[ValueTracking] Improve isKnownNonNaN() to recognize zero splats.

[jonpa]

Summary

Hi,

I noticed that this function:

define <2 x double> @f1(<2 x double> %val) {
  %cmp = fcmp ogt <2 x double> %val,  zeroinitializer
  %ret = select <2 x i1> %cmp, <2 x double> %val, <2 x double> zeroinitializer
  ret <2 x double> %ret
}

, which I expected to be implemented with a systemZ vfmaxdb (vector fp maximum), was not represented on the DAG with an fmaxnum node, and therefore not in the assembly output either.

I found this to be due to the fact that ValueTracking.cpp:isKnownNonNaN() is not recognizing the zero splat. This is because it is a ConstantAggregateZero which is a ConstantData and not a ConstantDataVector that the function is handling.

I wonder if my patch here is the right approach, or if it might be even better to make a utility function for this type of query to make this mistake less likely in other places? It is surprising to me to have a zero splat not recognized as a ConstantDataVector...

[tlively]

Looks good to me! I don't have enough context to know of any other locations where this could be missed, so I'm not sure whether a helper function would be useful. It could always be added later, anyways.

[arsenm]

llvm/lib/Analysis/ValueTracking.cpp

4781

I don't think you need the FP type check

[jonpa]

llvm/lib/Analysis/ValueTracking.cpp

4781

removed

llvm/lib/Analysis/ValueTracking.cpp

 static bool isKnownNonNaN(const Value *V, FastMathFlags FMF) {
   if (FMF.noNaNs())
     return true;
 
   if (auto *C = dyn_cast<ConstantFP>(V))
     return !C->isNaN();
 
   if (auto *C = dyn_cast<ConstantDataVector>(V)) {
     if (!C->getElementType()->isFloatingPointTy())
       return false;
     for (unsigned I = 0, E = C->getNumElements(); I < E; ++I) {
       if (C->getElementAsAPFloat(I).isNaN())
         return false;
     }
     return true;
   }
 
+  if (isa<ConstantAggregateZero>(V))
+    return true;
+
   return false;
 }
 
 static bool isKnownNonZero(const Value *V) {
   if (auto *C = dyn_cast<ConstantFP>(V))
     return !C->isZero();
 
   if (auto *C = dyn_cast<ConstantDataVector>(V)) {

llvm/test/CodeGen/SystemZ/vec-max-min-zerosplat.ll

+; Test vector maximum/minimum with a zero splat on z14.
+;
+; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z14 | FileCheck %s
+
+define <2 x double> @f1(<2 x double> %val) {
+; CHECK-LABEL: f1:
+; CHECK: vgbm %v0, 0
+; CHECK-NEXT: vfmaxdb %v24, %v24, %v0, 4
+; CHECK-NEXT: br %r14
+  %cmp = fcmp ogt <2 x double> %val,  zeroinitializer
+  %ret = select <2 x i1> %cmp, <2 x double> %val, <2 x double> zeroinitializer
+  ret <2 x double> %ret
+}
+
+define <2 x double> @f2(<2 x double> %val) {
+; CHECK-LABEL: f2:
+; CHECK: vgbm %v0, 0
+; CHECK-NEXT: vfmindb %v24, %v24, %v0, 4
+; CHECK-NEXT: br %r14
+  %cmp = fcmp olt <2 x double> %val,  zeroinitializer
+  %ret = select <2 x i1> %cmp, <2 x double> %val, <2 x double> zeroinitializer
+  ret <2 x double> %ret
+}
+
+define <4 x float> @f3(<4 x float> %val) {
+; CHECK-LABEL: f3:
+; CHECK: vgbm %v0, 0
+; CHECK-NEXT: vfmaxsb %v24, %v24, %v0, 4
+; CHECK-NEXT: br %r14
+  %cmp = fcmp ogt <4 x float> %val,  zeroinitializer
+  %ret = select <4 x i1> %cmp, <4 x float> %val, <4 x float> zeroinitializer
+  ret <4 x float> %ret
+}
+
+define <4 x float> @f4(<4 x float> %val) {
+; CHECK-LABEL: f4:
+; CHECK: vgbm %v0, 0
+; CHECK-NEXT: vfminsb %v24, %v24, %v0, 4
+; CHECK-NEXT: br %r14
+  %cmp = fcmp olt <4 x float> %val,  zeroinitializer
+  %ret = select <4 x i1> %cmp, <4 x float> %val, <4 x float> zeroinitializer
+  ret <4 x float> %ret
+}
+
+define <2 x double> @f5(<2 x double> %val) {
+; CHECK-LABEL: f5:
+; CHECK: vgbm %v0, 0
+; CHECK-NEXT: vfmaxdb %v24, %v24, %v0, 1
+; CHECK-NEXT: br %r14
+  %cmp = fcmp ugt <2 x double> %val,  zeroinitializer
+  %ret = select <2 x i1> %cmp, <2 x double> %val, <2 x double> zeroinitializer
+  ret <2 x double> %ret
+}
+
+define <2 x double> @f6(<2 x double> %val) {
+; CHECK-LABEL: f6:
+; CHECK: vgbm %v0, 0
+; CHECK-NEXT: vfmindb %v24, %v24, %v0, 1
+; CHECK-NEXT: br %r14
+  %cmp = fcmp ult <2 x double> %val,  zeroinitializer
+  %ret = select <2 x i1> %cmp, <2 x double> %val, <2 x double> zeroinitializer
+  ret <2 x double> %ret
+}
+
+define <4 x float> @f7(<4 x float> %val) {
+; CHECK-LABEL: f7:
+; CHECK: vgbm %v0, 0
+; CHECK-NEXT: vfmaxsb %v24, %v24, %v0, 1
+; CHECK-NEXT: br %r14
+  %cmp = fcmp ugt <4 x float> %val,  zeroinitializer
+  %ret = select <4 x i1> %cmp, <4 x float> %val, <4 x float> zeroinitializer
+  ret <4 x float> %ret
+}
+
+define <4 x float> @f8(<4 x float> %val) {
+; CHECK-LABEL: f8:
+; CHECK: vgbm %v0, 0
+; CHECK-NEXT: vfminsb %v24, %v24, %v0, 1
+; CHECK-NEXT: br %r14
+  %cmp = fcmp ult <4 x float> %val,  zeroinitializer
+  %ret = select <4 x i1> %cmp, <4 x float> %val, <4 x float> zeroinitializer
+  ret <4 x float> %ret
+}