Strengthen handling of GEPs and generic calls for all undef lanes

[reames]

Summary

If all lanes in a vector can be proven undef, then the result is undef. This patch implements and tests the GEP case. Once the masked.load patch (D57372) has landed, I'll update to include call tests as well.

[reames]

ping

[spatel]

lib/Transforms/InstCombine/InstCombineCalls.cpp

1884 ↗

(On Diff #184355)

I'm not sure how to expose this in a regression test, but would it make sense to add this check to the bottom of SimplifyDemandedVectorElts() itself, so callers don't have to deal with this possibility themselves?

[reames]

Address review comment

[spatel]

LGTM

lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp

1686 ↗

(On Diff #190846)

Interest -> Intersect?

[spatel]

Looks like this change can produce invalid IR:
https://bugs.llvm.org/show_bug.cgi?id=41624

[reames]

In D57468#1481385, @spatel wrote:

Looks like this change can produce invalid IR:
https://bugs.llvm.org/show_bug.cgi?id=41624

Fixed in https://reviews.llvm.org/rL359633.

llvm/trunk/lib/Transforms/InstCombine/InstCombineCalls.cpp

if (auto *MI = dyn_cast<AnyMemIntrinsic>(II)) {
     } else if (auto *MSI = dyn_cast<AnyMemSetInst>(MI)) {
       if (Instruction *I = SimplifyAnyMemSet(MSI))
         return I;
     }
 
     if (Changed) return II;
   }
 
-  // For vector result intrinsics, use the generic demanded vector support to
-  // simplify any operands before moving on to the per-intrinsic rules.    
+  // For vector result intrinsics, use the generic demanded vector support.
   if (II->getType()->isVectorTy()) {
     auto VWidth = II->getType()->getVectorNumElements();
     APInt UndefElts(VWidth, 0);
     APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
     if (Value *V = SimplifyDemandedVectorElts(II, AllOnesEltMask, UndefElts)) {
       if (V != II)
         return replaceInstUsesWith(*II, V);
       return II;

llvm/trunk/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp

Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
   default: break;
 
   case Instruction::GetElementPtr: {
     // Conservatively track the demanded elements back through any vector
     // operands we may have.  We know there must be at least one, or we
     // wouldn't have a vector result to get here. Note that we intentionally
     // merge the undef bits here since gepping with either an undef base or
     // index results in undef. 
-    for (unsigned i = 0; i < I->getNumOperands(); i++)
-      if (I->getOperand(i)->getType()->isVectorTy())
-        simplifyAndSetOp(I, i, DemandedElts, UndefElts);
+    for (unsigned i = 0; i < I->getNumOperands(); i++) {
+      if (isa<UndefValue>(I->getOperand(i))) {
+        // If the entire vector is undefined, just return this info.
+        UndefElts = EltMask;
+        return nullptr;
+      }
+      if (I->getOperand(i)->getType()->isVectorTy()) {
+        APInt UndefEltsOp(VWidth, 0);
+        simplifyAndSetOp(I, i, DemandedElts, UndefEltsOp);
+        UndefElts |= UndefEltsOp;
+      }
+    }
 
     break;
   }
   case Instruction::InsertElement: {
     // If this is a variable index, we don't know which element it overwrites.
     // demand exactly the same input as we produce.
     ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
     if (!Idx) {
if (match(I, m_BinOp(BO)) && !BO->isIntDivRem() && !BO->isShift()) {
     if (MadeChange)
       BO->dropPoisonGeneratingFlags();
 
     // Output elements are undefined if both are undefined. Consider things
     // like undef & 0. The result is known zero, not undef.
     UndefElts &= UndefElts2;
   }
 
+  // If we've proven all of the lanes undef, return an undef value.
+  // TODO: Intersect w/demanded lanes
+  if (UndefElts.isAllOnesValue())
+    return UndefValue::get(I->getType());;
+
   return MadeChange ? I : nullptr;
 }

llvm/trunk/lib/Transforms/InstCombine/InstructionCombining.cpp

 
 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
   SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
   Type *GEPType = GEP.getType();
   Type *GEPEltType = GEP.getSourceElementType();
   if (Value *V = SimplifyGEPInst(GEPEltType, Ops, SQ.getWithInstruction(&GEP)))
     return replaceInstUsesWith(GEP, V);
 
+  // For vector geps, use the generic demanded vector support.
+  if (GEP.getType()->isVectorTy()) {
+    auto VWidth = GEP.getType()->getVectorNumElements();
+    APInt UndefElts(VWidth, 0);
+    APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
+    if (Value *V = SimplifyDemandedVectorElts(&GEP, AllOnesEltMask,
+                                              UndefElts)) {
+      if (V != &GEP)
+        return replaceInstUsesWith(GEP, V);
+      return &GEP;
+    }
+  }
+
   Value *PtrOp = GEP.getOperand(0);
 
   // Eliminate unneeded casts for indices, and replace indices which displace
   // by multiples of a zero size type with zero.
   bool MadeChange = false;
 
   // Index width may not be the same width as pointer width.
   // Data layout chooses the right type based on supported integer types.

llvm/trunk/test/Transforms/InstCombine/vec_demanded_elts.ll

;
   %idxvec2 = shufflevector <2 x i64> %idxvec1, <2 x i64> undef, <2 x i32> zeroinitializer
   %gep = getelementptr i32, <2 x i32*> %basevec2, <2 x i64> %idxvec2
   %ee = extractelement <2 x i32*> %gep, i32 1
   ret i32* %ee
 }
 
 define <2 x i32*> @gep_all_lanes_undef(i32* %base, i64 %idx) {;
 ; CHECK-LABEL: @gep_all_lanes_undef(
-; CHECK-NEXT:    [[BASEVEC:%.*]] = insertelement <2 x i32*> undef, i32* [[BASE:%.*]], i32 0
-; CHECK-NEXT:    [[IDXVEC:%.*]] = insertelement <2 x i64> undef, i64 [[IDX:%.*]], i32 1
-; CHECK-NEXT:    [[GEP:%.*]] = getelementptr i32, <2 x i32*> [[BASEVEC]], <2 x i64> [[IDXVEC]]
-; CHECK-NEXT:    ret <2 x i32*> [[GEP]]
+; CHECK-NEXT:    ret <2 x i32*> undef
 ;
   %basevec = insertelement <2 x i32*> undef, i32* %base, i32 0
   %idxvec = insertelement <2 x i64> undef, i64 %idx, i32 1
   %gep = getelementptr i32, <2 x i32*> %basevec, <2 x i64> %idxvec
   ret <2 x i32*> %gep
 }
 
 define i32* @gep_demanded_lane_undef(i32* %base, i64 %idx) {
 ; CHECK-LABEL: @gep_demanded_lane_undef(
 ; CHECK-NEXT:    ret i32* undef
 ;
   %basevec = insertelement <2 x i32*> undef, i32* %base, i32 0
   %idxvec = insertelement <2 x i64> undef, i64 %idx, i32 1
   %gep = getelementptr i32, <2 x i32*> %basevec, <2 x i64> %idxvec
   %ee = extractelement <2 x i32*> %gep, i32 1
   ret i32* %ee
 }
-