uses of llvm.objectsize.*
       ConstantInt *RetVal =
           lowerObjectSizeCall(II, *DL, TLInfo, /*MustSucceed=*/true);
-      // Substituting this can cause recursive simplifications, which can
-      // invalidate our iterator.  Use a WeakTrackingVH to hold onto it in case
-      // this
-      // happens.
-      Value *CurValue = &*CurInstIterator;
-      WeakTrackingVH IterHandle(CurValue);
 
+      resetIteratorIfInvalidatedWhileCalling(BB, [&]() {
         replaceAndRecursivelySimplify(CI, RetVal, TLInfo, nullptr);
-
-      // If the iterator instruction was recursively deleted, start over at the
-      // start of the block.
-      if (IterHandle != CurValue) {
-        CurInstIterator = BB->begin();
-        SunkAddrs.clear();
-      }
+      });
+      return true;
+    case Intrinsic::is_constant: {
+      // If is_constant hasn't folded away yet, lower it to false now.
+      Constant *RetVal = ConstantInt::get(II->getType(), 0);
+      resetIteratorIfInvalidatedWhileCalling(BB, [&]() {
+        replaceAndRecursivelySimplify(CI, RetVal, TLInfo, nullptr);
+      });
       return true;
     }
     case Intrinsic::aarch64_stlxr:
     case Intrinsic::aarch64_stxr: {
       ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
       if (!ExtVal || !ExtVal->hasOneUse() ||
           ExtVal->getParent() == CI->getParent())
         return false;

llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp

bool IRTranslator::translateKnownIntrinsic(const CallInst &CI, Intrinsic::ID ID,
   }
   case Intrinsic::objectsize: {
     // If we don't know by now, we're never going to know.
     const ConstantInt *Min = cast<ConstantInt>(CI.getArgOperand(1));
 
     MIRBuilder.buildConstant(getOrCreateVReg(CI), Min->isZero() ? -1ULL : 0);
     return true;
   }
+  case Intrinsic::is_constant:
+    // If this wasn't constant-folded away by now, then it's not a
+    // constant.
+    MIRBuilder.buildConstant(getOrCreateVReg(CI), 0);
+    return true;
   case Intrinsic::stackguard:
     getStackGuard(getOrCreateVReg(CI), MIRBuilder);
     return true;
   case Intrinsic::stackprotector: {
     LLT PtrTy = getLLTForType(*CI.getArgOperand(0)->getType(), *DL);
     unsigned GuardVal = MRI->createGenericVirtualRegister(PtrTy);
     getStackGuard(GuardVal, MIRBuilder);
 

llvm/lib/CodeGen/SelectionDAG/FastISel.cpp

case Intrinsic::objectsize: {
     unsigned long long Res = CI->isZero() ? -1ULL : 0;
     Constant *ResCI = ConstantInt::get(II->getType(), Res);
     unsigned ResultReg = getRegForValue(ResCI);
     if (!ResultReg)
       return false;
     updateValueMap(II, ResultReg);
     return true;
   }
+  case Intrinsic::is_constant: {
+    Constant *ResCI = ConstantInt::get(II->getType(), 0);
+    unsigned ResultReg = getRegForValue(ResCI);
+    if (!ResultReg)
+      return false;
+    updateValueMap(II, ResultReg);
+    return true;
+  }
   case Intrinsic::launder_invariant_group:
   case Intrinsic::strip_invariant_group:
   case Intrinsic::expect: {
     unsigned ResultReg = getRegForValue(II->getArgOperand(0));
     if (!ResultReg)
       return false;
     updateValueMap(II, ResultReg);
     return true;

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

case Intrinsic::objectsize: {
     if (CI->isZero())
       Res = DAG.getConstant(-1ULL, sdl, Ty);
     else
       Res = DAG.getConstant(0, sdl, Ty);
 
     setValue(&I, Res);
     return nullptr;
   }
+
+  case Intrinsic::is_constant:
+    // If this wasn't constant-folded away by now, then it's not a
+    // constant.
+    setValue(&I, DAG.getConstant(0, sdl, MVT::i1));
+    return nullptr;
+
   case Intrinsic::annotation:
   case Intrinsic::ptr_annotation:
   case Intrinsic::launder_invariant_group:
   case Intrinsic::strip_invariant_group:
     // Drop the intrinsic, but forward the value
     setValue(&I, getValue(I.getOperand(0)));
     return nullptr;
   case Intrinsic::assume:

llvm/lib/Transforms/Scalar/SCCP.cpp

CallOverdefined:
 
     // Otherwise, if we have a single return value case, and if the function is
     // a declaration, maybe we can constant fold it.
     if (F && F->isDeclaration() && !I->getType()->isStructTy() &&
         canConstantFoldCallTo(CS, F)) {
       SmallVector<Constant*, 8> Operands;
       for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
            AI != E; ++AI) {
+        if (AI->get()->getType()->isStructTy())
+          return markOverdefined(I); // Can't handle struct args.
         LatticeVal State = getValueState(*AI);
 
         if (State.isUnknown())
           return;  // Operands are not resolved yet.
         if (State.isOverdefined())
           return (void)markOverdefined(I);
         assert(State.isConstant() && "Unknown state!");
         Operands.push_back(State.getConstant());

llvm/test/CodeGen/Generic/is-constant.ll

+; RUN: opt -O2 -S < %s  | FileCheck %s
+; RUN: llc -o /dev/null 2>&1 < %s
+; RUN: llc -O0 -o /dev/null 2>&1 < %s
+
+;; The llc runs above are just to ensure it doesn't blow up upon
+;; seeing an is_constant intrinsic.
+
+declare i1 @llvm.is.constant.i32(i32 %a)
+declare i1 @llvm.is.constant.i64(i64 %a)
+declare i1 @llvm.is.constant.i256(i256 %a)
+declare i1 @llvm.is.constant.v2i64(<2 x i64> %a)
+declare i1 @llvm.is.constant.f32(float %a)
+declare i1 @llvm.is.constant.sl_i32i32s({i32, i32} %a)
+declare i1 @llvm.is.constant.a2i32([2 x i32] %a)
+declare i1 @llvm.is.constant.p0i64(i64* %a)
+
+;; Basic test that optimization folds away the is.constant when given
+;; a constant.
+define i1 @test_constant() #0 {
+; CHECK-LABEL: @test_constant(
+; CHECK-NOT: llvm.is.constant
+; CHECK: ret i1 true
+%y = call i1 @llvm.is.constant.i32(i32 44)
+  ret i1 %y
+}
+
+;; And test that the intrinsic sticks around when given a
+;; non-constant.
+define i1 @test_nonconstant(i32 %x) #0 {
+; CHECK-LABEL: @test_nonconstant(
+; CHECK: @llvm.is.constant
+  %y = call i1 @llvm.is.constant.i32(i32 %x)
+  ret i1 %y
+}
+
+;; Ensure that nested is.constants fold.
+define i32 @test_nested() #0 {
+; CHECK-LABEL: @test_nested(
+; CHECK-NOT: llvm.is.constant
+; CHECK: ret i32 13
+  %val1 = call i1 @llvm.is.constant.i32(i32 27)
+  %val2 = zext i1 %val1 to i32
+  %val3 = add i32 %val2, 12
+  %1 = call i1 @llvm.is.constant.i32(i32 %val3)
+  %2 = zext i1 %1 to i32
+  %3 = add i32 %2, 12
+  ret i32 %3
+}
+
+@G = global [2 x i64] zeroinitializer
+define i1 @test_global() #0 {
+; CHECK-LABEL: @test_global(
+; CHECK: llvm.is.constant
+  %ret = call i1 @llvm.is.constant.p0i64(i64* getelementptr ([2 x i64], [2 x i64]* @G, i32 0, i32 0))
+  ret i1 %ret
+}
+
+define i1 @test_diff() #0 {
+; CHECK-LABEL: @test_diff(
+  %ret = call i1 @llvm.is.constant.i64(i64 sub (
+      i64 ptrtoint (i64* getelementptr inbounds ([2 x i64], [2 x i64]* @G, i64 0, i64 1) to i64),
+      i64 ptrtoint ([2 x i64]* @G to i64)))
+  ret i1 %ret
+}
+
+define i1 @test_various_types(i256 %int, float %float, <2 x i64> %vec, {i32, i32} %struct, [2 x i32] %arr, i64* %ptr) #0 {
+; CHECK-LABEL: @test_various_types(
+; CHECK: llvm.is.constant
+; CHECK: llvm.is.constant
+; CHECK: llvm.is.constant
+; CHECK: llvm.is.constant
+; CHECK: llvm.is.constant
+; CHECK: llvm.is.constant
+; CHECK-NOT: llvm.is.constant
+  %v1 = call i1 @llvm.is.constant.i256(i256 %int)
+  %v2 = call i1 @llvm.is.constant.f32(float %float)
+  %v3 = call i1 @llvm.is.constant.v2i64(<2 x i64> %vec)
+  %v4 = call i1 @llvm.is.constant.sl_i32i32s({i32, i32} %struct)
+  %v5 = call i1 @llvm.is.constant.a2i32([2 x i32] %arr)
+  %v6 = call i1 @llvm.is.constant.p0i64(i64* %ptr)
+
+  %c1 = call i1 @llvm.is.constant.i256(i256 -1)
+  %c2 = call i1 @llvm.is.constant.f32(float 17.0)
+  %c3 = call i1 @llvm.is.constant.v2i64(<2 x i64> <i64 -1, i64 44>)
+  %c4 = call i1 @llvm.is.constant.sl_i32i32s({i32, i32} {i32 -1, i32 32})
+  %c5 = call i1 @llvm.is.constant.a2i32([2 x i32] [i32 -1, i32 32])
+  %c6 = call i1 @llvm.is.constant.p0i64(i64* inttoptr (i32 42 to i64*))
+
+  %x1 = add i1 %v1, %c1
+  %x2 = add i1 %v2, %c2
+  %x3 = add i1 %v3, %c3
+  %x4 = add i1 %v4, %c4
+  %x5 = add i1 %v5, %c5
+  %x6 = add i1 %v6, %c6
+
+  %res2 = add i1 %x1, %x2
+  %res3 = add i1 %res2, %x3
+  %res4 = add i1 %res3, %x4
+  %res5 = add i1 %res4, %x5
+  %res6 = add i1 %res5, %x6
+
+  ret i1 %res6
+}
+
+define i1 @test_various_types2() #0 {
+; CHECK-LABEL: @test_various_types2(
+; CHECK: ret i1 false
+  %r = call i1 @test_various_types(i256 -1, float 22.0, <2 x i64> <i64 -1, i64 44>,
+                     {i32, i32} {i32 -1, i32 55}, [2 x i32] [i32 -1, i32 55],
+         i64* inttoptr (i64 42 to i64*))
+  ret i1 %r
+}
+
+attributes #0 = { nounwind uwtable }

llvm/test/CodeGen/X86/is-constant.ll

+; RUN: llc -O2 < %s | FileCheck %s --check-prefix=CHECK-O2 --check-prefix=CHECK
+; RUN: llc -O0 -fast-isel < %s | FileCheck %s --check-prefix=CHECK-O0 --check-prefix=CHECK
+; RUN: llc -O0 -fast-isel=0 < %s | FileCheck %s --check-prefix=CHECK-O0 --check-prefix=CHECK
+; RUN: llc -O0 -global-isel < %s | FileCheck %s --check-prefix=CHECK-O0 --check-prefix=CHECK
+
+;; Ensure that an unfoldable is.constant gets lowered reasonably in
+;; optimized codegen, in particular, that the "true" branch is
+;; eliminated.
+;;
+;; This isn't asserting any specific output from non-optimized runs,
+;; (e.g., currently the not-taken branch does not get eliminated). But
+;; it does ensure that lowering succeeds in all 3 codegen paths.
+
+target triple = "x86_64-unknown-linux-gnu"
+
+declare i1 @llvm.is.constant.i32(i32 %a) nounwind readnone
+declare i1 @llvm.is.constant.i64(i64 %a) nounwind readnone
+declare i64 @llvm.objectsize.i64.p0i8(i8*, i1, i1) nounwind readnone
+
+declare i32 @subfun_1()
+declare i32 @subfun_2()
+
+define i32 @test_branch(i32 %in) nounwind {
+; CHECK-LABEL:    test_branch:
+; CHECK-O2:       %bb.0:
+; CHECK-O2-NEXT:  jmp subfun_2
+  %v = call i1 @llvm.is.constant.i32(i32 %in)
+  br i1 %v, label %True, label %False
+
+True:
+  %call1 = tail call i32 @subfun_1()
+  ret i32 %call1
+
+False:
+  %call2 = tail call i32 @subfun_2()
+  ret i32 %call2
+}
+
+;; llvm.objectsize is another tricky case which gets folded to -1 very
+;; late in the game. We'd like to ensure that llvm.is.constant of
+;; llvm.objectsize is true.
+define i1 @test_objectsize(i8* %obj) nounwind {
+; CHECK-LABEL:    test_objectsize:
+; CHECK-O2:       %bb.0:
+; CHECK-O2:       movb $1, %al
+; CHECK-O2-NEXT:  retq
+  %os = call i64 @llvm.objectsize.i64.p0i8(i8* %obj, i1 false, i1 false)
+  %v = call i1 @llvm.is.constant.i64(i64 %os)
+  ret i1 %v
+}
+

llvm/test/Transforms/SCCP/ipsccp-basic.ll

define i64 @test11b() {
   ret i64 %call2
 ; CHECK-LABEL: define i64 @test11b
 ; CHECK: %[[call1:.*]] = call i64 @test11a()
 ; CHECK-NOT: call i64 @llvm.ctpop.i64
 ; CHECK-NEXT: ret i64 0
 }
 
 declare i64 @llvm.ctpop.i64(i64)
+
+;;======================== test12
+;; Ensure that a struct as an arg to a potentially constant-foldable
+;; function does not crash SCCP (for now it'll just ignores it)
+
+define i1 @test12() {
+  %c = call i1 @llvm.is.constant.sl_i32i32s({i32, i32} {i32 -1, i32 32})
+  ret i1 %c
+; CHECK-LABEL: define i1 @test12
+; CHECK: ret i1 %c
+}
+
+declare i1 @llvm.is.constant.sl_i32i32s({i32, i32} %a)