ValueTracking: Handle shufflevector in computeKnownFPClass

[RKSimon]

Add test coverage for demanded elts that include a undef shuffle mask value?

llvm/test/Transforms/Attributor/nofpclass.ll

1027

pre-commit these and rebase to show the diff

[arsenm]

Rebase on split tests

[foad]

LGTM with nit.

llvm/lib/Analysis/ValueTracking.cpp

4610

cast not dyn_cast

[RKSimon]

LGTM - cheers

[arsenm]

llvm/lib/Analysis/ValueTracking.cpp

4610

Shufflevector constant expressions do exist though I couldn’t figure out how to get one to reach here, so it’s better to leave it as is

[foad]

llvm/lib/Analysis/ValueTracking.cpp

4610

Fair enough - but maybe getShuffleDemandedElts should handle them too.

[arsenm]

dea4f37b7db5757dede09dccf3497fcfdc000809

llvm/lib/Analysis/ValueTracking.cpp

if (!!DemandedVecElts) {
       KnownFPClass Known2;
       computeKnownFPClass(Vec, DemandedVecElts, InterestedClasses, Known2,
                           Depth + 1, Q, TLI);
       Known |= Known2;
     }
 
     break;
   }
+  case Instruction::ShuffleVector: {
+    // For undef elements, we don't know anything about the common state of
+    // the shuffle result.
+    APInt DemandedLHS, DemandedRHS;
+    auto *Shuf = dyn_cast<ShuffleVectorInst>(Op);
+    if (!Shuf || !getShuffleDemandedElts(Shuf, DemandedElts, DemandedLHS, DemandedRHS))
+      return;
+
+    if (!!DemandedLHS) {
+      const Value *LHS = Shuf->getOperand(0);
+      computeKnownFPClass(LHS, DemandedLHS, InterestedClasses, Known,
+                          Depth + 1, Q, TLI);
+
+      // If we don't know any bits, early out.
+      if (Known.isUnknown())
+        break;
+    } else {
+      Known.KnownFPClasses = fcNone;
+    }
+
+    if (!!DemandedRHS) {
+      KnownFPClass Known2;
+      const Value *RHS = Shuf->getOperand(1);
+      computeKnownFPClass(RHS, DemandedRHS, InterestedClasses, Known2,
+                          Depth + 1, Q, TLI);
+      Known |= Known2;
+    }
+
+    break;
+  }
   case Instruction::ExtractValue: {
     computeKnownFPClass(Op->getOperand(0), DemandedElts, InterestedClasses,
                         Known, Depth + 1, Q, TLI);
     break;
   }
   default:
     break;
   }

llvm/test/Transforms/Attributor/nofpclass.ll

;
   %extract2 = extractelement <4 x float> %insert, i32 2
   %extract3 = extractelement <4 x float> %insert, i32 3
   %ins.0 = insertelement <2 x float> poison, float %extract3, i32 0
   %ins.1 = insertelement <2 x float> %ins.0, float %extract2, i32 1
   ret <2 x float> %ins.1
 }
 
 ; FIXME: Doesn't actually reach computeKnownFPClass
 define <4 x float> @shufflevector_constexpr() {
 ; CHECK: Function Attrs: nofree norecurse nosync nounwind willreturn memory(none)
 ; CHECK-LABEL: define <4 x float> @shufflevector_constexpr
 ; CHECK-SAME: () #[[ATTR2]] {
 ; CHECK-NEXT:    ret <4 x float> <float 1.000000e+00, float bitcast (i32 ptrtoint (ptr @shufflevector_constexpr to i32) to float), float 4.000000e+00, float 0.000000e+00>
 ;
   ret <4 x float> shufflevector (<2 x float> <float 1.0, float bitcast (i32 ptrtoint (ptr @shufflevector_constexpr to i32) to float)>, <2 x float> <float 4.0, float 0.0>, <4 x i32> <i32 0, i32 1, i32 2, i32 3>)
 }
 
 define <4 x float> @shufflevector_concat_disjoint(<2 x float> nofpclass(nan) %arg0, <2 x float> nofpclass(inf) %arg1) {
 ; CHECK: Function Attrs: nofree norecurse nosync nounwind willreturn memory(none)
 ; CHECK-LABEL: define <4 x float> @shufflevector_concat_disjoint
 ; CHECK-SAME: (<2 x float> nofpclass(nan) [[ARG0:%.*]], <2 x float> nofpclass(inf) [[ARG1:%.*]]) #[[ATTR2]] {
 ; CHECK-NEXT:    [[SHUFFLE:%.*]] = shufflevector <2 x float> [[ARG0]], <2 x float> [[ARG1]], <4 x i32> <i32 0, i32 1, i32 2, i32 3>
 ; CHECK-NEXT:    ret <4 x float> [[SHUFFLE]]
 ;
   %shuffle = shufflevector <2 x float> %arg0, <2 x float> %arg1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   ret <4 x float> %shuffle
 }
 
 define <4 x float> @shufflevector_concat_overlap(<2 x float> nofpclass(nan norm psub) %arg0, <2 x float> nofpclass(inf nan sub) %arg1) {
 ; CHECK: Function Attrs: nofree norecurse nosync nounwind willreturn memory(none)
-; CHECK-LABEL: define <4 x float> @shufflevector_concat_overlap
+; CHECK-LABEL: define nofpclass(nan psub) <4 x float> @shufflevector_concat_overlap
 ; CHECK-SAME: (<2 x float> nofpclass(nan psub norm) [[ARG0:%.*]], <2 x float> nofpclass(nan inf sub) [[ARG1:%.*]]) #[[ATTR2]] {
 ; CHECK-NEXT:    [[SHUFFLE:%.*]] = shufflevector <2 x float> [[ARG0]], <2 x float> [[ARG1]], <4 x i32> <i32 0, i32 1, i32 2, i32 3>
 ; CHECK-NEXT:    ret <4 x float> [[SHUFFLE]]
 ;
   %shuffle = shufflevector <2 x float> %arg0, <2 x float> %arg1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   ret <4 x float> %shuffle
 }
 

 ; CHECK-NEXT:    ret <4 x float> [[SHUFFLE]]
 ;
   %shuffle = shufflevector <2 x float> %arg0, <2 x float> %arg1, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
   ret <4 x float> %shuffle
 }
 
 define <4 x float> @shufflevector_only_demand_lhs(<2 x float> nofpclass(inf) %arg0, <2 x float> %arg1) {
 ; CHECK: Function Attrs: nofree norecurse nosync nounwind willreturn memory(none)
-; CHECK-LABEL: define <4 x float> @shufflevector_only_demand_lhs
+; CHECK-LABEL: define nofpclass(inf) <4 x float> @shufflevector_only_demand_lhs
 ; CHECK-SAME: (<2 x float> nofpclass(inf) [[ARG0:%.*]], <2 x float> [[ARG1:%.*]]) #[[ATTR2]] {
 ; CHECK-NEXT:    [[SHUFFLE:%.*]] = shufflevector <2 x float> [[ARG0]], <2 x float> [[ARG1]], <4 x i32> <i32 0, i32 1, i32 1, i32 0>
 ; CHECK-NEXT:    ret <4 x float> [[SHUFFLE]]
 ;
   %shuffle = shufflevector <2 x float> %arg0, <2 x float> %arg1, <4 x i32> <i32 0, i32 1, i32 1, i32 0>
   ret <4 x float> %shuffle
 }
 
 define <4 x float> @shufflevector_only_demand_rhs(<2 x float> %arg0, <2 x float> nofpclass(inf) %arg1) {
 ; CHECK: Function Attrs: nofree norecurse nosync nounwind willreturn memory(none)
-; CHECK-LABEL: define <4 x float> @shufflevector_only_demand_rhs
+; CHECK-LABEL: define nofpclass(inf) <4 x float> @shufflevector_only_demand_rhs
 ; CHECK-SAME: (<2 x float> [[ARG0:%.*]], <2 x float> nofpclass(inf) [[ARG1:%.*]]) #[[ATTR2]] {
 ; CHECK-NEXT:    [[SHUFFLE:%.*]] = shufflevector <2 x float> [[ARG0]], <2 x float> [[ARG1]], <4 x i32> <i32 2, i32 3, i32 3, i32 2>
 ; CHECK-NEXT:    ret <4 x float> [[SHUFFLE]]
 ;
   %shuffle = shufflevector <2 x float> %arg0, <2 x float> %arg1, <4 x i32> <i32 2, i32 3, i32 3, i32 2>
   ret <4 x float> %shuffle
 }
 

 ;
   %shuffle = shufflevector <2 x float> %arg0, <2 x float> %arg1, <4 x i32> <i32 1, i32 3, i32 0, i32 1>
   %extract = extractelement <4 x float> %shuffle, i32 0
   ret float %extract
 }
 
 define float @shufflevector_extractelt1(<2 x float> %arg0, <2 x float> nofpclass(inf) %arg1) {
 ; CHECK: Function Attrs: nofree norecurse nosync nounwind willreturn memory(none)
-; CHECK-LABEL: define float @shufflevector_extractelt1
+; CHECK-LABEL: define nofpclass(inf) float @shufflevector_extractelt1
 ; CHECK-SAME: (<2 x float> [[ARG0:%.*]], <2 x float> nofpclass(inf) [[ARG1:%.*]]) #[[ATTR2]] {
 ; CHECK-NEXT:    [[SHUFFLE:%.*]] = shufflevector <2 x float> [[ARG0]], <2 x float> [[ARG1]], <4 x i32> <i32 1, i32 3, i32 0, i32 1>
 ; CHECK-NEXT:    [[EXTRACT:%.*]] = extractelement <4 x float> [[SHUFFLE]], i32 1
 ; CHECK-NEXT:    ret float [[EXTRACT]]
 ;
   %shuffle = shufflevector <2 x float> %arg0, <2 x float> %arg1, <4 x i32> <i32 1, i32 3, i32 0, i32 1>
   %extract = extractelement <4 x float> %shuffle, i32 1
   ret float %extract