[ValueTracking, InstSimplify] extend isKnownNonZero() to handle vector constants

[spatel]

Summary

Similar in spirit to D20497 :
If all elements of a constant vector are known non-zero, then we can say that the whole vector is known non-zero.
It seems like we could extend this to FP scalar/vector too, but isKnownNonZero() says it only works for integers and pointers for now.

[majnemer]

LGTM with nits.

lib/Analysis/ValueTracking.cpp

1686–1689

Might be easier to read as:

Constant *Elt = C->getAggregateElement(i);
if (!Elt || Elt->isNullValue())
  return false;
if (!isa<UndefValue>(Elt) && !isa<ConstantInt>(Elt))
  return false;

While it is a little more code, I think it's a little easier to follow along.

[spatel]

lib/Analysis/ValueTracking.cpp

1686–1689

Sure - I actually had written it that way initially, but I thought the nit would be to make it one big 'if'. :)
I'll flip it back. Thanks!

lib/Analysis/ValueTracking.cpp

 /// integers.
 bool isKnownNonZero(Value *V, unsigned Depth, const Query &Q) {
   if (auto *C = dyn_cast<Constant>(V)) {
     if (C->isNullValue())
       return false;
     if (isa<ConstantInt>(C))
       // Must be non-zero due to null test above.
       return true;
-    // TODO: Handle vectors
+
+    // For constant vectors, check that all elements are undefined or known
+    // non-zero to determine that the whole vector is known non-zero.
+    if (auto *VTy = dyn_cast<VectorType>(V->getType())) {
+      for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
+        Constant *Elt = C->getAggregateElement(i);
+        if (!Elt || Elt->isNullValue() ||
+            (!isa<UndefValue>(Elt) && !isa<ConstantInt>(Elt)))
+          return false;
+      }
+      return true;
+    }
+
     return false;
   }
 
   if (auto *I = dyn_cast<Instruction>(V)) {
     if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) {
       // If the possible ranges don't contain zero, then the value is
       // definitely non-zero.
       if (auto *Ty = dyn_cast<IntegerType>(V->getType())) {

test/Transforms/InstSimplify/vec-cmp.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt < %s -instsimplify -S | FileCheck %s
 
-; FIXME: isKnownNonZero should work for integer vectors.
-
 define <2 x i1> @nonzero_vec_splat(<2 x i32> %x) {
 ; CHECK-LABEL: @nonzero_vec_splat(
-; CHECK-NEXT:    [[Y:%.*]] = or <2 x i32> %x, <i32 1, i32 1>
-; CHECK-NEXT:    [[C:%.*]] = icmp eq <2 x i32> [[Y]], zeroinitializer
-; CHECK-NEXT:    ret <2 x i1> [[C]]
+; CHECK-NEXT:    ret <2 x i1> zeroinitializer
 ;
   %y = or <2 x i32> %x, <i32 1, i32 1>
   %c = icmp eq <2 x i32> %y, zeroinitializer
   ret <2 x i1> %c
 }
 
 define <2 x i1> @nonzero_vec_nonsplat(<2 x i32> %x) {
 ; CHECK-LABEL: @nonzero_vec_nonsplat(
-; CHECK-NEXT:    [[Y:%.*]] = or <2 x i32> %x, <i32 2, i32 1>
-; CHECK-NEXT:    [[C:%.*]] = icmp ne <2 x i32> [[Y]], zeroinitializer
-; CHECK-NEXT:    ret <2 x i1> [[C]]
+; CHECK-NEXT:    ret <2 x i1> <i1 true, i1 true>
 ;
   %y = or <2 x i32> %x, <i32 2, i32 1>
   %c = icmp ne <2 x i32> %y, zeroinitializer
   ret <2 x i1> %c
 }
 
 define <2 x i1> @nonzero_vec_undef_elt(<2 x i32> %x) {
 ; CHECK-LABEL: @nonzero_vec_undef_elt(
-; CHECK-NEXT:    [[Y:%.*]] = or <2 x i32> %x, <i32 undef, i32 1>
-; CHECK-NEXT:    [[C:%.*]] = icmp eq <2 x i32> [[Y]], zeroinitializer
-; CHECK-NEXT:    ret <2 x i1> [[C]]
+; CHECK-NEXT:    ret <2 x i1> zeroinitializer
 ;
   %y = or <2 x i32> %x, <i32 undef, i32 1>
   %c = icmp eq <2 x i32> %y, zeroinitializer
   ret <2 x i1> %c
 }
 
 define <2 x i1> @may_be_zero_vec(<2 x i32> %x) {
 ; CHECK-LABEL: @may_be_zero_vec(
 ; CHECK-NEXT:    [[Y:%.*]] = or <2 x i32> %x, <i32 0, i32 1>
 ; CHECK-NEXT:    [[C:%.*]] = icmp ne <2 x i32> [[Y]], zeroinitializer
 ; CHECK-NEXT:    ret <2 x i1> [[C]]
 ;
   %y = or <2 x i32> %x, <i32 0, i32 1>
   %c = icmp ne <2 x i32> %y, zeroinitializer
   ret <2 x i1> %c
 }
 
 ; Multiplies of non-zero numbers are non-zero if there is no unsigned overflow.
 define <2 x i1> @nonzero_vec_mul_nuw(<2 x i32> %x, <2 x i32> %y) {
 ; CHECK-LABEL: @nonzero_vec_mul_nuw(
-; CHECK-NEXT:    [[XNZ:%.*]] = or <2 x i32> %x, <i32 1, i32 2>
-; CHECK-NEXT:    [[YNZ:%.*]] = or <2 x i32> %y, <i32 3, i32 undef>
-; CHECK-NEXT:    [[M:%.*]] = mul nuw <2 x i32> [[XNZ]], [[YNZ]]
-; CHECK-NEXT:    [[C:%.*]] = icmp eq <2 x i32> [[M]], zeroinitializer
-; CHECK-NEXT:    ret <2 x i1> [[C]]
+; CHECK-NEXT:    ret <2 x i1> zeroinitializer
 ;
   %xnz = or <2 x i32> %x, <i32 1, i32 2>
   %ynz = or <2 x i32> %y, <i32 3, i32 undef>
   %m = mul nuw <2 x i32> %xnz, %ynz
   %c = icmp eq <2 x i32> %m, zeroinitializer
   ret <2 x i1> %c
 }
 
 ; Multiplies of non-zero numbers are non-zero if there is no signed overflow.
 define <2 x i1> @nonzero_vec_mul_nsw(<2 x i32> %x, <2 x i32> %y) {
 ; CHECK-LABEL: @nonzero_vec_mul_nsw(
-; CHECK-NEXT:    [[XNZ:%.*]] = or <2 x i32> %x, <i32 undef, i32 2>
-; CHECK-NEXT:    [[YNZ:%.*]] = or <2 x i32> %y, <i32 3, i32 4>
-; CHECK-NEXT:    [[M:%.*]] = mul nsw <2 x i32> [[XNZ]], [[YNZ]]
-; CHECK-NEXT:    [[C:%.*]] = icmp ne <2 x i32> [[M]], zeroinitializer
-; CHECK-NEXT:    ret <2 x i1> [[C]]
+; CHECK-NEXT:    ret <2 x i1> <i1 true, i1 true>
 ;
   %xnz = or <2 x i32> %x, <i32 undef, i32 2>
   %ynz = or <2 x i32> %y, <i32 3, i32 4>
   %m = mul nsw <2 x i32> %xnz, %ynz
   %c = icmp ne <2 x i32> %m, zeroinitializer
   ret <2 x i1> %c
 }